请深入学习一下下面的推导过程,用中文文本的方式讲给我,并用符合IEEE TSP风格的推导过程用lat...

बनाया गया: 21 दिसंबर 2025

का उपयोग करके उत्तर दिया गया GPT-5.2 Thinking द्वारा Chat01

प्रश्न

请深入学习一下下面的推导过程,用中文文本的方式讲给我,并用符合IEEE TSP风格的推导过程用latex代码重新整理一下:
\documentclass[11pt]{article}

\usepackage[a4paper,margin=1in]{geometry}
\usepackage{amsmath,amssymb,amsthm,bm}
\usepackage{mathtools}
\usepackage{hyperref}

% ---------- Macros ----------
\newcommand{\C}{\mathbb{C}}
\newcommand{\R}{\mathbb{R}}
\newcommand{\E}{\mathbb{E}}
\newcommand{\ii}{\mathrm{i}}

\newcommand{\ww}{w}
\newcommand{\Cw}{\mathbf C_{\ww}}
\newcommand{\Hk}{\mathbf H_k}
\newcommand{\Mv}{\mathbf M_v}
\DeclareMathOperator{\Ree}{Re}
\DeclareMathOperator{\Imm}{Im}
\newcommand{\indic}[1]{\mathbb{1}{#1}}

\newtheorem{assumption}{Assumption}

\title{Fourth-Moment Analysis of CE-OFDM Signals in a Single-User ISAC System}
\author{}
\date{}

\begin{document}
\maketitle

\section{System Model and Assumptions}

Consider the fourth-moment analysis of Constant Envelope (CE)-OFDM signals in a single-user Integrated Sensing and Communication (ISAC) system.
The core objective is to evaluate the fourth moment (equivalently expressed later as \E{Fx44}\E\{\|F x\|_4^4\}).

\subsection{Transmitted Symbols and Constellation Characteristics}

Let the transmitted symbol vector s\CNd×1\bm{s}\in\C^{N_d\times 1} be independent and identically distributed (i.i.d.), and its constellation S\mathcal{S} satisfies:

\begin{assumption}[Unit power normalization]
Constellation points are normalized to unit power:
\begin{equation}
\E(|s|^2)=1,\quad \forall s\in\mathcal{S}.
\end{equation}
\end{assumption}

\begin{assumption}[Zero mean and zero pseudo-variance]
The mean and pseudo-variance of constellation points are both zero:
\begin{equation}
\E(s)=0,\qquad \E(s^2)=0,\quad \forall s\in\mathcal{S}.
\end{equation}
\end{assumption}

Define the (normalized) kurtosis of the constellation as
\begin{equation}
\mu_4 \triangleq
\frac{\E{|s-\E(s)|^4}}{\big(\E{|s-\E(s)|^2}\big)^2}
= \E(|s|^4),
\end{equation}
where the last equality uses Assumptions 1--2 and \E(s2)=1\E(|s|^2)=1.

\subsection{Frequency-/Time-Domain Signal Modeling}

\paragraph{1) Hermitian-symmetric frequency-domain vector.}
To obtain a real-valued time-domain OFDM symbol, map s\bm{s} to a Hermitian-symmetric frequency-domain vector v\CN×1\bm{v}\in\C^{N\times 1}:
\begin{equation}
\bm{v}

\big[,0,\ s_1,\ s_2,\ \ldots,\ s_{N_d},\ \bm{z}p^{\mathsf{T}},\ 0,\ s{N_d}^,\ \ldots,\ s_2^,\ s_1^,\big]^{\mathsf{T}},
\end{equation}
where zpRNz×1\bm{z}_p\in\R^{N_z\times 1} is the zero-padding vector, and the total length is
\begin{equation}
N = 2(N_d+1)+N_z.
\end{equation}
The Hermitian symmetry is
\begin{equation}
v[0]=v[N/2]=0,\qquad v[N/2+k]=v^
[N/2-k],\ \ 1\le k\le N/2-1.
\end{equation}

\paragraph{2) Time-domain signal generation.}
Let FN\CN×NF_N\in\C^{N\times N} be the normalized Discrete Fourier Transform (DFT) matrix. The time-domain signal is generated via inverse DFT:
\begin{equation}
\tilde{\bm{v}} = F_N^{\mathsf{H}}\bm{v}\in\R^{N\times 1}.
\end{equation}
Using Hermitian symmetry, the nn-th element can be written as
\begin{equation}
\tilde{v}[n]
= \frac{2}{N}\sum_{k=1}^{N/2-1}
\Big(
\Ree{s[k]}\cos!\big(\tfrac{2\pi k n}{N}\big)
-\Imm{s[k]}\sin!\big(\tfrac{2\pi k n}{N}\big)
\Big).
\end{equation}

\paragraph{3) CE-OFDM signal definition.}
Define the CE-OFDM signal (elementwise exponential) as

Let xCN×1\mathbf{x} \in \mathbb C^{N\times 1} be defined as
\begin{equation}
\mathbf{x} = \tilde{A}\exp!\left(j \tilde{h}\mathbf{f}n^{H}\bm v \right)\in\C^{N\times 1},
\end{equation}
where the exponential is applied element-wise.
The kk-th frequency-domain autocorrelation sample is defined as
\begin{align}
r_k&=\mathbf{x}^H\mathbf{J}k\mathbf{x}\notag\
& = \sqrt{N}\mathbf{x}^H\mathbf{F}N^H\text{Diag}(\mathbf{f}{N-k+1})\mathbf{F}N\mathbf{x}\notag\
&=\sum
{n=1}^N\big|\mathbf{f}n\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-1)}{N}\right)\notag\
&=\sum
{n=1}^N\left(\sum
{p=1}^N\sum
{q=1}^N\mathbf{f}_n[p]\mathbf{f}n^*[q]\exp\left(j\left(\mathbf{f}p^H\mathbf{v}-\mathbf{f}q^H\mathbf{v}\right)\right)\right)\exp!\left(j\frac{2\pi k (n-1)}{N}\right)\notag\
& = \frac{1}{N} \sum
{n=1}^N\exp!\left(j\frac{2\pi k (n-1)}{N}\right)\sum
{p=1}^N\sum
{q=1}^N\exp\bigg[j(\mathbf{f}_p^H-\mathbf{f}_q^H)\mathbf{v})\bigg]\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)
\end{align}

where fnH\mathbf{f}_n^{H} denotes the nn-th row of the unitary DFT matrix fn\mathbf{f}_n.

