你是无线通信方向的院士 我是大学教授,正在写一篇文章。下面是我们的一个推导的结果,现在我们需要一起对...
Created on: December 24, 2025
Answered using GPT-5.2 Thinking by Chat01
Created on: December 24, 2025
Answered using GPT-5.2 Thinking by Chat01
你是无线通信方向的院士 我是大学教授,正在写一篇文章。下面是我们的一个推导的结果,现在我们需要一起对齐整理一下思路,现在的有点乱。我们想要的推导结果是\mathbf{Q}k^{\mathrm{sp}}和\mathbf{Q}k^{\mathrm{bulk}}的矩阵形式解析解,即最重要的结果是:
\begin{equation}\label{eq:Qsp_compact_engl}
\mathbf{Q}k^{\mathrm{sp}}
=\Big(\tfrac{1}{N^{3/2}}\mathbf{I}+\tfrac{\sqrt{N}}{2N}\mathbf{A}k\Big)\odot \mathbf{S}(\mathbf{s}),
\qquad \mathbf{s}=\mathbf{F},\tilde{\boldsymbol{\kappa}}.
\end{equation}
和
\begin{align}
\mathbf Q_k^{\mathrm{bulk}}
=\frac{1}{N^2}\mathbf D{-k}\mathbf K{\mathrm{bulk}}\mathbf D_k
=\frac{1}{N^2}(\mathbf d_k\mathbf d_k^H)\odot \mathbf K{\mathrm{bulk}}.
\end{align}
下面是我们的推导过程。需要注意的是,我们在推导中尽量少定义新变量,避免混乱。另外,下面的推到中没有加入
\begin{equation}
\mathbb{E}{|r_k|^2}\approx
\frac{1}{N^2}\sum{n,m}\tilde{g}{n,k}\tilde{g}{m,k},e^{j\frac{2\pi}{N}k(n-m)}
\sum_{p,q,p',q'}e^{j\tilde{\phi}{p,q,p',q'}}
e^{-j\frac{2\pi}{N}\left((n-1)(q-p)+(m-1)(p'-q')\right)}
\exp!\big[-\frac{\tilde{h}^2}{N^2} S(p,q,p',q')\big],
\label{eq:Ek_rk2_start}
\end{equation}
中的e^{j\tilde{\phi}{p,q,p',q'}}这一项。这一项是我们自己可设计的相位, ,这一项最好在我们最终的解析解和其他部分独立出来,但是又不是仅由k点决定,而是由多个位置的点联合决定的。请用中文文本方式进行总结推导,用IEEE TSP的风格来写。下面是推导过程:
\subsection{Hyperplane Decomposition and the 1D Difference Weight }
After removing the pairing set, the remaining non-pairing index set can be further decomposed into four structured hyperplanes and a residual bulk complement:
\begin{equation}
\text{(H1)}\ p'=q',\qquad
\text{(H2)}\ p=q,\qquad
\text{(H3)}\ q=q',\qquad
\text{(H4)}\ p=p',\qquad
\text{Bulk: the remaining index combinations}.
\end{equation}
On each hyperplane, the shift invariance of reduces the four-argument quadratic form to a function of a single modular difference . To directly characterize this reduction, we introduce the one-dimensional weight:
\begin{equation}\label{eq:kappa_def_engl}
\kappa[d]\triangleq \exp!\left(-\alpha,\mathbf{b}^T\mathbf{K}_2(d,0,0,0)\mathbf{b}\right)
=\exp!\left(-2\alpha\big(N_d-\Psi(d)\big)\right),\qquad d\in\mathbb{Z}_N,
\end{equation}
which serves as a shorthand representation of the original kernel. Define and for , let , and compute the single DFT vector:
\begin{equation}
\mathbf{s}\triangleq \mathbf{F},\tilde{\boldsymbol{\kappa}}.
\end{equation}
Based on the above notation, the total contribution of the hyperplanes can be expressed as a Hermitian quadratic form:
\begin{equation}\label{eq:Sk_sp_quadratic_engl}
\mathcal{S}k^{\mathrm{sp}}(\mathbf{g})=\mathbf{g}^H\mathbf{Q}k^{\mathrm{sp}}\mathbf{g},
\qquad
\mathbf{Q}k^{\mathrm{sp}}=\mathbf{Q}{\mathrm{diag}}+\mathbf{Q}{k,ut}+\mathbf{Q}{k,v0}.
\end{equation}
Introduce the shared lifting'' matrix $\mathbf{S}(\mathbf{s})\triangleq \mathbf{1}\mathbf{s}^H+\mathbf{s}\mathbf{1}^T$ and $\mathbf{A}_k\triangleq \mathbf{f}_k\mathbf{f}_{-k}^H+\mathbf{f}_{-k}\mathbf{f}_k^H$. The diagonal layer induced by (H3) and (H4) then simplifies to: \begin{equation} \mathbf{Q}_{\mathrm{diag}} =\frac{2}{N^{3/2}}\operatorname{diag}(\operatorname{Re}\{\mathbf{s}\}) =\frac{1}{N^{3/2}}\big(\mathbf{I}\odot \mathbf{S}(\mathbf{s})\big), \end{equation} while the two (Hermitianized) low-rank layers induced by (H1) and (H2) combine into a compact Hadamard form: \begin{equation} \mathbf{Q}_{k,ut}+\mathbf{Q}_{k,v0} =\frac{\sqrt{N}}{2N}\big(\mathbf{A}_k\odot \mathbf{S}(\mathbf{s})\big). \end{equation} In summary, we obtain: \begin{equation}\label{eq:Qsp_compact_engl} \mathbf{Q}_k^{\mathrm{sp}} =\Big(\tfrac{1}{N^{3/2}}\mathbf{I}+\tfrac{\sqrt{N}}{2N}\mathbf{A}_k\Big)\odot \mathbf{S}(\mathbf{s}), \qquad \mathbf{s}=\mathbf{F}\,\tilde{\boldsymbol{\kappa}}. \end{equation} This result indicates that the structured baseline'' background is fully governed by the one-dimensional difference profile via a single DFT, and its matrix form exhibits DFT-friendliness (diagonal-plus-low-rank structure).
\subsection{Bulk Complement Analysis via Circulant Subspace Projection}
The remaining bulk complement excludes both perfect pairings and single-hyperplane structures, and primarily captures the residual high-dimensional coupling effects. Let and define:
\begin{equation}
[\mathbf{H}k]{p,q}\triangleq \gamma_{\langle k+p-q\rangle}.
\end{equation}
Since is a real-valued vector, it satisfies , such that the bulk contribution can be written as a Hermitian quadratic form:
\begin{equation}\label{eq:Sk_bulk_vec_engl}
\mathcal{S}_k^{(\mathrm{bulk})}(\mathbf{g})
=\frac{1}{N^2}\operatorname{vec}(\mathbf{H}k)^H\mathbf{M}{\mathrm{bulk}}\operatorname{vec}(\mathbf{H}_k).
\end{equation}
A key property is that lies in the circulant subspace, so there exists a vector such that . Using the circulant synthesis matrix , we have , which projects the -dimensional kernel matrix to an -dimensional matrix:
\begin{equation}
\mathcal{S}k^{(\mathrm{bulk})}(\mathbf{g})
=\frac{1}{N^2}\mathbf{c}k^H\mathbf{R}{\mathrm{bulk}}\mathbf{c}k,
\qquad
\mathbf{R}{\mathrm{bulk}}\triangleq \mathbf{B}^H\mathbf{M}{\mathrm{bulk}}\mathbf{B}\in\mathbb{C}^{N\times N}.
\end{equation}
Returning to the -domain and utilizing the modulation identity (which induces a Hadamard form in ), the bulk term can be expressed as:
\begin{equation}\label{eq:Sk_bulk_hadamard_engl}
\mathcal{S}_k^{(\mathrm{bulk})}(\mathbf{g})=\mathbf{g}^H\mathbf{Q}_k^{\mathrm{bulk}}\mathbf{g},
\qquad
\mathbf{Q}k^{\mathrm{bulk}}
=\frac{1}{N^2}(\mathbf{d}k\mathbf{d}k^H)\odot \mathbf{K}{\mathrm{bulk}},
\qquad
\mathbf{K}{\mathrm{bulk}}=\mathbf{F}\mathbf{R}{\mathrm{bulk}}\mathbf{F}^H,
\end{equation}
where .
\mathbf{g}^H\mathbf{Q}_k\mathbf{g},
\end{equation}
where:
\begin{equation}
\mathbf{Q}_k\triangleq
\left(2-\frac{1}{N}\right)\mathbf{f}_k\mathbf{f}_k^H+\mathbf{Q}_k^{\mathrm{sp}}+\mathbf{Q}_k^{\mathrm{bulk}}.
\end{equation}
The first term in \eqref{eq:final_matrix_form_engl} is a purely filter-dependent coherent peak, determined solely by the DFT bin projection ; the remaining terms constitute the background ``baseline''. Specifically, the structured background is controlled by the one-dimensional attenuation profile via , and its matrix form is DFT-friendly (diagonal-plus-low-rank structure); the bulk term characterizes the residual fine-grained modulation through the Hadamard factor .
We partition the non-pairing set into structured hyperplanes and a residual bulk:
On each hyperplane, reduces (by shift-invariance of ) to a function of a single modular difference . We encode this \emph{using itself} via the 1D weight
\begin{equation}\label{eq:kappa_def}
\kappa[d]\triangleq \exp!\left(-\alpha,\mathbf b^T\mathbf K_2(d,0,0,0)\mathbf b\right),\qquad d\in\mathbb{Z}N ,
\end{equation}
which is merely a shorthand (no new model) since it is generated by the original four-argument kernel.
\begin{equation}
\kappa[d]\triangleq \exp!\left(-\alpha,\mathbf b^T\mathbf K_2(d,0,0,0)\mathbf b\right)
=\exp!\left(-2\alpha\big(N_d-\Psi(d)\big)\right).
\end{equation}
% (Optional) elementwise definition
\emph{Low-rank layers:} the hyperplanes (H1) and (H2) lead to two (Hermitianized) rank- quadratic forms
% ===== Key identity: elementwise product of DFT columns =====
For normalized DFT columns, the elementwise products satisfy
Hence,
% ===== Compact forms for Q_{k,ut}^{(0)} and Q_{k,v0}^{(0)} without D_k =====
Using and ,
% ===== Final low-rank sum (no D_k) =====
\emph{Diagonal layer:} the hyperplanes (H3) and (H4) collapse (via the DFT shift/modulation property) to a diagonal matrix
% ===== Key idea: represent the diagonal operator by its diagonal vector =====
Since is diagonal, define its diagonal vector
% ===== Compute q_diag explicitly =====
For each ,
\begin{align*}
(\mathbf q_{\mathrm{diag}})n
&=\frac{1}{N^2}\sum{d=1}^{N-1}\kappa[d]\Big(e^{-j\omega d(n-1)}+e^{+j\omega d(n-1)}\Big)\
&=\frac{2}{N^2}\sum_{d=1}^{N-1}\kappa[d]\cos!\big(\omega d(n-1)\big).
\end{align*}
% ===== Pack kappa into a length-N vector and express via DFT columns =====
Define the length- vector by
Then
and
Therefore,
and the diagonal operator can be applied without any as
Thus the structured part is
\begin{equation}\label{eq:Sk_sp_quadratic}
\boxed{
\mathcal S_k^{\mathrm{sp}}(\mathbf g)=\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g,\qquad
\mathbf Q_k^{\mathrm{sp}}=\mathbf Q_{\mathrm{diag}}+\mathbf Q_{k,ut}+\mathbf Q_{k,v0}.
}
\end{equation}
% ===== Normalized DFT =====
and let be the -th column of (indices mod ).
% ===== kappa vector and one FFT =====
% ===== s-factor matrix shared by all terms =====
% ===== Q_diag without diag(·): diagonal picked by Hadamard with I =====
% ===== Low-rank sum expressed with the SAME S(s) =====
% Assume normalized DFT F and its columns f_k.
% Let z = F \tilde{kappa}. To make the Hadamard product well-defined, lift z to an NxN matrix:
Note that .
Define
If is Hermitian PSD, then
Let so that .
Define the two matrices
\begin{align}
[\mathbf H_k]{p,q} &\triangleq \gamma{\langle k+p-q\rangle}, \
[\mathbf G_k]{p',q'} &\triangleq \gamma{\langle -k+(q'-p')\rangle}.
\end{align}
Let the bulk kernel (including masking) be
\begin{align}
\mathbf M_{\mathrm{bulk}} \triangleq \operatorname{Herm}{\mathbf K\odot \mathbf P^{\mathrm{bulk}}},
\qquad
[\mathbf K]{(p,q),(p',q')} \triangleq \exp!\big(-\alpha S_2(p,q,p',q')\big),
\qquad \alpha\triangleq \frac{\tilde h^2}{N^2}.
\end{align}
The bulk contribution is (cf. the original vec--matrix--vec form)
\begin{align}
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
=\frac{1}{N^2},\operatorname{vec}(\mathbf H_k)^T,
\mathbf M{\mathrm{bulk}},
\operatorname{vec}(\mathbf G_k).
\label{eq:Sk_bulk_vec_start}
\end{align}
\paragraph*{Step 1: Hermitian quadratic form.}
Assume , hence .
Then for all ,
\begin{align}
[\mathbf G_k]{p',q'}
=\gamma{\langle -k-(p'-q')\rangle}
=\gamma_{\langle k+(p'-q')\rangle}^*
=[\mathbf H_k]{p',q'}^,
\end{align}
i.e.,
\begin{align}
\boxed{\mathbf G_k=\mathbf H_k^,\quad \operatorname{vec}(\mathbf G_k)=\operatorname{vec}(\mathbf H_k)^*.}
\label{eq:Gk_equals_Hk_conj}
\end{align}
Substituting \eqref{eq:Gk_equals_Hk_conj} into \eqref{eq:Sk_bulk_vec_start} yields
\begin{align}
\boxed{
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
=\frac{1}{N^2},\operatorname{vec}(\mathbf H_k)^H,
\mathbf M{\mathrm{bulk}},
\operatorname{vec}(\mathbf H_k).
}
\label{eq:Sk_bulk_quadratic_N2}
\end{align}
\paragraph*{Step 2: Circulant subspace parameterization.}
Define the circulant shift basis
\begin{align}
[\mathbf J_d]{p,q}\triangleq \mathbb 1{p-q\equiv d\ (\mathrm{mod}\ N)},
\qquad d\in\mathbb Z_N.
\end{align}
Any circulant matrix can be written as
\begin{align}
\operatorname{circ}(\mathbf c)=\sum{d=0}^{N-1} c[d]\mathbf J_d.
\end{align}
Let and .
Then
\begin{align}
\boxed{\mathbf H_k=\operatorname{circ}(\mathbf c_k).}
\label{eq:Hk_circ_ck}
\end{align}
Define the circulant synthesis matrix
\begin{align}
\mathbf B\triangleq \big[\operatorname{vec}(\mathbf J_0),\ldots,\operatorname{vec}(\mathbf J_{N-1})\big]\in\mathbb R^{N^2\times N}.
\end{align}
Using \eqref{eq:Hk_circ_ck},
\begin{align}
\boxed{\operatorname{vec}(\mathbf H_k)=\mathbf B,\mathbf c_k.}
\label{eq:vecHk_Bck}
\end{align}
\paragraph*{Step 3: Projection of the kernel.}
Substitute \eqref{eq:vecHk_Bck} into \eqref{eq:Sk_bulk_quadratic_N2}:
\begin{align}
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
&=\frac{1}{N^2},\mathbf c_k^H\underbrace{\big(\mathbf B^H\mathbf M_{\mathrm{bulk}}\mathbf B\big)}{\triangleq,\mathbf R{\mathrm{bulk}}\in\mathbb C^{N\times N}}\mathbf c_k.
\end{align}
Hence,
\begin{align}
\boxed{
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
=\frac{1}{N^2},\mathbf c_k^H \mathbf R_{\mathrm{bulk}},\mathbf c_k,
\qquad
\mathbf R_{\mathrm{bulk}}\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}\mathbf B.
}
\label{eq:Sk_bulk_ck_Rbulk}
\end{align}
\subsection{Explicit Difference-Domain Formula for }
Let .
Expanding the vec inner product gives
\begin{align}
[\mathbf R_{\mathrm{bulk}}]{d,d'}
&=\sum{p,q}\sum_{p',q'} \mathbb 1{p-q=d}\mathbb 1{p'-q'=d'},
[\mathbf M_{\mathrm{bulk}}]{(p,q),(p',q')}.
\label{eq:Rbulk_expand_pq}
\end{align}
Using the constraints and (all indices modulo ), we can rewrite
\begin{align}
[\mathbf R{\mathrm{bulk}}]{d,d'}
=\sum{q=0}^{N-1}\sum_{q'=0}^{N-1}
[\mathbf M_{\mathrm{bulk}}]_{(q+d,q),(q'+d',q')}.
\label{eq:Rbulk_qqprime}
\end{align}
\paragraph*{Shift-invariance reduction to a single modular difference .}
The term and the bulk mask depend only on modular differences
among , hence the summand in \eqref{eq:Rbulk_qqprime} depends on only through
.
For each fixed , there are exactly pairs such that .
Thus,
\begin{align}
\boxed{
[\mathbf R_{\mathrm{bulk}}]{d,d'}
= N\sum{t\in\mathbb Z_N}
\iota_{\mathrm{bulk}}(d,d',t),
\exp!\Big(-\alpha,S_2(t+d,t,d',0)\Big),
}
\label{eq:Rbulk_ddprime_tsum}
\end{align}
where is the indicator that removes pairings and single-pair hyperplanes.
\paragraph*{Bulk indicator in .}
With the parametrization
\begin{align}
(p,q,p',q')=(t+d,\ t,\ d',\ 0),
\end{align}
the excluded (structured) hyperplanes correspond to the following modular equalities:
\begin{align}
p=q \Leftrightarrow d=0,\qquad
p'=q' \Leftrightarrow d'=0,\qquad
q=q' \Leftrightarrow t=0,\qquad
p=q' \Leftrightarrow t=-d,\qquad
q=p' \Leftrightarrow t=d',\qquad
p=p' \Leftrightarrow t=d'-d.
\end{align}
Hence a convenient explicit choice is
\begin{align}
\boxed{
\iota_{\mathrm{bulk}}(d,d',t)
\triangleq
\mathbb 1{d\neq 0,\ d'\neq 0,\ t\notin{0,-d,d',d'-d}}.
}
\label{eq:iota_bulk}
\end{align}
\paragraph*{Closed form of .}
Recall the cosine kernel and the closed-form
expression for
\begin{align}
S_2(p,q,p',q')=(4N_d+2)\Big[
\Psi(p-q')+\Psi(q-p')
-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')
\Big].
\label{eq:S2_general_Psi}
\end{align}
Substituting into \eqref{eq:S2_general_Psi} yields
\begin{align}
\boxed{
S_2(t+d,t,d',0)
=(4N_d+2)\Big[
\Psi(t+d)+\Psi(t-d')-\Psi(d)-\Psi(d')-\Psi(t)-\Psi(t+d-d')
\Big].
}
\label{eq:S2_bulk_ddprime_t}
\end{align}
Combining \eqref{eq:Rbulk_ddprime_tsum}--\eqref{eq:S2_bulk_ddprime_t} gives a fully explicit
difference-domain construction of .
\subsection{Back to : sp-like Hadamard-modulated quadratic / norm form}
Let and define the circulant shift (permutation) matrix
such that (i.e., ).
Then from \eqref{eq:Sk_bulk_ck_Rbulk},
\begin{align}
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
&=\frac{1}{N^2},\boldsymbol\gamma^H\mathbf\Pi_k^H\mathbf R_{\mathrm{bulk}}\mathbf\Pi_k\boldsymbol\gamma
=\mathbf g^H\underbrace{\Big(\frac{1}{N^2}\mathbf F\mathbf\Pi_k^H\mathbf R_{\mathrm{bulk}}\mathbf\Pi_k\mathbf F^H\Big)}{\triangleq,\mathbf Q_k^{\mathrm{bulk}}}\mathbf g.
\end{align}
Therefore,
\begin{align}
\boxed{
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)=\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}\mathbf g,
\qquad
\mathbf Q_k^{\mathrm{bulk}}\triangleq
\frac{1}{N^2}\mathbf F\mathbf\Pi_k^H\mathbf R{\mathrm{bulk}}\mathbf\Pi_k\mathbf F^H.
}
\label{eq:Sk_bulk_g_Qbulk}
\end{align}
\paragraph*{Hadamard modulation in .}
Using with
,
let and
.
Then
\begin{align}
\mathbf Q_k^{\mathrm{bulk}}
=\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}\mathbf D_k
=\frac{1}{N^2}(\mathbf d_k\mathbf d_k^H)\odot \mathbf K_{\mathrm{bulk}}.
\end{align}
Hence,
\begin{align}
\boxed{
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)
=\mathbf g^H\Big(\frac{1}{N^2}(\mathbf d_k\mathbf d_k^H)\odot \mathbf K_{\mathrm{bulk}}\Big)\mathbf g,
\qquad
\mathbf K_{\mathrm{bulk}}=\mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^H,
}
\label{eq:Sk_bulk_hadamard}
\end{align}
where is given explicitly by \eqref{eq:Rbulk_ddprime_tsum}--\eqref{eq:S2_bulk_ddprime_t}.
\left|
\left(
\frac{1}{N^2}(\mathbf d_k\mathbf d_k^H)\odot \mathbf K_{\mathrm{bulk}}
\right)^{1/2}\mathbf g
\right|_2^2.
}
\label{eq:Sk_bulk_norm}
\end{align}
Otherwise one may use the standard stabilized surrogate
,
yielding .
下面给出一份**“思路对齐 + 推导脉络整理”**的中文总结稿(偏 IEEE TSP 写法),目标是把最终最重要的两个矩阵解析解
的来龙去脉说清楚,并把可设计相位项 作为“独立模块”从解析结构中分离出来。
在剔除“perfect pairing set”之后,剩余的非配对索引组合可进一步划分为四类结构化超平面与其余的 bulk 补集:
该分解的关键作用是:在每个超平面上,由于 的移不变性(shift-invariance),原本依赖四元组 的核指数项会退化为仅依赖一个模差 的函数。
为避免引入新的模型变量,我们直接用 本身定义一维权重
并设 、,令
这里的 是唯一一次 DFT得到的“差分谱向量”,后续结构化项全部共享它。
定义共享 lifting:
以及 -相关、但与 无关的结构矩阵
其中 为归一化 DFT 矩阵 的第 列(索引按模 计算)。
因此,结构化超平面总和可写为 Hermitian 二次型
且矩阵闭式解为(你们想要的关键式)
该式揭示:结构化“baseline”完全由一维差分衰减轮廓 决定(通过一次 DFT),并呈现明显的 DFT-friendly(对角 + 低秩) 结构。
令 ,并定义
若 ,则 ,从而 bulk 项可写成
其中 汇总了核与 bulk mask(不必再新引入变量解释物理含义)。
关键观察: 属于 circulant 子空间,存在 使
用循环基矩阵 表示
从而把 的核投影为 :
本质是 的循环移位,记该移位为 ,则(利用 )
其中
进一步用 Hadamard 恒等式得到最终形式(你们的关键式)
该式最重要的意义是:所有与 有关的变化被严格限制在一个 rank-1 的 Hadamard 调制因子 ;而“与环境/核有关”的复杂性全部被吸收进一个 -无关 的基矩阵 。
把 coherent pairing 主峰项与两类 baseline 叠加,可写为
其中
解释上:第一项是由滤波器投影决定的“相干峰”;其余两项构成背景 baseline,其中结构化部分由一维差分谱 控制(对角+低秩),bulk 则刻画剩余高维耦合(但 依赖仍可分离为 Hadamard 调制)。
你们指出原式里还应包含
并希望它在最终解析结构中作为“独立模块”出现,且强调它不是仅由单个 决定,而是由 与 两位置联合决定。这里有一个非常干净的矩阵化分离方式:
令相位编码矩阵 满足 。把二元索引 展平为单索引后,有
因此,对任意以 为行索引、 为列索引的核矩阵 ,加入该相位等价于
其逐元素恰为 ( K^{(\phi)}{(p,q),(p',q')}=e^{j\tilde{\phi}{p,q,p',q'}}K_{(p,q),(p',q')}. ) 这一步的优点是:相位完全以一个与 无关的“核矩阵同余变换”形式出现,天然满足你们希望的“独立出来但又是联合索引决定”的要求。
对 bulk:只需把 中的核替换为相位编码后的版本,即
则
注意这里 -依赖的 Hadamard 调制骨架完全不受影响;所有相位设计的影响都被压缩进一个 -无关矩阵 。
对 sp:你们现有的紧致式 之所以成立,本质依赖“超平面上仍可退化为单差分 ”这一 shift-invariance 归约。因此:
实务写法上,建议在论文中把 作为“coding layer”单独成段说明:先给出 时的闭式结构(即你们目前的两条关键式),随后一句话指出“加入 ”等价于对 -核做同余变换,bulk 的 -调制骨架保持不变,而 sp 的 1D 归约在满足某类相位结构时仍保持。
如果你愿意,我也可以把上述内容进一步改写成可直接粘进论文的 TSP 风格 LaTeX 段落(含“Proposition/Lemma + Proof sketch”的版式),并把“相位 coding layer”给出一个最小符号量的标准表述模板(避免新变量泛滥)。
我们现在不用K_2的结构,用s\begin{align}
S(p,q,p',q')&\triangleq(4N_d+2)[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)\notag\
&\quad-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')].
\end{align}
然后我觉得推导过程太简略了,还是要用DFT矩阵的列这种给公式看起来高级且solid。重新更新一版。然后Phi从最开始的推导就扔进去,一起算
下面给一版更“IEEE TSP 风格”、更展开、并且从一开始就把相位 扔进去一起算的整理稿。整体目标仍是把非配对部分写成
并给出 、 的矩阵解析形态(用 DFT 列向量推到最后)。
令 点归一化 DFT 矩阵
表示 的第 列(索引按 取模)。定义
把你们的可设计相位写成矩阵
后面我们只用 的行列求和/相关量,不额外引入复杂新模型。
另外把指数核写成你们现在指定的 :
并记
把相位项直接放回你们的起始式(此处只关心非配对部分的组织方式):
然后对“非配对索引集合”作分解(你们已固定):
结构化部分 就是四个超平面贡献的和;剩余为 。
Lemma (超平面单差分退化). 由于 的移不变/偶对称结构,四个超平面上 都可化为只依赖一个模差 的函数。等价地,可用同一个 1D 轮廓来表征:
并在 (H1)–(H4) 上分别把 取为对应那条超平面上“唯一自由差分”(例如 (H1) 里是 ,(H3) 里是 ,等等)。
这一步你们原来用 做过;现在只需把“核”换成 即可,结构不变。
对归一化 DFT 列,有两条最常用的元素乘积恒等式(索引按 ):
这就是你们“看起来高级且 solid”的关键:把所有超平面求和最后都压成 DFT 列的外积组合。
结论先说:带相位后, 的骨架仍是 “Hadamard 乘一个 lifting 矩阵”,但一般会出现两条 1D DFT 向量(分别对应“对角层”和“低秩层”),除非你们对 施加额外结构使其一致。
仍沿用你们最终希望的表达习惯,定义
为了尽量少定义新变量,我们只定义三个由 直接计算的 1D 序列/标量(它们都满足:若 ,则取值为 1):
并取
直观上: 把某条对角线“加权平均”成 1D;、 是相位编码在两个方向上的“循环自相关”。这些都不依赖 ,符合你们“相位应与 结构解耦”的初衷。
令 ,并做一次 DFT:
并同样做一次 DFT:
核心点:仍然只有“若干条 1D 序列 + 若干次 DFT”,没有把问题拉回 2D/4D 去。
沿你们原来的写法,(H1),(H2) 先生成两个(Hermitian 化前的)秩 ≤2 的外积层,形式上仍可写成
然后用“DFT 列元素乘积恒等式”把 变成 的形式,进而在 Hermitian 化并合并后得到你们熟悉的紧凑 Hadamard 形态:
同理,(H3),(H4) 在无相位时给出纯对角层;带相位后,对角层系数会被 修正,但结构仍保持为
综上,结构化超平面部分为
并且
退化检查:若 (即无相位),则 ,于是 ,你们原来的单 紧致式自然恢复:
若你们进一步设计 使得 ,也能回到同样“一条 ”的最紧凑版。
bulk 的推导与你们原来一致:先到 域,构造
利用 属于 circulant 子空间()以及合成矩阵 把 维核投影到 维:
相位从一开始进来时,bulk 的核矩阵只需替换为
这等价于在 展平空间对核作对角同余变换(如果你们想写得更“线性代数”一点):
回到 -域后, 的作用仍然可以完全剥离为一个 Hadamard 调制因子(这一点非常“干净”):
其中
也就是说:相位设计只改变 -无关的 ,而 -依赖骨架 完全不变。
把 coherent pairing 主峰项与两类 baseline 合起来仍是
其中
如果你希望,我可以把上面第 5 节(结构化部分)再进一步写成“Proof sketch”那种逐行变形版,把每个超平面从 的约束如何把相位求和压成 的过程完整展开(但不再引入额外符号)。
我们把上面的内容改写成一个proposition,主proposition写一下\mathcal{S}_k(\mathbf g)的解析解就行,剩下的推导部分用符合TSP的语言风格和结构的appendix来写。先给出中文版本的文本版本,然后appendix不要分点来写,直接用论文语言来表达。另外,proposition的结构参考下面:
\begin{proposition}\label{prop:Ik_closed_form}
Consider the transmission of random CE-OFDM symbols, the contribution to the squared ACF arising from the vanishing-kernel index configurations could be expressed as:
\begin{align}\label{eq:Ik_def_engll}
\mathcal{I}k(\mathbf{g})=(2N-1)\big|\mathbf{f}{k+1}^{H}\mathbf{\tilde{g}}_k\big|^2,
\end{align}
where .
\end{proposition}
\begin{IEEEproof}
Please refer to Appendix~\ref{app:proof_prop_Ik}.