\begin{align}
|r_k|^2 &= \sum_{n=1}^N\big|\mathbf{f}n\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-1)}{N}\right)\sum{m=1}^N\big|\mathbf{f}m\mathbf{x}\big|^2\exp!\left(-j\frac{2\pi k (m-1)}{N}\right)\notag\
& = \sum
{n=1}^N\sum_{m=1}^N \big|\mathbf{f}_n\mathbf{x}\big|^2\big|\mathbf{f}_m\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-m)}{N}\right)
\end{align}

where
\begin{align}
&\big|\mathbf{f}n\mathbf{x}\big|^2\big|\mathbf{f}m\mathbf{x}\big|^2 \notag\
&=\frac{1}{N^2}\sum
{p=1}^N\sum
{q=1}^N\exp\bigg[j(\mathbf{f}p^H-\mathbf{f}q^H)\mathbf{v})\bigg]\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\sum{p'=1}^N\sum{q'=1}^N\exp\bigg[-j(\mathbf{f}{p'}^H-\mathbf{f}{q'}^H)\mathbf{v})\bigg]\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\notag\
&= \frac{1}{N^2}\sum_{p,q,p'q'}\exp(j\mathbf{a}^H\mathbf{v})\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)
\end{align}
with
\begin{equation}
\mathbf{a}^H = \mathbf{f}p^H-\mathbf{f}q^H-\mathbf{f}{p'}^H+\mathbf{f}{q'}^H
\end{equation}

\begin{align}
|r_k|^2 & = \sum_{n=1}^N\sum_{m=1}^N \big|\mathbf{f}n\mathbf{x}\big|^2\big|\mathbf{f}m\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\notag\
& =\frac{1}{N^2}\sum
{n=1}^N\sum
{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\exp(j\mathbf{a}^H\mathbf{v})
\end{align}

\begin{align}
\mathbb{E}{|r_k|^2}& =\frac{1}{N^2}\mathbb{E}\left{\sum_{n=1}^N\sum_{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\exp(j\mathbf{a}^H\mathbf{v})\right}\notag\
& =\frac{1}{N^2}\sum_{n=1}^N\sum_{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\notag\
& \times \mathbb{E}\left{\exp(j\mathbf{a}^H\mathbf{v})\right}\notag\
\end{align}

\subsection{Characteristic function approximation (Lindeberg-CLT)}

We now replace the previous definition of a\bm a by
\begin{equation}
\bm a = \tilde h,(\bm f_p-\bm f_q-\bm f_{p'}+\bm f_{q'}),\qquad
\bm a^{\mathsf H}=\tilde h,(\bm f_p^{\mathsf H}-\bm f_q^{\mathsf H}-\bm f_{p'}^{\mathsf H}+\bm f_{q'}^{\mathsf H}).
\end{equation}
(If the new model does not include h~\tilde h, simply set h~=1\tilde h=1 in what follows.)

Using the Hermitian symmetry of v\bm v and the conjugate property of DFT columns, one still has
a[Nr]=a[r]a[N-r]=a^*[r]. Hence,
\begin{equation}
\bm a^{\mathsf H}\bm v=\sum_{r=1}^{N_d} z_r,
\qquad
z_r \triangleq a[r]^* s_r + a[r] s_r^*
=2,\Re{a[r]^* s_r}.
\end{equation}
The random variables {zr}\{z_r\} are independent but generally not identically distributed.

Under the Lindeberg condition
\begin{equation}
\max_{r}\frac{|a[r]|^2}{\sum_{r=1}^{N_d}|a[r]|^2}\to 0,
\qquad \mu_4<\infty,
\end{equation}
a cumulant expansion truncated to fourth order gives the approximation
\begin{equation}\label{eq:charfun_approx_new}
\E!\left[\exp!\big(\ii,\bm a^{\mathsf H}\bm v\big)\right]
\approx
\exp!\left(
-\sum_{r=1}^{N_d}|a[r]|^2
+\frac{\mu_4-2}{4}\sum_{r=1}^{N_d}|a[r]|^4
\right).
\end{equation}

After summing over kk (DFT orthogonality), the new sign pattern yields the modular constraint
\begin{equation}
p-q-p'+q' \equiv 0\ (\mathrm{mod}\ N)
\qquad\Leftrightarrow\qquad
p+q'\equiv q+p'\ (\mathrm{mod}\ N).
\end{equation}
A convenient re-parameterization is
\begin{equation}
(p,q,p',q')=(u+v,\ v,\ u,\ 0)\quad(\mathrm{mod}\ N).
\end{equation}

Let ω2π/N\omega\triangleq 2\pi/N. Define
\begin{align}
U_r(u,v)
&\triangleq e^{j\omega r(u+v)}-e^{j\omega rv}-e^{j\omega ru}+1 \nonumber\
&= (1-e^{j\omega ru})(1-e^{j\omega rv}).
\end{align}
With the normalized DFT convention, this implies
\begin{equation}
a[r]=\frac{\tilde h}{N}U_r(u,v).
\end{equation}
Introduce the deterministic sums
\begin{equation}
S_2(u,v)\triangleq \sum_{r=1}^{N_d}|U_r(u,v)|^2,\qquad
S_4(u,v)\triangleq \sum_{r=1}^{N_d}|U_r(u,v)|^4.
\end{equation}
Then
\begin{equation}
\sum_{r=1}^{N_d}|a[r]|^2=\frac{\tilde h^2}{N^2}S_2(u,v),
\qquad
\sum_{r=1}^{N_d}|a[r]|^4=\frac{\tilde h^4}{N^4}S_4(u,v),
\end{equation}
and \eqref{eq:charfun_approx_new} becomes
\begin{equation}
\E!\left[\exp!\big(j\bm a^{\mathsf H}\bm v\big)\right]
\approx
\exp!\left(
-\frac{\tilde h^2}{N^2}S_2(u,v)
+\frac{\mu_4-2}{4}\frac{\tilde h^4}{N^4}S_4(u,v)
\right).
\end{equation}

\section{Closed-Form Expressions for S2S_2 and S4S_4}

Define the auxiliary function
\begin{equation}
\Psi(d)\triangleq \sum_{r=1}^{N_d}\cos!\Big(\frac{2\pi}{N} r d\Big),
\qquad \Psi(d)=\Psi(-d).
\end{equation}
Then the standard closed form is
\begin{equation}
\boxed{
\Psi(d)=
\begin{cases}
N_d, & d\equiv 0\ (\mathrm{mod}\ N),\
\dfrac{\sin!\left(\frac{N_d\omega d}{2}\right)\cos!\left(\frac{(N_d+1)\omega d}{2}\right)}
{\sin!\left(\frac{\omega d}{2}\right)}, & \text{otherwise}.
\end{cases}}
\end{equation}

\subsection{Second-order sum (new index/sign pattern)}
For Ur(p,q,p,q)=ejωrpejωrqejωrp+ejωrqU_r(p,q,p',q') = e^{j\omega rp}-e^{j\omega rq}-e^{j\omega rp'}+e^{j\omega rq'}
(one may equivalently write Ur=ejωrp+ejωrqejωrqejωrpU_r=e^{j\omega rp}+e^{j\omega rq'}-e^{j\omega rq}-e^{j\omega rp'}),
the closed form becomes
\begin{equation}
\boxed{
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].}
\end{equation}

\subsection{Fourth-order sum (updated difference vector)}
Let (c0,c1,c2)=(32,2,12)(c_0,c_1,c_2)=\big(\tfrac{3}{2},-2,\tfrac{1}{2}\big) and define the difference vector
\begin{equation}
\bm g=\big[g_1,\ldots,g_6\big]^{\mathsf{T}}

\big[
p-q',\ q-p',\ p-q,\ p-p',\ q'-q,\ q'-p'
\big]^{\mathsf{T}}.
\end{equation}
A compact expression consistent with the provided formula is
\begin{align}
S_4(p,q,p',q')
&=
(16N_d+16)\Big[
\Psi(p-q')+\Psi(q-p')
-\Psi(p-q)-\Psi(p-p')-\Psi(q'-q)-\Psi(q'-p')
\Big]\nonumber\
&\quad
+2\sum_{m=0}^{2}\sum_{n=0}^{2} c_m c_n
\sum_{i=1}^{6}\sum_{j'=1}^{6}
\Big[
\Psi(m g_i - n g_{j'}) + \Psi(m g_i + n g_{j'})
\Big].
\end{align}

\subsection{Kernel-matrix form (updated b\mathbf b and n(u,v)\mathbf n(u,v))}
Define

b[1,1,1,1]T,n(u,v)[u+v, v, u, 0]T.\mathbf b\triangleq[1,-1,-1,1]^T,\qquad \mathbf n(u,v)\triangleq[u+v,\ v,\ u,\ 0]^T.