\end{IEEEproof}
下面给出按 IEEE TSP 论文习惯组织的一套写法:正文只放一个主 Proposition(给出 的解析闭式/矩阵形态),其余推导全部放到 Appendix(用连续论文语言叙述,不用分点)。
\begin{proposition}\label{prop:Sk_closed_form_cn} 考虑发送 个随机 CE-OFDM 符号。令 为给定的实值整形向量, 为归一化 -点 DFT 矩阵, 为其第 列(索引按 取模),并定义 、、。考虑包含可设计相位 的核项,其中 构成相位矩阵 。 令 \begin{align} S(p,q,p',q')\triangleq(4N_d+2)\big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\big], \end{align} 并记 。在“配对项 + 非配对项”分解以及非配对集合的“超平面 + bulk”分解下,有如下近似闭式矩阵解: \begin{align}\label{eq:Sk_main_prop_cn} \mathcal S_k(\mathbf g) &\approx \Big(2-\frac{1}{N}\Big)\big|\mathbf f_k^H\mathbf g\big|^2 +\mathbf g^H\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)\mathbf g +\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)\mathbf g =\mathbf g^H\mathbf Q_k(\mathbf\Phi)\mathbf g, \end{align} 其中 \begin{align} \mathbf Q_k(\mathbf\Phi)\triangleq \Big(2-\frac{1}{N}\Big)\mathbf f_k\mathbf f_k^H+\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)+\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi). \end{align} 结构化超平面项满足 \begin{align}\label{eq:Qsp_prop_cn} \mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big) +\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big), \end{align} 其中 , ; 并定义两条由 汇聚得到的一维权重序列 \begin{align} \eta_0&\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t},\qquad \eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q},\ \xi[d]&\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}\Phi_{p,q}^+\Phi_{p,q+d}\Phi_{p,q}^\Big), \end{align} 以及由核 诱导的一维衰减轮廓 \begin{align} \kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp!\big(-\alpha,\Delta_S(d)\big),\ \ d=1,\ldots,N-1, \end{align} 其中 为将 在单差分超平面上化简得到的仅关于差分 的标量函数(其显式化简见 Appendix~\ref{app:proof_prop_Sk})。 进一步令 \begin{align} \tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\ \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad \tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\ \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d], \end{align} 并记 \begin{align} \mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}{\mathrm{lr}},\qquad \mathbf s{\mathrm{diag}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}{\mathrm{diag}}. \end{align} 最后,bulk 补集项具有 -Hadamard 调制的解析形态 \begin{align}\label{eq:Qbulk_prop_cn} \mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi) =\frac{1}{N^2}\mathbf D{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k =\frac{1}{N^2}(\mathbf d_k\mathbf d_k^H)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi), \end{align} 其中 为与 无关的 Hermitian 基矩阵,由相位编码后的 bulk 核通过 circulant 子空间投影得到(其构造见 Appendix~\ref{app:proof_prop_Sk})。 当 时有 ,从而 ,并退化回无相位的单 紧致结构。 \end{proposition}
\begin{IEEEproof} 请参见 Appendix~\ref{app:proof_prop_Sk}。 \end{IEEEproof}
在带相位编码 的情形下, 的期望项可写成对四元组索引 的加权求和,其中权重由两部分组成:其一为 DFT 调制因子(由 与 的指数项决定),其二为核项 与相位项 的乘积。将索引集合按“perfect pairing(相干配对)”与“非配对”拆分后,前者给出 的相干主峰,这一步仅依赖归一化 DFT 正交性与配对计数,得到 \eqref{eq:Sk_main_prop_cn} 的第一项。
对非配对集合进一步分解为四个结构化超平面与其余 bulk 补集,即 、、、 以及剩余组合。由于 仅依赖差分且满足移不变结构,代入任一单超平面约束后,括号内的六项 会消去三维自由度,从而 在每个超平面上均退化为仅关于一个模差 的标量函数 。据此可引入一维衰减轮廓 (并令 以排除退化差分项),使得所有结构化超平面的核权重均可统一写成 的函数,从而将原本四维核的求和转化为一维差分求和。
为将这些一维差分求和进一步写成矩阵形式,引入归一化 DFT 列 的元素乘积恒等式: 与 (索引按 取模)。该恒等式使得形如 的求和可以等价改写为 ,其中 为 的单次 DFT 汇聚结果;这正是结构化闭式解呈现为“对角层 + 低秩层”的根本原因。
将相位项从一开始并入超平面求和时, 在不同超平面上以不同方式塌缩为一维统计量。对于 与 (产生低秩外积层的两类超平面),相位项分别包含 或 形式的因子,从而自然引出主对角均值 以及偏移对角均值 ,使得低秩层的有效一维权重由 给出。对 与 (产生对角层的两类超平面),相位项以 的局部相关形式进入,并在单差分约束下形成沿 与 两个方向的循环相关;为保持与无相位情形相同的 Hermitian 对称结构,可取两方向相关的均值 ,并得到对角层的有效权重 。分别令 与 ,并定义 lifting 以及 ,则通过将低秩外积层做 Hermitian 化并合并,可得到
而对角层可写为
从而得到 \eqref{eq:Qsp_prop_cn}。
对 bulk 补集,令 ,并定义 。当 时有 ,据此 bulk 项可整理为 Hermitian 二次型 。其中 由相位编码后的核与 bulk mask 共同决定,其元素形式为
关键在于 属于 circulant 子空间,因此存在 使 。用循环移位基 的向量化构造合成矩阵 ,可得 ,从而将 维核投影到 维:
再将 视为 的循环移位,并利用 的调制等价关系,记 ,即可得到
即 \eqref{eq:Qbulk_prop_cn}。最后将相干主峰、结构化超平面项与 bulk 项相加,即得 \eqref{eq:Sk_main_prop_cn},从而命题得证。
appendix的推导过程写的再详细一点,尽量少定义新矩阵,尽量用DFT和现有的去表示。然后给我优化后的latex代码的英文版本
latex% ========================= % Main text: Proposition % ========================= \begin{proposition}\label{prop:Sk_closed_form} Consider the transmission of $N$ random CE-OFDM symbols. Let $\mathbf g\in\mathbb R^{N}$ be a real-valued shaping vector and let $\mathbf F\in\mathbb C^{N\times N}$ denote the normalized $N$-point DFT matrix with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)\ell}$, $\omega\triangleq 2\pi/N$, and $\mathbf f_\ell$ its $\ell$th column (indices are taken modulo $N$). Define $\mathbf d_k=[e^{j\omega k(n-1)}]_{n=1}^{N}$ and $\mathbf D_k\triangleq \mathrm{diag}(e^{-j\omega k(n-1)})_{n=1}^{N}$. Let the designable phase be $e^{j\tilde\phi_{p,q,p',q'}}\triangleq e^{j\phi_{p,q}}e^{-j\phi_{p',q'}}$ and collect $\Phi_{p,q}\triangleq e^{j\phi_{p,q}}$ into $\mathbf\Phi\in\mathbb C^{N\times N}$. Let \begin{align} \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta),\qquad S(p,q,p',q')\triangleq(4N_d+2)\big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\big], \end{align} and $\alpha\triangleq \tilde h^2/N^2$. Then the (baseline) contribution to the squared ACF at bin $k$ admits the following matrix-form closed expression \begin{align}\label{eq:Sk_prop_main} \mathcal S_k(\mathbf g) ~\approx~ \Big(2-\frac{1}{N}\Big)\big|\mathbf f_k^H\mathbf g\big|^2 ~+~\mathbf g^H\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)\mathbf g ~+~\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)\mathbf g ~=~\mathbf g^H\mathbf Q_k(\mathbf\Phi)\mathbf g, \end{align} where \begin{align} \mathbf Q_k(\mathbf\Phi)\triangleq \Big(2-\frac{1}{N}\Big)\mathbf f_k\mathbf f_k^H+\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)+\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi). \end{align} Moreover, define the shared lifting operator and the $k$-dependent DFT structure factor \begin{align} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T,\qquad \mathbf A_k\triangleq \mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H, \end{align} and the following one-dimensional phase aggregates \begin{align} \eta_0~\triangleq~\frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t},\qquad \eta[d]~\triangleq~\frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q},\qquad \xi[d]~\triangleq~\frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}\Phi_{p,q}^*+\Phi_{p,q+d}\Phi_{p,q}^*\Big). \end{align} Let the (hyperplane-reduced) 1D difference attenuation profile be \begin{align}\label{eq:kappa_def_prop} \kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\quad d=1,\ldots,N-1, \end{align} where $\Delta_S(d)$ is the scalar difference-domain reduction induced by $S(\cdot)$ on the single-hyperplane configurations (an explicit reduction is given in Appendix~\ref{app:proof_prop_Sk}). Define \begin{align} \tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\quad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad \tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\quad \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d], \end{align} and stack them into $\tilde{\boldsymbol\kappa}_{\mathrm{lr}},\tilde{\boldsymbol\kappa}_{\mathrm{diag}}\in\mathbb C^{N}$. With \begin{align} \mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}},\qquad \mathbf s_{\mathrm{diag}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}, \end{align} the structured hyperplane (``sp'') term admits the compact Hadamard form \begin{align}\label{eq:Qsp_prop} \mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi) = \frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big) +\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big). \end{align} Finally, the bulk complement term admits the $k$-modulated Hadamard form \begin{align}\label{eq:Qbulk_prop} \mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k =\frac{1}{N^2}\big(\mathbf d_k\mathbf d_k^H\big)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi), \end{align} where $\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\in\mathbb C^{N\times N}$ is a $k$-independent Hermitian matrix induced by the bulk kernel after projecting onto the circulant subspace (see Appendix~\ref{app:proof_prop_Sk}). In particular, when $\mathbf\Phi\equiv \mathbf 1$, one has $\eta_0=\eta[d]=\xi[d]=1$ and hence $\mathbf s_{\mathrm{lr}}=\mathbf s_{\mathrm{diag}}=\mathbf F\tilde{\boldsymbol\kappa}$, recovering the phase-free single-DFT specialization. \end{proposition} \begin{IEEEproof} Please refer to Appendix~\ref{app:proof_prop_Sk}. \end{IEEEproof} % ========================= % Appendix: detailed proof (no itemized points) % ========================= \appendices \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We start from the phase-coded kernel factor $e^{j\tilde\phi_{p,q,p',q'}}\exp[-\alpha S(p,q,p',q')]$ with $e^{j\tilde\phi_{p,q,p',q'}}=\Phi_{p,q}\Phi_{p',q'}^*$ and the explicit $S(\cdot)$ given in the proposition. Following the standard ``pairing vs.\ non-pairing'' separation for the squared-ACF expansion, the perfect-pairing configurations yield a coherent contribution that depends only on the DFT-bin projection $\mathbf f_k^H\mathbf g$. Using the orthonormality of $\{\mathbf f_\ell\}$ and the pairing count, the coherent layer reduces to $\big(2-\frac{1}{N}\big)\lvert \mathbf f_k^H\mathbf g\rvert^2$, which establishes the first term in \eqref{eq:Sk_prop_main}. It remains to characterize the non-pairing part, which we further decompose into four structured hyperplanes and a bulk complement, namely $\text{(H1)}~p'=q'$, $\text{(H2)}~p=q$, $\text{(H3)}~q=q'$, $\text{(H4)}~p=p'$, and the remaining index combinations. The central simplification is that, on any single-hyperplane configuration, the exponent $S(p,q,p',q')$ collapses to a function of a single modular difference $d\in\mathbb Z_N$. This follows from the shift-invariance of $\Psi(\cdot)$: each term in $S(\cdot)$ is of the form $\Psi(\text{difference})$, hence imposing one linear modular constraint (such as $p'=q'$) eliminates all but one independent difference variable in the bracket. After simplification, the resulting scalar exponent can be written as $\Delta_S(d)$, which is common across the structured hyperplanes up to additive constants that do not depend on the free indices and thus only scale the overall baseline uniformly. A convenient explicit reduction is obtained by evaluating the net difference-dependent part as \begin{align}\label{eq:DeltaS_explicit} \Delta_S(d)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N, \end{align} where $\Psi(0)=N_d$ and $\Psi(d)\le \Psi(0)$ for all $d$, implying $\Delta_S(d)\ge 0$. Substituting \eqref{eq:DeltaS_explicit} into \eqref{eq:kappa_def_prop} yields the real, nonnegative 1D attenuation profile $\kappa[d]$ used by all structured layers. To exclude the degenerate zero-difference configurations already accounted for by the pairing layer and by the hyperplane bookkeeping, we set $\kappa[0]=0$ and only retain $d=1,\ldots,N-1$ in the following sums. We next show how the phase matrix $\mathbf\Phi$ enters the hyperplane sums without introducing any new high-dimensional objects. On (H1), $p'=q'=t$ and the phase factor becomes $\Phi_{p,q}\Phi_{t,t}^*$. Summing over $t$ produces $N\eta_0^*$, where $\eta_0=\frac{1}{N}\sum_t\Phi_{t,t}$ is the mean of the main diagonal of $\mathbf\Phi$. The remaining dependence on $(p,q)$ enters only through the single difference $d=p-q$, and averaging along the $d$th offset diagonal gives $\eta[d]=\frac{1}{N}\sum_q\Phi_{q+d,q}$. Consequently, the effective 1D weight for the (H1) low-rank layer is $\kappa[d]\eta[d]\eta_0^*$. On (H2), $p=q$ and the phase factor is again separable into a main-diagonal term and an offset-diagonal term after the same reduction to the single difference, leading to the same effective coefficient. Therefore, both low-rank hyperplanes share the same phase-corrected 1D profile $\tilde\kappa_{\mathrm{lr}}[d]=\kappa[d]\eta[d]\eta_0^*$. To express the resulting low-rank contribution in a DFT-column form, we invoke the normalized DFT identities (indices modulo $N$) \begin{align}\label{eq:DFT_hadamard_id} \mathbf f_a^*\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{b-a}, \qquad \mathbf f_a\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{a+b}. \end{align} Let $\tilde{\boldsymbol\kappa}_{\mathrm{lr}}=[\tilde\kappa_{\mathrm{lr}}[0],\ldots,\tilde\kappa_{\mathrm{lr}}[N-1]]^T$ and define $\mathbf s_{\mathrm{lr}}=\mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}}=\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_d$. The two hyperplanes (H1) and (H2) generate two rank-$\le 2$ matrices before Hermitian symmetrization, which can be written in the same DFT-column notation as in the phase-free case but with $\tilde\kappa_{\mathrm{lr}}[d]$: \begin{align} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{-k}^{*}\mathbf f_{k+d}^{H},\qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{k}^{*}\mathbf f_{d-k}^{H}. \end{align} Using $\mathbf f_{-k}^*=\mathbf f_k$ and $\mathbf f_k^*=\mathbf f_{-k}$, together with \eqref{eq:DFT_hadamard_id}, we obtain \begin{align}\label{eq:sum_shift_to_hadamard} \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d+k} =\sqrt N\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}}),\qquad \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d-k} =\sqrt N\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}}), \end{align} and hence \begin{align}\label{eq:Q0_lr_forms} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_k\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}})^H,\qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_{-k}\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}})^H. \end{align} Define the Hermitianized versions $\mathbf Q_{k,ut}(\mathbf\Phi)\triangleq \frac12(\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)+\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)^H)$ and similarly for $\mathbf Q_{k,v0}(\mathbf\Phi)$. To convert \eqref{eq:Q0_lr_forms} into a Hadamard form with the lifting matrix, we use the elementwise identity \begin{align}\label{eq:outer_to_hadamard} \mathbf a(\mathbf b\odot \mathbf c)^H = (\mathbf a\mathbf b^H)\odot (\mathbf 1\mathbf c^H), \end{align} which follows from $[\mathbf a(\mathbf b\odot \mathbf c)^H]_{n,m}=a_n(b_m c_m)^*=a_n b_m^* c_m^*=[\mathbf a\mathbf b^H]_{n,m}[\mathbf 1\mathbf c^H]_{n,m}$. Applying \eqref{eq:outer_to_hadamard} to \eqref{eq:Q0_lr_forms} and collecting the four symmetric terms yields \begin{align} \mathbf Q_{k,ut}(\mathbf\Phi)+\mathbf Q_{k,v0}(\mathbf\Phi) =\frac{\sqrt N}{2N}\Big(\mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H\Big)\odot \Big(\mathbf 1\mathbf s_{\mathrm{lr}}^H+\mathbf s_{\mathrm{lr}}\mathbf 1^T\Big) =\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big). \end{align} We now turn to the diagonal hyperplanes (H3) and (H4). On (H3), $q=q'=t$ and the phase factor becomes $\Phi_{p,t}\Phi_{p',t}^*$, which depends on the single difference $d=p-p'$ after averaging over $t$; this produces the 1D cyclic correlation along the $p$-direction. On (H4), $p=p'=t$ and the phase factor becomes $\Phi_{t,q}\Phi_{t,q'}^*$, yielding the analogous cyclic correlation along the $q$-direction. To keep the same Hermitian structure as in the phase-free case while using only 1D objects, we combine the two correlations via $\xi[d]=\frac{1}{2N^2}\sum_{p,q}(\Phi_{p+d,q}\Phi_{p,q}^*+\Phi_{p,q+d}\Phi_{p,q}^*)$ and define the diagonal-layer profile $\tilde\kappa_{\mathrm{diag}}[d]=\kappa[d]\xi[d]$, with $\tilde\kappa_{\mathrm{diag}}[0]=0$. Let $\mathbf s_{\mathrm{diag}}=\mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}$. The diagonal contribution can be written as a sum of diagonal modulation operators: \begin{align}\label{eq:Qdiag_sum} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(\mathbf D_d+\mathbf D_{-d}\big), \end{align} whose $n$th diagonal entry is \begin{align}\label{eq:Qdiag_entries} [\mathbf Q_{\mathrm{diag}}(\mathbf\Phi)]_{n,n} =\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(e^{-j\omega d(n-1)}+e^{j\omega d(n-1)}\big) =\frac{2}{N^2}\Re\!\left\{\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]e^{-j\omega d(n-1)}\right\} =\frac{2}{N^{3/2}}\Re\{(\mathbf s_{\mathrm{diag}})_n\}. \end{align} Since $[\mathbf S(\mathbf s_{\mathrm{diag}})]_{n,n}=(\mathbf s_{\mathrm{diag}})_n+(\mathbf s_{\mathrm{diag}})_n^*=2\Re\{(\mathbf s_{\mathrm{diag}})_n\}$, \eqref{eq:Qdiag_entries} is equivalently expressed as \begin{align} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re\{\mathbf s_{\mathrm{diag}}\}\big) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big). \end{align} Combining this diagonal layer with the low-rank layer derived earlier gives \eqref{eq:Qsp_prop} and thus establishes the structured hyperplane contribution in \eqref{eq:Sk_prop_main}. Finally, we derive the bulk complement. Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ so that $\gamma_\ell=\mathbf f_\ell^H\mathbf g$, and define the $N\times N$ matrix $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. Because $\mathbf g$ is real-valued, $\gamma_{\langle-\ell\rangle}=\gamma_{\langle\ell\rangle}^*$ and the bulk term can be written as a Hermitian quadratic form \begin{align}\label{eq:bulk_vec_form} \mathcal S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)$ is the bulk-masked, phase-coded kernel in the $(p,q)$-lifted space: \begin{align}\label{eq:Mbulk_entries} [\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)]_{(p,q),(p',q')} =\mathrm{Herm}\Big\{\Phi_{p,q}\Phi_{p',q'}^*\exp[-\alpha S(p,q,p',q')]\,[\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')}\Big\}. \end{align} A key property is that $\mathbf H_k$ lies in the circulant subspace: letting $\mathbf c_k\in\mathbb C^N$ with $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$, one has $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$. Using the standard circulant shift basis $\{\mathbf J_d\}_{d=0}^{N-1}$ with $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$, define the (fixed) circulant synthesis operator \begin{align} \mathbf B\triangleq [\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]\in\mathbb R^{N^2\times N}, \end{align} so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting into \eqref{eq:bulk_vec_form} yields the projected $N\times N$ kernel \begin{align}\label{eq:Rbulk_def} \mathcal S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathbf B. \end{align} To return to the $\mathbf g$-domain and isolate the $k$-dependence, note that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$; writing $\mathbf c_k=\mathbf\Pi_k\boldsymbol\gamma$ with the permutation $\mathbf\Pi_k$, and using the modulation identity $\mathbf\Pi_k=\mathbf F^H\mathbf D_k\mathbf F$, we obtain \begin{align} \mathcal S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\boldsymbol\gamma^H\mathbf\Pi_k^H\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf\Pi_k\boldsymbol\gamma =\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k\Big)\mathbf g, \qquad \mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf F^H. \end{align} The Hadamard-modulated form in \eqref{eq:Qbulk_prop} follows from the elementwise identity $\mathbf D_{-k}\mathbf X\mathbf D_k=(\mathbf d_k\mathbf d_k^H)\odot \mathbf X$. Collecting the coherent pairing term, the structured hyperplane term \eqref{eq:Qsp_prop}, and the bulk term \eqref{eq:Qbulk_prop} establishes \eqref{eq:Sk_prop_main} and completes the proof.
\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big)
+\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big).
\end{align}
\begin{align}\label{eq:Qbulk_prop}
\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)
=\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k
=\frac{1}{N^2}\big(\mathbf d_k\mathbf d_k^H\big)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi),
\end{align}
\end{proposition}
latex% ========================= % Main text: Proposition (as requested) % ========================= \begin{proposition}\label{prop:Sk_closed_form} Consider the transmission of $N$ random CE-OFDM symbols, the contribution to the squared ACF arising from the non-vanishing-kernel index configurations could be expressed as: \begin{align} S_k(\mathbf{g}) =\big\|(\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)+\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi))^{1/2}\mathbf g\big\|_2^2, \end{align} where \begin{align}\label{eq:Qsp_prop} \mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi) = \frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big) +\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big), \end{align} and \begin{align}\label{eq:Qbulk_prop} \mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k =\frac{1}{N^2}\big(\mathbf d_k\mathbf d_k^H\big)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi). \end{align} \end{proposition} \begin{IEEEproof} Please refer to Appendix~\ref{app:proof_prop_Sk}. \end{IEEEproof} % ========================= % Appendix: detailed proof (continuous TSP-style narrative; minimal new matrices) % ========================= \appendices \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)\ell}$, $\omega\triangleq 2\pi/N$, and denote by $\mathbf f_\ell$ its $\ell$th column (indices are taken modulo $N$). Define the modulation vector and matrix $\mathbf d_k\triangleq [e^{j\omega k(n-1)}]_{n=1}^N$ and $\mathbf D_k\triangleq \mathrm{diag}(e^{-j\omega k(n-1)})_{n=1}^N$. The designable phase is encoded by $\Phi_{p,q}\triangleq e^{j\phi_{p,q}}$, yielding $e^{j\tilde\phi_{p,q,p',q'}}=\Phi_{p,q}\Phi_{p',q'}^*$. The kernel exponent is specified by \begin{align}\label{eq:S_def_app} \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta),\qquad S(p,q,p',q')\triangleq(4N_d+2)\Big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\Big], \end{align} and we write $\alpha\triangleq \tilde h^2/N^2$. \subsection*{A) Hyperplane decomposition and 1D reduction of the exponent} The non-vanishing-kernel configurations are partitioned into four structured single-hyperplane families and a residual bulk complement: (H1) $p'=q'$, (H2) $p=q$, (H3) $q=q'$, (H4) $p=p'$, and the remaining index tuples (bulk). On each single-hyperplane family, the expression in \eqref{eq:S_def_app} reduces to a function of a single modular difference variable $d\in\mathbb Z_N$ because every $\Psi(\cdot)$ term depends only on a difference of its arguments. A convenient representative reduction is obtained by evaluating the difference-dependent part through the canonical slice $(p,q,p',q')=(d,0,0,0)$, which yields \begin{align}\label{eq:DeltaS_app} \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N, \end{align} where $\Psi(0)=\sum_{r=1}^{N_d}1=N_d$. Accordingly, all structured hyperplane layers share the same 1D attenuation profile \begin{align}\label{eq:kappa_app} \kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\qquad d=1,\ldots,N-1, \end{align} where $\kappa[0]=0$ is adopted to exclude the degenerate zero-difference configurations already handled by the bookkeeping of the index sets. \subsection*{B) DFT identities used throughout} The subsequent matrix forms rely on the elementwise product identities of normalized DFT columns (indices modulo $N$): \begin{align}\label{eq:DFT_id_app} \mathbf f_a^*\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{b-a}, \qquad \mathbf f_a\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{a+b}. \end{align} In addition, we repeatedly use the lifting operator \begin{align}\label{eq:S_lift_app} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T, \end{align} which is Hermitian for any $\mathbf s\in\mathbb C^N$ since $[\mathbf S(\mathbf s)]_{n,m}=s_m^*+s_n$ implies $[\mathbf S(\mathbf s)]_{n,m}=[\mathbf S(\mathbf s)]_{m,n}^*$. \subsection*{C) Derivation of $\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)$} We next show that the four single-hyperplane families produce a diagonal layer plus a Hermitianized low-rank layer, both admitting a compact Hadamard representation as in \eqref{eq:Qsp_prop}. The phase factor $\Phi_{p,q}\Phi_{p',q'}^*$ collapses to 1D aggregates under the same single-difference reduction. Specifically, for (H1) with $p'=q'=t$, the phase becomes $\Phi_{p,q}\Phi_{t,t}^*$; summation over $t$ yields the main-diagonal mean \begin{align}\label{eq:eta0_app} \eta_0\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}, \end{align} while the remaining dependence on $(p,q)$ under the single-difference parameterization $d=p-q$ is captured by the offset-diagonal mean \begin{align}\label{eq:eta_app} \eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}. \end{align} The same 1D collapse occurs for (H2) with $p=q$, leading to an identical effective phase weight. Consequently, the two low-rank hyperplanes share the phase-corrected 1D profile \begin{align}\label{eq:kappa_lr_app} \tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad d=1,\ldots,N-1, \end{align} and we define the associated single DFT vector \begin{align}\label{eq:s_lr_app} \mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}} =\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_d. \end{align} Before Hermitian symmetrization, the two low-rank contributions can be written (in the same DFT-column form as the phase-free case) as \begin{align}\label{eq:Q0_lr_app} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{-k}^{*}\mathbf f_{k+d}^{H}, \qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{k}^{*}\mathbf f_{d-k}^{H}. \end{align} Using $\mathbf f_{-k}^*=\mathbf f_k$, $\mathbf f_k^*=\mathbf f_{-k}$, and \eqref{eq:DFT_id_app}, we obtain the key shift-to-Hadamard relations \begin{align}\label{eq:sum_to_hadamard_app} \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d+k} =\sqrt N\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}}), \qquad \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d-k} =\sqrt N\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}}), \end{align} and hence \begin{align}\label{eq:Q0_lr_compact_app} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_k\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}})^H, \qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_{-k}\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}})^H. \end{align} Let $\mathbf Q_{k,ut}(\mathbf\Phi)\triangleq \frac12(\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)+\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)^H)$ and similarly for $\mathbf Q_{k,v0}(\mathbf\Phi)$. Converting \eqref{eq:Q0_lr_compact_app} into a Hadamard form uses the elementwise identity $\mathbf a(\mathbf b\odot\mathbf c)^H=(\mathbf a\mathbf b^H)\odot(\mathbf 1\mathbf c^H)$, which follows by entrywise inspection. Collecting the four symmetric terms then yields \begin{align}\label{eq:lr_hadamard_app} \mathbf Q_{k,ut}(\mathbf\Phi)+\mathbf Q_{k,v0}(\mathbf\Phi) =\frac{\sqrt N}{2N}\Big(\mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H\Big)\odot\Big(\mathbf 1\mathbf s_{\mathrm{lr}}^H+\mathbf s_{\mathrm{lr}}\mathbf 1^T\Big). \end{align} Defining $\mathbf A_k\triangleq \mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H$ and using \eqref{eq:S_lift_app} gives the low-rank term in \eqref{eq:Qsp_prop}. We now consider the diagonal hyperplanes (H3) and (H4). On (H3), $q=q'=t$ and the phase factor becomes $\Phi_{p,t}\Phi_{p',t}^*$, which depends on the single difference $d=p-p'$ after averaging over $t$; on (H4), $p=p'=t$ and the phase becomes $\Phi_{t,q}\Phi_{t,q'}^*$, depending on $d=q-q'$ after averaging over $t$. Using only 1D objects while preserving the same Hermitian symmetry, we combine the two directional cyclic correlations via \begin{align}\label{eq:xi_app} \xi[d]\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}\Phi_{p,q}^*+\Phi_{p,q+d}\Phi_{p,q}^*\Big), \end{align} and define the diagonal-layer profile \begin{align}\label{eq:kappa_diag_app} \tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d],\qquad d=1,\ldots,N-1, \end{align} together with $\mathbf s_{\mathrm{diag}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}$. The resulting diagonal contribution can be expressed using only $\mathbf D_d$ and the 1D weights as \begin{align}\label{eq:Qdiag_sum_app} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(\mathbf D_d+\mathbf D_{-d}\big). \end{align} Its $n$th diagonal entry equals \begin{align}\label{eq:Qdiag_entry_app} [\mathbf Q_{\mathrm{diag}}(\mathbf\Phi)]_{n,n} =\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(e^{-j\omega d(n-1)}+e^{j\omega d(n-1)}\big) =\frac{2}{N^{3/2}}\Re\{(\mathbf s_{\mathrm{diag}})_n\}. \end{align} Since $[\mathbf S(\mathbf s_{\mathrm{diag}})]_{n,n}=2\Re\{(\mathbf s_{\mathrm{diag}})_n\}$, we obtain \begin{align}\label{eq:Qdiag_hadamard_app} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{2}{N^{3/2}}\mathrm{diag}\!\big(\Re\{\mathbf s_{\mathrm{diag}}\}\big) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big), \end{align} which is the diagonal term in \eqref{eq:Qsp_prop}. Summing \eqref{eq:Qdiag_hadamard_app} and \eqref{eq:lr_hadamard_app} yields \eqref{eq:Qsp_prop}. \subsection*{D) Derivation of $\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)$ and the $k$-modulated Hadamard form} Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define the $N\times N$ matrix \begin{align}\label{eq:Hk_def_app} [\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}. \end{align} The bulk complement contribution can be written as a Hermitian quadratic form in $\mathrm{vec}(\mathbf H_k)$: \begin{align}\label{eq:bulk_vec_app} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)$ is the bulk-masked, phase-coded kernel in the $(p,q)$-lifted domain, i.e., \begin{align}\label{eq:Mbulk_entry_app} [\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)]_{(p,q),(p',q')} =\mathrm{Herm}\Big\{\Phi_{p,q}\Phi_{p',q'}^*\exp[-\alpha S(p,q,p',q')]\,[\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')}\Big\}. \end{align} A key observation is that $\mathbf H_k$ lies in the circulant subspace. Indeed, define $\mathbf c_k\in\mathbb C^{N}$ by $(\mathbf c_k)[d]\triangleq \gamma_{\langle k+d\rangle}$; then \eqref{eq:Hk_def_app} implies $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$. Using the standard circulant shift basis matrices $\{\mathbf J_d\}_{d=0}^{N-1}$ defined by $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$, any circulant matrix can be expanded as $\mathrm{circ}(\mathbf c_k)=\sum_{d=0}^{N-1}(\mathbf c_k)[d]\mathbf J_d$, and thus $\mathrm{vec}(\mathbf H_k)$ belongs to $\mathrm{span}\{\mathrm{vec}(\mathbf J_d)\}_{d=0}^{N-1}$. Introducing the fixed synthesis operator \begin{align}\label{eq:B_def_app} \mathbf B\triangleq [\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]\in\mathbb R^{N^2\times N}, \end{align} we have $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting into \eqref{eq:bulk_vec_app} yields the projected $N\times N$ kernel \begin{align}\label{eq:Rbulk_app} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathbf B. \end{align} To return to the $\mathbf g$-domain and isolate the $k$-dependence, note that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$, i.e., $\mathbf c_k=\mathbf\Pi_k\boldsymbol\gamma$ for a permutation $\mathbf\Pi_k$. The DFT modulation identity gives $\mathbf\Pi_k=\mathbf F^H\mathbf D_k\mathbf F$, and hence \begin{align}\label{eq:bulk_back_to_g_app} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\boldsymbol\gamma^H\mathbf\Pi_k^H\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf\Pi_k\boldsymbol\gamma =\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k\Big)\mathbf g, \qquad \mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf F^H. \end{align} Finally, the identity $\mathbf D_{-k}\mathbf X\mathbf D_k=(\mathbf d_k\mathbf d_k^H)\odot \mathbf X$ (verified entrywise) yields the Hadamard-modulated form in \eqref{eq:Qbulk_prop}. Combining the structured hyperplane contribution $\mathbf g^H\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)\mathbf g$ and the bulk complement contribution $\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)\mathbf g$ gives $S_k(\mathbf g)$ in Proposition~\ref{prop:Sk_closed_form}. \qed
这个d_k可以用DFT矩阵的列\mathbf{f}来表示么?
可以,而且写成 DFT 列向量会更“TSP”。
用你们的归一化 DFT 定义
则第 列
你们的
正好是“未归一化、共轭的 DFT 列”:
因为
顺带两个常用等价写法(如果你们想把 也统一成 DFT 符号):
因此你们命题里的 Hadamard 调制项也可写成
如果你愿意,我也可以把你们正文/appendix里所有出现 的地方统一替换成 (或 )并顺手把常数因子合并到 前面,让式子更干净。
把你们正文/appendix里所有出现
d
k
d
k
的地方统一替换成
N
f
−
k
N
f
−k
(或
N
f
k
∗
N
f
k
∗
)并顺手把常数因子合并到
1
N
2
N
2
1
前面,让式子更干净
latex% ========================= % Main text: Proposition (d_k eliminated; constants merged) % ========================= \begin{proposition}\label{prop:Sk_closed_form} Consider the transmission of $N$ random CE-OFDM symbols, the contribution to the squared ACF arising from the non-vanishing-kernel index configurations could be expressed as: \begin{align} S_k(\mathbf{g}) =\big\|(\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)+\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi))^{1/2}\mathbf g\big\|_2^2, \end{align} where \begin{align}\label{eq:Qsp_prop} \mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi) = \frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big) +\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big), \end{align} and \begin{align}\label{eq:Qbulk_prop_clean} \mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k =\frac{1}{N}\big(\mathbf f_{-k}\mathbf f_{-k}^{H}\big)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi). \end{align} \end{proposition} \begin{IEEEproof} Please refer to Appendix~\ref{app:proof_prop_Sk}. \end{IEEEproof} % ========================= % Appendix (updated): Proof with d_k removed everywhere % ========================= \appendices \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)\ell}$, $\omega\triangleq 2\pi/N$, and denote by $\mathbf f_\ell$ its $\ell$th column (indices are taken modulo $N$). Define the diagonal modulation operator \begin{align} \mathbf D_k\triangleq \mathrm{diag}(e^{-j\omega k(n-1)})_{n=1}^{N}. \end{align} The designable phase is encoded by $\Phi_{p,q}\triangleq e^{j\phi_{p,q}}$, yielding $e^{j\tilde\phi_{p,q,p',q'}}=\Phi_{p,q}\Phi_{p',q'}^*$. The kernel exponent is specified by \begin{align}\label{eq:S_def_app} \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta),\qquad S(p,q,p',q')\triangleq(4N_d+2)\Big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\Big], \end{align} and we write $\alpha\triangleq \tilde h^2/N^2$. The non-vanishing-kernel configurations are partitioned into four structured single-hyperplane families and a residual bulk complement: (H1) $p'=q'$, (H2) $p=q$, (H3) $q=q'$, (H4) $p=p'$, and the remaining index tuples (bulk). On each single-hyperplane family, the expression in \eqref{eq:S_def_app} reduces to a function of a single modular difference $d\in\mathbb Z_N$ because every $\Psi(\cdot)$ term depends only on a difference of its arguments. A convenient representative reduction is obtained by evaluating the canonical slice $(p,q,p',q')=(d,0,0,0)$, which yields \begin{align}\label{eq:DeltaS_app} \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N, \end{align} where $\Psi(0)=\sum_{r=1}^{N_d}1=N_d$. Accordingly, all structured hyperplane layers share the same 1D attenuation profile \begin{align}\label{eq:kappa_app} \kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\qquad d=1,\ldots,N-1, \end{align} where $\kappa[0]=0$ excludes degenerate zero-difference configurations as per the index-set bookkeeping. The subsequent matrix forms rely on the elementwise product identities of normalized DFT columns (indices modulo $N$): \begin{align}\label{eq:DFT_id_app} \mathbf f_a^*\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{b-a}, \qquad \mathbf f_a\odot \mathbf f_b=\frac{1}{\sqrt N}\mathbf f_{a+b}. \end{align} We also use the lifting operator \begin{align}\label{eq:S_lift_app} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T . \end{align} We next show that the four single-hyperplane families produce a diagonal layer plus a Hermitianized low-rank layer, both admitting the compact Hadamard representation in \eqref{eq:Qsp_prop}. The phase factor $\Phi_{p,q}\Phi_{p',q'}^*$ collapses to 1D aggregates under the same single-difference reduction. For (H1) with $p'=q'=t$, the phase becomes $\Phi_{p,q}\Phi_{t,t}^*$; summation over $t$ yields the main-diagonal mean \begin{align}\label{eq:eta0_app} \eta_0\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}, \end{align} while the remaining dependence on $(p,q)$ under the single-difference parameterization $d=p-q$ is captured by the offset-diagonal mean \begin{align}\label{eq:eta_app} \eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}. \end{align} The same 1D collapse occurs for (H2) with $p=q$, leading to an identical effective phase weight. Consequently, the two low-rank hyperplanes share the phase-corrected 1D profile \begin{align}\label{eq:kappa_lr_app} \tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad d=1,\ldots,N-1, \end{align} and we define \begin{align}\label{eq:s_lr_app} \mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}} =\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_d. \end{align} Before Hermitian symmetrization, the two low-rank contributions can be written as \begin{align}\label{eq:Q0_lr_app} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{-k}^{*}\mathbf f_{k+d}^{H}, \qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\ \mathbf f_{k}^{*}\mathbf f_{d-k}^{H}. \end{align} Using $\mathbf f_{-k}^*=\mathbf f_k$, $\mathbf f_k^*=\mathbf f_{-k}$, and \eqref{eq:DFT_id_app}, we obtain \begin{align}\label{eq:sum_to_hadamard_app} \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d+k} =\sqrt N\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}}), \qquad \sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d-k} =\sqrt N\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}}). \end{align} Hence, \begin{align}\label{eq:Q0_lr_compact_app} \mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_k\,(\mathbf f_{-k}\odot \mathbf s_{\mathrm{lr}})^H, \qquad \mathbf Q_{k,v0}^{(0)}(\mathbf\Phi)=\frac{\sqrt N}{N}\,\mathbf f_{-k}\,(\mathbf f_{k}\odot \mathbf s_{\mathrm{lr}})^H. \end{align} Let $\mathbf Q_{k,ut}(\mathbf\Phi)\triangleq \frac12(\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)+\mathbf Q_{k,ut}^{(0)}(\mathbf\Phi)^H)$ and similarly for $\mathbf Q_{k,v0}(\mathbf\Phi)$. Using the entrywise identity $\mathbf a(\mathbf b\odot\mathbf c)^H=(\mathbf a\mathbf b^H)\odot(\mathbf 1\mathbf c^H)$ and collecting the symmetric terms yields \begin{align}\label{eq:lr_hadamard_app} \mathbf Q_{k,ut}(\mathbf\Phi)+\mathbf Q_{k,v0}(\mathbf\Phi) =\frac{\sqrt N}{2N}\Big(\mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H\Big)\odot\Big(\mathbf 1\mathbf s_{\mathrm{lr}}^H+\mathbf s_{\mathrm{lr}}\mathbf 1^T\Big). \end{align} Defining $\mathbf A_k\triangleq \mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H$ and using \eqref{eq:S_lift_app} gives the low-rank term in \eqref{eq:Qsp_prop}. For the diagonal hyperplanes (H3) and (H4), the phase factor becomes a cyclic correlation along one index direction after averaging over the constrained index. To preserve the same Hermitian structure while using only 1D objects, we combine the two directional correlations via \begin{align}\label{eq:xi_app} \xi[d]\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}\Phi_{p,q}^*+\Phi_{p,q+d}\Phi_{p,q}^*\Big), \end{align} and define \begin{align}\label{eq:kappa_diag_app} \tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d],\qquad d=1,\ldots,N-1, \end{align} together with $\mathbf s_{\mathrm{diag}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}$. The diagonal contribution can be expressed using only $\mathbf D_d$ and the 1D weights as \begin{align}\label{eq:Qdiag_sum_app} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(\mathbf D_d+\mathbf D_{-d}\big), \end{align} whose $n$th diagonal entry equals \begin{align}\label{eq:Qdiag_entry_app} [\mathbf Q_{\mathrm{diag}}(\mathbf\Phi)]_{n,n} =\frac{2}{N^{3/2}}\Re\{(\mathbf s_{\mathrm{diag}})_n\}. \end{align} Since $[\mathbf S(\mathbf s_{\mathrm{diag}})]_{n,n}=2\Re\{(\mathbf s_{\mathrm{diag}})_n\}$, we obtain \begin{align}\label{eq:Qdiag_hadamard_app} \mathbf Q_{\mathrm{diag}}(\mathbf\Phi)=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big), \end{align} which is the diagonal term in \eqref{eq:Qsp_prop}. Summing \eqref{eq:Qdiag_hadamard_app} and \eqref{eq:lr_hadamard_app} yields \eqref{eq:Qsp_prop}. Finally, we derive the bulk complement and obtain \eqref{eq:Qbulk_prop_clean}. Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. The bulk contribution can be written as \begin{align}\label{eq:bulk_vec_app} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)$ contains the bulk mask and the phase-coded kernel with entries \begin{align}\label{eq:Mbulk_entry_app} [\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)]_{(p,q),(p',q')} =\mathrm{Herm}\Big\{\Phi_{p,q}\Phi_{p',q'}^*\exp[-\alpha S(p,q,p',q')]\,[\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')}\Big\}. \end{align} Since $\mathbf H_k$ is circulant, there exists $\mathbf c_k\in\mathbb C^{N}$ such that $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$ with $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$, and therefore $\mathrm{vec}(\mathbf H_k)$ lies in the span of $\{\mathrm{vec}(\mathbf J_d)\}_{d=0}^{N-1}$, where $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$. Let \begin{align} \mathbf B\triangleq [\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]\in\mathbb R^{N^2\times N}, \end{align} so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting into \eqref{eq:bulk_vec_app} yields \begin{align}\label{eq:Rbulk_app} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\mathbf\Phi)\mathbf B. \end{align} Let $\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}(\mathbf\Phi)\mathbf F^H$. Using that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$ and the modulation identity $\mathbf\Pi_k=\mathbf F^H\mathbf D_k\mathbf F$, we obtain \begin{align}\label{eq:bulk_back_to_g_app} S_k^{\mathrm{bulk}}(\mathbf g) =\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\mathbf\Phi)\mathbf D_k\Big)\mathbf g. \end{align} To eliminate $\mathbf d_k$ in the Hadamard-modulation step, note that \begin{align}\label{eq:D_as_DFTcol_app} \mathbf D_{-k}=\mathrm{diag}\big(e^{+j\omega k(n-1)}\big)_{n=1}^N=\mathrm{diag}\big(\sqrt N\,\mathbf f_{-k}\big),\qquad \mathbf D_{k}=\mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^N=\mathrm{diag}\big(\sqrt N\,\mathbf f_{k}\big). \end{align} Therefore, for any $\mathbf X\in\mathbb C^{N\times N}$, \begin{align}\label{eq:D_hadamard_f_app} \mathbf D_{-k}\mathbf X\mathbf D_k = N\big(\mathbf f_{-k}\mathbf f_{-k}^H\big)\odot \mathbf X, \end{align} which follows entrywise from $[\mathbf D_{-k}\mathbf X\mathbf D_k]_{n,m}=(\sqrt N f_{-k}[n])\,X_{n,m}\,(\sqrt N f_k[m]) = N\,f_{-k}[n]\,f_{-k}[m]^*\,X_{n,m}$. Applying \eqref{eq:D_hadamard_f_app} to \eqref{eq:bulk_back_to_g_app} immediately gives \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)=\frac{1}{N}\big(\mathbf f_{-k}\mathbf f_{-k}^H\big)\odot \mathbf K_{\mathrm{bulk}}(\mathbf\Phi), \end{align} which is exactly \eqref{eq:Qbulk_prop_clean}. Combining the structured hyperplane contribution $\mathbf g^H\mathbf Q_k^{\mathrm{sp}}(\mathbf\Phi)\mathbf g$ and the bulk complement contribution $\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}(\mathbf\Phi)\mathbf g$ yields $S_k(\mathbf g)$ in Proposition~\ref{prop:Sk_closed_form}, completing the proof.