Define the 4×44\times 4 kernel matrix
\begin{equation}
[\mathbf K_2(u,v)]{ij}\triangleq \Psi(n_i-n_j),
\qquad i,j\in{1,2,3,4},
\end{equation}
then
\begin{equation}
S_2(u,v)=\mathbf b^T\mathbf K_2(u,v)\mathbf b.
\end{equation}
Let wbbR16\mathbf w\triangleq \mathbf b\otimes \mathbf b\in\mathbb R^{16} and index pairs (i,j)(i,j) as one index in {1,,16}\{1,\dots,16\}.
Define the 16×1616\times16 kernel
\begin{equation}
[\mathbf K_4(u,v)]
{(i,j),(k,\ell)}\triangleq
\Psi\big((n_i-n_j)+(n_k-n_\ell)\big),
\end{equation}
then
\begin{equation}
S_4(u,v)=\mathbf w^T\mathbf K_4(u,v)\mathbf w.
\end{equation}
Finally,
\begin{equation}
\E!\left[\exp!\big(j\bm a^{\mathsf{H}}\bm v\big)\right]
\approx
\exp!\left(
-\frac{\tilde{h}^2}{N^2}\mathbf b^T\mathbf K_2(u,v)\mathbf b
+\frac{\mu_4-2}{4}\frac{\tilde{h}^4}{N^4}(\mathbf b\otimes \mathbf b)^T\mathbf K_4(u,v)(\mathbf b\otimes \mathbf b)
\right).
\end{equation}

% 2nd-order truncation (discard S4):
\begin{equation}
\mathbb E!\left[\exp(j\bm a^H\bm v)\right]\approx
\exp!\left(-\frac{\tilde h^2}{N^2}S_2(p,q,p',q')\right).
\end{equation}

\begin{equation}
\boxed{
\mathbb E{|r_k|^2}\approx
\frac{1}{N^3}\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m,e^{j\frac{2\pi}{N}k(n-m)}
\sum_{p,q,p',q'}
e^{-j\frac{2\pi}{N}\left((n-1)(q-p)+(m-1)(p'-q')\right)}
\exp!\left(-\frac{\tilde h^2}{N^2}S_2(p,q,p',q')\right).
}
\end{equation}

% ===== Iceberg extraction: terms with a = 0 (pairings) =====
% For a^H = f_p^H - f_q^H - f_{p'}^H + f_{q'}^H:
% Pairing-1: p=q and q'=p'
% Pairing-2: p=p' and q=q'
% These two pairings contribute (2N^2 - N) quadruples in (p,q,p',q').

\begin{equation}
\mathcal I_k(\bm g)

\frac{1}{N^3}(2N^2-N)\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m e^{j\frac{2\pi}{N}k(n-m)}

\left(2-\frac{1}{N}\right)\big|\bm f_k^H\bm g\big|^2.
\end{equation}

\begin{equation}
\boxed{
\mathbb E{|r_k|^2}\approx \mathcal I_k(\bm g)+\mathcal S_k(\bm g),
\qquad
\mathcal I_k(\bm g)=\left(2-\frac{1}{N}\right)|\bm f_k^H\bm g|^2,
}
\end{equation}

\begin{equation}
\mathcal S_k(\bm g)\triangleq
\frac{1}{N^3}\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m,e^{j\frac{2\pi}{N}k(n-m)}
\sum_{p,q,p',q'\notin \text{pairings}}
e^{-j\frac{2\pi}{N}\left((n-1)(q-p)+(m-1)(p'-q')\right)}
\exp!\left(-\frac{\tilde h^2}{N^2}S_2(p,q,p',q')\right).
\end{equation}

% ============================================================
% TSP-style derivation starting from the 4-index summation
% (no (u,v,t) re-parameterization), and reusing K_2(·) in 4 args
% ============================================================

\paragraph{Starting point (4-index form).}
Let ω2π/N\omega \triangleq 2\pi/N and αh~2/N2\alpha \triangleq \tilde h^2/N^2. From
\begin{equation}\label{eq:Ek_rk2_start}
\boxed{
\mathbb E{|r_k|^2}\approx
\frac{1}{N^3}\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m,e^{j\omega k(n-m)}
\sum_{p,q,p',q'}
e^{-j\omega\left((n-1)(q-p)+(m-1)(p'-q')\right)}
\exp!\left(-\alpha,S_2(p,q,p',q')\right),
}
\end{equation}
we re-introduce the \emph{four-parameter} kernel representation of S2S_2.
Define the index vector and kernel matrix

n(p,q,p,q)[p, q, p, q]T (mod N),[K2(p,q,p,q)]ijΨ(ninj), i,j{1,2,3,4},\mathbf n(p,q,p',q')\triangleq [p,\ q,\ p',\ q']^T\ (\mathrm{mod}\ N),\qquad [\mathbf K_2(p,q,p',q')]_{ij}\triangleq \Psi(n_i-n_j),\ i,j\in\{1,2,3,4\},

with Ψ(d)r=1Ndcos(ωrd)\Psi(d)\triangleq \sum_{r=1}^{N_d}\cos(\omega r d), and set

b[1,1,1,1]T.\mathbf b\triangleq[1,-1,-1,1]^T.

Then
\begin{equation}\label{eq:S2_K2_4args}
\boxed{
S_2(p,q,p',q')=\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b,
}
\end{equation}
so that the weight in \eqref{eq:Ek_rk2_start} becomes exp ⁣(αbTK2(p,q,p,q)b)\exp\!\big(-\alpha\,\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b\big) without introducing any new 1D kernel.

\paragraph{Pairings and iceberg term Ik(g)\mathcal I_k(\mathbf g).}
Under the CE-OFDM model xj=exp(jh~fjHv)x_j=\exp(-j\tilde h\,\mathbf f_j^H\mathbf v), the 4th-order moment entering the derivation has the characteristic-function form
E[xqxpxpxq]=E[exp(jaHv)]\mathbb E[x_qx_{p'}x_p^*x_{q'}^*]=\mathbb E[\exp(j\mathbf a^H\mathbf v)] with

aH=h~(fpHfqHfpH+fqH).\mathbf a^H=\tilde h\big(\mathbf f_p^H-\mathbf f_q^H-\mathbf f_{p'}^H+\mathbf f_{q'}^H\big).

Hence a=0\mathbf a=\mathbf 0 (equivalently S2=0S_2=0 and exp(αS2)=1\exp(-\alpha S_2)=1) occurs if and only if one of the following \emph{pairings} holds:

(A) p=q and p=q,(B) p=p and q=q.\text{(A)}\ p=q\ \text{and}\ p'=q',\qquad \text{(B)}\ p=p'\ \text{and}\ q=q'.

These contribute A+BAB=N2+N2N=2N2N|\text{A}|+|\text{B}|-|\text{A}\cap\text{B}|=N^2+N^2-N=2N^2-N quadruples. Extracting them from \eqref{eq:Ek_rk2_start} yields
\begin{equation}\label{eq:Ik_def}
\mathcal I_k(\mathbf g)

\frac{1}{N^3}(2N^2-N)\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m e^{j\omega k(n-m)}.
\end{equation}
Using

n=1Nm=1Ngngmejωk(nm)=n=1Ngnejωk(n1)2=NfkHg2,\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m e^{j\omega k(n-m)} =\left|\sum_{n=1}^N g_n e^{j\omega k(n-1)}\right|^2 = N\,|\mathbf f_k^H\mathbf g|^2,

we obtain the closed form
\begin{equation}\label{eq:Ik_closed}
\boxed{
\mathcal I_k(\mathbf g)=\left(2-\frac{1}{N}\right)|\mathbf f_k^H\mathbf g|^2.
}
\end{equation}

\paragraph{Sea-level term Sk(g)\mathcal S_k(\mathbf g) and elimination of the (n,m)(n,m) sums.}
Let the pairing set be

P{(p,q,p,q): (p=qp=q) or (p=pq=q)}.\mathcal P \triangleq \{(p,q,p',q'):\ (p=q\land p'=q')\ \text{or}\ (p=p'\land q=q')\}.

Define

\mathbb E|r_k|^2 \approx \mathcal I_k(\mathbf g)+\mathcal S_k(\mathbf g), \qquad \mathcal S_k(\mathbf g)\ \text{is \eqref{eq:Ek_rk2_start} restricted to}\ (p,q,p',q')\notin\mathcal P.

Next, combine the exponential factors in \eqref{eq:Ek_rk2_start} to separate the nn- and mm-dependent parts:

ejωk(nm)ejω(n1)(qp)=ejω(n1)(k+pq),ejωk(nm)ejω(m1)(pq)=ejω(m1)(kp+q).e^{j\omega k(n-m)}e^{-j\omega(n-1)(q-p)}=e^{j\omega(n-1)(k+p-q)},\qquad e^{-j\omega k(n-m)}e^{-j\omega(m-1)(p'-q')}=e^{j\omega(m-1)(-k-p'+q')}.

Therefore,

n=1Ngnejω(n1)(k+pq)=Nf\modNk+pqHg,m=1Ngmejω(m1)(kp+q)=Nf\modNkp+qHg,\sum_{n=1}^N g_n e^{j\omega(n-1)(k+p-q)}=\sqrt N\,\mathbf f_{\modN{k+p-q}}^H\mathbf g,\qquad \sum_{m=1}^N g_m e^{j\omega(m-1)(-k-p'+q')}=\sqrt N\,\mathbf f_{\modN{-k-p'+q'}}^H\mathbf g,

where \modN\modN{\cdot} denotes reduction modulo NN into {0,,N1}\{0,\ldots,N-1\}. Substituting back gives the \emph{pure 4-index} sea-level representation
\begin{equation}\label{eq:Sk_4idx_DFT}
\boxed{
\mathcal S_k(\mathbf g)=
\frac{1}{N^2}!!\sum_{(p,q,p',q')\notin\mathcal P}!!
\big(\mathbf f_{\modN{k+p-q}}^H\mathbf g\big),
\big(\mathbf f_{\modN{-k-p'+q'}}^H\mathbf g\big),
\exp!\left(-\alpha,\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b\right).
}
\end{equation}

\begin{equation}
\boxed{
\mathcal Q_k=
\frac{1}{N^2}!!\sum_{(p,q,p',q')\notin\mathcal P}!!
\big(\mathbf f_{\modN{k+p-q}}^H\big),
\big(\mathbf f_{\modN{-k-p'+q'}}^H\big),
\exp!\left(-\alpha,\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b\right).
}
\end{equation}

\begin{equation}
\boxed{
\mathcal Q_k=\mathcal Q_k^{SP}+\mathcal Q_k^{bulk}
}
\end{equation}

\paragraph{Structured hyperplanes (single-pair layers) and matrix forms.}
We partition the non-pairing set into structured hyperplanes and a residual bulk:

(H1) p=q,(H2) p=q,(H3) q=q,(H4) p=p,bulk: complement.\text{(H1)}\ p'=q',\qquad \text{(H2)}\ p=q,\qquad \text{(H3)}\ q=q',\qquad \text{(H4)}\ p=p',\qquad \text{bulk: complement}.

On each hyperplane, bTK2(p,q,p,q)b\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b reduces (by shift-invariance of Ψ()\Psi(\cdot)) to a function of a single modular difference d\ZNd\in\ZN. We encode this \emph{using K2K_2 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\ZN ,
\end{equation}
which is merely a shorthand (no new model) since it is generated by the original four-argument kernel.

% ===== kappa[d] in closed form and vectorization =====

% Parameters

ω2πN,αh~2N2,b[1,1,1,1]T.\omega \triangleq \frac{2\pi}{N},\qquad \alpha\triangleq \frac{\tilde h^2}{N^2},\qquad \mathbf b \triangleq [\,1,-1,-1,1\,]^T.

% Psi(d)

Ψ(d)r=1Ndcos(ωrd),Ψ(d)=Ψ(d).\Psi(d)\triangleq \sum_{r=1}^{N_d}\cos(\omega r d),\qquad \Psi(d)=\Psi(-d).

% Key simplification for the special representative (d,0,0,0)

bTK2(d,0,0,0)b=2(Ψ(0)Ψ(d))=2(NdΨ(d)).\boxed{ \mathbf b^T \mathbf K_2(d,0,0,0)\mathbf b =2\big(\Psi(0)-\Psi(d)\big) =2\big(N_d-\Psi(d)\big). }

% kappa[d] explicit (sum form)

κ[d]exp ⁣(αbTK2(d,0,0,0)b)=exp ⁣(2α(NdΨ(d)))=exp ⁣(2h~2N2(Ndr=1Ndcos(ωrd))).\boxed{ \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) =\exp\!\left( -\frac{2\tilde h^2}{N^2}\Big(N_d-\sum_{r=1}^{N_d}\cos(\omega r d)\Big) \right). }

% ===== kappa as a length-N vector =====

κ[κ[0],κ[1],,κ[N1]]TRN,(κ)d=κ[d],  dZN.\boldsymbol\kappa \triangleq \big[\kappa[0],\kappa[1],\ldots,\kappa[N-1]\big]^T \in \mathbb R^{N}, \qquad (\boldsymbol\kappa)_d = \kappa[d],\ \ d\in\mathbb Z_N.