如果你更偏好用 也可以:因为 (归一化 DFT 列、模 索引),于是
二者完全等价。
我们写的标准是\begin{align}\label{eq:Qbulk_prop}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})
=\frac{1}{N}\big(\mathbf{f}{k+1}^*\mathbf{f}{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}),
\end{align},因为我们的DFT是从1~N这么写的,都更新一下,按照我的偏好来写
下面把正文 + Appendix里所有与调制向量相关的写法,统一改成你偏好的 1-based DFT 列记法,并把常数合并到 前面,最终得到你给的标准形式
(这里 ,列下标 。)
下面是整段优化后的英文 LaTeX 代码(已按你的偏好更新:不再出现 ,并且所有地方都用 ):
latex% ========================= % Main text: Proposition (your preferred Qbulk form) % ========================= \begin{proposition}\label{prop:Sk_closed_form} Consider the transmission of $N$ random CE-OFDM symbols, the contribution to the squared ACF arising from the non-vanishing-kernel index configurations could be expressed as: \begin{align} S_k(\mathbf{g}) =\big\|(\mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta})+\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta}))^{1/2}\mathbf g\big\|_2^2, \end{align} where \begin{align}\label{eq:Qsp_prop} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta}) = \frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big) +\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big), \end{align} and \begin{align}\label{eq:Qbulk_prop} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta}) =\frac{1}{N}\big(\mathbf{f}_{k+1}^*\mathbf{f}_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}). \end{align} \end{proposition} \begin{IEEEproof} Please refer to Appendix~\ref{app:proof_prop_Sk}. \end{IEEEproof} % ========================= % Appendix: detailed proof (d_k removed; 1-based DFT columns) % ========================= \appendices \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with \begin{align} [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\qquad \omega\triangleq\frac{2\pi}{N}, \end{align} and denote by $\mathbf f_{\ell}$ its $\ell$th column, $\ell=1,\ldots,N$. Hence, the bin index $k\in\{0,\ldots,N-1\}$ corresponds to the column $\mathbf f_{k+1}$. Define the diagonal modulation operator \begin{align}\label{eq:Dk_def_app} \mathbf D_k\triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}. \end{align} The designable phase is encoded by $\Phi_{p,q}(\boldsymbol{\theta})\triangleq e^{j\phi_{p,q}(\boldsymbol{\theta})}$, yielding $e^{j\tilde\phi_{p,q,p',q'}}=\Phi_{p,q}(\boldsymbol{\theta})\Phi_{p',q'}(\boldsymbol{\theta})^*$. The kernel exponent is specified by \begin{align}\label{eq:S_def_app} \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta),\qquad S(p,q,p',q')\triangleq(4N_d+2)\Big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\Big], \end{align} and we write $\alpha\triangleq \tilde h^2/N^2$. The non-vanishing-kernel configurations are partitioned into four structured single-hyperplane families and a residual bulk complement: (H1) $p'=q'$, (H2) $p=q$, (H3) $q=q'$, (H4) $p=p'$, and the remaining index tuples (bulk). On each single-hyperplane family, the expression in \eqref{eq:S_def_app} reduces to a function of a single modular difference $d\in\mathbb Z_N$ because every $\Psi(\cdot)$ term depends only on a difference of its arguments. A convenient representative reduction is obtained by evaluating the canonical slice $(p,q,p',q')=(d,0,0,0)$, which yields \begin{align}\label{eq:DeltaS_app} \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N, \end{align} where $\Psi(0)=\sum_{r=1}^{N_d}1=N_d$. Accordingly, all structured hyperplane layers share the same 1D attenuation profile \begin{align}\label{eq:kappa_app} \kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\qquad d=1,\ldots,N-1. \end{align} The subsequent matrix forms rely on the elementwise product identities of normalized DFT columns (indices modulo $N$ in the frequency argument): \begin{align}\label{eq:DFT_id_app} \mathbf f_{a+1}^*\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(b-a)+1}, \qquad \mathbf f_{a+1}\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(a+b)+1}. \end{align} We also use the lifting operator \begin{align}\label{eq:S_lift_app} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T . \end{align} We next show that the four single-hyperplane families produce a diagonal layer plus a Hermitianized low-rank layer, both admitting the compact Hadamard representation in \eqref{eq:Qsp_prop}. The phase factor $\Phi_{p,q}(\boldsymbol{\theta})\Phi_{p',q'}(\boldsymbol{\theta})^*$ collapses to 1D aggregates under the same single-difference reduction. For (H1) with $p'=q'=t$, the phase becomes $\Phi_{p,q}(\boldsymbol{\theta})\Phi_{t,t}(\boldsymbol{\theta})^*$; summation over $t$ yields the main-diagonal mean \begin{align}\label{eq:eta0_app} \eta_0\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}(\boldsymbol{\theta}), \end{align} while the remaining dependence on $(p,q)$ under the single-difference parameterization $d=p-q$ is captured by the offset-diagonal mean \begin{align}\label{eq:eta_app} \eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}(\boldsymbol{\theta}). \end{align} The same 1D collapse occurs for (H2) with $p=q$, leading to an identical effective phase weight. Consequently, the two low-rank hyperplanes share the phase-corrected 1D profile \begin{align}\label{eq:kappa_lr_app} \tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad d=1,\ldots,N-1, \end{align} and we define \begin{align}\label{eq:s_lr_app} \mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}} =\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d+1}. \end{align} Following the same DFT-column manipulations as in the phase-free case and invoking \eqref{eq:DFT_id_app}, the Hermitianized low-rank contribution can be written as \begin{align}\label{eq:lr_hadamard_app} \mathbf Q_{k,ut}(\boldsymbol{\theta})+\mathbf Q_{k,v0}(\boldsymbol{\theta}) =\frac{\sqrt N}{2N}\,\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}), \end{align} where, under the 1-based column convention, a convenient Hermitian choice is \begin{align}\label{eq:Ak_1based} \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H. \end{align} For the diagonal hyperplanes (H3) and (H4), the phase factor becomes a cyclic correlation along one index direction after averaging over the constrained index. To preserve the Hermitian structure while using only 1D objects, we combine the two directional correlations via \begin{align}\label{eq:xi_app} \xi[d]\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}(\boldsymbol{\theta})\Phi_{p,q}(\boldsymbol{\theta})^*+\Phi_{p,q+d}(\boldsymbol{\theta})\Phi_{p,q}(\boldsymbol{\theta})^*\Big), \end{align} and define \begin{align}\label{eq:kappa_diag_app} \tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d],\qquad d=1,\ldots,N-1, \end{align} together with $\mathbf s_{\mathrm{diag}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}$. As in the phase-free case, the diagonal contribution can be expressed using only $\mathbf D_d$ and the 1D weights: \begin{align}\label{eq:Qdiag_sum_app} \mathbf Q_{\mathrm{diag}}(\boldsymbol{\theta})=\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(\mathbf D_d+\mathbf D_{-d}\big) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big). \end{align} Combining \eqref{eq:lr_hadamard_app} with \eqref{eq:Qdiag_sum_app} yields \eqref{eq:Qsp_prop}. Finally, we derive the bulk complement and obtain \eqref{eq:Qbulk_prop}. Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. The bulk contribution can be written as \begin{align}\label{eq:bulk_vec_app} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})$ contains the bulk mask and the phase-coded kernel with entries \begin{align}\label{eq:Mbulk_entry_app} [\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})]_{(p,q),(p',q')} =\mathrm{Herm}\Big\{\Phi_{p,q}(\boldsymbol{\theta})\Phi_{p',q'}(\boldsymbol{\theta})^*\exp[-\alpha S(p,q,p',q')]\,[\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')}\Big\}. \end{align} Since $\mathbf H_k$ is circulant, there exists $\mathbf c_k\in\mathbb C^{N}$ such that $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$ with $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$. Let $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$ and $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]$, so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting into \eqref{eq:bulk_vec_app} yields \begin{align}\label{eq:Rbulk_app} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\boldsymbol{\theta})\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf B. \end{align} Define $\mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf F^H$. Using that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$ and $\mathbf\Pi_k=\mathbf F^H\mathbf D_k\mathbf F$, we obtain \begin{align}\label{eq:bulk_back_to_g_app} S_k^{\mathrm{bulk}}(\mathbf g) =\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf D_k\Big)\mathbf g. \end{align} Under the 1-based DFT convention, the diagonal modulators satisfy \begin{align}\label{eq:D_as_fkp1_app} \mathbf D_{-k}=\mathrm{diag}\big(e^{+j\omega k(n-1)}\big)_{n=1}^N=\mathrm{diag}\big(\sqrt N\,\mathbf f_{k+1}^*\big),\qquad \mathbf D_{k}=\mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^N=\mathrm{diag}\big(\sqrt N\,\mathbf f_{k+1}\big). \end{align} Therefore, for any $\mathbf X\in\mathbb C^{N\times N}$, \begin{align}\label{eq:D_hadamard_fkp1_app} \mathbf D_{-k}\mathbf X\mathbf D_k = N\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf X, \end{align} which follows entrywise from $[\mathbf D_{-k}\mathbf X\mathbf D_k]_{n,m}=(\sqrt N f_{k+1}^*[n])\,X_{n,m}\,(\sqrt N f_{k+1}[m]) = N\,f_{k+1}^*[n]\,f_{k+1}[m]\,X_{n,m}$. Applying \eqref{eq:D_hadamard_fkp1_app} to \eqref{eq:bulk_back_to_g_app} yields \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta}) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf D_k =\frac{1}{N}\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}), \end{align} which is exactly \eqref{eq:Qbulk_prop}. Combining the structured hyperplane contribution $\mathbf g^H\mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta})\mathbf g$ and the bulk complement contribution $\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf g$ yields $S_k(\mathbf g)$ in Proposition~\ref{prop:Sk_closed_form}, completing the proof.
如果你们的 在正文里仍然想保留 “” 的样子(只不过换成 1-based),也可以把 \eqref{eq:Ak_1based} 改写成
并在文中注明 。但我上面给的 更贴合你现在对 Qbulk 的偏好、也更统一。
\frac{1}{N^{3/2}}\big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big]
+\frac{\sqrt N}{2N}\big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big].
\end{align}
这两个 其实是同一个 lifting 结构,只不过由两条不同的 DFT 向量 和 生成。把它们写成逐元素解析式如下(按你们 1-based 的 DFT 记号)。
对任意 ,你们定义
因此其矩阵元素为
它天然是 Hermitian:。
因为 Hadamard 乘 会把非对角全部置零,所以
即
所以结构化项的“对角层”就是
你们当前 1-based 约定下(,对应列 ),取
先写出 的元素:
于是
因此
再乘上 的元素式 ,得到
因此结构化项的“低秩层”整体系数合并后为
由
可直接写成
两者都是“一维权重序列做一次 DFT”的结果(你们前面定义的 、):
其中 ,其余 由你们的 与相位汇聚量()给出。
如果你希望把 写得更“一个 就够”的最紧凑形式,那么需要你们的相位设计满足让 的条件(例如某类结构化 ),这样两层就能共享同一个 。
不用合成一个 分两个写 但是写出一个解析解 然后把theta的部分可以抽出来
可以。把 拆成两项分别给出解析解最清晰:一项来自 (纯对角层),另一项来自 (低秩层)。同时把 的影响完全抽到两个 1D 权重向量里,DFT 只作用在“基准衰减 ”与“相位汇聚向量”的 Hadamard 乘积上。
对任意 ,
因此你式子里两个 分别是
先定义与 无关的 1D “基准衰减”向量(只由 决定):
其中(你们目前用的那套 下的标准化简)
令 。
接着把所有 的影响只放进两个 1D 汇聚向量(长度 ):
其中相位矩阵 给出
这样两条 DFT 向量就是一行:
到这里为止, 完全不含 , 全部被“抽到” 里。
因为 只保留对角元,
因此
你们 1-based DFT 列(第 个 bin 对应 )下,取
且
又因为 ,所以
若代入 的余弦形式,可写成更“展开”的闭式:
如果你想把这段直接放到论文里,我建议在 后面紧跟一句“elementwise form”,把上面 (a)(b) 两个 box 式子贴进去即可;同时把 的抽取用 、 这两行强调一下,读者会一眼看到“核的物理衰减”和“相位设计”是怎么解耦的。
我们能不能把你现在定义的两个s的部分还是用一个公用的s,和不同的两种对theta矩阵的取法来相乘组成
可以的,而且写出来会更“干净”:把核本身产生的那条公共谱向量先固定为一个 ,然后把 的作用做成两种不同的(但结构完全相同的)对角加权算子去“作用”在 上。
先定义与 无关的 1D 基准差分权重(你们原来 那条)组成向量
并定义公共谱向量
在“差分域”定义两条仅由 决定的权重向量(长度 ,按差分索引 排列):
那么你原来两条 , 可以统一改写成**“公共 + 两个 算子”**:
这里 与 都是 DFT-对角-DFT 形式(本质是循环算子/循环卷积核),所以非常符合你们“用 DFT 列/DFT 结构表达”的偏好;同时 完全被抽在 (或等价的 )里。
于是你式子保持不变,只是把两条 用上面的公共 表达替换:
如果你希望“更少引入新矩阵”,甚至可以不显式定义 ,只写成 、,同时强调公共 ;但从论文观感上, 这种“算子作用于公共谱”会更像你想要的那种高级/solid 写法。
u的两种定义能不能里面不用两个字母?里面具体的解析解是啥?可不可以用theta和我们的dft的列来表示
能的:把「两种 」都写成同一个字母 (只用上标区分),并且把它们的解析式写成
而且这两种取法都能用 + 你们的 DFT 列 来表示。
保留公共的基准差分谱(与 无关):
然后把两条用于 的向量统一记为
这样 的影响全部被抽到 里(它们都是长度 的 1D 向量,差分域权重)。
对应地你们的结构式可以写成(不合并、仍然两项分开):
其中 。
下面给出两套等价的解析式:(a) 直接用 ;(b) 用 DFT 列 (你想要的“高级且 solid”的写法)。
记
这是你们原来 那条,只是现在统一叫 :
(a) 直接差分求和(最直观)
(b) 用 DFT 列写(用移位矩阵的特征分解消掉二维求和)
令 为你们 1-based DFT 的第 列(),则
这是你们原来 那条,只是现在统一叫 :
(a) 直接相关求和(最贴近原定义)
(b) 用 DFT 列写(把相关写成二次型的 DFT 加权和)
利用移位矩阵的 DFT 特征分解,可以把上式等价写成
这条式子非常“DFT-friendly”: 是对每个 DFT 模态 的能量响应做指数加权求和, 只通过 进入。
严格来说:
是先在差分域乘权,再做 DFT,所以一般不能直接写成“某个标量/逐元素权重 ”(除非 是常数或很特殊)。
但你要的“共享一个 + 两种对 的取法去作用它”可以用一个标准 DFT-对角-DFT 算子来写(仍然只用 DFT):
这在论文里通常被认为是“干净的解耦”:核的 公用, 只改变一个频域对角算子。
如果你希望我把这套统一记号直接嵌回你们的 Proposition/Appendix LaTeX(把 全部替换成 ,并把 的 DFT 列表达作为“closed-form phase aggregates”写进去),我也可以直接给你一段可粘贴的替换版代码。
先用latex代码把你上面的两个s分别写一下具体解析解,以及每一部分对应的解析解
latex% ========================================================= % Elementwise closed forms for the two liftings S(s_diag), S(s_lr) % and the corresponding two parts of Q_k^{sp}(\theta) % (1-based DFT columns: f_{k+1}) % ========================================================= % --- Lifting operator (shared) --- \begin{align}\label{eq:S_lifting_elem} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T \quad\Longrightarrow\quad \big[\mathbf S(\mathbf s)\big]_{n,m}=s_n+s_m^*,\ \ n,m=1,\ldots,N. \end{align} % --------------------------------------------------------- % (A) s_diag(\theta): explicit definition + elementwise DFT expansion % --------------------------------------------------------- \begin{align}\label{eq:sdiag_def} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) ~\triangleq~ \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{diag}}(\boldsymbol\theta) ~=~ \mathbf F\!\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big), \end{align} where \begin{align}\label{eq:udiag_def} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta) ~\triangleq~ \big[\,0,\ \xi[1](\boldsymbol\theta),\ldots,\xi[N\!-\!1](\boldsymbol\theta)\,\big]^T, \qquad \tilde{\boldsymbol\kappa}\triangleq[\,0,\ \tilde\kappa[1],\ldots,\tilde\kappa[N\!-\!1]\,]^T, \end{align} and (difference-domain attenuation induced by $S(\cdot)$) \begin{align}\label{eq:kappa_base_def} \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\ \ d=1,\ldots,N-1,\qquad \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\ \ \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta). \end{align} The phase aggregate $\xi[d](\boldsymbol\theta)$ is \begin{align}\label{eq:xi_def} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big( \Phi_{p+d,q}(\boldsymbol\theta)\Phi_{p,q}^*(\boldsymbol\theta) +\Phi_{p,q+d}(\boldsymbol\theta)\Phi_{p,q}^*(\boldsymbol\theta)\Big), \qquad \Phi_{p,q}(\boldsymbol\theta)\triangleq e^{j\phi_{p,q}(\boldsymbol\theta)}. \end{align} Under the 1-based DFT convention $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)}$, the elementwise expansion is \begin{align}\label{eq:sdiag_elem} \big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ \xi[d](\boldsymbol\theta)\ e^{-j\omega (n-1)d},\qquad n=1,\ldots,N. \end{align} % --------------------------------------------------------- % (B) s_lr(\theta): explicit definition + elementwise DFT expansion % --------------------------------------------------------- \begin{align}\label{eq:slr_def} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) ~\triangleq~ \mathbf F\tilde{\boldsymbol\kappa}_{\mathrm{lr}}(\boldsymbol\theta) ~=~ \mathbf F\!\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big), \end{align} where \begin{align}\label{eq:ulr_def} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta) ~\triangleq~ \big[\,0,\ \eta[1](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta),\ldots,\eta[N\!-\!1](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta)\,\big]^T, \end{align} and \begin{align}\label{eq:eta_def} \eta_0(\boldsymbol\theta)\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}(\boldsymbol\theta), \qquad \eta[d](\boldsymbol\theta)\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}(\boldsymbol\theta),\ \ d\in\mathbb Z_N. \end{align} The elementwise expansion is \begin{align}\label{eq:slr_elem} \big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta)\ e^{-j\omega (n-1)d},\qquad n=1,\ldots,N. \end{align} % ========================================================= % Each part of Q_k^{sp}(\theta): elementwise closed forms % ========================================================= % --------------------------------------------------------- % Part-1 (diagonal layer): (1/N^{3/2}) [ I \odot S(s_diag) ] % --------------------------------------------------------- \begin{align}\label{eq:Qsp_part1_elem} \Big[\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big)\Big]_{n,m} =\frac{1}{N^{3/2}}\delta_{n,m}\Big(\big(\mathbf s_{\mathrm{diag}}\big)_n+\big(\mathbf s_{\mathrm{diag}}\big)_n^*\Big) =\frac{2}{N^{3/2}}\delta_{n,m}\Re\!\Big\{\big(\mathbf s_{\mathrm{diag}}\big)_n\Big\}. \end{align} % --------------------------------------------------------- % Part-2 (low-rank Hadamard layer): (sqrt(N)/(2N)) [ A_k \odot S(s_lr) ] % with 1-based DFT column f_{k+1} % --------------------------------------------------------- \begin{align}\label{eq:Ak_def_1based} \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H, \qquad \big[\mathbf f_{k+1}\big]_n=\frac{1}{\sqrt N}e^{-j\omega (n-1)k}. \end{align} Hence \begin{align}\label{eq:Ak_elem_cos} \big[\mathbf A_k\big]_{n,m} =\frac{1}{N}e^{-j\omega k(n+m-2)}+\frac{1}{N}e^{+j\omega k(n+m-2)} =\frac{2}{N}\cos\!\big(\omega k(n+m-2)\big). \end{align} Using \eqref{eq:S_lifting_elem}, the elementwise form of the second part is \begin{align}\label{eq:Qsp_part2_elem} \Big[\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big)\Big]_{n,m} =\frac{\sqrt N}{2N}\big[\mathbf A_k\big]_{n,m}\Big(\big(\mathbf s_{\mathrm{lr}}\big)_n+\big(\mathbf s_{\mathrm{lr}}\big)_m^*\Big) =\frac{\sqrt N}{N^2}\cos\!\big(\omega k(n+m-2)\big)\Big(\big(\mathbf s_{\mathrm{lr}}\big)_n+\big(\mathbf s_{\mathrm{lr}}\big)_m^*\Big). \end{align} % (Optional) If you want the full elementwise Q_k^{sp} entry without "combining" into one boxed formula, % you may keep \eqref{eq:Qsp_part1_elem} and \eqref{eq:Qsp_part2_elem} side-by-side as the two closed forms.
把每一个矩阵和响亮的维度和是R还是C体现出来
latex% ========================================================= % Dimensions / field annotations % ========================================================= % N: DFT size (positive integer). % Indexing: k \in {0,1,...,N-1}, DFT columns are 1-based: \mathbf f_{k+1}. % \mathbf g \in \mathbb R^{N\times 1} (real). % \mathbf F \in \mathbb C^{N\times N} (complex, unitary). % \mathbf f_\ell \in \mathbb C^{N\times 1} for \ell=1,\dots,N. % \mathbf s_{\mathrm{diag}}(\boldsymbol\theta), \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) \in \mathbb C^{N\times 1}. % \mathbf S(\cdot), \mathbf A_k, \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta) \in \mathbb C^{N\times N} (Hermitian). % Scalars \tilde\kappa[d] \in \mathbb R_{+}, \eta_0(\boldsymbol\theta),\eta[d](\boldsymbol\theta),\xi[d](\boldsymbol\theta)\in\mathbb C. % \mathbf u_{\mathrm{diag}}(\boldsymbol\theta), \mathbf u_{\mathrm{lr}}(\boldsymbol\theta) \in \mathbb C^{N\times 1}. % ========================================================= % Normalized DFT (1-based columns) % ========================================================= \begin{align} \mathbf F \in \mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad \omega\triangleq\frac{2\pi}{N}. \end{align} \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell=1,\ldots,N. \end{align} % ========================================================= % Lifting operator S(\cdot): definition, dimensions, entries % ========================================================= \begin{align}\label{eq:S_lifting_def_dim} \mathbf S:\mathbb C^{N\times 1}\to\mathbb C^{N\times N},\qquad \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T, \quad \mathbf 1\in\mathbb R^{N\times 1}. \end{align} \begin{align}\label{eq:S_lifting_elem_dim} [\mathbf S(\mathbf s)]_{n,m}=s_n+s_m^*,\qquad n,m=1,\ldots,N. \end{align} % Note: \mathbf S(\mathbf s) is Hermitian for any \mathbf s \in \mathbb C^{N\times 1}. % ========================================================= % Baseline 1D attenuation (theta-free): \tilde{\kappa} % ========================================================= \begin{align} \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1},\qquad \tilde\kappa[0]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big)\in\mathbb R_{+},\ \ d=1,\ldots,N-1, \end{align} \begin{align} \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta)\in\mathbb R,\qquad \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big)\in\mathbb R_{+}. \end{align} % ========================================================= % Phase-coded aggregates (theta-dependent): eta0, eta[d], xi[d] % ========================================================= \begin{align} \mathbf \Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad \Phi_{p,q}(\boldsymbol\theta)\triangleq e^{j\phi_{p,q}(\boldsymbol\theta)}\in\mathbb C,\quad p,q\in\mathbb Z_N. \end{align} \begin{align} \eta_0(\boldsymbol\theta)\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}(\boldsymbol\theta)\in\mathbb C, \qquad \eta[d](\boldsymbol\theta)\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}(\boldsymbol\theta)\in\mathbb C. \end{align} \begin{align} \xi[d](\boldsymbol\theta)\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big( \Phi_{p+d,q}(\boldsymbol\theta)\Phi_{p,q}^*(\boldsymbol\theta)+ \Phi_{p,q+d}(\boldsymbol\theta)\Phi_{p,q}^*(\boldsymbol\theta)\Big)\in\mathbb C. \end{align} % ========================================================= % The two theta-dependent 1D weights u_diag, u_lr (same dimension) % ========================================================= \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{diag}}[0](\boldsymbol\theta)\triangleq 0,\quad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta),\ d=1,\ldots,N-1. \end{align} \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{lr}}[0](\boldsymbol\theta)\triangleq 0,\quad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta),\ d=1,\ldots,N-1. \end{align} % ========================================================= % The two s-vectors: explicit DFT forms + elementwise expansions % ========================================================= \begin{align}\label{eq:sdiag_dim} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big). \end{align} \begin{align}\label{eq:sdiag_elem_dim} \big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ \xi[d](\boldsymbol\theta)\ e^{-j\omega (n-1)d},\qquad n=1,\ldots,N. \end{align} \begin{align}\label{eq:slr_dim} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big). \end{align} \begin{align}\label{eq:slr_elem_dim} \big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta)\ e^{-j\omega (n-1)d},\qquad n=1,\ldots,N. \end{align} % ========================================================= % Each part of Q_k^{sp}(\theta): dimensions and elementwise closed forms % ========================================================= \begin{align} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)\in\mathbb C^{N\times N}\ \text{(Hermitian)}. \end{align} % ---- Part 1: diagonal layer ---- \begin{align}\label{eq:Qsp_part1_dim} \mathbf Q_{k,\mathrm{diag}}^{\mathrm{sp}}(\boldsymbol\theta) \triangleq \frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big) \in\mathbb C^{N\times N}. \end{align} \begin{align}\label{eq:Qsp_part1_elem_dim} \big[\mathbf Q_{k,\mathrm{diag}}^{\mathrm{sp}}(\boldsymbol\theta)\big]_{n,m} =\frac{1}{N^{3/2}}\delta_{n,m}\Big(\big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)_n+\big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)_n^*\Big) =\frac{2}{N^{3/2}}\delta_{n,m}\Re\!\Big\{\big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)_n\Big\}. \end{align} % ---- Part 2: low-rank Hadamard layer ---- \begin{align}\label{eq:Ak_dim_def} \mathbf A_k\in\mathbb C^{N\times N}\ \text{(real-valued entries, Hermitian)},\qquad \mathbf A_k\triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H. \end{align} \begin{align}\label{eq:Ak_entries_dim} [\mathbf f_{k+1}]_n=\frac{1}{\sqrt N}e^{-j\omega (n-1)k}\in\mathbb C,\qquad [\mathbf A_k]_{n,m} =\frac{2}{N}\cos\!\big(\omega k(n+m-2)\big)\in\mathbb R. \end{align} \begin{align}\label{eq:Qsp_part2_dim} \mathbf Q_{k,\mathrm{lr}}^{\mathrm{sp}}(\boldsymbol\theta) \triangleq \frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big) \in\mathbb C^{N\times N}. \end{align} \begin{align}\label{eq:Qsp_part2_elem_dim} \big[\mathbf Q_{k,\mathrm{lr}}^{\mathrm{sp}}(\boldsymbol\theta)\big]_{n,m} =\frac{\sqrt N}{2N}[\mathbf A_k]_{n,m}\Big(\big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_n+\big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_m^*\Big) =\frac{\sqrt N}{N^2}\cos\!\big(\omega k(n+m-2)\big)\Big(\big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_n+\big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)_m^*\Big). \end{align} % ---- Sum (kept separate, but you can state both parts define Q_k^{sp}) ---- \begin{align} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta) =\mathbf Q_{k,\mathrm{diag}}^{\mathrm{sp}}(\boldsymbol\theta)+\mathbf Q_{k,\mathrm{lr}}^{\mathrm{sp}}(\boldsymbol\theta) \in\mathbb C^{N\times N}. \end{align}
如果我们之前已经定义过一个四阶的相位变量,可不可以引入以简化:。我们可以把这个四阶相位变成一个N^2*N^2的矩阵,然后从里面取我们需要的元素?
可以,而且这是一个很标准的“lifting / Gram matrix”写法:把四阶相位项变成一个 的秩一矩阵,然后你需要的任何 元素、或者你们在推导里出现的“沿对角线求和/相关求和”,都能用取元素或**线性泛函(trace / 二次型)**从这张大矩阵里读出来。
下面给一套和你们 1-based 记号兼容、可直接放 LaTeX 的定义与“怎么取”。
先把二维相位矩阵写出来(由 参数化也行):
令
(默认 column-wise vec,即按列堆叠)。定义 lifting 相位矩阵
那么对任意 ,令
就有
所以你说的“从里面取我们需要的元素”就是:要哪一项四阶相位,直接取 的对应元素即可。
同时 是 Hermitian 且 PSD,并且秩一。
你们推导里常见的两类“从 塌缩到 1D 的量”(例如 )也能用 更统一地写。
先定义循环移位矩阵(1-based):
主对角均值:
偏移对角均值:
沿 方向的循环相关:
沿 方向的循环相关:
你们之前用的对称平均就是
这样写的好处:所有“需要沿某个方向移位再做内积”的东西,都变成
或 ,形式特别统一。
能。关键是循环移位矩阵 被 DFT 对角化:
(你们的 1-based DFT 列 ;本质是 shift 的特征值是 )。
于是像 这类量也可以进一步写成“对 DFT 模态能量的加权和”(不过会比上面 trace 形式更长)。如果你目标是简化推导/统一符号,我建议正文/appendix 用上面这种 、 就已经非常“DFT-structured”了,而且不引入额外字母。
你完全可以在 appendix 里加一句:
是秩一 lifting,因此不需要真的构造 矩阵即可计算所需量;但用于推导和记号统一很方便。
如果你愿意,我可以把你们现在 appendix 里 的定义全部替换成上面这种 “+Kronecker/trace” 的统一写法,并保持你们的 1-based DFT 记号不变。
\frac{1}{N^{3/2}}\big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big]
+\frac{\sqrt N}{2N}\big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}})\big].