% (Optional) elementwise definition

(κ)d=exp ⁣(2α(NdΨ(d))),d=0,1,,N1.\boxed{ (\boldsymbol\kappa)_d =\exp\!\left(-2\alpha\big(N_d-\Psi(d)\big)\right), \qquad d=0,1,\ldots,N-1. }

\emph{Low-rank layers:} the hyperplanes (H1) and (H2) lead to two (Hermitianized) rank-2\le2 quadratic forms

Qk,ut(0)1Nd=1N1κ[d] fkfk+dH,Qk,ut12(Qk,ut(0)+Qk,ut(0)H),\mathbf Q_{k,ut}^{(0)}\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\kappa[d]\ \mathbf f_{-k}^{*}\mathbf f_{k+d}^{H}, \qquad \mathbf Q_{k,ut}\triangleq \tfrac12\big(\mathbf Q_{k,ut}^{(0)}+\mathbf Q_{k,ut}^{(0)H}\big), Qk,v0(0)1Nd=1N1κ[d] fkfdkH,Qk,v012(Qk,v0(0)+Qk,v0(0)H).\mathbf Q_{k,v0}^{(0)}\triangleq \frac{1}{N}\sum_{d=1}^{N-1}\kappa[d]\ \mathbf f_{k}^{*}\mathbf f_{d-k}^{H}, \qquad \mathbf Q_{k,v0}\triangleq \tfrac12\big(\mathbf Q_{k,v0}^{(0)}+\mathbf Q_{k,v0}^{(0)H}\big).

% ===== Normalized DFT convention =====

ω2πN,[F]n,d=1Nejω(n1)d,n=1,,N, d=0,,N1,\omega \triangleq \frac{2\pi}{N},\qquad [\mathbf F]_{n,d}=\frac{1}{\sqrt N}e^{-j\omega (n-1)d}, \quad n=1,\ldots,N,\ d=0,\ldots,N-1,

and let fd\mathbf f_d denote the dd-th column of F\mathbf F (indices are taken mod NN).

% ===== Pack kappa into a length-N vector and take one FFT =====

κ~[0]0,κ~[d]κ[d], d=1,,N1,κ~[κ~[0],,κ~[N1]]TRN,\tilde\kappa[0]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \kappa[d],\ d=1,\ldots,N-1, \qquad \tilde{\boldsymbol\kappa}\triangleq[\tilde\kappa[0],\ldots,\tilde\kappa[N-1]]^T\in\mathbb R^N, sFκ~=d=1N1κ[d]fdCN.\mathbf s \triangleq \mathbf F\,\tilde{\boldsymbol\kappa} =\sum_{d=1}^{N-1}\kappa[d]\mathbf f_d\in\mathbb C^N.

% ===== Key identity: elementwise product of DFT columns =====
For normalized DFT columns, the elementwise products satisfy

fkfd=1Nfdk,fkfd=1Nfd+k,(mod N).\mathbf f_k^*\odot \mathbf f_d=\frac{1}{\sqrt N}\mathbf f_{d-k}, \qquad \mathbf f_k\odot \mathbf f_d=\frac{1}{\sqrt N}\mathbf f_{d+k}, \qquad (\,\mathrm{mod}\ N\,).

Hence,

d=1N1κ[d]fd+k=N(fks),d=1N1κ[d]fdk=N(fks).\sum_{d=1}^{N-1}\kappa[d]\mathbf f_{d+k} =\sqrt N\,(\mathbf f_{-k}\odot \mathbf s), \qquad \sum_{d=1}^{N-1}\kappa[d]\mathbf f_{d-k} =\sqrt N\,(\mathbf f_{k}\odot \mathbf s).

% ===== Compact forms for Q_{k,ut}^{(0)} and Q_{k,v0}^{(0)} without D_k =====
Using fk=fk\mathbf f_{-k}^*=\mathbf f_k and fk=fk\mathbf f_k^*=\mathbf f_{-k},

Qk,ut(0)=1Nfk(d=1N1κ[d]fd+k)H=NNfk(fks)H,\boxed{ \mathbf Q_{k,ut}^{(0)} =\frac{1}{N}\,\mathbf f_k\Big(\sum_{d=1}^{N-1}\kappa[d]\mathbf f_{d+k}\Big)^H =\frac{\sqrt N}{N}\,\mathbf f_k\,(\mathbf f_{-k}\odot \mathbf s)^H, } Qk,v0(0)=1Nfk(d=1N1κ[d]fdk)H=NNfk(fks)H,s=Fκ~.\boxed{ \mathbf Q_{k,v0}^{(0)} =\frac{1}{N}\,\mathbf f_{-k}\Big(\sum_{d=1}^{N-1}\kappa[d]\mathbf f_{d-k}\Big)^H =\frac{\sqrt N}{N}\,\mathbf f_{-k}\,(\mathbf f_{k}\odot \mathbf s)^H, } \qquad \mathbf s=\mathbf F\tilde{\boldsymbol\kappa}.

% ===== Hermitian symmetrization =====

Qk,ut12(Qk,ut(0)+Qk,ut(0)H),Qk,v012(Qk,v0(0)+Qk,v0(0)H).\mathbf Q_{k,ut}\triangleq \tfrac12\big(\mathbf Q_{k,ut}^{(0)}+\mathbf Q_{k,ut}^{(0)H}\big), \qquad \mathbf Q_{k,v0}\triangleq \tfrac12\big(\mathbf Q_{k,v0}^{(0)}+\mathbf Q_{k,v0}^{(0)H}\big).

% ===== Final low-rank sum (no D_k) =====

Qk,ut+Qk,v0=N2N(fk(fks)H+(fks)fkH+fk(fks)H+(fks)fkH),s=Fκ~.\boxed{ \mathbf Q_{k,ut}+\mathbf Q_{k,v0} = \frac{\sqrt N}{2N}\Big( \mathbf f_k(\mathbf f_{-k}\odot \mathbf s)^H + (\mathbf f_{-k}\odot \mathbf s)\mathbf f_k^H + \mathbf f_{-k}(\mathbf f_{k}\odot \mathbf s)^H + (\mathbf f_{k}\odot \mathbf s)\mathbf f_{-k}^H \Big), \quad \mathbf s=\mathbf F\tilde{\boldsymbol\kappa}. }

\emph{Diagonal layer:} the hyperplanes (H3) and (H4) collapse (via the DFT shift/modulation property) to a diagonal matrix

Dddiag ⁣(ejωd(n1))n=1N,Qdiag1N2d=1N1κ[d](Dd+Dd).\mathbf D_d \triangleq \mathrm{diag}\!\big(e^{-j\omega d(n-1)}\big)_{n=1}^N,\qquad \mathbf Q_{\mathrm{diag}}\triangleq \frac{1}{N^2}\sum_{d=1}^{N-1}\kappa[d]\big(\mathbf D_d+\mathbf D_{-d}\big).