\end{align}
\begin{align}\label{eq:Qbulk_prop}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})
=\frac{1}{N}\big(\mathbf{f}{k+1}^*\mathbf{f}{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}),
\end{align}
这样的话我们可不可以用\kappa、DFT矩阵和二维或者四维的相位矩阵,这三个量来表示
latex% ========================================================= % Goal: express the two liftings S(s_diag), S(s_lr) % using only (i) the 1D attenuation kappa, (ii) the DFT matrix F, % and (iii) the 2D/4D phase object (Phi or lifted Phi-tilde). % Convention: 1-based DFT columns (f_{k+1}), k=0,...,N-1. % ========================================================= % ----------------------------- % Normalized DFT (1-based columns) % ----------------------------- \begin{align} \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad \omega\triangleq\frac{2\pi}{N}, \end{align} \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell=1,\ldots,N. \end{align} % ----------------------------- % 2D phase matrix and 4th-order phase lifting (N^2 x N^2) % ----------------------------- \begin{align} \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad [\mathbf\Phi(\boldsymbol\theta)]_{p,q}\triangleq \Phi_{p,q}(\boldsymbol\theta)=e^{j\phi_{p,q}(\boldsymbol\theta)}. \end{align} \begin{align} \mathbf v(\boldsymbol\theta)\triangleq \mathrm{vec}\big(\mathbf\Phi(\boldsymbol\theta)\big)\in\mathbb C^{N^2\times 1},\qquad \widetilde{\mathbf\Phi}(\boldsymbol\theta)\triangleq \mathbf v(\boldsymbol\theta)\mathbf v(\boldsymbol\theta)^H\in\mathbb C^{N^2\times N^2}. \end{align} % Then: [\widetilde{\mathbf\Phi}]_{i(p,q),i(p',q')} = \Phi_{p,q}\Phi_{p',q'}^* = e^{j\tilde\phi_{p,q,p',q'}}. % ----------------------------- % Circulant shift matrices (used in linear/quadratic functionals) % ----------------------------- \begin{align} \mathbf J_d\in\{0,1\}^{N\times N},\qquad [\mathbf J_d]_{p,q}=\mathbbm 1\{\,p\equiv q+d\ (\mathrm{mod}\ N)\,\},\qquad d=0,\ldots,N-1. \end{align} % ----------------------------- % Theta-free 1D attenuation profile: \tilde{kappa} % (collect into a length-N real vector, indexed by d=0,...,N-1) % ----------------------------- \begin{align} \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1},\qquad \tilde\kappa[0]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d)\big),\ \ d=1,\ldots,N-1, \end{align} \begin{align} \Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad \Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta). \end{align} % ========================================================= % (I) Two theta-dependent 1D weights u_diag(theta), u_lr(theta) % written as: (a) linear functional / (b) quadratic functional of Phi-tilde % ========================================================= % -------- u_lr(theta): "linear x linear" -> bilinear in v, linear in Phi-tilde % eta[d] = (1/N) tr(J_d^T Phi), eta0 = (1/N) tr(Phi) \begin{align} \eta[d](\boldsymbol\theta)\triangleq \frac{1}{N}\mathrm{tr}\!\big(\mathbf J_d^T\mathbf\Phi(\boldsymbol\theta)\big) =\frac{1}{N}\mathrm{vec}(\mathbf J_d)^T \mathbf v(\boldsymbol\theta)\in\mathbb C, \qquad \eta_0(\boldsymbol\theta)\triangleq \frac{1}{N}\mathrm{tr}\!\big(\mathbf\Phi(\boldsymbol\theta)\big) =\frac{1}{N}\mathrm{vec}(\mathbf I)^T \mathbf v(\boldsymbol\theta)\in\mathbb C. \end{align} \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{lr}}[0](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta),\ \ d=1,\ldots,N-1. \end{align} % Equivalent "lifted" (linear in Phi-tilde) form: \begin{align} u_{\mathrm{lr}}[d](\boldsymbol\theta) =\frac{1}{N^2}\,\mathrm{vec}(\mathbf J_d)^T\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\,\mathrm{vec}(\mathbf I)^*, \qquad d=1,\ldots,N-1. \end{align} % -------- u_diag(theta): "quadratic" correlation -> trace(Kronecker * Phi-tilde) \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{diag}}[0](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta),\ \ d=1,\ldots,N-1, \end{align} \begin{align} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\mathrm{tr}\!\Big(\big(\mathbf I\otimes \mathbf J_d+\mathbf J_d\otimes \mathbf I\big)\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\Big), \qquad d=1,\ldots,N-1. \end{align} % ========================================================= % (II) Two theta-dependent s-vectors (both are one DFT of kappa-weighted u) % ========================================================= \begin{align}\label{eq:sdiag_lifted} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big). \end{align} \begin{align}\label{eq:slr_lifted} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big). \end{align} % Elementwise expansions (optional): \begin{align} (\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ u_{\mathrm{diag}}[d](\boldsymbol\theta)\ e^{-j\omega (n-1)d},\qquad (\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))_n =\frac{1}{\sqrt N}\sum_{d=1}^{N-1}\tilde\kappa[d]\ u_{\mathrm{lr}}[d](\boldsymbol\theta)\ e^{-j\omega (n-1)d}. \end{align} % ========================================================= % (III) The two liftings S(s_diag), S(s_lr) (explicit closed forms) % ========================================================= \begin{align} \mathbf S(\cdot):\mathbb C^{N\times 1}\to \mathbb C^{N\times N},\qquad \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T, \quad \mathbf 1\in\mathbb R^{N\times 1}. \end{align} \begin{align} [\mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))]_{n,m} =(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))_n+(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))_m^*, \qquad [\mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))]_{n,m} =(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))_n+(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))_m^*. \end{align} % ========================================================= % (IV) Plug into Q_sp and Q_bulk (your preferred forms) % ========================================================= \begin{align}\label{eq:Qsp_prop_with_lifted} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)\in\mathbb C^{N\times N} = \frac{1}{N^{3/2}}\Big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\Big] +\frac{\sqrt N}{2N}\Big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\Big], \end{align} with (1-based DFT bin column) \begin{align} \mathbf A_k\triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H\in\mathbb C^{N\times N}. \end{align} \begin{align}\label{eq:Qbulk_prop_pref} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N\times N} =\frac{1}{N}\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta), \end{align} % Optional: show that K_bulk(theta) is induced by the lifted phase matrix (4D phase) linearly. % (Here K_S is the fixed exp(-alpha S) kernel in (p,q)-(p',q') lifted indexing; P_bulk is the bulk mask.) \begin{align} \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta) =\mathcal L_{\mathrm{bulk}}\!\big(\widetilde{\mathbf\Phi}(\boldsymbol\theta)\big), \qquad \mathcal L_{\mathrm{bulk}}(\mathbf Z)\triangleq \mathbf F\,\mathbf B^H\,\mathrm{Herm}\!\Big(\mathbf Z\odot \mathbf K_S\odot \mathbf P^{\mathrm{bulk}}\Big)\mathbf B\,\mathbf F^H, \end{align} where $\mathbf Z\in\mathbb C^{N^2\times N^2}$, $[\mathbf K_S]_{(p,q),(p',q')}\triangleq \exp\!\big(-\alpha S(p,q,p',q')\big)$, and $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]\in\mathbb R^{N^2\times N}$ is fixed.
结论(对应你最后一句):在这套写法下,
那个J_d我们直接用DFT给对角化来表示,不要现在的方式!然后记住所有的索引都是1N,不是0N-1
下面给你一份完全 1-based(索引全是 )且把原来的 全部改成 DFT 对角化表达(不再用指示函数定义)的 LaTeX 代码。核心思路是:用一个“循环移位算子”
其中 表示“移位量 ”。这样你原来所有涉及 的线性型/二次型都能直接改写成只含 、 的形式。
latex% ========================================================= % 1-based indexing everywhere: all indices in {1,...,N} % Replace J_d by DFT-diagonalized circulant shift: Pi_d = F Lambda_d F^H % ========================================================= % ---------- DFT ---------- \begin{align} \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad \omega\triangleq\frac{2\pi}{N}, \end{align} \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell=1,\ldots,N. \end{align} % ---------- 2D phase matrix and 4th-order lifting ---------- \begin{align} \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad [\mathbf\Phi(\boldsymbol\theta)]_{p,q}\triangleq \Phi_{p,q}(\boldsymbol\theta)=e^{j\phi_{p,q}(\boldsymbol\theta)},\qquad p,q\in\{1,\ldots,N\}. \end{align} \begin{align} \mathbf v(\boldsymbol\theta)\triangleq \mathrm{vec}\big(\mathbf\Phi(\boldsymbol\theta)\big)\in\mathbb C^{N^2\times 1},\qquad \widetilde{\mathbf\Phi}(\boldsymbol\theta)\triangleq \mathbf v(\boldsymbol\theta)\mathbf v(\boldsymbol\theta)^H\in\mathbb C^{N^2\times N^2}. \end{align} % Then: [\widetilde{\mathbf\Phi}]_{i(p,q),i(p',q')} = \Phi_{p,q}\Phi_{p',q'}^* = e^{j\tilde\phi_{p,q,p',q'}}, % with i(p,q)=p+(q-1)N. % ---------- Mod-N wrap to {1,...,N} ---------- \begin{align} \langle x\rangle_N \triangleq 1+\mathrm{mod}(x-1,N)\in\{1,\ldots,N\}. \end{align} % ========================================================= % Circulant shift operator via DFT diagonalization (NO J_d) % d \in {1,...,N} corresponds to shift amount (d-1) % ========================================================= \begin{align}\label{eq:Pi_d_def} \mathbf\Lambda_d \triangleq \mathrm{diag}\!\Big(e^{-j\omega(\ell-1)(d-1)}\Big)_{\ell=1}^{N}\in\mathbb C^{N\times N},\qquad \mathbf\Pi_d \triangleq \mathbf F\,\mathbf\Lambda_d\,\mathbf F^H\in\mathbb C^{N\times N}. \end{align} % ========================================================= % theta-free 1D attenuation profile (1-based, d=1 is zero-difference) % ========================================================= \begin{align} \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1},\qquad \tilde\kappa[1]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d-1)\big),\ \ d=2,\ldots,N. \end{align} % ========================================================= % Linear functional (eta_d) and bilinear u_lr[d] using Pi_d = F Lambda_d F^H % ========================================================= \begin{align}\label{eq:eta0_lin} \eta_0(\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{t=1}^N \Phi_{t,t}(\boldsymbol\theta) = \frac{1}{N}\mathrm{tr}\big(\mathbf\Phi(\boldsymbol\theta)\big) = \frac{1}{N}\,\mathrm{vec}(\mathbf I)^T\mathbf v(\boldsymbol\theta)\in\mathbb C. \end{align} \begin{align}\label{eq:eta_d_lin_Pi} \eta[d](\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{q=1}^N \Phi_{\langle q+d-1\rangle_N,\,q}(\boldsymbol\theta) = \frac{1}{N}\mathrm{tr}\!\Big(\mathbf\Pi_d^T\,\mathbf\Phi(\boldsymbol\theta)\Big) = \frac{1}{N}\mathrm{tr}\!\Big(\big(\mathbf F\mathbf\Lambda_d\mathbf F^H\big)^T\mathbf\Phi(\boldsymbol\theta)\Big)\in\mathbb C. \end{align} % Optional: lifted (linear in Phi-tilde) version of u_lr[d] \begin{align}\label{eq:u_lr_lifted} u_{\mathrm{lr}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta) =\frac{1}{N^2}\,\mathrm{vec}(\mathbf\Pi_d)^T\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\,\mathrm{vec}(\mathbf I)^*,\quad d=2,\ldots,N, \end{align} \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq [u_{\mathrm{lr}}[1],\ldots,u_{\mathrm{lr}}[N]]^T\in\mathbb C^{N\times 1}. \end{align} % ========================================================= % Quadratic functional (xi_d) using Pi_d (DFT-diagonalized) and Phi-tilde % ========================================================= \begin{align}\label{eq:u_diag_lifted} u_{\mathrm{diag}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta),\quad d=2,\ldots,N, \end{align} \begin{align}\label{eq:xi_quad_Pi} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\mathrm{tr}\!\Big(\big(\mathbf I\otimes \mathbf\Pi_d+\mathbf\Pi_d\otimes \mathbf I\big)\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\Big) =\frac{1}{2N^2}\mathrm{tr}\!\Big(\big(\mathbf I\otimes \mathbf F\mathbf\Lambda_d\mathbf F^H+\mathbf F\mathbf\Lambda_d\mathbf F^H\otimes \mathbf I\big)\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\Big), \quad d=2,\ldots,N, \end{align} \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq [u_{\mathrm{diag}}[1],\ldots,u_{\mathrm{diag}}[N]]^T\in\mathbb C^{N\times 1}. \end{align} % ========================================================= % The two s-vectors (only kappa, DFT, and 2D/4D phase objects) % ========================================================= \begin{align}\label{eq:sdiag_final} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big), \end{align} \begin{align}\label{eq:slr_final} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big). \end{align} % Optional (elementwise, 1-based): \begin{align} (\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))_n =\frac{1}{\sqrt N}\sum_{d=2}^{N}\tilde\kappa[d]\ u_{\mathrm{diag}}[d](\boldsymbol\theta)\ e^{-j\omega (n-1)(d-1)},\qquad (\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))_n =\frac{1}{\sqrt N}\sum_{d=2}^{N}\tilde\kappa[d]\ u_{\mathrm{lr}}[d](\boldsymbol\theta)\ e^{-j\omega (n-1)(d-1)}. \end{align} % ========================================================= % Liftings and Q_sp % ========================================================= \begin{align} \mathbf S:\mathbb C^{N\times 1}\to\mathbb C^{N\times N},\qquad \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T,\quad \mathbf 1\in\mathbb R^{N\times 1}, \end{align} \begin{align} [\mathbf S(\mathbf s)]_{n,m}=s_n+s_m^*,\qquad n,m\in\{1,\ldots,N\}. \end{align} \begin{align}\label{eq:Qsp_prop_final} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)\in\mathbb C^{N\times N} = \frac{1}{N^{3/2}}\Big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\Big] +\frac{\sqrt N}{2N}\Big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\Big], \end{align} \begin{align} \mathbf A_k\triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H\in\mathbb C^{N\times N}, \qquad k\in\{1,\ldots,N\}\ \text{(bin index is }k-1\text{)}. \end{align} % ========================================================= % Q_bulk (your preferred standard) % ========================================================= \begin{align}\label{eq:Qbulk_prop_final} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})\in\mathbb C^{N\times N} =\frac{1}{N}\big(\mathbf{f}_{k}^*\mathbf{f}_{k}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}), \qquad k\in\{1,\ldots,N\}. \end{align} % NOTE: here the bin-k uses column f_k (1-based). If you want "f_{k+1}" in the main text, % keep it as in your original equation and let k be 0..N-1. Since you requested all indices 1..N, % the cleanest is Q_bulk uses f_k with k=1..N. % If you insist to keep "f_{k+1}" while also saying k=1..N, then interpret it modulo N: % f_{k+1} := f_{\langle k+1\rangle_N}.
你刚才坚持“索引都是 ”,那最一致的写法其实是
如果你们正文固定写成 (并且 ),那也没问题;但两者要二选一才不会在读者那里产生“到底 从 0 还是从 1” 的歧义。上面代码里我已经把这点用 NOTE 标出来了。
如果你告诉我你们最终想把 bin index 记成 还是 ,我可以把整份代码里所有出现的 、 再做一次全局一致化(保证没有任何混用)。
不用新字母 还用f来表示
可以。你要“不要 、不要再引入新的移位矩阵字母”,那就把“移位算子”直接用 DFT 特征分解展开成 的加权和,然后把线性型 和二次型 都写成只含 、、/ 的形式。
下面是一段可直接粘贴的 LaTeX(所有索引都在 ;移位量用 ,因此 表示零移位/零差分):
latex% ========================================================= % 1-based indexing throughout: all indices in {1,...,N} % No J_d / no new shift-matrix letter; use only DFT columns f_\ell. % ========================================================= % --- DFT and its columns --- \begin{align} \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad \omega\triangleq\frac{2\pi}{N},\quad n,\ell\in\{1,\ldots,N\}, \end{align} \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell=1,\ldots,N. \end{align} % --- Mod-N wrap to {1,...,N} --- \begin{align} \langle x\rangle_N \triangleq 1+\mathrm{mod}(x-1,N)\in\{1,\ldots,N\}. \end{align} % --- 2D phase and 4th-order lifting phase --- \begin{align} \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad \Phi_{p,q}(\boldsymbol\theta)\triangleq e^{j\phi_{p,q}(\boldsymbol\theta)},\quad p,q\in\{1,\ldots,N\}, \end{align} \begin{align} \mathbf v(\boldsymbol\theta)\triangleq \mathrm{vec}(\mathbf\Phi(\boldsymbol\theta))\in\mathbb C^{N^2\times 1},\qquad \widetilde{\mathbf\Phi}(\boldsymbol\theta)\triangleq \mathbf v(\boldsymbol\theta)\mathbf v(\boldsymbol\theta)^H\in\mathbb C^{N^2\times N^2}. \end{align} % --- theta-free 1D attenuation profile (d=1 is zero-difference) --- \begin{align} \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1},\qquad \tilde\kappa[1]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d-1)\big),\ \ d=2,\ldots,N. \end{align} % ========================================================= % Linear functional: eta0(theta), eta[d](theta) using ONLY f_\ell % ========================================================= \begin{align}\label{eq:eta0_lin_f} \eta_0(\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{t=1}^N \Phi_{t,t}(\boldsymbol\theta) =\frac{1}{N}\mathrm{tr}\big(\mathbf\Phi(\boldsymbol\theta)\big)\in\mathbb C. \end{align} \begin{align}\label{eq:eta_d_lin_f} \eta[d](\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{q=1}^N \Phi_{\langle q+d-1\rangle_N,\,q}(\boldsymbol\theta) =\frac{1}{N}\sum_{\ell=1}^{N} e^{+j\omega(d-1)(\ell-1)}\ \mathbf f_\ell^{H}\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell \in\mathbb C, \qquad d=1,\ldots,N. \end{align} % --- u_lr(theta) (length-N), 1-based: u_lr[1]=0 corresponds to d=1 (zero shift) --- \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{lr}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta),\ \ d=2,\ldots,N. \end{align} % ========================================================= % Quadratic functional: xi[d](theta) using ONLY f_\ell and Phi-tilde % (no shift-matrix symbol; expand shift via DFT spectral projectors f_\ell f_\ell^H) % ========================================================= \begin{align}\label{eq:xi_quad_f} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\, \mathrm{tr}\!\Big( \big(\mathbf I\otimes \mathbf f_\ell\mathbf f_\ell^H+\mathbf f_\ell\mathbf f_\ell^H\otimes \mathbf I\big)\, \widetilde{\mathbf\Phi}(\boldsymbol\theta) \Big)\in\mathbb C, \qquad d=1,\ldots,N. \end{align} % --- u_diag(theta) (length-N), 1-based: u_diag[1]=0 corresponds to d=1 (zero shift) --- \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad u_{\mathrm{diag}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta),\ \ d=2,\ldots,N. \end{align} % ========================================================= % Two s-vectors: only kappa + DFT + (Phi or Phi-tilde) % ========================================================= \begin{align}\label{eq:sdiag_final_f} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big). \end{align} \begin{align}\label{eq:slr_final_f} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big). \end{align} % ========================================================= % Lift operator and Q_sp (unchanged form) % ========================================================= \begin{align} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T\in\mathbb C^{N\times N},\qquad [\mathbf S(\mathbf s)]_{n,m}=s_n+s_m^*,\quad n,m\in\{1,\ldots,N\}. \end{align} \begin{align}\label{eq:Qsp_prop_f_based} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta})\in\mathbb C^{N\times N} = \frac{1}{N^{3/2}}\big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big] +\frac{\sqrt N}{2N}\big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big], \end{align} \begin{align} \mathbf A_k\triangleq \mathbf f_{k}\mathbf f_{k}^{T}+\mathbf f_{k}^{*}\mathbf f_{k}^{H}\in\mathbb C^{N\times N},\qquad k\in\{1,\ldots,N\}. \end{align} % ========================================================= % Q_bulk (your standard; 1-based bin index) % ========================================================= \begin{align}\label{eq:Qbulk_prop_1based} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})\in\mathbb C^{N\times N} =\frac{1}{N}\big(\mathbf{f}_{k}^*\mathbf{f}_{k}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}), \qquad k\in\{1,\ldots,N\}. \end{align}
是的。在这套写法里:
并且:
如果你还想把 也只用二维 (不用 )来写,我也可以给出等价的“”那种形式,不过你前面说希望用四阶相位矩阵来取元素/简化记号,用 会更统一。
先别写latex代码 我们先讨论推导过程
可以,而且推导上也很顺:把“沿对角线取样/移位相关”这类东西全部用 DFT 模态的谱展开来做,这样就不需要再写 (或者任何新的“移位矩阵字母”),只用 就能把线性型/二次型都写出来。并且你要的三个量:
确实足够把 和两个 (以及进而 )都表达掉。
下面按 全 1-based(所有下标 )说清楚过程。
你们差分/移位的“真实移位量”是 ,其中
循环移位算子(移位量 )在 DFT 基下是对角的,所以它等价于一个 DFT 投影子的加权和:
如果你需要的是它的转置(你们 那个“偏移对角线求和”对应的那个),因为移位矩阵是置换矩阵,转置等于逆移位,所以特征值取倒数/共轭:
这一步就是你要的:没有 ,没有新字母,只有 。
先定义二维相位矩阵 ,以及每个 DFT 模态上的“对角二次型”:
把第 1 部分的谱展开代进去,得到完全用 的形式:
于是低秩层权重(你们原来 )就是
如果你想把它进一步写成只含四维相位 lifting ,也行,而且仍然只用 :注意
所以
从而
这里没有任何 或新移位字母,只出现 (以及 vec、lifting)。
你们的
本质是“沿 方向移位相关 + 沿 方向移位相关”的平均。
把它写成 vec 二次型会是
然后把 shift 用第 1 部分的谱展开替换(只用 )即可。
更“物理直观”的等价结果是(同样只用 和二维 ):
如果你坚持统一用四维 lifting ,那就把两个能量项写成 trace:
于是
最后对角层权重就是
你们的核衰减只负责一个 (-free),相位只通过 进入:
然后两个 lifting:
代回你们的
就完成了:整个 依赖只在 、 里,而这两个又分别由上面的“DFT 列加权和(线性/二次型)”给出。
你给的标准式
如果你现在坚持 ,那最不歧义的做法是把它写成
因为“bin ”就对应“第 列”。
如果你们一定要保留 ,那就必须在文中声明 (模 wrap),否则读者会困惑。
所以答案是:可以。并且推导上你可以这样组织:
你如果认可这个推导路线,我再把它“翻译”成你们 appendix 里的一段连贯证明文字(不写代码版本),再下一步再落回 LaTeX。
把\sume^(jomega(d-1)(l-1)),用DFT矩阵的列来表示
可以。因为你们 1-based 的 DFT 列定义是
所以把指数写成关于“求和下标 ”的向量
就有(注意 )
因此:
(也等价于 时等于 ,其余 时为 0。)
它就是
并且等价于 (因为 )。
把这个带到我们刚才分析的S的两部分里面
好,我们就把你刚才那个“带权指数求和 = DFT 列内积”的恒等式,直接塞回我们之前给出的两条聚合量(线性型 和二次型 ),再把它们带进两条 ,最后就得到两部分 、 的“只含 、DFT、相位(2D/4D)”表达。
下面全程 1-based:,移位量永远是 。
你们的 1-based DFT 列满足
因此对任意 :
我们之前有
用上面的恒等式(正指数):
同时
于是你们 low-rank 用到的那条权重(之前叫 )可以写成
( 那项反正会被 消掉,设不设 0 都不影响。)
我们之前有
其中(两种等价写法,二选一):
把负指数求和用 DFT 列吃掉(第二条恒等式):
于是 diagonal 那支用到的权重就是
你们的两条向量本来就是
其中 是核衰减(1-based,,其余由 给出)。
代入上面两个 的 DFT-列形式,就得到非常“你要的三件套”表达:
更紧一点就是“把那一列内积打包成 ”:
同样可以打包成 :
因为
所以
其中 取上面第 4 步那条(由 、、 决定,而 由 或 给出)。
其中 由 、、 决定,而 只用二维 的模态二次型 。
按这个替换,你的目标“只用 、DFT(矩阵/列)、二维或四维相位矩阵”就能把 里两部分的 完整写出来,而且推导中没有 或任何新的移位矩阵符号,全靠 的谱展开在背后工作。
如果你同意这条推导链条,下一步我可以把它整理成你们 Appendix 里一段“从 到 再到 ”的连贯证明文字(不写代码版本),然后你再决定哪些句子放正文/哪些放 appendix。
把上面你分析的公式推导用latex代码给我 然后标注出每一个向量/矩阵的维度以及是实数还是复数
latex% ========================================================= % Derivation: substitute the DFT-column identity into the two parts of S(\cdot) % 1-based indexing ONLY: all indices in {1,...,N} % Annotate dimensions and field (R/C) for every object. % ========================================================= % ----------------------------- % Basic objects and dimensions % ----------------------------- % N \in \mathbb N: blocklength / DFT size. % \omega \triangleq 2\pi/N \in \mathbb R. % Normalized DFT: % \mathbf F \in \mathbb C^{N\times N}, unitary. \begin{align} \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\qquad n,\ell\in\{1,\ldots,N\}, \end{align} % DFT columns: % \mathbf f_\ell \in \mathbb C^{N\times 1}. \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell=1,\ldots,N. \end{align} % All-one vector and identity: % \mathbf 1 \in \mathbb R^{N\times 1}, \mathbf I \in \mathbb R^{N\times N}. \begin{align} \mathbf 1\in\mathbb R^{N\times 1},\qquad \mathbf I\in\mathbb R^{N\times N}. \end{align} % 2D phase matrix: % \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N}, with unit-modulus entries. \begin{align} \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad [\mathbf\Phi(\boldsymbol\theta)]_{p,q}\triangleq \Phi_{p,q}(\boldsymbol\theta)=e^{j\phi_{p,q}(\boldsymbol\theta)}\in\mathbb C, \qquad p,q\in\{1,\ldots,N\}. \end{align} % 4th-order lifted phase (N^2 x N^2): % \mathbf v(\boldsymbol\theta)=vec(\Phi) \in \mathbb C^{N^2\times 1}, % \widetilde{\mathbf\Phi}(\boldsymbol\theta)=\mathbf v\mathbf v^H \in \mathbb C^{N^2\times N^2}. \begin{align} \mathbf v(\boldsymbol\theta)\triangleq \mathrm{vec}\!\big(\mathbf\Phi(\boldsymbol\theta)\big)\in\mathbb C^{N^2\times 1},\qquad \widetilde{\mathbf\Phi}(\boldsymbol\theta)\triangleq \mathbf v(\boldsymbol\theta)\mathbf v(\boldsymbol\theta)^H\in\mathbb C^{N^2\times N^2}. \end{align} % Theta-free attenuation profile (1-based in d): % \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1}, \tilde\kappa[1]=0. \begin{align} \tilde{\boldsymbol\kappa}\triangleq[\tilde\kappa[1],\ldots,\tilde\kappa[N]]^T\in\mathbb R^{N\times 1},\qquad \tilde\kappa[1]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \exp\!\big(-\alpha\,\Delta_S(d-1)\big)\in\mathbb R_{+},\ \ d=2,\ldots,N, \end{align} \begin{align} \alpha\in\mathbb R_{+},\qquad \Delta_S(\cdot)\in\mathbb R. \end{align} % --------------------------------------------------------- % Key DFT-column identity (weighted exponential sum) % --------------------------------------------------------- % For any \mathbf a=[a_1,...,a_N]^T \in \mathbb C^{N\times 1} and any d\in{1,...,N}: % \sum_{\ell} e^{+j\omega(d-1)(\ell-1)} a_\ell = \sqrt N f_d^H a, % \sum_{\ell} e^{-j\omega(d-1)(\ell-1)} a_\ell = \sqrt N f_d^T a. \begin{align}\label{eq:DFT_weighted_sum_id} \sum_{\ell=1}^{N} e^{+j\omega(d-1)(\ell-1)}\,a_\ell =\sqrt N\,\mathbf f_d^{H}\mathbf a, \qquad \sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\,a_\ell =\sqrt N\,\mathbf f_d^{T}\mathbf a, \end{align} \begin{align} \mathbf a\in\mathbb C^{N\times 1},\qquad d\in\{1,\ldots,N\}. \end{align} % ========================================================= % Part A: linear functional -> eta[d](theta) (low-rank branch) % ========================================================= % Define modal quadratic forms: % q_\ell(\theta) \in \mathbb C, collect into \mathbf q(\theta)\in \mathbb C^{N\times 1}. \begin{align} q_\ell(\boldsymbol\theta)\triangleq \mathbf f_\ell^{H}\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell \in\mathbb C, \qquad \mathbf q(\boldsymbol\theta)\triangleq [q_1(\boldsymbol\theta),\ldots,q_N(\boldsymbol\theta)]^T\in\mathbb C^{N\times 1}. \end{align} % Define eta0 and eta[d] using 1-based shift amount (d-1): % eta0 \in \mathbb C, eta[d]\in \mathbb C. \begin{align} \eta_0(\boldsymbol\theta)\triangleq \frac{1}{N}\mathrm{tr}\!\big(\mathbf\Phi(\boldsymbol\theta)\big) =\frac{1}{N}\sum_{\ell=1}^{N} q_\ell(\boldsymbol\theta)\in\mathbb C. \end{align} \begin{align}\label{eq:eta_d_specsum} \eta[d](\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{\ell=1}^{N} e^{+j\omega(d-1)(\ell-1)}\,q_\ell(\boldsymbol\theta)\in\mathbb C, \qquad d\in\{1,\ldots,N\}. \end{align} % Apply the weighted-sum identity with \mathbf a=\mathbf q(\theta): \begin{align}\label{eq:eta_d_fHq} \eta[d](\boldsymbol\theta)=\frac{1}{N}\Big(\sqrt N\,\mathbf f_d^H\mathbf q(\boldsymbol\theta)\Big) =\frac{1}{\sqrt N}\,\mathbf f_d^H\mathbf q(\boldsymbol\theta),\qquad d\in\{1,\ldots,N\}. \end{align} % Define u_lr (length N, complex): \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq [u_{\mathrm{lr}}[1](\boldsymbol\theta),\ldots,u_{\mathrm{lr}}[N](\boldsymbol\theta)]^T\in\mathbb C^{N\times 1}, \end{align} \begin{align}\label{eq:u_lr_def} u_{\mathrm{lr}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta)\in\mathbb C,\quad d=2,\ldots,N. \end{align} % Substitute \eta[d] into u_lr: \begin{align}\label{eq:u_lr_fHq} u_{\mathrm{lr}}[d](\boldsymbol\theta)=\frac{\eta_0^*(\boldsymbol\theta)}{\sqrt N}\,\mathbf f_d^H\mathbf q(\boldsymbol\theta),\qquad d=2,\ldots,N. \end{align} % ========================================================= % Part B: quadratic functional -> xi[d](theta) (diagonal branch) % ========================================================= % Define modal energy responses: % t_\ell(\theta)\in\mathbb R_{+}, collect into \mathbf t(\theta)\in\mathbb R^{N\times 1}. \begin{align} t_\ell(\boldsymbol\theta)\triangleq \|\mathbf\Phi(\boldsymbol\theta)^H\mathbf f_\ell\|_2^2+\|\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell\|_2^2 \in\mathbb R_{+}, \qquad \mathbf t(\boldsymbol\theta)\triangleq [t_1(\boldsymbol\theta),\ldots,t_N(\boldsymbol\theta)]^T\in\mathbb R^{N\times 1}. \end{align} % (Optional) the same t_\ell(\theta) in lifted form (uses only f_\ell and Phi-tilde): \begin{align} t_\ell(\boldsymbol\theta) =\mathrm{tr}\!\Big(\big(\mathbf I\otimes \mathbf f_\ell\mathbf f_\ell^H+\mathbf f_\ell\mathbf f_\ell^H\otimes \mathbf I\big)\,\widetilde{\mathbf\Phi}(\boldsymbol\theta)\Big)\in\mathbb R_{+}. \end{align} % Define xi[d] via a negative-exponential spectral sum: \begin{align}\label{eq:xi_specsum} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\,t_\ell(\boldsymbol\theta)\in\mathbb C, \qquad d\in\{1,\ldots,N\}. \end{align} % Apply the weighted-sum identity with \mathbf a=\mathbf t(\theta) (real vector, but complex sum): \begin{align}\label{eq:xi_fdTt} \xi[d](\boldsymbol\theta) =\frac{1}{2N^2}\Big(\sqrt N\,\mathbf f_d^T\mathbf t(\boldsymbol\theta)\Big) =\frac{1}{2N^{3/2}}\,\mathbf f_d^T\mathbf t(\boldsymbol\theta), \qquad d\in\{1,\ldots,N\}. \end{align} % Define u_diag (length N, complex): \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq [u_{\mathrm{diag}}[1](\boldsymbol\theta),\ldots,u_{\mathrm{diag}}[N](\boldsymbol\theta)]^T\in\mathbb C^{N\times 1}, \end{align} \begin{align}\label{eq:u_diag_def} u_{\mathrm{diag}}[1](\boldsymbol\theta)\triangleq 0,\qquad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta)\in\mathbb C,\quad d=2,\ldots,N. \end{align} % ========================================================= % Two s-vectors (both are one DFT of kappa-weighted u) % ========================================================= % \mathbf s_lr(\theta), \mathbf s_diag(\theta) are complex N-vectors. \begin{align}\label{eq:s_lr_def} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) \triangleq \mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\Big) \in\mathbb C^{N\times 1}, \end{align} \begin{align}\label{eq:s_diag_def} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) \triangleq \mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\Big) \in\mathbb C^{N\times 1}. \end{align} % ========================================================= % Lifting operator S(\cdot) and the two explicit parts used in Q_sp % ========================================================= % \mathbf S(\mathbf s)\in\mathbb C^{N\times N}, Hermitian. \begin{align} \mathbf S:\mathbb C^{N\times 1}\to\mathbb C^{N\times N},\qquad \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T, \qquad [\mathbf S(\mathbf s)]_{n,m}=s_n+s_m^*. \end{align} % Two explicit liftings: \begin{align} \mathbf S\big(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times N}, \qquad \mathbf S\big(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times N}. \end{align} % The two contributions inside Q_k^{sp} (kept separate): \begin{align} \mathbf Q^{\mathrm{sp}}_{k,\mathrm{diag}}(\boldsymbol\theta) \triangleq \frac{1}{N^{3/2}}\Big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\Big) \in\mathbb C^{N\times N}, \end{align} \begin{align} \mathbf Q^{\mathrm{sp}}_{k,\mathrm{lr}}(\boldsymbol\theta) \triangleq \frac{\sqrt N}{2N}\Big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\Big) \in\mathbb C^{N\times N}. \end{align} % Here \mathbf A_k is constructed from the 1-based DFT column \mathbf f_k: % (so k \in {1,...,N} is the bin index) \begin{align} \mathbf A_k\triangleq \mathbf f_{k}\mathbf f_{k}^{T}+\mathbf f_{k}^{*}\mathbf f_{k}^{H}\in\mathbb C^{N\times N}, \qquad k\in\{1,\ldots,N\}. \end{align} % Finally: \begin{align} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta) =\mathbf Q^{\mathrm{sp}}_{k,\mathrm{diag}}(\boldsymbol\theta)+\mathbf Q^{\mathrm{sp}}_{k,\mathrm{lr}}(\boldsymbol\theta) \in\mathbb C^{N\times N}. \end{align} % ========================================================= % Q_bulk (1-based indexing, using DFT column f_k) % ========================================================= \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta}) =\frac{1}{N}\big(\mathbf{f}_{k}^*\mathbf{f}_{k}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}), \qquad \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})\in\mathbb C^{N\times N},\ \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\in\mathbb C^{N\times N},\ k\in\{1,\ldots,N\}. \end{align}
下面这个公式里面 实际上是、Phi矩阵的diag么?还是啥?q_\ell(\boldsymbol\theta)\triangleq \mathbf f_\ell^{H}\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell \in\mathbb C
不是 在原始坐标系里的 。
本质是 在向量 上的 Rayleigh 商/二次型。它等价于把 变换到 DFT 基底后得到矩阵的对角元素:
也就是说, 组成的向量就是
这是“ 在傅里叶域(DFT 基)里的对角线”,而不是 。
把它完全展开还能看到它一般会混进 的非对角项:
所以除非 本身有很强的结构(比如本来就对角、或循环/可被 DFT 对角化等),否则 不等于 。
两个特例帮助直觉:
如果你把你们 的具体结构(例如是否 或某种可分离形式)告诉我,我可以进一步把 化简成更短的闭式。
下面这两个,能不能改成用向量形式来写,相当于改成矩阵计算:
\begin{align}\label{eq:xi_fdTt}
\xid
=\frac{1}{2N^2}\Big(\sqrt N,\mathbf f_d^T\mathbf t(\boldsymbol\theta)\Big)
=\frac{1}{2N^{3/2}},\mathbf f_d^T\mathbf t(\boldsymbol\theta),
\qquad d\in{1,\ldots,N}.
\end{align}
\begin{align}\label{eq:u_lr_fHq}
u_{\mathrm{lr}}d=\frac{\eta_0^*(\boldsymbol\theta)}{\sqrt N},\mathbf f_d^H\mathbf q(\boldsymbol\theta),\qquad d=2,\ldots,N.
\end{align}
能,直接把“对每个 的标量内积”一次性堆成长度 的向量,就是一个 DFT 矩阵乘法(或转置/共轭转置):
你现在有(对所有 ):
把 定义出来,利用
就得到
(其中 ,所以 一般是复数向量。)
你现在有(对 ):
先把“全 ”都写出来:定义
于是
然后如果你们仍坚持把零差分项置零(1-based 的 ),就再乘一个选择矩阵/掩码向量即可:
如果你想同样把 也向量化(),它就是 的同一条,只是少了 的系数。
我把下面的q和t的结构也代入进上面的解析解中,看看能不能化简:\begin{align}
q_\ell(\boldsymbol\theta)\triangleq \mathbf f_\ell^{H}\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell \in\mathbb C,
\qquad
\mathbf q(\boldsymbol\theta)\triangleq [q_1(\boldsymbol\theta),\ldots,q_N(\boldsymbol\theta)]^T\in\mathbb C^{N\times 1}.