% ===== Key idea: represent the diagonal operator by its diagonal vector =====
Since Qdiag\mathbf Q_{\mathrm{diag}} is diagonal, define its diagonal vector

qdiagQdiag1RN,\mathbf q_{\mathrm{diag}}\triangleq \mathbf Q_{\mathrm{diag}}\mathbf 1 \in \mathbb R^N,

% ===== Compute q_diag explicitly =====
For each n=1,,Nn=1,\ldots,N,
\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-NN vector κ~RN\tilde{\boldsymbol\kappa}\in\mathbb R^N by

κ~[0]0,κ~[d]κ[d], d=1,,N1,κ~=[κ~[0],,κ~[N1]]T.\tilde\kappa[0]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \kappa[d],\ d=1,\ldots,N-1, \qquad \tilde{\boldsymbol\kappa}=[\tilde\kappa[0],\ldots,\tilde\kappa[N-1]]^T.

Then

sFκ~=d=1N1κ[d]fd,\mathbf s \triangleq \mathbf F\tilde{\boldsymbol\kappa} = \sum_{d=1}^{N-1}\kappa[d]\mathbf f_d,

and

(qdiag)n=1N2(Nsn+Nsn)=2N3/2{sn}.(\mathbf q_{\mathrm{diag}})_n =\frac{1}{N^2}\Big(\sqrt N\,s_n+\sqrt N\,s_n^*\Big) =\frac{2}{N^{3/2}}\Re\{s_n\}.

Therefore,

qdiag=2N3/2{Fκ~}=2N3/2{d=1N1κ[d]fd},\boxed{ \mathbf q_{\mathrm{diag}} =\frac{2}{N^{3/2}}\Re\{\mathbf F\tilde{\boldsymbol\kappa}\} =\frac{2}{N^{3/2}}\Re\left\{\sum_{d=1}^{N-1}\kappa[d]\mathbf f_d\right\}, }

and the diagonal operator can be applied without any diag()\mathrm{diag}(\cdot) as

Qdiag=(2N3/2Fκ~).\mathbf Q_{\mathrm{diag}}= \left(\frac{2}{N^{3/2}}\mathbf F\tilde{\boldsymbol\kappa}\right).

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 =====

ω2πN,[F]n,d=1Nejω(n1)d,n=1,,N, d=0,,N1,\omega \triangleq \frac{2\pi}{N},\qquad [\mathbf F]_{n,d}=\frac{1}{\sqrt N}e^{-j\omega (n-1)d}, \quad n=1,\ldots,N,\ d=0,\ldots,N-1,

and let fd\mathbf f_d be the dd-th column of F\mathbf F (indices mod NN).

% ===== kappa vector and one FFT =====

κ~[0]0,κ~[d]κ[d], d=1,,N1,κ~[κ~[0],,κ~[N1]]T,\tilde\kappa[0]\triangleq 0,\qquad \tilde\kappa[d]\triangleq \kappa[d],\ d=1,\ldots,N-1, \qquad \tilde{\boldsymbol\kappa}\triangleq[\tilde\kappa[0],\ldots,\tilde\kappa[N-1]]^T, sFκ~.\mathbf s\triangleq \mathbf F\tilde{\boldsymbol\kappa}.

% ===== s-factor matrix shared by all terms =====

S(s)1sH+s1T,so [S(s)]n,n=sn+sn=2{sn}.\mathbf S(\mathbf s)\triangleq \mathbf 1\,\mathbf s^H+\mathbf s\,\mathbf 1^T, \qquad \text{so } [\mathbf S(\mathbf s)]_{n,n}=s_n^*+s_n=2\Re\{s_n\}.

% ===== Q_diag without diag(·): diagonal picked by Hadamard with I =====

Qdiag=2N3/2diag({s})=1N3/2(IS(s)).\boxed{ \mathbf Q_{\mathrm{diag}} =\frac{2}{N^{3/2}}\operatorname{diag}(\Re\{\mathbf s\}) =\frac{1}{N^{3/2}}\big(\mathbf I\odot \mathbf S(\mathbf s)\big). }

% ===== Low-rank sum expressed with the SAME S(s) =====

AkfkfkH+fkfkH,\mathbf A_k\triangleq \mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H, Qk,ut+Qk,v0=N2N(AkS(s)).\boxed{ \mathbf Q_{k,ut}+\mathbf Q_{k,v0} =\frac{\sqrt N}{2N}\big(\mathbf A_k\odot \mathbf S(\mathbf s)\big). }

% 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:

zFκ~,S~1zH+z1T.\mathbf z \triangleq \mathbf F\tilde{\boldsymbol\kappa},\qquad \tilde{\mathbf S}\triangleq \mathbf 1\,\mathbf z^H+\mathbf z\,\mathbf 1^T .

Note that diag(S~)=2{z}=2{Fκ~}\mathrm{diag}(\tilde{\mathbf S})=2\Re\{\mathbf z\}=2\Re\{\mathbf F\tilde{\boldsymbol\kappa}\}.

Define

AkfkfkH+fkfkH,Qksp=(1N3/2I+N2NAk)S~.\mathbf A_k\triangleq \mathbf f_k\mathbf f_{-k}^H+\mathbf f_{-k}\mathbf f_k^H, \qquad \mathbf Q_k^{\mathrm{sp}} =\Big(\tfrac{1}{N^{3/2}}\mathbf I+\tfrac{\sqrt N}{2N}\mathbf A_k\Big)\odot \tilde{\mathbf S}.

If Qksp\mathbf Q_k^{\mathrm{sp}} is Hermitian PSD, then

Sksp(g)=gHQkspg=(Qksp)1/2g22=((1N3/2I+N2NAk)S~)1/2g22.\boxed{ \mathcal S_k^{\mathrm{sp}}(\mathbf g) =\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g =\big\|(\mathbf Q_k^{\mathrm{sp}})^{1/2}\mathbf g\big\|_2^2 = \left\| \left( \Big(\tfrac{1}{N^{3/2}}\mathbf I+\tfrac{\sqrt N}{2N}\mathbf A_k\Big)\odot \tilde{\mathbf S} \right)^{1/2}\mathbf g \right\|_2^2 . }

\paragraph{Bulk term as a vec--matrix--vec product (still 4 indices).}
Define the DFT coefficient vector γFHg\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g so that γ=fHg\gamma_\ell=\mathbf f_\ell^H\mathbf g.
Introduce the two N×NN\times N circulant-by-difference matrices

[Hk]p,qf\modNk+pqHg=γ\modNk+pq,[Gk]p,qf\modNkp+qHg=γ\modNk+(qp).[\mathbf H_k]_{p,q}\triangleq \mathbf f_{\modN{k+p-q}}^H\mathbf g=\gamma_{\modN{k+p-q}}, \qquad [\mathbf G_k]_{p',q'}\triangleq \mathbf f_{\modN{-k-p'+q'}}^H\mathbf g=\gamma_{\modN{-k+(q'-p')}}.

Let ck[d]γ\modNk+d\mathbf c_k[d]\triangleq \gamma_{\modN{k+d}} and ck[d]γ\modNk+d\mathbf c_{-k}[d]\triangleq \gamma_{\modN{-k+d}}. Then

Hk=circ(ck),Gk=circ(ck)T.\mathbf H_k=\mathrm{circ}(\mathbf c_k),\qquad \mathbf G_k=\mathrm{circ}(\mathbf c_{-k})^{T}.

Next define the N2×N2N^2\times N^2 kernel matrix and an explicit bulk mask:

[K](p,q),(p,q)exp ⁣(αbTK2(p,q,p,q)b),[\mathbf K]_{(p,q),(p',q')}\triangleq \exp\!\left(-\alpha\,\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b\right), [Pbulk](p,q),(p,q)1{pq, pq, pp, qq}(mod N),[\mathbf P^{\mathrm{bulk}}]_{(p,q),(p',q')}\triangleq \mathbf 1\{p\neq q,\ p'\neq q',\ p\neq p',\ q\neq q'\}\qquad(\mathrm{mod}\ N),

which removes all pairings and the structured hyperplanes. The bulk contribution becomes
\begin{equation}\label{eq:Sk_bulk_vec}
\boxed{
\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)=
\frac{1}{N^2},
\operatorname{vec}(\mathbf H_k)^{T},
\big(\mathbf K\odot \mathbf P^{\mathrm{bulk}}\big),
\operatorname{vec}(\mathbf G_k).
}
\end{equation}

\paragraph{Final TSP-style decomposition.}
Combining \eqref{eq:Ik_closed}, \eqref{eq:Sk_sp_quadratic}, and \eqref{eq:Sk_bulk_vec}, we obtain

Erk2(21N)fkHg2Ik(g)+gHQkspgSksp(g)+1N2vec(Hk)T(KPbulk)vec(Gk)Sk(bulk)(g).\boxed{ \mathbb E|r_k|^2 \approx \underbrace{\left(2-\frac1N\right)|\mathbf f_k^H\mathbf g|^2}_{\mathcal I_k(\mathbf g)} + \underbrace{\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g}_{\mathcal S_k^{\mathrm{sp}}(\mathbf g)} + \underbrace{\frac{1}{N^2}\operatorname{vec}(\mathbf H_k)^{T}\big(\mathbf K\odot \mathbf P^{\mathrm{bulk}}\big)\operatorname{vec}(\mathbf G_k)}_{\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)}. }

\end{document}

विचार

Crafting step-by-step LaTeX explanation in Chinese

For LaTeX, I'll provide a detailed breakdown:

  • First, define the s vector with assumptions: iid, E(s)=0, E(s²)=0, E(|s|²)=1, and kurtosis μ4 = E(|s|⁴).
  • Next, build a Hermitian symmetric v in the frequency domain for real-time-domain IDFT, setting v[0]=v[N/2]=0, and constructing the rest with s and conjugates, adding zero padding.
  • Then, apply Time-domain tilde v = F^H v ∈ R^N, and use cos/sin for the formula.
  • For CE-OFDM, express x as \tilde A exp(j \tilde h \tilde v) or exp(j \tilde h f_n^H v) elementwise, and x[n] = \tilde A exp(j \tilde h \tilde v[n]) with F row.

Breaking down constant envelope amplitude and expansions

First up, we define A~\tilde{A} for the constant envelope amplitude and go into detail with rk=xHJkxr_k = x^H J_k x, a cyclic shift correlation. It's closely tied to the Fourier transform.

Then, expanding fnHx2|f_n^H x|^2 gives us another sum involving exponents and Fourier terms. When you expand rk2|r_k|^2, a double sum emerges, and we connect it with an expectation factor involving a Hermitian-symmetric random vector vv.

Exploring approximation and sums of independent terms

First, we sum the independent terms zrz_r under the Lindeberg condition and apply a cumulant expansion (Edgeworth) up to the 4th order for the characteristic function.

We then constrain the summation over kk by the DFT orthogonality, leading to modular constraints. By parameterizing with uu and vv, we define Ur(u,v)U_r(u,v) and a[r]a[r]. Next, we compute S2S2 and S4S4, then insert them into the characteristic function to finalize the expressions.

Calculating and clarifying S4 with kernel matrices and weights

To compute S2S2 and S4S4, we use a kernel matrix approach where S2 = bTK2bb^T K2 b and S4=wTK4wS4 = w^T K4 w with w=bbw = b \otimes b. For Erk2E|r_k|^2, I approximate the original quadruple sum with a second-order truncation and identify pairings for the iceberg term IkI_k. The vector gng_n needs clarification – it appears as deterministic and dependent on fnxf_n x.

Clarifying equations and preparing deliverables

In the equation, the term gngmg_n g_m could be 1 if there's no weighting. We’ll interpret gng_n as fnx2|f_n x|^2, or some frequency weighting. For the sea-level term SkS_k, we eliminate sums with DFT identity, resulting in a pure representation involving exponential terms.

Next, I’ll break down the decomposition of QkQ_k into low-rank terms. I'll provide definitions, derived equations, and formulas, using 0-based indexing for the DFT. For the final deliverables, I need clear, sequential Chinese narrative commentary and a LaTeX code block with restructured derivations.

Organizing final response in LaTeX sections and explanations

I’ll structure the response into clear sections: goal and core quantity, frequency to time domain, the CE nonlinearity, rkr_k and energy spectral density, expansions, expectation, CLT approximation, definitions for S2S2 and S4S4, and

Thought for 45s

उत्तर

下面我按“读论文/写推导”的方式,把你给的整段推导用中文讲清楚(讲逻辑、讲每一步为什么这么写),然后再给一份更像 IEEE TSP(Transactions on Signal Processing)论文正文风格的 LaTeX 推导稿(把符号、引理/命题、关键等式整理成一条主线,避免你原稿里那种“到处散落的等式”)。


一、中文讲解:这段推导到底在干什么、怎么一步步走到最终式

1) 目标量:你真正要算的是一个“四阶量”

在 CE-OFDM(Constant-Envelope OFDM)里,时域信号是把 OFDM 的实值时域样本 v~[n]\tilde v[n] 送进指数:

x[n]=A~ejh~v~[n].x[n]=\tilde A\,e^{j\tilde h\,\tilde v[n]}.

由于指数是非线性,频域/相关量的高阶统计会变得很复杂。你这里关心的核心对象是频域自相关(或等价的谱域二次型)样本

rk=xHJkx,r_k=\mathbf x^{H}\mathbf J_k\mathbf x,

并进一步需要它的二阶功率(四阶矩):

E{rk2}.\mathbb E\{|r_k|^2\}.

这就是“第四矩分析”的落点:rk2|r_k|^2 本质上把 x\mathbf x 乘了四次。


2) 为什么要做 Hermitian 对称:为了让 v~\tilde{\mathbf v} 是实值

你从独立同分布的数据信号 sCNd\mathbf s\in\mathbb C^{N_d} 出发,并假设星座满足:

  • Es2=1\mathbb E|s|^2=1(单位功率)
  • E[s]=0\mathbb E[s]=0E[s2]=0\mathbb E[s^2]=0(零均值、零伪方差——“圆对称/ proper”的关键条件)
  • μ4=Es4\mu_4=\mathbb E|s|^4(归一化峰度/四阶矩)

然后把 s\mathbf s 填到一个长度为 NN 的频域向量 v\mathbf v 并做 Hermitian 对称,这样 IDFT 后

v~=FNHvRN\tilde{\mathbf v}= \mathbf F_N^{H}\mathbf v \in \mathbb R^{N}

就自然是实值,这对 CE-OFDM 的相位调制模型 ejh~v~[n]e^{j\tilde h \tilde v[n]} 是标准设定。


3) 把 rkr_k 写成“某个序列的 DFT 系数”

你给的推导里使用了一个常见恒等式:循环移位矩阵 Jk\mathbf J_k 可以被 DFT 对角化,于是

rk=xHJkx=n=1NfnHx2ej2πNk(n1).r_k = \mathbf x^H\mathbf J_k\mathbf x = \sum_{n=1}^{N}\big|\mathbf f_n^{H}\mathbf x\big|^2 e^{j\frac{2\pi}{N}k(n-1)}.