\end{align}
\begin{align}
t_\ell(\boldsymbol\theta)
=\mathrm{tr}!\Big(\big(\mathbf I\otimes \mathbf f_\ell\mathbf f_\ell^H+\mathbf f_\ell\mathbf f_\ell^H\otimes \mathbf I\big),\widetilde{\mathbf\Phi}(\boldsymbol\theta)\Big)\in\mathbb R_{+}.
\end{align}
可以化简,而且会变得更“矩阵运算化”。关键是:你给的 和 都是某个矩阵在 DFT 基底下的对角元素,所以 都能写成一个 diag(·)。
注意 ,所以
因此整条向量就是
并且
用 DC 列 还能写成
diag你给的
利用 的 rank-1 结构,等价于(我们之前用过的那条)
再把范数写成二次型:
令
(它是 Hermitian,且 ),则
所以
我们之前已经有
把 代入:
我们之前向量化是
把 代入:
(也可用上面那条 把 也写成 的线性泛函。)
在不额外假设 结构(例如循环、可分离、Toeplitz 等)的情况下,基本就到这一步了:你已经把一切压缩成
diag(·) 操作再往下要更短,通常得利用你们相位设计的具体形式(比如 这类会让 变成 rank-1,从而 还能再塌缩)。如果你们的 具体结构你给我一句,我可以把 和 再进一步闭式化。
我们的,看看能不能继续化简。就是我们这个Phi就是这个二阶的矩阵
能,而且在你们这个结构 下会大幅化简: 变成 rank-1 外积,于是 其实携带的是同一份信息。
令(全 1-based)
则
并且 .
你们定义
代入 :
因此整向量就是
这里 表示逐元素模平方。
顺便
你给的
等价于我们之前用的能量形式
在 下:
同理 。所以
结论:在你们这个相位结构下, 不再是独立对象,就是 乘常数。
你们已有
而 ,并且 ,所以
若你仍坚持“零差分项不计”,就再做一次掩码:,
等价的“时域直观”是:对 ,令移位量 ,
也就是 的归一化循环自相关。
你们已有
代入 得
但注意:在你们这个结构下, 的两项相关(沿 与沿 )互为共轭,所以 实际上等于相关的实部:
而 ,因此
(同样,若你要置零第 1 项,用 掩一下即可:。)
有了
其中 (1-based,),并且
那么两部分 lift 就是
如果你愿意,我可以把这套“”的推导,整理成你们 Appendix 里一段连续的证明文字(保持全 1-based,不引入新字母矩阵)。
把上面的公式给我latex代码
latex% ========================================================= % Special phase structure: Phi_{p,q} = e^{j(\theta_p-\theta_q)} % All indices are 1-based: {1,...,N}. No 0,...,N-1 anywhere. % Dimensions / fields are annotated inline. % ========================================================= % ---------- DFT ---------- \begin{align} \omega \triangleq \frac{2\pi}{N}\in\mathbb R,\qquad \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad n,\ell\in\{1,\ldots,N\}. \end{align} \begin{align} \mathbf f_\ell \triangleq \mathbf F(:,\ell)\in\mathbb C^{N\times 1},\qquad \ell\in\{1,\ldots,N\}. \end{align} % ---------- Wrapped index (1..N) ---------- \begin{align} \langle x\rangle_N \triangleq 1+\mathrm{mod}(x-1,N)\in\{1,\ldots,N\}. \end{align} % ---------- Phase vector and rank-1 phase matrix ---------- \begin{align} \boldsymbol\theta=[\theta_1,\ldots,\theta_N]^T\in\mathbb R^{N\times 1},\qquad \mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T\in\mathbb C^{N\times 1}. \end{align} \begin{align} \mathbf\Phi(\boldsymbol\theta)\in\mathbb C^{N\times N},\qquad \Phi_{p,q}(\boldsymbol\theta)\triangleq e^{j(\theta_p-\theta_q)}=u_p(\boldsymbol\theta)\,u_q(\boldsymbol\theta)^*, \qquad p,q\in\{1,\ldots,N\}. \end{align} \begin{align} \boxed{\ \mathbf\Phi(\boldsymbol\theta)=\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H\ }\quad\Rightarrow\quad \mathrm{rank}(\mathbf\Phi)=1,\ \ \|\mathbf u(\boldsymbol\theta)\|_2^2=\sum_{p=1}^N |e^{j\theta_p}|^2=N. \end{align} % ---------- Lifted 4th-order phase (optional; if you still reference it) ---------- \begin{align} \mathbf v(\boldsymbol\theta)\triangleq \mathrm{vec}(\mathbf\Phi(\boldsymbol\theta))\in\mathbb C^{N^2\times 1},\qquad \widetilde{\mathbf\Phi}(\boldsymbol\theta)\triangleq \mathbf v(\boldsymbol\theta)\mathbf v(\boldsymbol\theta)^H\in\mathbb C^{N^2\times N^2}. \end{align} % ========================================================= % q_\ell(\theta) and q(\theta) % ========================================================= \begin{align} q_\ell(\boldsymbol\theta)\triangleq \mathbf f_\ell^{H}\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell =\mathbf f_\ell^H\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H\mathbf f_\ell =\big|\mathbf u(\boldsymbol\theta)^H\mathbf f_\ell\big|^2 =\big|\mathbf f_\ell^H\mathbf u(\boldsymbol\theta)\big|^2 \in\mathbb R_{+}, \qquad \ell\in\{1,\ldots,N\}. \end{align} \begin{align} \mathbf q(\boldsymbol\theta)\triangleq [q_1(\boldsymbol\theta),\ldots,q_N(\boldsymbol\theta)]^T\in\mathbb R_{+}^{N\times 1}, \qquad \boxed{\ \mathbf q(\boldsymbol\theta)=\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\ }. \end{align} % ========================================================= % eta0(\theta) collapses to 1 % ========================================================= \begin{align} \eta_0(\boldsymbol\theta)\triangleq \frac{1}{N}\mathrm{tr}\big(\mathbf\Phi(\boldsymbol\theta)\big) =\frac{1}{N}\mathrm{tr}\big(\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H\big) =\frac{1}{N}\|\mathbf u(\boldsymbol\theta)\|_2^2 =\boxed{\,1\,}\in\mathbb R. \end{align} % ========================================================= % Weighted-sum identity (DFT columns) % ========================================================= \begin{align} \forall\,\mathbf a\in\mathbb C^{N\times 1},\ \forall\,d\in\{1,\ldots,N\}:\quad \sum_{\ell=1}^{N} e^{+j\omega(d-1)(\ell-1)}a_\ell=\sqrt N\,\mathbf f_d^H\mathbf a,\qquad \sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}a_\ell=\sqrt N\,\mathbf f_d^T\mathbf a. \end{align} % ========================================================= % eta[d](theta) in terms of q(theta) (linear functional) % ========================================================= \begin{align} \eta[d](\boldsymbol\theta) \triangleq \frac{1}{N}\sum_{q=1}^N e^{j(\theta_{\langle q+d-1\rangle_N}-\theta_q)} =\frac{1}{N}\sum_{\ell=1}^{N} e^{+j\omega(d-1)(\ell-1)}\,q_\ell(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf f_d^H\mathbf q(\boldsymbol\theta) \in\mathbb C, \qquad d\in\{1,\ldots,N\}. \end{align} % ========================================================= % u_lr[d](theta) collapses (since eta0=1) % ========================================================= \begin{align} \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq [u_{\mathrm{lr}}[1](\boldsymbol\theta),\ldots,u_{\mathrm{lr}}[N](\boldsymbol\theta)]^T\in\mathbb C^{N\times 1},\qquad u_{\mathrm{lr}}[d](\boldsymbol\theta)\triangleq \eta[d](\boldsymbol\theta)\eta_0^*(\boldsymbol\theta)=\eta[d](\boldsymbol\theta). \end{align} \begin{align} \boxed{\ \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)=\frac{1}{\sqrt N}\mathbf F^H\mathbf q(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf F^H\Big(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\Big)\ }\in\mathbb C^{N\times 1}. \end{align} % (Optional) if you still want to enforce "zero-difference removed" at d=1: \begin{align} \mathbf m\triangleq [0,1,\ldots,1]^T\in\mathbb R^{N\times 1},\qquad \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\leftarrow \mathbf m\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta). \end{align} % ========================================================= % t_\ell(\theta) collapses to 2N q_\ell(\theta) % ========================================================= \begin{align} t_\ell(\boldsymbol\theta) \triangleq \|\mathbf\Phi(\boldsymbol\theta)^H\mathbf f_\ell\|_2^2+\|\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell\|_2^2 \in\mathbb R_{+}. \end{align} \begin{align} \mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell=\mathbf u(\boldsymbol\theta)\big(\mathbf u(\boldsymbol\theta)^H\mathbf f_\ell\big) \ \Rightarrow\ \|\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell\|_2^2 =\|\mathbf u(\boldsymbol\theta)\|_2^2\big|\mathbf u(\boldsymbol\theta)^H\mathbf f_\ell\big|^2 =N\,q_\ell(\boldsymbol\theta). \end{align} \begin{align} \|\mathbf\Phi(\boldsymbol\theta)^H\mathbf f_\ell\|_2^2 =\|\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell\|_2^2 =N\,q_\ell(\boldsymbol\theta) \quad\Rightarrow\quad \boxed{\ t_\ell(\boldsymbol\theta)=2N\,q_\ell(\boldsymbol\theta)\ }. \end{align} \begin{align} \mathbf t(\boldsymbol\theta)\triangleq [t_1(\boldsymbol\theta),\ldots,t_N(\boldsymbol\theta)]^T\in\mathbb R_{+}^{N\times 1}, \qquad \boxed{\ \mathbf t(\boldsymbol\theta)=2N\,\mathbf q(\boldsymbol\theta)\ }. \end{align} % ========================================================= % xi[d](theta) becomes Re{eta[d](theta)} % ========================================================= \begin{align} \xi[d](\boldsymbol\theta) \triangleq \frac{1}{2N^2}\sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\,t_\ell(\boldsymbol\theta) =\frac{1}{2N^2}\sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\,(2N q_\ell(\boldsymbol\theta)) =\frac{1}{N}\sum_{\ell=1}^{N} e^{-j\omega(d-1)(\ell-1)}\,q_\ell(\boldsymbol\theta). \end{align} \begin{align} \xi[d](\boldsymbol\theta) =\frac{1}{N}\Big(\sqrt N\,\mathbf f_d^T\mathbf q(\boldsymbol\theta)\Big) =\frac{1}{\sqrt N}\mathbf f_d^T\mathbf q(\boldsymbol\theta) =\Re\!\left\{\frac{1}{\sqrt N}\mathbf f_d^H\mathbf q(\boldsymbol\theta)\right\} =\boxed{\ \Re\{\eta[d](\boldsymbol\theta)\}\ }\in\mathbb R, \qquad d\in\{1,\ldots,N\}. \end{align} % ========================================================= % Vector form of xi and u_diag % ========================================================= \begin{align} \boldsymbol\xi(\boldsymbol\theta)\triangleq[\xi[1](\boldsymbol\theta),\ldots,\xi[N](\boldsymbol\theta)]^T\in\mathbb R^{N\times 1}, \qquad \boxed{\ \boldsymbol\xi(\boldsymbol\theta)=\Re\!\left\{\frac{1}{\sqrt N}\mathbf F^H\mathbf q(\boldsymbol\theta)\right\} =\Re\!\left\{\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\right\}\ }. \end{align} \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq [u_{\mathrm{diag}}[1](\boldsymbol\theta),\ldots,u_{\mathrm{diag}}[N](\boldsymbol\theta)]^T\in\mathbb R^{N\times 1},\qquad u_{\mathrm{diag}}[d](\boldsymbol\theta)\triangleq \xi[d](\boldsymbol\theta). \end{align} % (Optional) remove d=1 term similarly: \begin{align} \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\leftarrow \mathbf m\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta),\qquad \mathbf m=[0,1,\ldots,1]^T\in\mathbb R^{N\times 1}. \end{align} % ========================================================= % Finally, s_lr and s_diag (kappa + DFT + theta via u) % ========================================================= \begin{align} \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1},\qquad \tilde\kappa[1]=0,\quad \tilde\kappa[d]=\exp\!\big(-\alpha\,\Delta_S(d-1)\big)\in\mathbb R_{+},\ d=2,\ldots,N. \end{align} \begin{align} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1},\qquad \mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1}. \end{align} % Lifting operator S(\cdot) (for completeness): \begin{align} \mathbf 1\in\mathbb R^{N\times 1},\qquad \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T\in\mathbb C^{N\times N},\qquad [\mathbf S(\mathbf s)]_{n,m}=s_n+s_m^*. \end{align}
我们现在用解析解而不是latex的方式讨论下,下面的两个u分别带入s,看看能不能让s的矩阵结构化简一下:
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1},\qquad
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1}.
\end{align}
可以,而且在你们这个特殊相位结构
下, 和 都能改写成“结构化矩阵 一个实非负向量”的形式,结构会比原来清楚很多。
下面都默认 1-based 索引 。
令相位向量
则
于是你之前定义的
直接化简为
把它堆起来:
同时
并且你之前的 会塌缩成
你刚才已经有向量化结果:
(若你们坚持“零差分项”置零,就是再做一次掩码 ,不影响下面结构结论。)
代入 :
这里的关键点:
等价地你也可以说:。
这已经是非常“结构化”的矩阵表达了:只剩一个 circulant 矩阵。
同样:
用 ,且 是实向量,所以
因此
这里出现了第二个结构矩阵
它的矩阵元素(你可以自己快速推一下)是
只依赖 (模 意义下),所以它是典型的**Hankel/反循环(anti-diagonal / sum-structured)**那一类结构(“按和索引结构化”),和 的“按差索引结构化(circulant)”正好互补。
所以结论是:
因为 对 是线性的,所以:
如果你想进一步“把 也变成纯 circulant 结构”,那就需要额外条件让 能被吸收到 里(比如对 或 施加某种共轭/对称性约束)。你告诉我你们 在 上有没有对称性(例如 这种 1-based 的“偶对称”),我可以继续把 和 的关系推到更紧。
我们之前的定义如下: A convenient representative reduction is obtained by evaluating the canonical slice , which yields
\begin{align}\label{eq:DeltaS_app}
\Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N,
\end{align}
where .
Accordingly, all structured hyperplane layers share the same 1D attenuation profile
\begin{align}\label{eq:kappa_app}
\kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp!\big(-\alpha,\Delta_S(d)\big),\qquad d=1,\ldots,N-1.
\end{align}
The subsequent matrix forms rely on the elementwise product identities of normalized DFT columns (indices modulo in the frequency argument):
\begin{align}\label{eq:DFT_id_app}
\mathbf f_{a+1}^*\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(b-a)+1},
\qquad
\mathbf f_{a+1}\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(a+b)+1}.
\end{align}
We also use the lifting operator
\begin{align}\label{eq:S_lift_app}
\mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T .
\end{align}
We next show that the four single-hyperplane families produce a diagonal layer plus a Hermitianized low-rank layer, both admitting the compact Hadamard representation in \eqref{eq:Qsp_prop}. The phase factor collapses to 1D aggregates under the same single-difference reduction. For (H1) with , the phase becomes ; summation over yields the main-diagonal mean
\begin{align}\label{eq:eta0_app}
\eta_0\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}(\boldsymbol{\theta}),
\end{align}
while the remaining dependence on under the single-difference parameterization is captured by the offset-diagonal mean
\begin{align}\label{eq:eta_app}
\eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}(\boldsymbol{\theta}).
\end{align}
The same 1D collapse occurs for (H2) with , leading to an identical effective phase weight. Consequently, the two low-rank hyperplanes share the phase-corrected 1D profile
\begin{align}\label{eq:kappa_lr_app}
\tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^*,\qquad d=1,\ldots,N-1,
\end{align}》另外,我们能不能用kron的方式,把S(s)这种结构给拆开,相当于不要这个套层,就用咱们现在写的s来直接表示
可以,而且用 Kronecker + vec 写完之后,你甚至可以把整个 这层“套层”完全消掉,变成一个对 与 的线性算子。
你们的 lifting:
用向量化恒等式 ,注意 ,有:
所以
其中 ,。
更“矩阵计算”的写法是把 与 拼起来:
这就等价于:你可以不再写 ,而把它看成一个固定的线性算子作用在 上。
你们有
用一个标准事实:Hadamard 乘法在 vec 下等价于对角矩阵乘法
于是把上面的 kron 拆解直接代入:
那么两部分分别变成
因此整条 的“无 版本”就是
这就是你说的“不要套层,就用我们现在写的 直接表示”,而且是纯矩阵乘法链路。
你们本来就有
写成矩阵乘法就是
然后上面第 2 步的 也能继续展开成
所以如果你愿意, 可以完全写成
“” 这种固定算子 的形式( 全部只在 里)。
你们 appendix 现在用的是 1D 塌缩的 ,这对应的就是把 写成 的那个向量;而上面 kron 拆解只是在“最后一步把 展开”,不影响你们前面的 推导,完全兼容。
如果你想进一步把 也直接简化成对角算子(避免 vec 维度 ),也可以:它就是 。但你刚才问的是“用 kron 拆开、不要套层”,上面这套是最统一、最矩阵化的一种写法。
我们不kron了 还是用矩阵形式来写,这样的话我们是否还是可以把S给拆掉,用里面的内容来直接表示,同时看看能否化简
可以,不用 kron 也能把 彻底拆掉,而且两部分都会明显“结构化”:
对任意 、,
这是纯矩阵形式(不需要 vec/kron)。
注意 。令 (因为 ),得到
所以
这一步已经把 完全消掉,并且告诉你:第一部分永远就是一个实对角层。
同理拆:
用上面的恒等式(取 或 ):
所以
又因为你们的 是 Hermitian(实际上元素是实数 cos),于是第二项等价写成
这就对应你们文字里说的 “Hermitianized low-rank layer”(因为 本身是 rank-2,乘个对角矩阵仍是低秩,再做 Hermitianize)。
把上面两条直接代回你给的式子:
得到
这就是你要的:仍然是矩阵形式,但已经不出现 这个“套层”。
用你们的 ,以及恒等式 ,有
所以第二项是一个**显式的“rank-2 + Hermitianize”**结构:
这和你们 appendix 里 DFT 列的 Hadamard 恒等式 非常一致:因为这里的“”就是调制/频移结构。
你现在的 还是
这一步一般不会把 再变成更简单的标准矩阵(因为“取对角”破坏了循环结构)。但在你们特例 下, 本身可以写成“circulant 核作用在一个实非负向量上”,这已经是很强的结构化(你前面我们推到的 )。
如果你想,我可以下一步把:
把上面分析的结果用latex代码的形式先给我
latex% ========================================================= % Remove the lifting operator S(\cdot) in pure matrix form (no kron/vec). % All indices are 1-based. Dimensions/fields are annotated inline. % ========================================================= % --- Objects --- % \mathbf s_{\mathrm{diag}}(\theta), \mathbf s_{\mathrm{lr}}(\theta) \in \mathbb C^{N\times 1} % \mathbf 1 \in \mathbb R^{N\times 1}, \mathbf I \in \mathbb R^{N\times N} % \mathbf A_k \in \mathbb C^{N\times N} (Hermitian; in fact real-valued entries for DFT construction) % \mathrm{diag}(\cdot): vector -> diagonal matrix % --- Lifting --- \begin{align} \mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T\in\mathbb C^{N\times N},\qquad \mathbf s\in\mathbb C^{N\times 1},\ \mathbf 1\in\mathbb R^{N\times 1}. \end{align} % ========================================================= % Identity: Hadamard with an outer product equals diagonal scaling % For any \mathbf M\in\mathbb C^{N\times N}, \mathbf a,\mathbf b\in\mathbb C^{N\times 1}: % \mathbf M \odot (\mathbf a\mathbf b^T)=\mathrm{diag}(\mathbf a)\mathbf M\mathrm{diag}(\mathbf b). % ========================================================= \begin{align} \mathbf M\odot(\mathbf a\mathbf b^T)=\mathrm{diag}(\mathbf a)\,\mathbf M\,\mathrm{diag}(\mathbf b). \end{align} % ========================================================= % Part 1: I \odot S(s_diag) becomes a real diagonal layer % ========================================================= \begin{align} \mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}) &=\mathbf I\odot(\mathbf 1\mathbf s_{\mathrm{diag}}^H)+\mathbf I\odot(\mathbf s_{\mathrm{diag}}\mathbf 1^T) \\ &=\mathrm{diag}(\mathbf s_{\mathrm{diag}}^*)+\mathrm{diag}(\mathbf s_{\mathrm{diag}}) \\ &=\mathrm{diag}(\mathbf s_{\mathrm{diag}}+\mathbf s_{\mathrm{diag}}^*) \\ &=2\,\mathrm{diag}\!\big(\Re\{\mathbf s_{\mathrm{diag}}\}\big)\in\mathbb R^{N\times N}. \end{align} % ========================================================= % Part 2: A_k \odot S(s_lr) becomes "Hermitianized diagonal-scaled low-rank" % ========================================================= \begin{align} \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) &=\mathbf A_k\odot(\mathbf 1\mathbf s_{\mathrm{lr}}^H)+\mathbf A_k\odot(\mathbf s_{\mathrm{lr}}\mathbf 1^T) \\ &=\mathbf A_k\,\mathrm{diag}(\mathbf s_{\mathrm{lr}}^*)+\mathrm{diag}(\mathbf s_{\mathrm{lr}})\,\mathbf A_k \in\mathbb C^{N\times N}. \end{align} If $\mathbf A_k$ is Hermitian (true for $\mathbf A_k$ built from DFT columns), then equivalently \begin{align} \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) =\mathrm{Herm}\!\big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k\big) \triangleq \mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k\big)^H. \end{align} % ========================================================= % Substitute into Q_k^{sp} and eliminate S(\cdot) completely % ========================================================= \begin{align} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta})\in\mathbb C^{N\times N} &= \frac{1}{N^{3/2}}\big[\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big] +\frac{\sqrt N}{2N}\big[\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big] \\ &= \frac{2}{N^{3/2}}\mathrm{diag}\!\big(\Re\{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\}\big) +\frac{\sqrt N}{2N}\Big( \mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\,\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)^*) \Big). \end{align} % ========================================================= % Further low-rank structure using A_k from DFT columns (1-based) % ========================================================= % \mathbf f_k \in \mathbb C^{N\times 1}, k\in{1,...,N} % \mathbf A_k = \mathbf f_k\mathbf f_k^T + \mathbf f_k^*\mathbf f_k^H \begin{align} \mathbf f_k\in\mathbb C^{N\times 1},\qquad \mathbf A_k\triangleq \mathbf f_k\mathbf f_k^T+\mathbf f_k^*\mathbf f_k^H\in\mathbb C^{N\times N},\qquad k\in\{1,\ldots,N\}. \end{align} Using $\mathrm{diag}(\mathbf s)\mathbf f=\mathbf s\odot \mathbf f$, we have \begin{align} \mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k =(\mathbf s_{\mathrm{lr}}\odot \mathbf f_k)\mathbf f_k^T+(\mathbf s_{\mathrm{lr}}\odot \mathbf f_k^*)\mathbf f_k^H, \end{align} and thus \begin{align} \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) =\mathrm{Herm}\!\Big((\mathbf s_{\mathrm{lr}}\odot \mathbf f_k)\mathbf f_k^T+(\mathbf s_{\mathrm{lr}}\odot \mathbf f_k^*)\mathbf f_k^H\Big), \end{align} which is explicitly a Hermitianized (at most) rank-2 construction.
不不不 你只需要把附件中图片里的公式用latex代码给我就好
pythonfrom PIL import Image, ImageEnhance img=Image.open('/mnt/data/image.png') img.size (733, 1099)
pythonimport matplotlib.pyplot as plt, numpy as np plt.figure(figsize=(7,10)) plt.imshow(img) plt.axis('off') plt.show()
latex% 1) Hadamard 乘外积 = 左右对角缩放 \begin{align} \text{对任意 }\mathbf M\in\mathbb C^{N\times N},\ \mathbf a,\mathbf b\in\mathbb C^{N},\qquad \mathbf M \odot (\mathbf a\mathbf b^{T})=\mathrm{diag}(\mathbf a)\,\mathbf M\,\mathrm{diag}(\mathbf b). \end{align} % 2) 直接把 \mathbf S(\mathbf s)=\mathbf 1\mathbf s^{H}+\mathbf s\mathbf 1^{T} 拆掉 \begin{align} \mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}) &=\mathbf I\odot(\mathbf 1\mathbf s_{\mathrm{diag}}^{H})+\mathbf I\odot(\mathbf s_{\mathrm{diag}}\mathbf 1^{T}),\\ \mathbf I\odot(\mathbf 1\mathbf s_{\mathrm{diag}}^{H}) &=\mathrm{diag}(\mathbf s_{\mathrm{diag}}^{*}),\qquad \mathbf I\odot(\mathbf s_{\mathrm{diag}}\mathbf 1^{T}) =\mathrm{diag}(\mathbf s_{\mathrm{diag}}),\\ \mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}) &=\mathrm{diag}(\mathbf s_{\mathrm{diag}}+\mathbf s_{\mathrm{diag}}^{*}) =2\,\mathrm{diag}\!\big(\Re\{\mathbf s_{\mathrm{diag}}\}\big). \end{align} % (b) 第二项:\mathbf A_k \odot \mathbf S(\mathbf s_{\mathrm{lr}}) \begin{align} \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) &=\mathbf A_k\odot(\mathbf 1\mathbf s_{\mathrm{lr}}^{H})+\mathbf A_k\odot(\mathbf s_{\mathrm{lr}}\mathbf 1^{T}),\\ \mathbf A_k\odot(\mathbf 1\mathbf s_{\mathrm{lr}}^{H}) &=\mathrm{diag}(\mathbf 1)\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^{*}) =\mathbf A_k\,\mathrm{diag}(\mathbf s_{\mathrm{lr}}^{*}),\\ \mathbf A_k\odot(\mathbf s_{\mathrm{lr}}\mathbf 1^{T}) &=\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k\mathrm{diag}(\mathbf 1) =\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k,\\ \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) &=\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\,\mathrm{diag}(\mathbf s_{\mathrm{lr}}^{*}). \end{align} \begin{align} \mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}) &=\mathrm{Herm}\!\big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k\big) \ \triangleq\ \mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k\big)^{H}. \end{align}
我们用公式的角度来讨论,s_diag是否可以用s_ul来表示:
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1},\qquad
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\big)\in\mathbb C^{N\times 1}.
\end{align}
\begin{align}
\ \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)
=\frac{1}{\sqrt N}\mathbf F^H\Big(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\Big)\ \in\mathbb C^{N\times 1}.
\end{align}
\begin{align}
u_{\mathrm{diag}}(\boldsymbol\theta)=\Re!\left{\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\right}\ .
\end{align}
pythonimport numpy as np, cmath def Fmat(N): w=2*np.pi/N n=np.arange(N)[:,None] l=np.arange(N)[None,:] return (1/np.sqrt(N))*np.exp(-1j*w*n*l) def Pi_freq(N): # permutation k->(-k mod N) in 0-based P=np.zeros((N,N)) for k in range(N): k2=(-k)%N P[k2,k]=1 return P for N in [4,5,8]: F=Fmat(N) P=Pi_freq(N) C=F@P@F.conj().T # Check if C is permutation of time reversal n->(-n mod N) R=np.zeros((N,N)) for n in range(N): n2=(-n)%N R[n2,n]=1 print(N, np.allclose(C,R))
可以,而且在你们当前这组定义下(尤其是 为实向量、以及 关于 偶对称), 确实可以只用 写出来,并且会出现一个非常干净的“时域反转取平均”结构。
由你给的定义
用 :
所以一般情况下,要从 得到 ,你还得处理那项 。
你们这里
注意 是 逐元素模平方,因此它是实向量(且非负)。设
那么 。对“IDFT 作用在实向量”有标准性质:输出满足共轭对称,等价地存在一个“频域反转”置换矩阵 使得
其中(1-based) 对应索引映射
你们 是由 (余弦和)构成, 是偶函数,所以 也是偶的,从而
换成你们现在 1-based 的向量 (对应移位量 )就是
这等价于
回到第 1 步:
代入 且 与 可交换:
又因为 ,所以
而 恰好是时域的循环反转置换矩阵(不是共轭!),满足
因此更直观的分量形式就是
在你们当前的设定下(,并且 偶对称), 就是 的“循环反转后取平均”(偶对称化),所以可以完全用 来表示,而且结构非常简单。
如果你希望我把这里用到的两个关键性质(“ 作用在实向量 共轭对称”以及“ 偶对称”)用你们 appendix 的 记号(0-based)严谨地对应到 1-based 写法里,我也可以继续把那段推导补齐。
下面公式中,\tilde{\boldsymbol\kappa}的解析解是什么?可以给出其中第d个元素,然后用向量构成
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big),\quad
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\big).
\end{align}
本质上就是你们在附录里定义的那条 1D 衰减谱 (把它收集成一个长度为 的向量),其中每个元素由 的解析式给出。
令 ,并定义
则
你们原始定义是(0-based):
把 代入得到显式:
(因此 ,且通常是偶对称的: 在 意义下。)
把它组成向量就是
如果你现在所有索引都用 ,那就对应一个 1-based 向量 :
于是
把这条 代回你给的
就得到完全解析的“核衰减向量”。
如果你告诉我你现在的正文最终是用 0-based()还是 1-based(),我可以把两者在全文里的符号统一成你们最终版本(避免 到处出现)。
我们现在是1-based。请把第三部分的两个公式给我latex代码的形式
latex% 1-based attenuation profile vector \tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1} % with \omega = 2\pi/N and \Psi(0)=N_d. \begin{align} \tilde{\kappa}[1]\triangleq 0,\qquad \tilde{\kappa}[d]\triangleq \exp\!\left( -\alpha(4N_d+2)\left( N_d-\sum_{r=1}^{N_d}\cos\!\big(\omega r (d-1)\big) \right) \right),\quad d=2,\ldots,N. \end{align} \begin{align} \tilde{\boldsymbol{\kappa}} \triangleq \big[\tilde{\kappa}[1],\tilde{\kappa}[2],\ldots,\tilde{\kappa}[N]\big]^T \in\mathbb R^{N\times 1}. \end{align}
下面这个解析解,能化简么?在不引入别的变量情况下,先用文本结构分析:\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F\big(\tilde{\boldsymbol\kappa}\odot (\mathbf F^H(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}))\big),
\end{align}
能化简,而且可以把它看成“一个固定的(由 决定的)线性滤波矩阵”作用在 上。
你的式子是
把“逐点乘”写成对角矩阵就是:
所以整条可以直接合并成一个矩阵乘法链(不引入新变量名也行):
这一步的意义是:原来“先 ,再逐点乘 ,再 ”这三步,被合并成了一个DFT 可对角化(循环)算子 。
你们这里 是逐元素模平方,所以它是实非负向量。
另外 由余弦和构成(你们的 里就是 ),因此 是实数且(循环意义下)偶对称。这会推出
是一个实对称的循环矩阵。
于是它作用在实向量 上,结果仍是实向量,因此
这个“实值性”在你后面拆 的时候会进一步省掉很多共轭号(例如 )。
在不引入额外新符号的前提下,你这条解析解最简洁、最结构化的等价写法就是:
如果你接下来想用这个结果去进一步压缩 (例如写成 的“反转平均”那条),我们就可以沿着“实值 + 偶对称”继续往下推。
再看看下面这个解析解怎么化简:
\begin{align}
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf \Re!\left{\mathbf \frac{1}{\sqrt N}\mathbf F^H(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}),, u_{\mathrm{diag}}(\boldsymbol\theta)\right}\big).
\end{align}
你这个式子里 \Re{ ... , u_{\mathrm{diag}} } 这一坨看起来像是排版误写(\Re{·} 里不该再塞一个 u_{\mathrm{diag}})。按你们前面一致的定义,它应当就是
在不引入新“符号变量”的前提下,化简可以分两步做:
又因为 是实向量,所以
从而得到一个很干净的矩阵式:
你已经有
所以立刻得到
这已经是“结构化”最明显的一步:第一项就是 ,第二项是把同一个实谱 经过 分支再走一遍同样的 滤波。
由于你们的 来自余弦和,通常满足 1-based 的偶对称
,这会把上面那项“”分支变成“对 做循环反转”。于是可以写成分量级的超简洁形式:
则
如果你想把这一步也严格写进文稿,我建议你先确认/声明 的“偶对称(模 )”性质;从你给的 (余弦和)来看这通常是成立的。
把第一步压缩出来的东西,逐点乘写成对角矩阵然后给real展开后的解析解给我latex代码
latex% Assume the consistent definition: % u_{\mathrm{diag}}(\boldsymbol\theta)=\Re\left\{\frac{1}{\sqrt N}\mathbf F^H\left(\left|\mathbf F^H\mathbf u(\boldsymbol\theta)\right|^{\odot 2}\right)\right\}, % and s_{\mathrm{diag}}(\boldsymbol\theta)=\mathbf F\left(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\right). \begin{align} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) &=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \Re\Big\{\frac{1}{\sqrt N}\mathbf F^H\big(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\big)\Big\}\Big) \\ &=\mathbf F\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\; \Re\Big\{\frac{1}{\sqrt N}\mathbf F^H\big(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\big)\Big\}. \end{align} % Since \big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\in\mathbb R^{N\times 1}, % we use Re{ \mathbf F^H \mathbf x } = 1/2 (\mathbf F^H + \mathbf F^T)\mathbf x for real \mathbf x: \begin{align} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) =\frac{1}{2\sqrt N}\, \mathbf F\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\big(\mathbf F^H+\mathbf F^T\big)\, \Big(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2}\Big). \end{align}
下面这个部分我们可以不可以单独定义下,就是一个和theta相关额度矩阵或者向量,可以给他一个定义,比如什么频域相位这种,然后给一个矩阵形式,同时说明其中单个元素的解析解,先用文本形式;
(\big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2})
可以单独拿出来定义,而且它确实就是一个很自然的“频域相位能量/功率谱”量:把相位向量 做 DFT(更准确是 ),再取逐点模平方。
我们定义一个与 相关的 频域相位能量向量(或“频域相位功率谱向量”):
它的物理含义: 是相位序列 的离散频谱系数(归一化 DFT), 则是其逐频点能量(功率)分布。
记 ,。则对任意 ,
因此
(顺带一个常用性质:,因为 是酉矩阵。)
把上面的公式分析,用latex代码给我
latex% ========================================================= % Define the theta-dependent "frequency-domain phase energy (power spectrum)" vector % 1-based indexing: n,\ell \in {1,...,N} % ========================================================= \begin{align} \mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T\in\mathbb C^{N\times 1},\qquad \omega\triangleq \frac{2\pi}{N}\in\mathbb R. \end{align} % Normalized DFT (for reference) \begin{align} \mathbf F\in\mathbb C^{N\times N},\qquad [\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\quad n,\ell\in\{1,\ldots,N\}. \end{align} % Frequency-domain phase energy / power spectrum vector \begin{align} \mathbf p(\boldsymbol\theta)\triangleq \big|\mathbf F^H\mathbf u(\boldsymbol\theta)\big|^{\odot 2} \in\mathbb R_{+}^{N\times 1}. \end{align} % Equivalent matrix/elementwise forms \begin{align} \mathbf p(\boldsymbol\theta) =\big(\mathbf F^H\mathbf u(\boldsymbol\theta)\big)\odot\big(\mathbf F^H\mathbf u(\boldsymbol\theta)\big)^*. \end{align} % Single-element closed form (1-based) \begin{align} \big[\mathbf F^H\mathbf u(\boldsymbol\theta)\big]_{\ell} =\frac{1}{\sqrt N}\sum_{n=1}^{N} e^{+j\omega (n-1)(\ell-1)}\,e^{j\theta_n}, \qquad \ell\in\{1,\ldots,N\}, \end{align} \begin{align} p_{\ell}(\boldsymbol\theta)\triangleq \big[\mathbf p(\boldsymbol\theta)\big]_{\ell} =\left|\frac{1}{\sqrt N}\sum_{n=1}^{N} e^{j\left(\theta_n+\omega (n-1)(\ell-1)\right)}\right|^{2} \in\mathbb R_{+}. \end{align} % (Optional) energy conservation (Parseval / unitarity of F) \begin{align} \sum_{\ell=1}^{N} p_{\ell}(\boldsymbol\theta)=\|\mathbf u(\boldsymbol\theta)\|_2^2=N. \end{align}
把\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F \text{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta),
\end{align}
带入
\begin{align}
\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol{\theta})=
\frac{2}{N^{3/2}}\mathrm{diag}\big[\Re{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)}\big]
\end{align}
看看能不能再化简
可以化简到“只剩一个对角算子作用在 ”的形式;但因为外面有 diag(·)(取向量做对角矩阵),所以再往里把 等完全乘开意义不大——diag 会把所有非对角结构都“截断”掉。
下面给你两种等价的化简路线(第二种会把 用 表示,这样就真的是“把 带入 ”)。
你之前已经有():
代入
得到系数合并:
在你们这里 实且偶对称、 实,最终 通常也是实向量,因此很多时候可以进一步写成
你给了
并且在你们当前设定( 实、 偶对称)下,我们前面推过:
其中 是“1-based 循环反转”算子(分量上就是 ,, for )。
于是
代入 :
再把你的 代进去,系数同样合并成 :
如果你更想看“单个对角元”的解析式(这通常最清楚):
在不再引入新结构(比如把 命名成一个循环卷积矩阵)且保留 diag(·) 的前提下,上面基本就是“能化到的最干净形式”了:diag 会阻止你把内部矩阵乘法进一步合并成更短的闭式(因为最终只取结果向量的分量)。
如果你愿意允许给 起一个名字(比如“核矩阵/循环平滑矩阵”),那表达还能更短一行。
我们分别把下面的s带入Q,尝试化简:
\begin{align}
\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol{\theta})=
\frac{2}{N^{3/2}}\mathrm{diag}\big[\Re{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)}\big]
\end{align}
\begin{equation}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta})= \frac{\sqrt N}{2N}\Big[\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)^*)\Big].
\end{equation}
需要被带入的:
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F \text{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta),
\end{align}
\begin{align}
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)
=\frac{1}{2\sqrt N},
\mathbf F,\mathrm{diag}(\tilde{\boldsymbol\kappa})\big(\mathbf F^H+\mathbf F^T\big)\tilde{\mathbf p}(\boldsymbol\theta).
\end{align}
可以,把常数先合并、再把 的解析式直接代进去,两项都会更紧凑。
已知
系数合并:
所以
(如果你们后面明确 ,那就可把 直接去掉。)
已知
注意 ,因此
外面系数合并:
所以
如果进一步使用你们的常见事实( 实且偶对称、 实)从而推出 ,则还可再缩一行:
如果你下一步想把 合并成一个更“统一”的结构(例如显式 Hermitianize 形式、或显式 rank-2 形式),我也可以继续往下推。
把上面的公式结果给我latex代码,然后再合成一个更统一的结果,这一步用文本给我
latex% ========================= % Substitute s_diag into Q_k^{sp1} % ========================= \begin{align} \mathbf Q_k^{\mathrm{sp1}}(\boldsymbol{\theta}) &=\frac{2}{N^{3/2}}\mathrm{diag}\!\big(\Re\{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\}\big),\\ \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) &=\frac{1}{2\sqrt N}\, \mathbf F\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\big(\mathbf F^H+\mathbf F^T\big)\tilde{\mathbf p}(\boldsymbol\theta), \end{align} hence \begin{align} \boxed{ \mathbf Q_k^{\mathrm{sp1}}(\boldsymbol{\theta}) =\frac{1}{N^{2}}\; \mathrm{diag}\!\left( \Re\!\left\{ \mathbf F\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\big(\mathbf F^H+\mathbf F^T\big)\tilde{\mathbf p}(\boldsymbol\theta) \right\} \right). } \end{align} % ========================= % Substitute s_lr into Q_k^{sp2} % ========================= \begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta}) &= \frac{\sqrt N}{2N}\Big[\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)^*)\Big],\\ \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) &= \frac{1}{\sqrt N}\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta), \end{align} hence \begin{align} \boxed{ \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta}) =\frac{1}{2N}\Big[ \mathrm{diag}\!\Big(\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta)\Big)\mathbf A_k +\mathbf A_k\,\mathrm{diag}\!\Big(\big(\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta)\big)^*\Big) \Big]. } \end{align}
统一写法(文本说明)
把两项都改写成“对角缩放 + Hermitianize”就最统一:
diag(·) 把它们注入到一个对角层 + 一个 Hermitianized 低秩层”里。下面这个公式能不能写的再简短一些,现在这个太长啦,两个和能放一起么?\begin{align}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta})
=\frac{1}{2N}\Big[
\mathrm{diag}!\Big(\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta)\Big)\mathbf A_k
+\mathbf A_k,\mathrm{diag}!\Big(\big(\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf{F}^H\tilde{\mathbf p}(\boldsymbol\theta)\big)^*\Big)
\Big].
\end{align}
可以,把两个和**合并成一次 Hermitian 化(Hermitianization)**就行;前提是你们这里的 是 Hermitian(按你们定义 确实是 Hermitian)。
记
因为 ,有
所以
latex\begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta}) =\frac{\sqrt N}{2N}\,\mathrm{Herm}\!\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}}(\boldsymbol{\theta}))\,\mathbf A_k\Big). \end{align}
令里面重复出现的向量就是 ,则
latex\begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol{\theta}) =\frac{1}{2N}\,\mathrm{Herm}\!\Big( \mathrm{diag}\!\big(\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta)\big)\,\mathbf A_k \Big). \end{align}
如果你希望连 这个记号都不引入,那就只能保留“”的形式(本质还是同一个合并)。
下面这部分我想用一个变量来表示,能再化简一些么\mathbf F \mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta)
可以,而且这块本身就是一个**“用 做谱加权的循环滤波器”**作用在 上——所以用一个变量表示会非常自然、也会让后面的 写法短很多。
直接定义
这样你想简化的那一段就变成一句:
并且单个元素是
解释: 是“相位序列的频域能量”, 就是把这份能量用 做谱加权后再变回时域得到的“滤波输出”。
因为
只由 决定、与 无关,你也可以定义
这里 是一个 circulant(循环)矩阵(特征值就是 ),所以它等价于对 做一次循环卷积/平滑。
如果你告诉我你更倾向“方案 1(直接定义 )”还是“方案 2(先定义固定核矩阵 )”,我就可以顺手把后面 、 都改写成最短的统一形式。
我们用方案2
好,用方案 2 的话就非常干净了:把这坨当成一个固定(与 无关)的循环核矩阵去作用在 。
定义一个仅由 决定的矩阵
那么你要简化的部分就直接变成
单个元素(矩阵形式展开)就是
因此
结构含义: 只依赖 (模 ),所以 是 circulant(循环)矩阵;等价于对 做一个由 指定的循环卷积/平滑。
kappa是如下定义的,看看能不能化简下,然后我们把这个C和p的化简结果带入两个·Q
可以化简,关键点有两个:
你们有(0-based)
, 。
这条余弦和有标准闭式(Dirichlet kernel 形式):
对应到你们现在 1-based 的 (差分量是 ):
其中 用上面的闭式即可(当 时你已经直接置零了,不会遇到分母为零的问题)。
重要结构: 因为 是偶函数,所以 偶对称 偶对称
你选的方案 2:
其元素为
由于 偶对称且实数,上式的虚部成对抵消,所以还能写成纯余弦型:
从而 是 实对称,且只依赖 (mod )所以是 circulant。
再因为你定义的
所以
你给的
因此 也是实向量(因为 实)。
原式
代入 且 ,得到
(若你们用 记号,还可写成 。)
你已有
在 偶对称 + 实 的条件下,这一项等价于对 做“反转偶对称化”,也就是
于是
其中 。
代回
系数合并后得到最短的版本:
如果你愿意,我可以把这两条再合并成一个“对角层 + Hermitianized 层”的统一表达,只用 (以及可选的反转操作 )来写。
把这两条再合并成一个“对角层 + Hermitianized 层”的统一表达
在你们当前的假设( 且在 意义下偶对称;)下,可以把 统一写成“对角层 + Hermitianized 层”,只用同一个中间向量 。
再定义 1-based 的“循环反转”算子(写成置换矩阵也行):
令 。则
其中 。
如果你不想引入 记号,也可等价写成
这就是你要的“合并成一个更统一的结果”:同一条 同时驱动对角层与 Hermitianized 层。
把上面推导的公式给我latex代码
latex% ========================================================= % Unified form: diagonal layer + Hermitianized low-rank layer % (1-based indexing) % ========================================================= % --- Definitions --- \begin{align} \mathbf C_{\tilde{\kappa}} &\triangleq \mathbf F\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\,\mathbf F^H \in \mathbb C^{N\times N},\\ \mathbf v(\boldsymbol\theta) &\triangleq \mathbf C_{\tilde{\kappa}}\,\tilde{\mathbf p}(\boldsymbol\theta)\in\mathbb R^{N\times 1}. \end{align} % --- 1-based circular reversal operator (permutation) --- \begin{align} (\mathbf R\mathbf x)[n] \triangleq \mathbf x[\mathrm{rev}(n)],\qquad \mathrm{rev}(1)=1,\ \ \mathrm{rev}(n)=N-n+2,\quad n=2,\ldots,N, \end{align} for any $\mathbf x\in\mathbb C^{N\times 1}$. % --- Hermitianization operator --- \begin{align} \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H. \end{align} % --- Unified structured form --- \begin{align} \boxed{ \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta) = \underbrace{\frac{1}{N^2}\,\mathrm{diag}\!\big((\mathbf I+\mathbf R)\mathbf v(\boldsymbol\theta)\big)}_{\text{diagonal layer}} \;+\; \underbrace{\frac{1}{2N}\,\mathrm{Herm}\!\Big(\mathrm{diag}\big(\mathbf v(\boldsymbol\theta)\big)\,\mathbf A_k\Big)}_{\text{Hermitianized low-rank layer}} } \end{align} % --- Equivalent expanded (no Herm(\cdot)) form --- \begin{align} \mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta) = \frac{1}{N^2}\,\mathrm{diag}\!\big((\mathbf I+\mathbf R)\mathbf v(\boldsymbol\theta)\big) +\frac{1}{2N}\Big(\mathrm{diag}(\mathbf v(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\,\mathrm{diag}(\mathbf v(\boldsymbol\theta))\Big). \end{align}
\mathbf 是什么矩阵
这里的 不是随便的矩阵,它是一个 固定的置换矩阵(permutation matrix),表示 1-based 的“循环反转 / 索引翻转”操作。
对任意向量 , 的作用定义为
其中(1-based)
也就是说:
,其元素为:
这是一个典型的 反对角置换矩阵(带 1-based 的特殊固定点)。
所以:
作用在 上:
因为你们构造里
会导致输出向量 也满足一种“反转对称”结构,于是
自然出现 来做“反转取平均”。
如果你愿意,我也可以给你一个 不引入 的等价写法(用 直接写每个元素),方便你写论文时更“透明”。
\underbrace{\frac{1}{N^2},\mathrm{diag}!\big((\mathbf I+\mathbf R)\mathbf v(\boldsymbol\theta)\big)}{\text{diagonal layer}}
;+;
\underbrace{\frac{1}{2N},\mathrm{Herm}!\Big(\mathrm{diag}\big(\mathbf v(\boldsymbol\theta)\big),\mathbf A_k\Big)}{\text{Hermitianized low-rank layer}}
}
\end{align}这里面的herm有没有在不写的更长的情况下换一种用正常矩阵计算的表达
可以的,而且不用把式子写得更长。关键是把
换成一个标准矩阵恒等式,而不是展开成两项。
因为你们这里
有恒等式
于是你的统一表达可以 直接改写为:
diag(·)、矩阵乘法、加法Herm(·) 换成了一对自然对称项Herm(·)Herm 很清楚如果你愿意,我也可以帮你给论文里两种写法各放一次(主文用 Herm,附录给展开式),保持可读性和严谨性兼顾。
用符合TSP的语言风格,用一句话介绍一下下面这个向量。其中theta是我们可设计的时域相位:
\mathbf{p}(\boldsymbol{\theta}) = \mathbf F^H\mathbf u(\boldsymbol\theta)\in\mathbb{C}^{N\times1}
可以这样表述(符合 TSP / 信号处理论文 的语言风格,简洁且标准):
denotes the frequency-domain representation of the time-domain phase sequence , where is the designable time-domain phase vector.
如果你想再偏“设计导向”一点,也可以用这一句(更偏 TSP/optimization):
represents the DFT-domain coefficients induced by the designable time-domain phase vector .
需要我帮你把这一句话和后面 的“能量谱”解释自然衔接成一段正文吗?
我们直接介绍下面这个能量谱的p:,
当然可以。下面给你一句 TSP / 信号处理论文风格、简洁且标准的表述(直接介绍能量谱 ):
denotes the frequency-domain energy (power) spectrum of the phase-coded sequence induced by the designable time-domain phase vector .
如果你想更强调“逐频点能量”这一点,可以用更明确的一句:
collects the per-frequency energy of , i.e., .
(顺带提醒: 实际上是 ,不是 ,如果你愿意我也可以帮你把这句写得更严谨。)
润色优化一下下面的表达,另外告诉我为什么是R不是C?and denotes the frequency power spectrum of the phase vector to be optimized.
下面我先给你一个更符合 TSP/信号处理论文风格、更顺畅且更严谨的改写版本;然后解释你问的:** 为什么应该写 而不是 **。
版本 A(最常用、最像论文的一句话):
and denotes the frequency-domain power spectrum induced by the designable time-domain phase vector .
版本 B(强调逐频点能量收集,更清楚一点):
and collects the per-frequency power of the DFT coefficients of the phase-coded sequence , where is the design variable.
你原句写的是
但实际上应当是
原因很简单:
所以:
and denotes the frequency-domain power spectrum induced by the designable phase vector .
如果你希望我把这句话和你们后续定义的 连成一小段“过渡段落”,让论文读起来更自然,我也可以顺手帮你写一段。
按照上面的要求补全一下下面的denotes后面的内容:
and denotes, where .
可以按与前面对 一致的 TSP/信号处理论文风格这样补全并润色(同时把语义说清楚):
and denotes the frequency-domain attenuation (weighting) profile associated with the structured hyperplane distance, whose -th entry is given by
如果你想稍微多强调一点“它的作用”,也可以用这一版(更偏建模解释):
and denotes a frequency-dependent attenuation kernel determined by the structured hyperplane distance, with entries , .
这两种都和你前面对“frequency-domain power spectrum”的表述在风格和抽象层次上是完全一致的。
基于上面的推导思路 我们做了如下的proposition,现在我们要对这个proposition写一个proof 先用文本和解析解的方法写一下中文版本的, proposition:
\begin{proposition}\label{prop:Sk_closed_form}
Consider the transmission of random CE-OFDM symbols, the contribution to the squared ACF arising from the non-vanishing-kernel index configurations could be expressed as:
\begin{align}
S_k(\mathbf{g}) =\big|[\mathbf Q_k^{\mathrm{sp}}(\boldsymbol{\theta})+\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})]^{1/2}\mathbf g\big|2^2,
\end{align}
where
\begin{align}
\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)&=\frac{1}{2N}[\mathrm{diag}(\mathbf C{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k
+\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))]\notag\
&+\frac{1}{N^2}\mathrm{diag}[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)].
\end{align}
\begin{align}\label{eq:Qbulk_prop}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})=\frac{1}{N}\big(\mathbf{f}{k+1}^*\mathbf{f}{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}),
\end{align}
with
\begin{equation}
\mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\quad \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H,
\end{equation}
while denotes the frequency power spectrum of the designable phase vector , and denotes the attenuation kernel determined by the structured hyperplane distance, with .
\end{proposition}
可参考的内容如下:
\section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk}
We adopt the normalized -point DFT matrix with
\begin{align}
[\mathbf F]{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)},\qquad \omega\triangleq\frac{2\pi}{N},
\end{align}
and denote by its th column, . Hence, the bin index corresponds to the column .
Define the diagonal modulation operator
\begin{align}\label{eq:Dk_def_app}
\mathbf D_k\triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big){n=1}^{N}.
\end{align}
The designable phase is encoded by , yielding
.
The kernel exponent is specified by
\begin{align}\label{eq:S_def_app}
\Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta),\qquad
S(p,q,p',q')\triangleq(4N_d+2)\Big[\Psi(p-q')+\Psi(q-p')-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')\Big],
\end{align}
and we write .
The non-vanishing-kernel configurations are partitioned into four structured single-hyperplane families and a residual bulk complement:
(H1) , (H2) , (H3) , (H4) , and the remaining index tuples (bulk).
On each single-hyperplane family, the expression in \eqref{eq:S_def_app} reduces to a function of a single modular difference because every term depends only on a difference of its arguments. A convenient representative reduction is obtained by evaluating the canonical slice , which yields
\begin{align}\label{eq:DeltaS_app}
\Delta_S(d)\triangleq S(d,0,0,0)=(4N_d+2)\big(\Psi(0)-\Psi(d)\big),\qquad d\in\mathbb Z_N,
\end{align}
where .
Accordingly, all structured hyperplane layers share the same 1D attenuation profile
\begin{align}\label{eq:kappa_app}
\kappa[0]\triangleq 0,\qquad \kappa[d]\triangleq \exp!\big(-\alpha,\Delta_S(d)\big),\qquad d=1,\ldots,N-1.
\end{align}
The subsequent matrix forms rely on the elementwise product identities of normalized DFT columns (indices modulo in the frequency argument):
\begin{align}\label{eq:DFT_id_app}
\mathbf f_{a+1}^*\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(b-a)+1},
\qquad
\mathbf f_{a+1}\odot \mathbf f_{b+1}=\frac{1}{\sqrt N}\mathbf f_{(a+b)+1}.
\end{align}
We also use the lifting operator
\begin{align}\label{eq:S_lift_app}
\mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T .
\end{align}
We next show that the four single-hyperplane families produce a diagonal layer plus a Hermitianized low-rank layer, both admitting the compact Hadamard representation in \eqref{eq:Qsp_prop}. The phase factor collapses to 1D aggregates under the same single-difference reduction. For (H1) with , the phase becomes ; summation over yields the main-diagonal mean
\begin{align}\label{eq:eta0_app}
\eta_0\triangleq \frac{1}{N}\sum_{t\in\mathbb Z_N}\Phi_{t,t}(\boldsymbol{\theta}),
\end{align}
while the remaining dependence on under the single-difference parameterization is captured by the offset-diagonal mean
\begin{align}\label{eq:eta_app}
\eta[d]\triangleq \frac{1}{N}\sum_{q\in\mathbb Z_N}\Phi_{q+d,q}(\boldsymbol{\theta}).
\end{align}
The same 1D collapse occurs for (H2) with , leading to an identical effective phase weight. Consequently, the two low-rank hyperplanes share the phase-corrected 1D profile
\begin{align}\label{eq:kappa_lr_app}
\tilde\kappa_{\mathrm{lr}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{lr}}[d]\triangleq \kappa[d]\eta[d]\eta_0^,\qquad d=1,\ldots,N-1,
\end{align}
and we define
\begin{align}\label{eq:s_lr_app}
\mathbf s_{\mathrm{lr}}\triangleq \mathbf F\tilde{\boldsymbol\kappa}{\mathrm{lr}}
=\sum{d=1}^{N-1}\tilde\kappa_{\mathrm{lr}}[d]\mathbf f_{d+1}.
\end{align}
Following the same DFT-column manipulations as in the phase-free case and invoking \eqref{eq:DFT_id_app}, the Hermitianized low-rank contribution can be written as
\begin{align}\label{eq:lr_hadamard_app}
\mathbf Q_{k,ut}(\boldsymbol{\theta})+\mathbf Q_{k,v0}(\boldsymbol{\theta})
=\frac{\sqrt N}{2N},\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}),
\end{align}
where, under the 1-based column convention, a convenient Hermitian choice is
\begin{align}\label{eq:Ak_1based}
\mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^\mathbf f_{k+1}^H.
\end{align}
For the diagonal hyperplanes (H3) and (H4), the phase factor becomes a cyclic correlation along one index direction after averaging over the constrained index. To preserve the Hermitian structure while using only 1D objects, we combine the two directional correlations via
\begin{align}\label{eq:xi_app}
\xi[d]\triangleq \frac{1}{2N^2}\sum_{p,q\in\mathbb Z_N}\Big(\Phi_{p+d,q}(\boldsymbol{\theta})\Phi_{p,q}(\boldsymbol{\theta})^+\Phi_{p,q+d}(\boldsymbol{\theta})\Phi_{p,q}(\boldsymbol{\theta})^\Big),
\end{align}
and define
\begin{align}\label{eq:kappa_diag_app}
\tilde\kappa_{\mathrm{diag}}[0]\triangleq 0,\qquad \tilde\kappa_{\mathrm{diag}}[d]\triangleq \kappa[d]\xi[d],\qquad d=1,\ldots,N-1,
\end{align}
together with .
As in the phase-free case, the diagonal contribution can be expressed using only and the 1D weights:
\begin{align}\label{eq:Qdiag_sum_app}
\mathbf Q_{\mathrm{diag}}(\boldsymbol{\theta})=\frac{1}{N^2}\sum_{d=1}^{N-1}\tilde\kappa_{\mathrm{diag}}[d]\big(\mathbf D_d+\mathbf D_{-d}\big)
=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}})\big).
\end{align}
Combining \eqref{eq:lr_hadamard_app} with \eqref{eq:Qdiag_sum_app} yields \eqref{eq:Qsp_prop}.
Finally, we derive the bulk complement and obtain \eqref{eq:Qbulk_prop}. Let and define
. The bulk contribution can be written as
\begin{align}\label{eq:bulk_vec_app}
S_k^{\mathrm{bulk}}(\mathbf g)
=\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})\mathrm{vec}(\mathbf H_k),
\end{align}
where contains the bulk mask and the phase-coded kernel with entries
\begin{align}\label{eq:Mbulk_entry_app}
[\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})]{(p,q),(p',q')}
=\mathrm{Herm}\Big{\Phi{p,q}(\boldsymbol{\theta})\Phi_{p',q'}(\boldsymbol{\theta})^\exp[-\alpha S(p,q,p',q')],[\mathbf P^{\mathrm{bulk}}]{(p,q),(p',q')}\Big}.
\end{align}
Since is circulant, there exists such that with .
Let and
, so that .
Substituting into \eqref{eq:bulk_vec_app} yields
\begin{align}\label{eq:Rbulk_app}
S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf c_k,\qquad
\mathbf R_{\mathrm{bulk}}(\boldsymbol{\theta})\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf B.
\end{align}
Define .
Using that is a circular shift of and , we obtain
\begin{align}\label{eq:bulk_back_to_g_app}
S_k^{\mathrm{bulk}}(\mathbf g)
=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf D_k\Big)\mathbf g.
\end{align}
Under the 1-based DFT convention, the diagonal modulators satisfy
\begin{align}\label{eq:D_as_fkp1_app}
\mathbf D_{-k}=\mathrm{diag}\big(e^{+j\omega k(n-1)}\big){n=1}^N=\mathrm{diag}\big(\sqrt N,\mathbf f{k+1}^\big),\qquad
\mathbf D_{k}=\mathrm{diag}\big(e^{-j\omega k(n-1)}\big){n=1}^N=\mathrm{diag}\big(\sqrt N,\mathbf f{k+1}\big).
\end{align}
Therefore, for any ,
\begin{align}\label{eq:D_hadamard_fkp1_app}
\mathbf D_{-k}\mathbf X\mathbf D_k
= N\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf X,
\end{align}
which follows entrywise from
.
Applying \eqref{eq:D_hadamard_fkp1_app} to \eqref{eq:bulk_back_to_g_app} yields
\begin{align}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol{\theta})
=\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta})\mathbf D_k
=\frac{1}{N}\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol{\theta}),
\end{align}
which is exactly \eqref{eq:Qbulk_prop}. Combining the structured hyperplane contribution and the bulk complement contribution yields in Proposition~\ref{prop:Sk_closed_form}, completing the proof.
下面给出一个**中文(文本 + 解析式)**的 proof 写法(整体逻辑与你提供的附录一致,但把我们前面推到的 、 等化简结果嵌进去,从而得到 Proposition 里的最终闭式)。
对 点 CE-OFDM,平方 ACF 的非零核项来自四重索引 的加权求和:权重由
接下来把 按索引结构分成 structured part 与 bulk part:
与你给的附录一致,非零核配置可分为四个“单超平面族”:
以及其余的 bulk complement。
在 (H1)–(H4) 上,由于 只由各类“差分”进入 ,因此均可塌缩为单一模差 的函数。取 canonical slice 得
从而得到统一的 1D 衰减谱 (你们文中给的是等价形式,这里只强调“仅依赖 ”这一点)。对应到 1-based 的向量 即可用于后续 DFT 表示。
你们相位矩阵取
令
并定义频域能量谱(逐频点功率)
接下来关键恒等式是:在该结构下
而右侧正是 的第 个分量,因此
于是你们在 (H1)/(H2) 的 1D 平均相位权重(offset-diagonal mean)最终可写成 DFT 形式,从而得到
由定义
定义(方案 2)
则
同理(用你们推到的 展开式)可得到 的等价表达,并最终用于对角层。
再用我们已经证明过的恒等式
并代入 ,系数合并后就得到 Proposition 里 的第一行:
其中 是 1-based 的循环反转置换矩阵(, for ),用于把两条“方向相关”的对角贡献合并成一个 Hermitian/实对角表达。这就给出 Proposition 中 的第二行。
综上,四个 structured 家族合并得到你们声明的
bulk 部分沿用你附录的推导:把 bulk 的四重索引贡献写成 ,利用 的循环结构降到 ,再回到 域并用调制对角矩阵与 Hadamard 的关系,最终得到
即 Proposition 的 \eqref{eq:Qbulk_prop}。
把 structured 与 bulk 两部分相加,即得
从而命题成立。
如果你希望,我可以把这份中文 proof 进一步“论文口吻化”(比如每一步开头一句话说明“我们先证明…/接下来证明…”),并且把“ 出现的原因(两方向相关平均)”单独写成一句非常短的解释,方便你放进附录。
推导过程写的再详细一些 现在有点看起来奇怪
下面给一个更细、更顺的中文 proof(按你们附录的路线走,但把我们前面“”以及“”这条线嵌进去,从而得到 Proposition 里的闭式)。
取归一化 -点 DFT 矩阵 ,第 列记为 (你们 bin 索引 对应 )。我们要证明:非零核配置(non-vanishing-kernel index configurations)对平方 ACF 的贡献可以写成二次型
其中 具有你 proposition 给出的闭式。
平方 ACF(或等价的二阶统计量)展开后会出现对四重索引 的加权求和。其权重包含两部分:
其中 由 组成(只依赖各类“差分”)。
把所有这些项按 的线性组合整理(并对随机符号取期望,使得只剩关于 的二次项),可写成
其中 是 Hermitian 且半正定(因为它来自某个能量型统计量的期望)。因此也可写成范数形式 。
接下来只需把 的结构具体化,并分成 structured 与 bulk 两部分。
与你附录一致,非零核配置可分成四个“单超平面族”:
其余为 bulk complement。
关键点:因为 中每一项 都只依赖“差分”,所以在每个单超平面族上, 会塌缩为某个单一模差 的函数。取 canonical slice 得到 1D 化简:
因此所有 structured 族共享同一个 1D 衰减谱
在 1-based 记号下,把它收集成 (第 1 个元素为 0,其余是 对应的 值)。
你们现在采用可设计时域相位模型
令
并定义频域能量谱(逐频点功率)
这里的关键桥梁是(我们前面已推过):
而 ,所以
进一步,附录里定义的 offset-diagonal mean(0-based 的 )在 1-based 对应下,可以写成对 的 DFT 线性泛函,从而得到(仍是我们前面推过的结论)
这一步把所有 的依赖都压缩进了 。
按你们定义
把逐点乘写成对角矩阵,并代入 ,得
定义(你选的方案 2)
则
同理,把 展开并利用 为实向量,可得到
这正是你 proposition 里使用的 解析式来源。
附录给出 structured 部分可写成两层叠加(低秩层 + 对角层),典型形式为
接下来用两个纯矩阵恒等式把 lifting 拆掉:
(i) 对角层恒等式
因此
在你们当前 (余弦和)带来的偶对称结构下,上式进一步等价于你们最终写法里的
其中 是 1-based 循环反转置换矩阵,用来把 (H3) 与 (H4) 两个方向的相关贡献合并成一个 Hermitian/实对角表达(这一步本质就是“两个方向的循环相关取平均”,对应我们前面推到的“反转偶对称化”)。
(ii) 低秩层恒等式(把 拆成常规矩阵运算) 对任意 与向量 ,有
因此
代入 并合并系数 ,得到
这正是 proposition 中 的第一行。
把对角层与低秩层相加,就得到 proposition 给出的
bulk 部分完全沿用你附录最后一段推导即可:
即 proposition 的 \eqref{eq:Qbulk_prop}。
综上,
并且 的闭式正是 proposition 中所给。命题得证。
如果你愿意,我可以再把上面第 5 步里“为什么对角层能写成 ”那一段补成两行更硬的等价推导(用 的分量式来直接推出),这样审稿人读起来会更“不可反驳”。
把你提出的建议补充上去 然后每一步都是一个完整的推导。不要把之前推过的就不写了
下面给出一个更完整、逐步展开的中文证明稿(按你附录的推导主线走,但不跳步,把我们前面用到的关键等价关系都在 proof 里重新推一遍),目标是严格推出 Proposition 中的 与 的闭式。
考虑第 个 ACF bin(或等价的频移/循环移位 bin)。将随机 CE-OFDM 符号的平方 ACF 展开后,会出现对四重索引 的求和项,其权重由两部分构成:
其中(与附录一致) 由 组合而成,且 。
由于最终统计量是能量型量(平方 ACF 的贡献),对随机符号取期望并整理后,所有保留下来的项都可写成
其中 为 Hermitian 且半正定,因此等价写成
接下来我们要把 拆成 structured 与 bulk 两部分,并将 structured 部分进一步化为 Proposition 中的闭式。
与附录一致,非零核配置(non-vanishing-kernel index configurations)可分成四个“单超平面族”:
以及其余的 bulk complement。因此
附录中定义
并给出
由于每一项 只依赖“差分”,在每个单超平面族上, 都会塌缩为某个单一模差 的函数。取 canonical slice ,得到(附录式 )
于是得到统一的 1D 衰减谱(0-based)
在你们现在的 1-based 记号下,把它收集成向量 (第 1 个元素对应 )即可;Proposition 中给出的闭式条目 是对上述 代入并吸收常数后的等价写法(本质仍是“仅依赖差分 ”的 1D 衰减核)。
定义可设计时域相位向量与其 DFT:
并定义频域功率谱(逐频点能量)
下面说明为什么它会进入 structured 闭式:
由相位模型
可写成 rank-1 形式
因此对任意 DFT 列 ,有
而 ,所以
在你们的 structured 推导里,(H1)/(H2) 会出现某种“差分平均”(本质是循环自相关),其频域表达由 Wiener–Khinchin 型关系给出;在你们的记号下对应(你们已采用的表达)
据此定义
把逐点乘写成对角矩阵并代入 :
于是定义(Proposition 里的定义)
就得到
由定义
因此
第一项就是 。下面处理第二项。
(i) 计算 的显式形式
因为 ,有
利用归一化 DFT 的列反转恒等式(可由指数取模直接验证)
其中 是 1-based 的“频域反转”置换矩阵,满足
于是
(ii) 利用 的偶对称与置换交换性
由于 是余弦和,因此 为偶函数,进而 (模 意义下)满足偶对称,从而
(iii) 把第二项化为对 的反转
第二项为
用交换性把 移到左侧:
定义时域反转置换矩阵
于是
综上得到关键等式
再代入 ,可得
(这一步就是你们 两个方向的对角相关合并后出现 的严格来源。)
structured 部分由两层组成:
附录中 (H1)+(H2) 给出 Hadamard-lifting 形式
其中 。用恒等式(逐元素即可验证)
得到
再代入 ,并使用 ,系数合并 ,得
附录中 (H3)+(H4) 的对角层可写成
同样逐元素可验证
因此
利用 Step 6 已严格推出的
并注意右侧为实向量(因此 可去掉),系数合并 ,得到
于是
即得到 Proposition 中给出的 闭式。
这一段与附录一致,这里也不跳步地写出关键链条。
其中 含 bulk mask 与相位核(见附录 )。
代入得
利用 是 的循环移位并引入调制对角矩阵 ,可得到
从而对任意 有
代入即得
与 Proposition 的 一致。
将 structured 与 bulk 两部分合并:
并且 的闭式由 Step 7 得出,故命题成立。
如果你愿意,我还能把这份中文 proof 改成你们论文附录那种英文 TSP 风格(段落结构、术语与符号对齐),并把“”和“”这两个关键恒等式加一个一行的指数验证,让审稿人更舒服。
我们现在改成TSP风格的中文的proof先来写一下
下面给出一个偏 TSP(IEEE Trans. Signal Processing)写法风格的中文证明稿:段落化、用“我们首先/接着/最后”组织,并且每一步都给出完整推导链条(不依赖“之前推过”)。
证明: 采用归一化 -点 DFT 矩阵 ,其元素为
并记第 列为 。因此 bin 索引 对应列 。
平方 ACF 在展开并对随机 CE-OFDM 符号取期望后,其非零核配置(non-vanishing kernel configurations)对第 个 bin 的贡献可写为关于 的 Hermitian 二次型
其中 。故等价地
进一步,按索引集合将非零核配置划分为四个单超平面族 (H1)–(H4) 与其余 bulk 补集,从而
其中 由 (H1)–(H4) 给出, 由 bulk 补集给出。
定义(与附录一致)
以及核指数
由于每一项 仅依赖自变量差分,故在每个单超平面族 (H1)–(H4) 上, 均可塌缩为单一模差 的函数。取 canonical slice ,得
于是四个 structured 家族共享同一条 1D 衰减谱 ,并将其收集为 1-based 向量 (与命题表述一致),其中条目给定为
(该式与 的写法仅差常数吸收与索引约定,本质均为 1D 差分塌缩。)
采用命题中的可设计时域相位模型
令
并定义频域功率谱
下面说明该量确为 structured 权重的频域聚合对象:由于
可写 。对任意 DFT 列 ,有
因此逐元素即得到上述 。
按你们的 structured 推导(Wiener–Khinchin 型关系对应的 1D 聚合),定义
将逐点乘写为对角矩阵并代入 ,得
定义
则
定义
由 ,有
接着处理第二项。由于 ,有
而由 DFT 指数结构可验证存在 1-based 的频域反转置换矩阵 使
从而
又因为 由余弦和生成,满足模 意义下的偶对称,从而
因此
定义时域反转算子
于是
综上得到
首先回顾附录中的 lifting 算子
附录给出 (H1)+(H2) 的 structured 贡献可写为
利用恒等式(逐元素可直接验证)
得
代入 并合并系数 ,且注意 从而共轭可去,得到
附录中对角贡献写为
由
得到
代入上一节严格推出的 ,并注意右侧为实向量,合并系数 ,得
将两层相加,即得命题中 structured 部分闭式
令 ,并定义循环矩阵
bulk 部分可写为
其中 由 bulk mask 与相位核构成(附录 )。利用 的 circulant 结构,可进一步化到
其中 ,且 (与附录一致)。
在 1-based DFT 约定下有
从而对任意 ,
代入即得
这与命题 一致。
综上,
并且 与 的闭式分别由上述 F、G 两节推得,命题得证。
如果你下一步希望我把这份中文 proof 进一步压缩到“IEEE 附录可接受的长度”(保留关键等式、删去冗余解释、每段只保留一两句过渡),我也可以直接给一个“投稿版精简稿”。
压缩到“IEEE 附录可接受的长度,然后不用分section 用一个proof来写
Proof: 取归一化 -点 DFT 矩阵 , 且 ,记其第 列为 ,因此 bin 对应 。平方 ACF 的非零核配置项在对随机 CE-OFDM 符号取期望后可写为关于 的 Hermitian 二次型
其中 ,故
将非零核索引配置按四个单超平面族 (H1) 、(H2) 、(H3) 、(H4) 与其余 bulk 补集划分,可得
核指数由 组合而成,且仅依赖差分;因此在每个单超平面族上均可塌缩为单一模差 的函数。取 canonical slice 得
从而定义 1D 衰减谱并按 1-based 收集为 ,其条目与命题一致:
相位采用 。令
并定义频域功率谱
由于 ,对任意 有
故 structured 部分的 1D 相位聚合可由 表示,并得到(与正文/前述推导一致)
据此
定义
则
(H1)+(H2) 族产生 Hermitianized 低秩层。令
由逐元素恒等式 与 ,可将附录中的 Hadamard–lifting 形式化为
代入 并合并系数得
(H3)+(H4) 族产生对角层。由 ,可写
又因为 且 实、 为余弦核而满足模 偶对称,上述两方向相关可合并为反转对称化,存在 1-based 循环反转置换矩阵 (,, )使
代回并合并系数 得
因此
与命题中的 一致。
最后推导 bulk 补集项。令 并构造 circulant 矩阵 。bulk 贡献可写为 的二次型,并利用 circulant 结构降维得到 (见附录同构推导),从而
其中 。在 1-based DFT 约定下 、,故对任意 ,
从而
即 \eqref{eq:Qbulk_prop}。综上
命题得证。
对于K_{bulk}得有个说明。我们需要在最开始就对p、q、p‘、q’的取值分类一下,然后给出K_{bulk}里面剩下的是啥
下面给你一段可以直接放在 proof 一开始的“索引分类 + 含义/组成说明”(TSP 风格中文,带必要公式,且把 bulk 里“剩下的是什么”说清楚)。
latex% --------------------------------------------------------- % Index classification and what remains in K_bulk(theta) % --------------------------------------------------------- \textbf{Proof (index partition and definition of $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)$):} Let $p,q,p',q'\in\mathbb Z_N$ and denote the full index set by \begin{align} \Omega \triangleq \{(p,q,p',q'):\; p,q,p',q'\in\mathbb Z_N\}. \end{align} We partition $\Omega$ into four structured single-hyperplane families and the residual bulk complement: \begin{align} \Omega_{\mathrm{H1}} &\triangleq \{(p,q,p',q')\in\Omega:\; p'=q'\},\nonumber\\ \Omega_{\mathrm{H2}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=q\},\nonumber\\ \Omega_{\mathrm{H3}} &\triangleq \{(p,q,p',q')\in\Omega:\; q=q'\},\nonumber\\ \Omega_{\mathrm{H4}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=p'\},\nonumber\\ \Omega_{\mathrm{bulk}} &\triangleq \Omega\setminus\Big(\Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}\Big). \end{align} Accordingly, the \emph{bulk} part collects \emph{all remaining} index tuples that do not satisfy any of the above single-hyperplane constraints. To encode this complement, we introduce a bulk mask $\mathbf P^{\mathrm{bulk}}\in\{0,1\}^{N^2\times N^2}$ indexed by row $(p,q)$ and column $(p',q')$ such that \begin{align} [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \triangleq \mathbbm 1\big\{(p,q,p',q')\in\Omega_{\mathrm{bulk}}\big\}. \end{align} Then $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N\times N}$ is the $N\times N$ PSD matrix that aggregates, \emph{over the bulk complement}, the phase-coded kernel weights $\Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^*\exp[-\alpha S(p,q,p',q')]$ after projecting onto the circulant subspace induced by $\mathbf H_k$. Concretely, define $\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N^2\times N^2}$ entrywise by \begin{align} [\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]_{(p,q),(p',q')} \triangleq \mathrm{Herm}\!\Big\{ \Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^* \exp[-\alpha S(p,q,p',q')]\, [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \Big\}, \end{align} and let $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]$ with $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\;(\mathrm{mod}\;N)\}$. The bulk projection is \begin{align} \mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B, \qquad \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F\,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\,\mathbf F^H. \end{align} In words, $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)$ contains exactly the residual contribution of all $(p,q,p',q')\in\Omega_{\mathrm{bulk}}$ after removing the four structured families, and it inherits the dependence on $\boldsymbol\theta$ only through the phase factors $\Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^*$.
如果你愿意,我也可以把这段**无缝嵌入你当前“压缩版 proof”**的开头(把符号重复定义删到最少),并把后面出现的 的文字过渡再压一压,让整体长度更像 IEEE 附录最终稿。
可以 按照你说的来弄 给出优化后的整体latex代码版本,符合TSP风格的附录
latex\begin{proof} We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)}$ and $\omega\triangleq 2\pi/N$, and denote by $\mathbf f_{\ell}$ its $\ell$th column ($\ell=1,\ldots,N$). Hence, the bin index $k\in\{0,\ldots,N-1\}$ corresponds to $\mathbf f_{k+1}$. Define the diagonal modulator \begin{align} \mathbf D_k \triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}. \end{align} \textit{(Index partition and bulk complement.)} Let $p,q,p',q'\in\mathbb Z_N$ and define the full index set $\Omega \triangleq \{(p,q,p',q'):\; p,q,p',q'\in\mathbb Z_N\}$. We partition $\Omega$ into four structured single-hyperplane families and the residual bulk complement: \begin{align} \Omega_{\mathrm{H1}} &\triangleq \{(p,q,p',q')\in\Omega:\; p'=q'\},\nonumber\\ \Omega_{\mathrm{H2}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=q\},\nonumber\\ \Omega_{\mathrm{H3}} &\triangleq \{(p,q,p',q')\in\Omega:\; q=q'\},\nonumber\\ \Omega_{\mathrm{H4}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=p'\},\nonumber\\ \Omega_{\mathrm{bulk}} &\triangleq \Omega\setminus\Big(\Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}\Big). \end{align} Introduce the bulk mask $\mathbf P^{\mathrm{bulk}}\in\{0,1\}^{N^2\times N^2}$ indexed by row $(p,q)$ and column $(p',q')$ as \begin{align} [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \triangleq \mathbbm 1\big\{(p,q,p',q')\in\Omega_{\mathrm{bulk}}\big\}. \end{align} Thus, the \emph{bulk} part collects exactly the remaining index tuples that do not satisfy any single-hyperplane constraint. \textit{(Kernel reduction and attenuation profile.)} Let $\Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta)$ and $S(p,q,p',q')$ be defined as in \eqref{eq:S_def_app}, with $\alpha\triangleq \tilde h^2/N^2$. Since each $\Psi(\cdot)$ depends only on modular differences, on each single-hyperplane family the exponent reduces to a 1D function of one modular difference. This yields a shared attenuation profile $\tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1}$ with entries $\kappa[d]=\exp\!\left[-2\alpha\big(N_d-\Psi(d)\big)\right]$, $d=1,\ldots,N$ (with the convention $\kappa[1]=0$). \textit{(Phase model and frequency-domain power spectrum.)} With $\Phi_{p,q}(\boldsymbol\theta)=e^{j(\theta_p-\theta_q)}$, define \begin{align} \mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T,\qquad \mathbf p(\boldsymbol\theta)\triangleq \mathbf F^H\mathbf u(\boldsymbol\theta),\qquad \tilde{\mathbf p}(\boldsymbol\theta)\triangleq \mathbf p(\boldsymbol\theta)\odot \mathbf p^*(\boldsymbol\theta)\in\mathbb R_+^{N\times 1}. \end{align} (Indeed, $\mathbf\Phi(\boldsymbol\theta)=\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H$ and $\mathbf f_\ell^H\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell = |\mathbf f_\ell^H\mathbf u(\boldsymbol\theta)|^2 =|[\mathbf F^H\mathbf u(\boldsymbol\theta)]_\ell|^2$, hence $\tilde{\mathbf p}$ collects these per-frequency powers.) \textit{(Structured part: low-rank layer.)} Define \begin{align} \mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\qquad \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta),\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big). \end{align} Then \begin{align} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Let $\mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T$ and \begin{align} \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H . \end{align} Using the Hadamard identities $\mathbf A\odot(\mathbf s\mathbf 1^T)=\mathrm{diag}(\mathbf s)\mathbf A$ and $\mathbf A\odot(\mathbf 1\mathbf s^H)=\mathbf A\,\mathrm{diag}(\mathbf s^*)$, the (H1)+(H2) contribution satisfies \begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta) =\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big) =\frac{\sqrt N}{2N}\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^*)\Big), \end{align} and substituting $\mathbf s_{\mathrm{lr}}=\frac{1}{\sqrt N}\mathbf C_{\tilde\kappa}\tilde{\mathbf p}$ yields \begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta) =\frac{1}{2N}\Big[\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\Big]. \end{align} \textit{(Structured part: diagonal layer and the reversal operator.)} Define $\mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \Re\{\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\}$ and $\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta))$. Since $\Re\{\mathbf x\}=\frac12(\mathbf x+\mathbf x^*)$ and $\tilde{\mathbf p}(\boldsymbol\theta)\in\mathbb R^{N\times 1}$, we have $\mathbf u_{\mathrm{lr}}^*=\frac{1}{\sqrt N}\mathbf F^T\tilde{\mathbf p}$. Moreover, there exists a 1-based reversal permutation $\mathbf\Pi$ such that $\mathbf F^T=\mathbf F^H\mathbf\Pi$, with $(\mathbf\Pi\mathbf x)[1]=x[1]$ and $(\mathbf\Pi\mathbf x)[n]=x[N-n+2]$ for $n=2,\ldots,N$. Since $\tilde{\boldsymbol\kappa}$ is generated from cosine sums, it is (mod-$N$) even-symmetric, implying $\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi=\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})$. Define the (time-domain) reversal operator \begin{align} \mathbf R \triangleq \mathbf F\,\mathbf\Pi\,\mathbf F^H, \qquad\text{equivalently}\qquad (\mathbf R\mathbf x)[1]=x[1],\ \ (\mathbf R\mathbf x)[n]=x[N-n+2],\ n=2,\ldots,N . \end{align} Then \begin{align} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) &=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \tfrac12(\mathbf u_{\mathrm{lr}}+\mathbf u_{\mathrm{lr}}^*)\Big) =\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}^*\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi\,\mathbf u_{\mathrm{lr}} =\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}} =\tfrac12(\mathbf I+\mathbf R)\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\nonumber\\ &=\frac{1}{2\sqrt N}(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Using $\mathbf I\odot \mathbf S(\mathbf s)=\mathrm{diag}(\mathbf s)+\mathrm{diag}(\mathbf s^*)=2\,\mathrm{diag}(\Re\{\mathbf s\})$, the (H3)+(H4) diagonal contribution becomes \begin{align} \mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big) =\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re\{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\}\big) =\frac{1}{N^2}\mathrm{diag}\big[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)\big]. \end{align} Therefore, $\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)=\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)$ coincides with the claimed closed form. \textit{(Bulk complement and $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)$.} Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. The bulk contribution can be written as \begin{align} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N^2\times N^2}$ is defined entrywise by \begin{align} [\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]_{(p,q),(p',q')} \triangleq \mathrm{Herm}\!\Big\{ \Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^*\exp[-\alpha S(p,q,p',q')]\, [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \Big\}, \qquad \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H. \end{align} Since $\mathbf H_k$ is circulant, there exists $\mathbf c_k\in\mathbb C^{N}$ such that $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$ and $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$. Let $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$ and $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]$, so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting yields \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B. \end{align} Define \begin{align} \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F\,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\,\mathbf F^H, \end{align} which contains exactly the residual (bulk-complement) contribution of all $(p,q,p',q')\in\Omega_{\mathrm{bulk}}$, and depends on $\boldsymbol\theta$ only through the phase factors in $\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)$. Using that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$ (hence $\mathbf c_k=\mathbf F^H\mathbf D_k\mathbf F\,\boldsymbol\gamma$), one obtains \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k\Big)\mathbf g. \end{align} Finally, under the 1-based DFT convention, \begin{align} \mathbf D_{-k}=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}^*),\qquad \mathbf D_k=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}), \end{align} so for any $\mathbf X\in\mathbb C^{N\times N}$, \begin{align} \mathbf D_{-k}\mathbf X\mathbf D_k = N\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf X, \end{align} and hence \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k =\frac{1}{N}\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta), \end{align} which is \eqref{eq:Qbulk_prop}. Combining the structured and bulk parts yields $S_k(\mathbf g)=\big\|[\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)]^{1/2}\mathbf g\big\|_2^2$, completing the proof. \end{proof}
我们结合一下前面匹配输出的结果,把这两个proof写成一个我看看:
\section{Proof of Proposition~\ref{prop:Ik_closed_form}}\label{app:proof_prop_Ik}
In this appendix, We derive a closed-form expression for the symbol-independent term in \eqref{eq:Ek_rk2_start}, where the contribution is induced by the index quadruples that satisfy one of the following pairing conditions:
Let and denote the sets of quadruples satisfying (A) and (B), respectively, and define .
Under (A), both and are diagonal pairs, yielding . under (B), the quadruple is uniquely determined by , hence .
Their intersection corresponds to fully diagonal quadruples , so .
By inclusion--exclusion:
\begin{equation}
|\mathcal{Q}|=|\mathcal{Q}_A|+|\mathcal{Q}_B|-|\mathcal{Q}_A\cap\mathcal{Q}_B|=2N^2-N.
\end{equation}
Extracting from \eqref{eq:Ek_rk2_start} the terms indexed by yields
\begin{align}\label{eq:Ik_def_engl}
\mathcal{I}k(\mathbf{g})=&\frac{1}{N^2}(2N^2-N)\sum{n=1}^{N}\sum_{m=1}^{N} \tilde{g}{n,k} \tilde{g}{m,k} e^{j\frac{2\pi}{N} k(n-m)}\notag\
=& \frac{1}{N^2}(2N^2-N)\bigg|\sum_{n=1}^{N}\tilde{g}{n,k}e^{j\frac{2\pi}{N}k(n-1)}\bigg|^2\notag\
=& \frac{1}{N^2}(2N^2-N)N\left|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k\right|^2=(2N-1)|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k|^2,
\end{align}
where
\begin{equation}
\mathbf{\tilde{g}}k=[\tilde{g}{1,k},\tilde{g}{2,k},\dots,\tilde{g}_{N,k}]^T\in\mathbb{R}^{N\times1}.
\end{equation}
which complete the proof of Proposition~\ref{prop:Ik_closed_form}.
\section{Proof of Proposition~\ref{prop:Ik_closed_form}}\label{app:proof_prop_Ik}
In this appendix, We derive a closed-form expression for the symbol-independent term in \eqref{eq:Ek_rk2_start}, where the contribution is induced by the index quadruples that satisfy one of the following pairing conditions:
Let and denote the sets of quadruples satisfying (A) and (B), respectively, and define .
Under (A), both and are diagonal pairs, yielding . under (B), the quadruple is uniquely determined by , hence .
Their intersection corresponds to fully diagonal quadruples , so .
By inclusion--exclusion:
\begin{equation}
|\mathcal{Q}|=|\mathcal{Q}_A|+|\mathcal{Q}_B|-|\mathcal{Q}_A\cap\mathcal{Q}_B|=2N^2-N.
\end{equation}
Extracting from \eqref{eq:Ek_rk2_start} the terms indexed by yields
\begin{align}\label{eq:Ik_def_engl}
\mathcal{I}k(\mathbf{g})=&\frac{1}{N^2}(2N^2-N)\sum{n=1}^{N}\sum_{m=1}^{N} \tilde{g}{n,k} \tilde{g}{m,k} e^{j\frac{2\pi}{N} k(n-m)}\notag\
=& \frac{1}{N^2}(2N^2-N)\bigg|\sum_{n=1}^{N}\tilde{g}{n,k}e^{j\frac{2\pi}{N}k(n-1)}\bigg|^2\notag\
=& \frac{1}{N^2}(2N^2-N)N\left|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k\right|^2=(2N-1)|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k|^2,
\end{align}
where
\begin{equation}
\mathbf{\tilde{g}}k=[\tilde{g}{1,k},\tilde{g}{2,k},\dots,\tilde{g}_{N,k}]^T\in\mathbb{R}^{N\times1}.
\end{equation}
which complete the proof of Proposition~\ref{prop:Ik_closed_form}.
\section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk}
We adopt the normalized -point DFT matrix with
and , and denote by
its th column (). Hence, the bin index corresponds to
. Define the diagonal modulator
\begin{align}
\mathbf D_k \triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}.
\end{align}
\textit{(Index partition and bulk complement.)}
Let and define the full index set
.
We partition into four structured single-hyperplane families and the residual bulk complement:
\begin{align}
\Omega_{\mathrm{H1}} &\triangleq {(p,q,p',q')\in\Omega:; p'=q'},\nonumber\
\Omega_{\mathrm{H2}} &\triangleq {(p,q,p',q')\in\Omega:; p=q},\nonumber\
\Omega_{\mathrm{H3}} &\triangleq {(p,q,p',q')\in\Omega:; q=q'},\nonumber\
\Omega_{\mathrm{H4}} &\triangleq {(p,q,p',q')\in\Omega:; p=p'},\nonumber\
\Omega_{\mathrm{bulk}} &\triangleq \Omega\setminus\Big(\Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}\Big).
\end{align}
Introduce the bulk mask indexed by row and column as
\begin{align}
[\mathbf P^{\mathrm{bulk}}]{(p,q),(p',q')}
\triangleq \mathbbm 1\big{(p,q,p',q')\in\Omega{\mathrm{bulk}}\big}.
\end{align}
Thus, the \emph{bulk} part collects exactly the remaining index tuples that do not satisfy any single-hyperplane constraint.
\textit{(Kernel reduction and attenuation profile.)}
Let and be defined as in \eqref{eq:S_def_app},
with . Since each depends only on modular differences, on each single-hyperplane
family the exponent reduces to a 1D function of one modular difference. This yields a shared attenuation profile
with entries
, (with the convention ).
\textit{(Phase model and frequency-domain power spectrum.)}
With , define
\begin{align}
\mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T,\qquad
\mathbf p(\boldsymbol\theta)\triangleq \mathbf F^H\mathbf u(\boldsymbol\theta),\qquad
\tilde{\mathbf p}(\boldsymbol\theta)\triangleq \mathbf p(\boldsymbol\theta)\odot \mathbf p^*(\boldsymbol\theta)\in\mathbb R_+^{N\times 1}.
\end{align}
(Indeed, and
, hence collects these per-frequency powers.)
\textit{(Structured part: low-rank layer.)}
Define
\begin{align}
\mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\qquad
\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta),\qquad
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big).
\end{align}
Then
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)
=\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)
=\frac{1}{\sqrt N}\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta)
=\frac{1}{\sqrt N}\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta).
\end{align}
Let and
\begin{align}
\mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^\mathbf f_{k+1}^H .
\end{align}
Using the Hadamard identities and
, the (H1)+(H2) contribution satisfies
\begin{align}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)
=\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big)
=\frac{\sqrt N}{2N}\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^)\Big),
\end{align}
and substituting yields
\begin{align}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)
=\frac{1}{2N}\Big[\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k
+\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\Big].
\end{align}
\textit{(Structured part: diagonal layer and the reversal operator.)}
Define and
.
Since and , we have
. Moreover, there exists a 1-based reversal permutation
such that , with and
for . Since is generated from cosine sums, it is
(mod-) even-symmetric, implying .
Define the (time-domain) reversal operator
\begin{align}
\mathbf R \triangleq \mathbf F,\mathbf\Pi,\mathbf F^H,
\qquad\text{equivalently}\qquad
(\mathbf R\mathbf x)[1]=x[1],\ \ (\mathbf R\mathbf x)[n]=x[N-n+2],\ n=2,\ldots,N .
\end{align}
Then
\begin{align}
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)
&=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \tfrac12(\mathbf u_{\mathrm{lr}}+\mathbf u_{\mathrm{lr}}^)\Big)
=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}^\nonumber\
&=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi,\mathbf u_{\mathrm{lr}}
=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathbf\Pi,\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}
=\tfrac12(\mathbf I+\mathbf R)\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\nonumber\
&=\frac{1}{2\sqrt N}(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta).
\end{align}
Using , the (H3)+(H4)
diagonal contribution becomes
\begin{align}
\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta)
=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big)
=\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)}\big)
=\frac{1}{N^2}\mathrm{diag}\big[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)\big].
\end{align}
Therefore,
coincides with the claimed closed form.
\textit{(Bulk complement and .}
Let and define .
The bulk contribution can be written as
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)
=\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathrm{vec}(\mathbf H_k),
\end{align}
where is defined entrywise by
\begin{align}
[\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]{(p,q),(p',q')}
\triangleq
\mathrm{Herm}!\Big{
\Phi{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^\exp[-\alpha S(p,q,p',q')],
[\mathbf P^{\mathrm{bulk}}]{(p,q),(p',q')}
\Big},
\qquad \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H.
\end{align}
Since is circulant, there exists such that and
. Let and
, so that .
Substituting yields
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R{\mathrm{bulk}}(\boldsymbol\theta)\mathbf c_k,\qquad
\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B.
\end{align}
Define
\begin{align}
\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta),\mathbf F^H,
\end{align}
which contains exactly the residual (bulk-complement) contribution of all , and depends on
only through the phase factors in .
Using that is a circular shift of (hence ),
one obtains
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k\Big)\mathbf g.
\end{align}
Finally, under the 1-based DFT convention,
\begin{align}
\mathbf D_{-k}=\mathrm{diag}(\sqrt N,\mathbf f_{k+1}^),\qquad \mathbf D_k=\mathrm{diag}(\sqrt N,\mathbf f_{k+1}),
\end{align}
so for any ,
\begin{align}
\mathbf D_{-k}\mathbf X\mathbf D_k
= N\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf X,
\end{align}
and hence
\begin{align}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)
=\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k
=\frac{1}{N}\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta),
\end{align}
which is \eqref{eq:Qbulk_prop}. Combining the structured and bulk parts yields
,
completing the proof.
latex%========================================================== % Appendix: Proofs of Propositions (closed forms) %========================================================== \section{Proof of Proposition~\ref{prop:Ik_closed_form}}\label{app:proof_prop_Ik} In this appendix, we derive a closed-form expression for the symbol-independent term $\mathcal{I}_k(\mathbf g)$ in \eqref{eq:Ek_rk2_start}. The contribution is induced by the index quadruples $(p,q,p',q')\in\mathbb Z_N^4$ that satisfy at least one of the following pairing conditions: $$ \text{(A)}\ p=q\ \text{and}\ p'=q', \qquad \text{(B)}\ p=p'\ \text{and}\ q=q'. $$ Let $\mathcal{Q}_A$ and $\mathcal{Q}_B$ denote the sets of quadruples satisfying (A) and (B), respectively, and define $\mathcal{Q}\triangleq \mathcal{Q}_A\cup\mathcal{Q}_B$. Under (A), both $(p,q)$ and $(p',q')$ are diagonal pairs, yielding $|\mathcal{Q}_A|=N^2$. Under (B), the quadruple is uniquely determined by $(p,q)$, hence $|\mathcal{Q}_B|=N^2$. Their intersection corresponds to fully diagonal quadruples $p=q=p'=q'$, so $|\mathcal{Q}_A\cap\mathcal{Q}_B|=N$. By inclusion--exclusion, \begin{equation} |\mathcal{Q}|=|\mathcal{Q}_A|+|\mathcal{Q}_B|-|\mathcal{Q}_A\cap\mathcal{Q}_B|=2N^2-N. \end{equation} Extracting from \eqref{eq:Ek_rk2_start} the terms indexed by $\mathcal{Q}$ yields \begin{align}\label{eq:Ik_def_engl} \mathcal{I}_k(\mathbf{g}) &=\frac{1}{N^2}(2N^2-N)\sum_{n=1}^{N}\sum_{m=1}^{N} \tilde{g}_{n,k} \tilde{g}_{m,k} e^{j\frac{2\pi}{N} k(n-m)}\notag\\ &= \frac{1}{N^2}(2N^2-N)\bigg|\sum_{n=1}^{N}\tilde{g}_{n,k}e^{j\frac{2\pi}{N}k(n-1)}\bigg|^2\notag\\ &= \frac{1}{N^2}(2N^2-N)N\left|\mathbf{f}_{k+1}^H\mathbf{\tilde{g}}_k\right|^2 =(2N-1)\left|\mathbf{f}_{k+1}^H\mathbf{\tilde{g}}_k\right|^2, \end{align} where $\mathbf{f}_{k+1}$ is the $(k+1)$th column of the normalized DFT matrix $\mathbf F$, and \begin{equation} \mathbf{\tilde{g}}_k=[\tilde{g}_{1,k},\tilde{g}_{2,k},\dots,\tilde{g}_{N,k}]^T\in\mathbb{R}^{N\times1}. \end{equation} This completes the proof of Proposition~\ref{prop:Ik_closed_form}. \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)}$ and $\omega\triangleq 2\pi/N$, and denote by $\mathbf f_{\ell}$ its $\ell$th column ($\ell=1,\ldots,N$). Hence, the bin index $k\in\{0,\ldots,N-1\}$ corresponds to $\mathbf f_{k+1}$. Define the diagonal modulator \begin{align} \mathbf D_k \triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}. \end{align} \textit{(Index partition and bulk complement.)} Let $p,q,p',q'\in\mathbb Z_N$ and define the full index set $\Omega \triangleq \{(p,q,p',q'):\; p,q,p',q'\in\mathbb Z_N\}$. We partition $\Omega$ into four structured single-hyperplane families and the residual bulk complement: \begin{align} \Omega_{\mathrm{H1}} &\triangleq \{(p,q,p',q')\in\Omega:\; p'=q'\},\nonumber\\ \Omega_{\mathrm{H2}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=q\},\nonumber\\ \Omega_{\mathrm{H3}} &\triangleq \{(p,q,p',q')\in\Omega:\; q=q'\},\nonumber\\ \Omega_{\mathrm{H4}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=p'\},\nonumber\\ \Omega_{\mathrm{bulk}} &\triangleq \Omega\setminus\Big(\Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}\Big). \end{align} Introduce the bulk mask $\mathbf P^{\mathrm{bulk}}\in\{0,1\}^{N^2\times N^2}$ indexed by row $(p,q)$ and column $(p',q')$ as \begin{align} [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \triangleq \mathbbm 1\big\{(p,q,p',q')\in\Omega_{\mathrm{bulk}}\big\}. \end{align} Thus, the \emph{bulk} part collects exactly the remaining index tuples that do not satisfy any single-hyperplane constraint. \textit{(Kernel reduction and attenuation profile.)} Let $\Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta)$ and $S(p,q,p',q')$ be defined as in \eqref{eq:S_def_app}, with $\alpha\triangleq \tilde h^2/N^2$. Since each $\Psi(\cdot)$ depends only on modular differences, on each single-hyperplane family the exponent reduces to a 1D function of one modular difference. This yields a shared attenuation profile $\tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1}$ with entries \begin{align} \kappa[d]=\exp\!\left[-2\alpha\big(N_d-\Psi(d)\big)\right],\qquad d=1,\ldots,N, \end{align} with the convention $\kappa[1]=0$. \textit{(Phase model and frequency-domain power spectrum.)} With $\Phi_{p,q}(\boldsymbol\theta)=e^{j(\theta_p-\theta_q)}$, define \begin{align} \mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T,\qquad \mathbf p(\boldsymbol\theta)\triangleq \mathbf F^H\mathbf u(\boldsymbol\theta),\qquad \tilde{\mathbf p}(\boldsymbol\theta)\triangleq \mathbf p(\boldsymbol\theta)\odot \mathbf p^*(\boldsymbol\theta)\in\mathbb R_+^{N\times 1}. \end{align} Indeed, $\mathbf\Phi(\boldsymbol\theta)=\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H$ and, for any $\ell$, $$ \mathbf f_\ell^H\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell =\big|\mathbf f_\ell^H\mathbf u(\boldsymbol\theta)\big|^2 =\big|[\mathbf F^H\mathbf u(\boldsymbol\theta)]_\ell\big|^2, $$ hence $\tilde{\mathbf p}(\boldsymbol\theta)$ collects these per-frequency powers. \textit{(Structured part: low-rank layer.)} Define \begin{align} \mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\qquad \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta),\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big). \end{align} Then \begin{align} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Let $\mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T$ and \begin{align} \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H . \end{align} Using $\mathbf A\odot(\mathbf s\mathbf 1^T)=\mathrm{diag}(\mathbf s)\mathbf A$ and $\mathbf A\odot(\mathbf 1\mathbf s^H)=\mathbf A\,\mathrm{diag}(\mathbf s^*)$, the (H1)+(H2) contribution satisfies \begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta) &=\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big) =\frac{\sqrt N}{2N}\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^*)\Big)\nonumber\\ &=\frac{1}{2N}\Big[\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\Big]. \end{align} \textit{(Structured part: diagonal layer and the reversal operator.)} Define $\mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \Re\{\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\}$ and $\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta))$. Since $\Re\{\mathbf x\}=\frac12(\mathbf x+\mathbf x^*)$ and $\tilde{\mathbf p}(\boldsymbol\theta)$ is real-valued, we have $\mathbf u_{\mathrm{lr}}^*=\frac{1}{\sqrt N}\mathbf F^T\tilde{\mathbf p}$. Moreover, there exists a 1-based reversal permutation $\mathbf\Pi$ such that $\mathbf F^T=\mathbf F^H\mathbf\Pi$, with $(\mathbf\Pi\mathbf x)[1]=x[1]$ and $(\mathbf\Pi\mathbf x)[n]=x[N-n+2]$ for $n=2,\ldots,N$. Since $\tilde{\boldsymbol\kappa}$ is generated from cosine sums, it is (mod-$N$) even-symmetric, implying $\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi=\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})$. Define the (time-domain) reversal operator \begin{align} \mathbf R \triangleq \mathbf F\,\mathbf\Pi\,\mathbf F^H, \qquad\text{equivalently}\qquad (\mathbf R\mathbf x)[1]=x[1],\ \ (\mathbf R\mathbf x)[n]=x[N-n+2],\ n=2,\ldots,N . \end{align} Then \begin{align} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) &=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \tfrac12(\mathbf u_{\mathrm{lr}}+\mathbf u_{\mathrm{lr}}^*)\Big)\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}^*\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi\,\mathbf u_{\mathrm{lr}}\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}} =\tfrac12(\mathbf I+\mathbf R)\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\nonumber\\ &=\frac{1}{2\sqrt N}(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Using $\mathbf I\odot \mathbf S(\mathbf s)=\mathrm{diag}(\mathbf s)+\mathrm{diag}(\mathbf s^*)=2\,\mathrm{diag}(\Re\{\mathbf s\})$, the (H3)+(H4) diagonal contribution becomes \begin{align} \mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta) &=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big) =\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re\{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\}\big)\nonumber\\ &=\frac{1}{N^2}\mathrm{diag}\big[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)\big]. \end{align} Therefore, $\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)=\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)$ coincides with the claimed closed form. \textit{(Bulk complement and $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)$.} Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. The bulk contribution can be written as \begin{align} S_k^{\mathrm{bulk}}(\mathbf g) =\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathrm{vec}(\mathbf H_k), \end{align} where $\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N^2\times N^2}$ is defined entrywise by \begin{align} [\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]_{(p,q),(p',q')} \triangleq \mathrm{Herm}\!\Big\{ \Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^*\exp[-\alpha S(p,q,p',q')]\, [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \Big\}, \qquad \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H. \end{align} Since $\mathbf H_k$ is circulant, there exists $\mathbf c_k\in\mathbb C^{N}$ such that $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$ and $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$. Let $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$ and $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]$, so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting yields \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B. \end{align} Define \begin{align} \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F\,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\,\mathbf F^H, \end{align} which contains exactly the residual contribution of all $(p,q,p',q')\in\Omega_{\mathrm{bulk}}$ after removing the four structured families. Using that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$, one obtains \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k\Big)\mathbf g. \end{align} Finally, under the 1-based DFT convention, \begin{align} \mathbf D_{-k}=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}^*),\qquad \mathbf D_k=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}), \end{align} so for any $\mathbf X\in\mathbb C^{N\times N}$, \begin{align} \mathbf D_{-k}\mathbf X\mathbf D_k = N\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf X, \end{align} and hence \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k =\frac{1}{N}\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta), \end{align} which is \eqref{eq:Qbulk_prop}. Combining the structured and bulk parts yields $S_k(\mathbf g)=\big\|[\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)]^{1/2}\mathbf g\big\|_2^2$, completing the proof.
如果你希望我再进一步做两点“投稿级”统一:
算了 就按照现在这样写吧。但是我们在后面这个proof中的分类,可以把他放在前面的proof里面,在那里面分出来完全匹配的、H1-4和完全不匹配的,然后下一个appendix里面就不用再赘述了,就直接用appendixC里面的就好,优化一下下面两个appendix里面的proof;
\section{Proof of Proposition~\ref{prop:Ik_closed_form}}\label{app:proof_prop_Ik}
In this appendix, We derive a closed-form expression for the symbol-independent term in \eqref{eq:Ek_rk2_start}, where the contribution is induced by the index quadruples that satisfy one of the following pairing conditions:
Let and denote the sets of quadruples satisfying (A) and (B), respectively, and define .
Under (A), both and are diagonal pairs, yielding . under (B), the quadruple is uniquely determined by , hence .
Their intersection corresponds to fully diagonal quadruples , so .
By inclusion--exclusion:
\begin{equation}
|\mathcal{Q}|=|\mathcal{Q}_A|+|\mathcal{Q}_B|-|\mathcal{Q}_A\cap\mathcal{Q}_B|=2N^2-N.
\end{equation}
Extracting from \eqref{eq:Ek_rk2_start} the terms indexed by yields
\begin{align}\label{eq:Ik_def_engl}
\mathcal{I}k(\mathbf{g})=&\frac{1}{N^2}(2N^2-N)\sum{n=1}^{N}\sum_{m=1}^{N} \tilde{g}{n,k} \tilde{g}{m,k} e^{j\frac{2\pi}{N} k(n-m)}\notag\
=& \frac{1}{N^2}(2N^2-N)\bigg|\sum_{n=1}^{N}\tilde{g}{n,k}e^{j\frac{2\pi}{N}k(n-1)}\bigg|^2\notag\
=& \frac{1}{N^2}(2N^2-N)N\left|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k\right|^2=(2N-1)|\mathbf{f}{k+1}^H\mathbf{\tilde{g}}k|^2,
\end{align}
where
\begin{equation}
\mathbf{\tilde{g}}k=[\tilde{g}{1,k},\tilde{g}{2,k},\dots,\tilde{g}_{N,k}]^T\in\mathbb{R}^{N\times1}.
\end{equation}
which complete the proof of Proposition~\ref{prop:Ik_closed_form}.
\section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk}
We adopt the normalized -point DFT matrix with
and , and denote by
its th column (). Hence, the bin index corresponds to
. Define the diagonal modulator
\begin{align}
\mathbf D_k \triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}.
\end{align}
\textit{(Index partition and bulk complement.)}
Let and define the full index set
.
We partition into four structured single-hyperplane families and the residual bulk complement:
\begin{align}
\Omega_{\mathrm{H1}} &\triangleq {(p,q,p',q')\in\Omega:; p'=q'},\nonumber\
\Omega_{\mathrm{H2}} &\triangleq {(p,q,p',q')\in\Omega:; p=q},\nonumber\
\Omega_{\mathrm{H3}} &\triangleq {(p,q,p',q')\in\Omega:; q=q'},\nonumber\
\Omega_{\mathrm{H4}} &\triangleq {(p,q,p',q')\in\Omega:; p=p'},\nonumber\
\Omega_{\mathrm{bulk}} &\triangleq \Omega\setminus\Big(\Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}\Big).
\end{align}
Introduce the bulk mask indexed by row and column as
\begin{align}
[\mathbf P^{\mathrm{bulk}}]{(p,q),(p',q')}
\triangleq \mathbbm 1\big{(p,q,p',q')\in\Omega{\mathrm{bulk}}\big}.
\end{align}
Thus, the \emph{bulk} part collects exactly the remaining index tuples that do not satisfy any single-hyperplane constraint.
\textit{(Kernel reduction and attenuation profile.)}
Let and be defined as in \eqref{eq:S_def_app},
with . Since each depends only on modular differences, on each single-hyperplane
family the exponent reduces to a 1D function of one modular difference. This yields a shared attenuation profile
with entries
, (with the convention ).
\textit{(Phase model and frequency-domain power spectrum.)}
With , define
\begin{align}
\mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T,\qquad
\mathbf p(\boldsymbol\theta)\triangleq \mathbf F^H\mathbf u(\boldsymbol\theta),\qquad
\tilde{\mathbf p}(\boldsymbol\theta)\triangleq \mathbf p(\boldsymbol\theta)\odot \mathbf p^*(\boldsymbol\theta)\in\mathbb R_+^{N\times 1}.
\end{align}
(Indeed, and
, hence collects these per-frequency powers.)
\textit{(Structured part: low-rank layer.)}
Define
\begin{align}
\mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\qquad
\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta),\qquad
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big).
\end{align}
Then
\begin{align}
\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)
=\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)
=\frac{1}{\sqrt N}\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta)
=\frac{1}{\sqrt N}\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta).
\end{align}
Let and
\begin{align}
\mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^\mathbf f_{k+1}^H .
\end{align}
Using the Hadamard identities and
, the (H1)+(H2) contribution satisfies
\begin{align}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)
=\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big)
=\frac{\sqrt N}{2N}\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^)\Big),
\end{align}
and substituting yields
\begin{align}
\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)
=\frac{1}{2N}\Big[\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k
+\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\Big].
\end{align}
\textit{(Structured part: diagonal layer and the reversal operator.)}
Define and
.
Since and , we have
. Moreover, there exists a 1-based reversal permutation
such that , with and
for . Since is generated from cosine sums, it is
(mod-) even-symmetric, implying .
Define the (time-domain) reversal operator
\begin{align}
\mathbf R \triangleq \mathbf F,\mathbf\Pi,\mathbf F^H,
\qquad\text{equivalently}\qquad
(\mathbf R\mathbf x)[1]=x[1],\ \ (\mathbf R\mathbf x)[n]=x[N-n+2],\ n=2,\ldots,N .
\end{align}
Then
\begin{align}
\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)
&=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \tfrac12(\mathbf u_{\mathrm{lr}}+\mathbf u_{\mathrm{lr}}^)\Big)
=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}^\nonumber\
&=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi,\mathbf u_{\mathrm{lr}}
=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathbf\Pi,\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}
=\tfrac12(\mathbf I+\mathbf R)\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\nonumber\
&=\frac{1}{2\sqrt N}(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta).
\end{align}
Using , the (H3)+(H4)
diagonal contribution becomes
\begin{align}
\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta)
=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big)
=\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)}\big)
=\frac{1}{N^2}\mathrm{diag}\big[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)\big].
\end{align}
Therefore,
coincides with the claimed closed form.
\textit{(Bulk complement and .}
Let and define .
The bulk contribution can be written as
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)
=\frac{1}{N^2}\mathrm{vec}(\mathbf H_k)^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathrm{vec}(\mathbf H_k),
\end{align}
where is defined entrywise by
\begin{align}
[\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]{(p,q),(p',q')}
\triangleq
\mathrm{Herm}!\Big{
\Phi{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^\exp[-\alpha S(p,q,p',q')],
[\mathbf P^{\mathrm{bulk}}]{(p,q),(p',q')}
\Big},
\qquad \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H.
\end{align}
Since is circulant, there exists such that and
. Let and
, so that .
Substituting yields
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R{\mathrm{bulk}}(\boldsymbol\theta)\mathbf c_k,\qquad
\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B.
\end{align}
Define
\begin{align}
\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta),\mathbf F^H,
\end{align}
which contains exactly the residual (bulk-complement) contribution of all , and depends on
only through the phase factors in .
Using that is a circular shift of (hence ),
one obtains
\begin{align}
S_k^{\mathrm{bulk}}(\mathbf g)=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k\Big)\mathbf g.
\end{align}
Finally, under the 1-based DFT convention,
\begin{align}
\mathbf D_{-k}=\mathrm{diag}(\sqrt N,\mathbf f_{k+1}^),\qquad \mathbf D_k=\mathrm{diag}(\sqrt N,\mathbf f_{k+1}),
\end{align}
so for any ,
\begin{align}
\mathbf D_{-k}\mathbf X\mathbf D_k
= N\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf X,
\end{align}
and hence
\begin{align}
\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)
=\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k
=\frac{1}{N}\big(\mathbf f_{k+1}^\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta),
\end{align}
which is \eqref{eq:Qbulk_prop}. Combining the structured and bulk parts yields
,
completing the proof.
latex%========================================================== % Appendix B: Proof of Proposition~\ref{prop:Ik_closed_form} % (also defines the global index partition used later) %========================================================== \section{Proof of Proposition~\ref{prop:Ik_closed_form}}\label{app:proof_prop_Ik} In this appendix, we derive a closed-form expression for the symbol-independent term $\mathcal{I}_k(\mathbf g)$ in \eqref{eq:Ek_rk2_start}. We also fix a global partition of the index quadruples $(p,q,p',q')$ that will be reused in Appendix~\ref{app:proof_prop_Sk}. Let $p,q,p',q'\in\mathbb Z_N$ and define the full index set \begin{align} \Omega \triangleq \{(p,q,p',q'):\; p,q,p',q'\in\mathbb Z_N\}. \end{align} We distinguish three disjoint classes: (i) \emph{fully paired} (exact matches) which induce $\mathcal{I}_k(\mathbf g)$; (ii) \emph{single-hyperplane} configurations (H1)--(H4) excluding the fully paired ones, which induce the structured term in Proposition~\ref{prop:Sk_closed_form}; (iii) the \emph{fully mismatched} residual complement (bulk), also handled in Appendix~\ref{app:proof_prop_Sk}. Specifically, the fully paired quadruples satisfy at least one of the pairing conditions $$ \text{(A)}\ p=q\ \text{and}\ p'=q', \qquad \text{(B)}\ p=p'\ \text{and}\ q=q'. $$ Let $\mathcal{Q}_A$ and $\mathcal{Q}_B$ denote the sets of quadruples satisfying (A) and (B), respectively, and define \begin{align} \Omega_{\mathcal I}\triangleq \mathcal{Q}\triangleq \mathcal{Q}_A\cup\mathcal{Q}_B . \end{align} Moreover, define the four single-hyperplane families \begin{align} \Omega_{\mathrm{H1}} &\triangleq \{(p,q,p',q')\in\Omega:\; p'=q'\},\nonumber\\ \Omega_{\mathrm{H2}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=q\},\nonumber\\ \Omega_{\mathrm{H3}} &\triangleq \{(p,q,p',q')\in\Omega:\; q=q'\},\nonumber\\ \Omega_{\mathrm{H4}} &\triangleq \{(p,q,p',q')\in\Omega:\; p=p'\}, \end{align} and set $\Omega_{\mathrm{H}}\triangleq \Omega_{\mathrm{H1}}\cup\Omega_{\mathrm{H2}}\cup\Omega_{\mathrm{H3}}\cup\Omega_{\mathrm{H4}}$. Then we define the residual partitions \begin{align} \Omega_{\mathrm{sp}} \triangleq \Omega_{\mathrm{H}}\setminus \Omega_{\mathcal I},\qquad \Omega_{\mathrm{bulk}} \triangleq \Omega\setminus \Omega_{\mathrm{H}}. \end{align} Note that $\Omega_{\mathcal I}\subseteq \Omega_{\mathrm{H}}$ (indeed, (A) implies (H1) and (H2), while (B) implies (H3) and (H4)), hence $\Omega=\Omega_{\mathcal I}\,\dot\cup\,\Omega_{\mathrm{sp}}\,\dot\cup\,\Omega_{\mathrm{bulk}}$. We now evaluate the contribution of $\Omega_{\mathcal I}$ to obtain $\mathcal I_k(\mathbf g)$. Under (A), both $(p,q)$ and $(p',q')$ are diagonal pairs, hence $|\mathcal{Q}_A|=N^2$. Under (B), the quadruple is uniquely determined by $(p,q)$, hence $|\mathcal{Q}_B|=N^2$. Their intersection corresponds to fully diagonal quadruples $p=q=p'=q'$, so $|\mathcal{Q}_A\cap\mathcal{Q}_B|=N$. By inclusion--exclusion, \begin{align} |\mathcal{Q}|=|\mathcal{Q}_A|+|\mathcal{Q}_B|-|\mathcal{Q}_A\cap\mathcal{Q}_B|=2N^2-N. \end{align} Extracting from \eqref{eq:Ek_rk2_start} the terms indexed by $\mathcal{Q}$ yields \begin{align}\label{eq:Ik_def_engl} \mathcal{I}_k(\mathbf{g}) &=\frac{1}{N^2}(2N^2-N)\sum_{n=1}^{N}\sum_{m=1}^{N} \tilde{g}_{n,k} \tilde{g}_{m,k} e^{j\frac{2\pi}{N} k(n-m)}\notag\\ &= \frac{1}{N^2}(2N^2-N)\bigg|\sum_{n=1}^{N}\tilde{g}_{n,k}e^{j\frac{2\pi}{N}k(n-1)}\bigg|^2\notag\\ &= \frac{1}{N^2}(2N^2-N)N\left|\mathbf{f}_{k+1}^H\mathbf{\tilde{g}}_k\right|^2 =(2N-1)\left|\mathbf{f}_{k+1}^H\mathbf{\tilde{g}}_k\right|^2, \end{align} where $\mathbf f_{k+1}$ is the $(k+1)$th column of the normalized DFT matrix $\mathbf F$, and \begin{align} \mathbf{\tilde{g}}_k=[\tilde{g}_{1,k},\tilde{g}_{2,k},\dots,\tilde{g}_{N,k}]^T\in\mathbb{R}^{N\times1}. \end{align} This completes the proof of Proposition~\ref{prop:Ik_closed_form}. %========================================================== % Appendix C: Proof of Proposition~\ref{prop:Sk_closed_form} % (reuses the index partition from Appendix B, no repetition) %========================================================== \section{Proof of Proposition~\ref{prop:Sk_closed_form}}\label{app:proof_prop_Sk} We adopt the same notation as Appendix~\ref{app:proof_prop_Ik}. In particular, the index set $\Omega$ is partitioned as $\Omega=\Omega_{\mathcal I}\,\dot\cup\,\Omega_{\mathrm{sp}}\,\dot\cup\,\Omega_{\mathrm{bulk}}$, where $\Omega_{\mathrm{sp}}=\Omega_{\mathrm{H}}\setminus\Omega_{\mathcal I}$ collects the single-hyperplane configurations (H1)--(H4) excluding the fully paired ones, and $\Omega_{\mathrm{bulk}}=\Omega\setminus\Omega_{\mathrm{H}}$ is the fully mismatched (bulk) complement. We derive the closed form for the non-symbol-independent contribution $S_k(\mathbf g)$ induced by $\Omega_{\mathrm{sp}}\cup\Omega_{\mathrm{bulk}}$. We adopt the normalized $N$-point DFT matrix $\mathbf F\in\mathbb C^{N\times N}$ with $[\mathbf F]_{n,\ell}=\frac{1}{\sqrt N}e^{-j\omega (n-1)(\ell-1)}$ and $\omega\triangleq 2\pi/N$, and denote by $\mathbf f_{\ell}$ its $\ell$th column ($\ell=1,\ldots,N$). Hence, the bin index $k\in\{0,\ldots,N-1\}$ corresponds to $\mathbf f_{k+1}$. Define the diagonal modulator \begin{align} \mathbf D_k \triangleq \mathrm{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^{N}. \end{align} \textit{(Kernel reduction and attenuation profile.)} Let $\Psi(\Delta)\triangleq \sum_{r=1}^{N_d}\cos(\omega r\Delta)$ and $S(p,q,p',q')$ be defined as in \eqref{eq:S_def_app}, with $\alpha\triangleq \tilde h^2/N^2$. Since each $\Psi(\cdot)$ depends only on modular differences, on each single-hyperplane family (H1)--(H4) the exponent reduces to a 1D function of one modular difference. This yields a shared attenuation profile $\tilde{\boldsymbol\kappa}\in\mathbb R^{N\times 1}$ with entries \begin{align} \kappa[d]=\exp\!\left[-2\alpha\big(N_d-\Psi(d)\big)\right],\qquad d=1,\ldots,N, \end{align} with the convention $\kappa[1]=0$. \textit{(Phase model and frequency-domain power spectrum.)} With $\Phi_{p,q}(\boldsymbol\theta)=e^{j(\theta_p-\theta_q)}$, define \begin{align} \mathbf u(\boldsymbol\theta)\triangleq [e^{j\theta_1},\ldots,e^{j\theta_N}]^T,\qquad \mathbf p(\boldsymbol\theta)\triangleq \mathbf F^H\mathbf u(\boldsymbol\theta),\qquad \tilde{\mathbf p}(\boldsymbol\theta)\triangleq \mathbf p(\boldsymbol\theta)\odot \mathbf p^*(\boldsymbol\theta)\in\mathbb R_+^{N\times 1}. \end{align} Indeed, $\mathbf\Phi(\boldsymbol\theta)=\mathbf u(\boldsymbol\theta)\mathbf u(\boldsymbol\theta)^H$ and, for any $\ell$, $$ \mathbf f_\ell^H\mathbf\Phi(\boldsymbol\theta)\mathbf f_\ell =\big|\mathbf f_\ell^H\mathbf u(\boldsymbol\theta)\big|^2 =\big|[\mathbf F^H\mathbf u(\boldsymbol\theta)]_\ell\big|^2, $$ hence $\tilde{\mathbf p}(\boldsymbol\theta)$ collects these per-frequency powers. \textit{(Structured part over $\Omega_{\mathrm{sp}}$: low-rank layer.)} Define \begin{align} \mathbf C_{\tilde{\kappa}}\triangleq \mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H,\qquad \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \frac{1}{\sqrt N}\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta),\qquad \mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\triangleq \mathbf F\big(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\big). \end{align} Then \begin{align} \mathbf s_{\mathrm{lr}}(\boldsymbol\theta) =\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf F^H\tilde{\mathbf p}(\boldsymbol\theta) =\frac{1}{\sqrt N}\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Let $\mathbf S(\mathbf s)\triangleq \mathbf 1\mathbf s^H+\mathbf s\mathbf 1^T$ and \begin{align} \mathbf A_k \triangleq \mathbf f_{k+1}\mathbf f_{k+1}^T+\mathbf f_{k+1}^*\mathbf f_{k+1}^H . \end{align} Using $\mathbf A\odot(\mathbf s\mathbf 1^T)=\mathrm{diag}(\mathbf s)\mathbf A$ and $\mathbf A\odot(\mathbf 1\mathbf s^H)=\mathbf A\,\mathrm{diag}(\mathbf s^*)$, the (H1)+(H2) contribution satisfies \begin{align} \mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta) &=\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s_{\mathrm{lr}}(\boldsymbol\theta))\big) =\frac{\sqrt N}{2N}\Big(\mathrm{diag}(\mathbf s_{\mathrm{lr}})\mathbf A_k+\mathbf A_k\mathrm{diag}(\mathbf s_{\mathrm{lr}}^*)\Big)\nonumber\\ &=\frac{1}{2N}\Big[\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\mathbf A_k +\mathbf A_k\mathrm{diag}(\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta))\Big]. \end{align} \textit{(Structured part over $\Omega_{\mathrm{sp}}$: diagonal layer and the reversal operator.)} Define $\mathbf u_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \Re\{\mathbf u_{\mathrm{lr}}(\boldsymbol\theta)\}$ and $\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\triangleq \mathbf F(\tilde{\boldsymbol\kappa}\odot \mathbf u_{\mathrm{diag}}(\boldsymbol\theta))$. Since $\Re\{\mathbf x\}=\frac12(\mathbf x+\mathbf x^*)$ and $\tilde{\mathbf p}(\boldsymbol\theta)\in\mathbb R^{N\times 1}$, we have $\mathbf u_{\mathrm{lr}}^*=\frac{1}{\sqrt N}\mathbf F^T\tilde{\mathbf p}$. Moreover, there exists a 1-based reversal permutation $\mathbf\Pi$ such that $\mathbf F^T=\mathbf F^H\mathbf\Pi$, with $(\mathbf\Pi\mathbf x)[1]=x[1]$ and $(\mathbf\Pi\mathbf x)[n]=x[N-n+2]$ for $n=2,\ldots,N$. Since $\tilde{\boldsymbol\kappa}$ is generated from cosine sums, it is (mod-$N$) even-symmetric, implying $\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi=\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})$. Define the (time-domain) reversal operator \begin{align} \mathbf R \triangleq \mathbf F\,\mathbf\Pi\,\mathbf F^H, \qquad\text{equivalently}\qquad (\mathbf R\mathbf x)[1]=x[1],\ \ (\mathbf R\mathbf x)[n]=x[N-n+2],\ n=2,\ldots,N . \end{align} Then \begin{align} \mathbf s_{\mathrm{diag}}(\boldsymbol\theta) &=\mathbf F\Big(\tilde{\boldsymbol\kappa}\odot \tfrac12(\mathbf u_{\mathrm{lr}}+\mathbf u_{\mathrm{lr}}^*)\Big)\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}}^*\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf\Pi\,\mathbf u_{\mathrm{lr}}\nonumber\\ &=\tfrac12\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)+\tfrac12\mathbf F\mathbf\Pi\,\mathrm{diag}(\tilde{\boldsymbol\kappa})\mathbf u_{\mathrm{lr}} =\tfrac12(\mathbf I+\mathbf R)\mathbf s_{\mathrm{lr}}(\boldsymbol\theta)\nonumber\\ &=\frac{1}{2\sqrt N}(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta). \end{align} Using $\mathbf I\odot \mathbf S(\mathbf s)=\mathrm{diag}(\mathbf s)+\mathrm{diag}(\mathbf s^*)=2\,\mathrm{diag}(\Re\{\mathbf s\})$, the (H3)+(H4) diagonal contribution becomes \begin{align} \mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta) &=\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s_{\mathrm{diag}}(\boldsymbol\theta))\big) =\frac{2}{N^{3/2}}\mathrm{diag}\big(\Re\{\mathbf s_{\mathrm{diag}}(\boldsymbol\theta)\}\big)\nonumber\\ &=\frac{1}{N^2}\mathrm{diag}\big[(\mathbf I+\mathbf R)\mathbf C_{\tilde{\kappa}}\tilde{\mathbf p}(\boldsymbol\theta)\big]. \end{align} Therefore, $\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)=\mathbf Q_k^{\mathrm{sp1}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{sp2}}(\boldsymbol\theta)$ coincides with the claimed closed form. \textit{(Bulk complement over $\Omega_{\mathrm{bulk}}$ and $\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)$.)} Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ and define $[\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle}$. Define a bulk mask $\mathbf P^{\mathrm{bulk}}\in\{0,1\}^{N^2\times N^2}$ (row $(p,q)$, column $(p',q')$) by \begin{align} [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \triangleq \mathbbm 1\big\{(p,q,p',q')\in\Omega_{\mathrm{bulk}}\big\}, \end{align} and form $\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\in\mathbb C^{N^2\times N^2}$ entrywise as \begin{align} [\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)]_{(p,q),(p',q')} \triangleq \mathrm{Herm}\!\Big\{ \Phi_{p,q}(\boldsymbol\theta)\Phi_{p',q'}(\boldsymbol\theta)^*\exp[-\alpha S(p,q,p',q')]\, [\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')} \Big\}, \qquad \mathrm{Herm}(\mathbf X)\triangleq \mathbf X+\mathbf X^H. \end{align} Since $\mathbf H_k$ is circulant, there exists $\mathbf c_k\in\mathbb C^{N}$ such that $\mathbf H_k=\mathrm{circ}(\mathbf c_k)$ and $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$. Let $[\mathbf J_d]_{p,q}=\mathbbm 1\{p-q\equiv d\ (\mathrm{mod}\ N)\}$ and $\mathbf B=[\mathrm{vec}(\mathbf J_0),\ldots,\mathrm{vec}(\mathbf J_{N-1})]$, so that $\mathrm{vec}(\mathbf H_k)=\mathbf B\mathbf c_k$. Substituting yields \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\frac{1}{N^2}\mathbf c_k^H\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf c_k,\qquad \mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf B^H\mathbf M_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf B. \end{align} Define \begin{align} \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\triangleq \mathbf F\,\mathbf R_{\mathrm{bulk}}(\boldsymbol\theta)\,\mathbf F^H. \end{align} Using that $\mathbf c_k$ is a circular shift of $\boldsymbol\gamma$, one obtains \begin{align} S_k^{\mathrm{bulk}}(\mathbf g)=\mathbf g^H\Big(\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k\Big)\mathbf g. \end{align} Finally, under the 1-based DFT convention, \begin{align} \mathbf D_{-k}=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}^*),\qquad \mathbf D_k=\mathrm{diag}(\sqrt N\,\mathbf f_{k+1}), \end{align} so for any $\mathbf X\in\mathbb C^{N\times N}$, \begin{align} \mathbf D_{-k}\mathbf X\mathbf D_k = N\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf X, \end{align} and hence \begin{align} \mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta) =\frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}(\boldsymbol\theta)\mathbf D_k =\frac{1}{N}\big(\mathbf f_{k+1}^*\mathbf f_{k+1}^T\big)\odot \mathbf K_{\mathrm{bulk}}(\boldsymbol\theta), \end{align} which is \eqref{eq:Qbulk_prop}. Combining the structured and bulk parts yields $S_k(\mathbf g)=\big\|[\mathbf Q_k^{\mathrm{sp}}(\boldsymbol\theta)+\mathbf Q_k^{\mathrm{bulk}}(\boldsymbol\theta)]^{1/2}\mathbf g\big\|_2^2$, completing the proof.
这版改动点:
我们这样 把这几类的分析直接不放在附录里了。放到正文里。我们先不写latex代码 先用中文正文的TSP风格来写。就是在下面这一段前面,先把S里面的p,q,p',q'给分类,在一个公式里面或者部分里面分成完全匹配,部分匹配的四种以及不匹配的剩下的,这样比较好的引出这个公式,分成匹配的和不匹配的:
Motivated by the above observation, we further partition the summation therein according to whether the index quadruple satisfies the vanishing-kernel condition . Specifically, we denote the contribution from the quadruples satisfying this condition by , and the remaining contribution by , which yields:
\begin{equation}
\mathbb{E}{|r_k|^2} = \mathcal{I}_k(\mathbf{g})+\mathcal{S}_k(\mathbf g,\boldsymbol{\theta}).
\end{equation}
下面是一段可直接放正文、偏 TSP 风格的中文表述:在你给的那段 “Motivated by …” 之前,先把 里的 做一个一次性分类,把“完全匹配 / 部分匹配(四类)/ 完全不匹配”说清楚,从而自然引出后面的 与 分解。
为便于刻画四重求和中核指数 的结构,我们首先对索引四元组 进行分类。记全体索引集合为
我们将 按“匹配关系”划分为三类:完全匹配、部分匹配(四种单超平面)以及完全不匹配(剩余补集),具体为
其中“”表示不交并,且各子集定义为
以及
上述划分的关键在于: 对应“完全配对”的四元组,其使得核指数满足vanishing-kernel 条件 (从而该部分对符号统计不敏感);– 中剔除 后得到的集合对应“仅满足单一配对关系”的部分匹配项; 则是既不满足任何单一配对关系的完全不匹配剩余项。基于这一结构分解,我们进一步将四重求和按是否满足 vanishing-kernel 条件 进行分组:令满足该条件的贡献记为 ,其余贡献记为 ,从而得到
如果你愿意,我还能把这段再做两个“投稿版微调”选项(不加长多少字):
可以 按你说的调整
可以,下面给一段正文可直接用的 TSP 风格中文表述(把两点都做了:1)把“”与“完全匹配”关系说清楚;2)把“部分匹配/完全不匹配”分别命名为 structured / bulk complement,和后文 对齐)。你可以把它放在你那段 “Motivated by …” 之前。
为刻画四重求和中核指数 的非零结构,我们首先对索引四元组 进行统一分类。记全体索引集合为
我们将 划分为三类:完全匹配(fully paired)、**部分匹配的 structured(四个单超平面族,剔除完全匹配)**以及 bulk complement(完全不匹配剩余项),即
其中“”表示不交并,并定义
以及四个单超平面族
并令
该划分与核指数的 vanishing-kernel 条件直接对应:在本文的核结构下, 中的完全配对会使 的各差分项成对抵消,因此
而 中的其余项不再具备这种完全抵消结构,从而对应非零核(即 )。因此,我们可将四重求和按是否满足 vanishing-kernel 条件 进行分组:将满足该条件的贡献记为 (对应完全匹配项),其余贡献记为 (对应 structured 与 bulk complement),从而得到
如果你希望更“硬核”一点(更像 TSP),我还能把上面那句“”改成更保守的表述(例如 “generically” 或 “in our setting”),以避免审稿人纠缠“是否存在非平凡解也能令 ”这种边界情况。
我们这个段落里面,SP的部分不分H1-4,把1-4写成一个集合就行
好的,下面把 structured(sp)那部分不再展开写 H1–H4,而是直接用一个“单超平面并集”集合来表述(仍然保留 fully paired / structured / bulk complement 三类的引出关系,且和后文 对齐)。
为刻画四重求和中核指数 的非零结构,我们首先对索引四元组 进行统一分类。记全体索引集合为
我们将 划分为三类:完全匹配(fully paired)、**structured(部分匹配)**以及 bulk complement(完全不匹配剩余项),即
其中“”表示不交并。完全匹配集合定义为
进一步,将“满足任一单一配对关系”的四元组集合(即四个单超平面的并集)记为
则 structured(部分匹配)与 bulk complement(完全不匹配)分别定义为
该划分与 vanishing-kernel 条件直接对应:在本文的核结构下, 中的完全配对使 的差分项成对抵消,因此
而其余的 structured 与 bulk complement 项不再具备这种完全抵消结构,从而对应非零核(即 )。因此,我们可将四重求和按是否满足 vanishing-kernel 条件 进行分组:将满足该条件的贡献记为 (对应完全匹配项),其余贡献记为 (对应 structured 与 bulk complement),从而得到
omega_sp直接用集合来定义 而不是写成omega_sp = omega_H/oomega_M
为刻画四重求和中核指数 的非零结构,我们首先对索引四元组 进行统一分类。记全体索引集合为
我们将 划分为三类:完全匹配(fully paired)、**structured(部分匹配)**以及 bulk complement(完全不匹配剩余项),即
其中“”表示不交并。完全匹配集合定义为
structured(部分匹配)集合直接定义为“满足恰好一个单一配对关系、但不属于完全匹配”的四元组集合:
其中 表示逻辑异或, 表示括号内命题为真(等价地说:四个条件中恰有一个成立)。bulk complement 则定义为
该划分与 vanishing-kernel 条件直接对应:在本文的核结构下, 中的完全配对使 的差分项成对抵消,因此
而 中的其余项不再具备这种完全抵消结构,从而对应非零核(即 )。因此,我们可将四重求和按是否满足 vanishing-kernel 条件 进行分组:将满足该条件的贡献记为 (对应完全匹配项),其余贡献记为 (对应 structured 与 bulk complement),从而得到
(如果你不想用异或符号,我也可以把 改写成四个“并集且两两互斥”的子集形式,但依然不通过 来定义。)
让我们参考下面的分类方式来定义集合
pythonfrom PIL import Image, ImageEnhance img = Image.open('/mnt/data/image.png') img.size (382, 133)
pythonimport matplotlib.pyplot as plt plt.figure(figsize=(9,3)) plt.imshow(img) plt.axis('off') plt.show()
好的,我们就按你图里那种“分段/分类”的写法来做一个互不相交的索引分类(完全匹配 3 类 + structured 的 4 类 + 其余 bulk),而且每一类都用集合本身直接定义(不再写成差集)。
为刻画四重求和中核指数 的结构,我们将索引四元组 按“配对关系”作如下互不相交的划分。记
我们将 分为三大块:完全匹配、**structured(部分匹配,四类)**以及 bulk(完全不匹配剩余项)。其中“完全匹配”再按是否全对角拆成与你截图一致的三种互斥情形:
其中
其中四个互斥子类分别对应“仅有一个等式成立”:
(这样写的好处是:四类天然互斥,也自动排除了上面“完全匹配”的情形。)
于是我们得到一个与截图风格一致的分段式集合划分:
在本文的核结构下,完全匹配集合 对应差分项的成对抵消,从而满足 vanishing-kernel 条件
而 则不具备这种完全抵消结构,对应非零核项并携带 的影响。基于此,我们将四重求和按 与否分组:把 的贡献记为 ,其余()的贡献记为 ,从而自然得到
如果你觉得上面 structured 的四个子集写得还是略长,我也可以把它们压成一行“恰好一个成立”的逻辑式,但仍保持“集合直接定义”的要求。
不要这个版本了 还是要上面一个版本。写的更像TSP的中文语言风格
为刻画四重求和中核指数 的结构,我们先对索引四元组 进行分类。记全体索引集合为
我们将其划分为三类:完全匹配(fully paired)、部分匹配的 structured 以及 完全不匹配的 bulk complement。其中完全匹配集合定义为
即两种“成对配对”模式(对角配对或一一对应配对)。接着,structured(部分匹配)集合定义为:满足至少一个单一配对关系,但又不属于上述完全匹配两种成对配对模式的四元组,即
最后,bulk complement(完全不匹配)集合则直接定义为四个“单一配对关系”均不成立的剩余项:
在本文考虑的核结构下,完全匹配项 对应差分项的成对抵消,从而满足 vanishing-kernel 条件 ;而 不再具备完全抵消结构,构成其余(一般为非零核)贡献。基于此,我们将四重求和按是否满足 分组:将 的贡献记为 ,其余贡献记为 ,从而得到
把上面的部分改写成符合TSP风格的英文latex代码给我
latex% --- Index-classification paragraph (TSP-style) --- To expose the structure of the kernel exponent $S(p,q,p',q')$ in the ensuing quadruple summation, we first classify the index quadruples $(p,q,p',q')\in\mathbb Z_N^4$. Let $$ \Omega \triangleq \{(p,q,p',q'):\; p,q,p',q'\in\mathbb Z_N\}. $$ We partition $\Omega$ into three disjoint classes: \emph{fully paired} quadruples, \emph{structured} (partially matched) quadruples, and the \emph{bulk complement} (fully mismatched) quadruples, i.e., $$ \Omega \;=\; \Omega_{\mathrm{M}} \;\dot\cup\; \Omega_{\mathrm{sp}} \;\dot\cup\; \Omega_{\mathrm{bulk}}, $$ where ``$\dot\cup$'' denotes disjoint union. The fully paired set is defined as $$ \Omega_{\mathrm{M}} \triangleq \Big\{(p,q,p',q'):\ (p=q \wedge p'=q') \ \vee\ (p=p' \wedge q=q')\Big\}. $$ Next, the structured set collects the quadruples that satisfy \emph{at least one} single pairing relation, but do \emph{not} fall into either of the two fully paired patterns above: $$ \Omega_{\mathrm{sp}} \triangleq \Big\{(p,q,p',q'):\ (p=q \vee p'=q' \vee p=p' \vee q=q') \ \wedge\ \neg\big[(p=q \wedge p'=q') \ \vee\ (p=p' \wedge q=q')\big]\Big\}. $$ Finally, the bulk complement is defined as the remaining quadruples for which none of the single pairing relations holds: $$ \Omega_{\mathrm{bulk}} \triangleq \Big\{(p,q,p',q'):\ p\neq q,\ \ p'\neq q',\ \ p\neq p',\ \ q\neq q'\Big\}. $$ Under the kernel structure considered in this work, the fully paired class $\Omega_{\mathrm{M}}$ induces exact cancellation among the difference terms in $S(p,q,p',q')$, and hence satisfies the vanishing-kernel condition, $$ (p,q,p',q')\in\Omega_{\mathrm{M}} \ \Longrightarrow\ S(p,q,p',q')=0, $$ whereas the remaining indices in $\Omega_{\mathrm{sp}}\cup\Omega_{\mathrm{bulk}}$ do not admit such complete cancellation and contribute to the non-vanishing kernel part. Motivated by this observation, we further partition the summation therein according to whether the index quadruple satisfies $S(p,q,p',q')=0$. Specifically, we denote the contribution from the quadruples satisfying this condition by $\mathcal{I}_k(\mathbf g)$, and the remaining contribution by $\mathcal{S}_k(\mathbf g,\boldsymbol{\theta})$, which yields: \begin{equation} \mathbb{E}\{|r_k|^2\} = \mathcal{I}_k(\mathbf{g})+\mathcal{S}_k(\mathbf g,\boldsymbol{\theta}). \end{equation}
\begin{equation}
\Omega_{\mathrm{sp}} \triangleq \Big{(p,q,p',q') \mid (p=q \lor p'=q' \lor p=p' \lor q=q') \land \neg\big[(p=q \land p'=q') \lor (p=p' \land q=q')\big]\Big}.
\end{equation}把这个还是拆成四个集合相加吧,现在这样不好看
latex% --- replace Omega_sp definition by a union of four disjoint subsets --- \begin{equation} \Omega_{\mathrm{sp}} \triangleq \Omega_{\mathrm{sp}}^{(1)} \,\dot\cup\, \Omega_{\mathrm{sp}}^{(2)} \,\dot\cup\, \Omega_{\mathrm{sp}}^{(3)} \,\dot\cup\, \Omega_{\mathrm{sp}}^{(4)}, \end{equation} where \begin{align} \Omega_{\mathrm{sp}}^{(1)} &\triangleq \Big\{(p,q,p',q'):\ p=q,\ \ p'\neq q',\ \ (p\neq p'\ \vee\ q\neq q')\Big\}, \label{eq:Omega_sp1}\\ \Omega_{\mathrm{sp}}^{(2)} &\triangleq \Big\{(p,q,p',q'):\ p'=q',\ \ p\neq q,\ \ (p\neq p'\ \vee\ q\neq q')\Big\}, \label{eq:Omega_sp2}\\ \Omega_{\mathrm{sp}}^{(3)} &\triangleq \Big\{(p,q,p',q'):\ p=p',\ \ q\neq q',\ \ (p\neq q\ \vee\ p'\neq q')\Big\}, \label{eq:Omega_sp3}\\ \Omega_{\mathrm{sp}}^{(4)} &\triangleq \Big\{(p,q,p',q'):\ q=q',\ \ p\neq p',\ \ (p\neq q\ \vee\ p'\neq q')\Big\}. \label{eq:Omega_sp4} \end{align}
这四个子集的设计点是:
如果你希望四个集合写得更“对称且更短”,我也可以把每个集合末尾的“排除项”改成统一的 (但那样每行会稍长一点)。