这里 fnH\mathbf f_n^{H} 是单位ary DFT 矩阵的第 nn 行。直观上就是:

y[n]fnHx2,y[n]\triangleq |\mathbf f_n^H\mathbf x|^2,

rkr_k 就是序列 y[n]y[n] 的 DFT(差个归一化/符号约定),所以 rk2|r_k|^2 就是 yy 的频域能量耦合。


4) 展开 rk2|r_k|^2:必然出现四重索引 (p,q,p,q)(p,q,p',q')

先写

rk2=nmy[n]y[m]ejωk(nm),ω=2πN.|r_k|^2=\sum_{n}\sum_{m} y[n]y[m]e^{j\omega k(n-m)},\quad \omega=\tfrac{2\pi}{N}.

关键是 y[n]=fnHx2y[n]=|\mathbf f_n^H\mathbf x|^2 还要展开成 x\mathbf x 的二次型:

fnHx2=1Np=1Nq=1Nx[p]x[q]  ejω(n1)(pq).|\mathbf f_n^H\mathbf x|^2 =\frac{1}{N}\sum_{p=1}^{N}\sum_{q=1}^{N} x[p]x^*[q]\;e^{-j\omega(n-1)(p-q)}.

乘上另一个 y[m]y[m] 就自然出现四个索引 (p,q,p,q)(p,q,p',q')。把指数项合并后,你得到一个非常典型的结构:

exp ⁣(jaHv),aH=h~(fpHfqHfpH+fqH).\exp\!\big(j\,\mathbf a^{H}\mathbf v\big), \qquad \mathbf a^{H}=\tilde h(\mathbf f_p^H-\mathbf f_q^H-\mathbf f_{p'}^H+\mathbf f_{q'}^H).

于是

E{rk2}\mathbb E\{|r_k|^2\}

就变成:一个大求和 ×\times 一个关键因子 E[ejaHv]\mathbb E[e^{j\mathbf a^H\mathbf v}]
所以真正的“难点”压缩成一句话:

要近似计算 aHv\mathbf a^H\mathbf v 的特征函数(characteristic function)。


5) 用 Hermitian 结构把 aHv\mathbf a^H\mathbf v 化成独立和:进入 Lindeberg-CLT

因为 v\mathbf v 的非零部分由 {sr}\{s_r\} 和其共轭成对构造,你可以利用 DFT 列的共轭对称性质得到

a[Nr]=a[r].a[N-r]=a^*[r].

于是

aHv=r=1Nd(a[r]sr+a[r]sr)=r=1Ndzr,zr=2{a[r]sr}.\mathbf a^H\mathbf v =\sum_{r=1}^{N_d}\big(a[r]^*s_r+a[r]s_r^*\big) =\sum_{r=1}^{N_d} z_r, \quad z_r=2\Re\{a[r]^*s_r\}.

这里 {zr}\{z_r\} 相互独立(因为 {sr}\{s_r\} 独立),但一般不同分布(系数 a[r]a[r] 不同)。这就是 Lindeberg CLT 的典型使用场景:大量独立项的加权和。

在满足 Lindeberg 条件

maxra[r]2ra[r]20,μ4<\max_r\frac{|a[r]|^2}{\sum_{r}|a[r]|^2}\to 0,\qquad \mu_4<\infty

时,用“累积量/Edgeworth”的四阶截断,你得到你文中最关键的近似:

E ⁣[ejaHv]exp ⁣(ra[r]2+μ424ra[r]4).\mathbb E\!\left[e^{j\mathbf a^H\mathbf v}\right] \approx \exp\!\left( -\sum_r |a[r]|^2 +\frac{\mu_4-2}{4}\sum_r |a[r]|^4 \right).

解释一下这个结构为什么合理:

  • a[r]2-\sum|a[r]|^2 对应“近似高斯”的二阶项(方差项)
  • (μ42)a[r]4(\mu_4-2)\sum|a[r]|^4 是对“非高斯性”的四阶修正;对 proper 复高斯有 μ4=2\mu_4=2,此时修正项消失

6) 约束与重参数化:把四重索引收敛到 (u,v)(u,v)

你后面提到“对 kk 求和(DFT 正交性)”会导出模约束:

pqp+q0(modN)    p+qq+p(modN).p-q-p'+q'\equiv 0\pmod N \;\Longleftrightarrow\; p+q'\equiv q+p'\pmod N.

然后用一个非常方便的参数化

(p,q,p,q)(u+v,  v,  u,  0)(modN),(p,q,p',q')\equiv(u+v,\;v,\;u,\;0)\pmod N,

把结构写得更像“两个差分”的乘积。定义

Ur(u,v)=ejωr(u+v)ejωrvejωru+1=(1ejωru)(1ejωrv),U_r(u,v)=e^{j\omega r(u+v)}-e^{j\omega rv}-e^{j\omega ru}+1 =(1-e^{j\omega ru})(1-e^{j\omega rv}),

并由 DFT 归一化得到

a[r]=h~NUr(u,v).a[r]=\frac{\tilde h}{N}U_r(u,v).

于是二阶/四阶和变成两个确定性核:

S2(u,v)=r=1NdUr(u,v)2,S4(u,v)=r=1NdUr(u,v)4,S_2(u,v)=\sum_{r=1}^{N_d}|U_r(u,v)|^2,\qquad S_4(u,v)=\sum_{r=1}^{N_d}|U_r(u,v)|^4,

并带回特征函数近似,得到一个非常干净的“核指数权重”。


7) 计算 S2,S4S_2,S_4:引入 Ψ(d)\Psi(d) 并给闭式

你定义

Ψ(d)=r=1Ndcos(ωrd),\Psi(d)=\sum_{r=1}^{N_d}\cos(\omega r d),

它有标准闭式(Dirichlet kernel 的截断形式)。然后 S2S_2 可以写成若干个 Ψ()\Psi(\cdot) 的线性组合;S4S_4 则是更复杂的二重求和组合(你用差分向量 g\mathbf g 和系数 c0,c1,c2c_0,c_1,c_2 给了紧凑表达)。

这一段的本质是:

由于 Ur2|U_r|^2 展开后会出现 cos(ωrΔ)\cos(\omega r\Delta) 的和,因此所有东西都能用 Ψ(Δ)\Psi(\Delta) 表达。


8) 用“冰山/海平面”分解 Erk2\mathbb E|r_k|^2:把最强的配对项先抠出来

a=0\mathbf a=\mathbf 0 时,指数因子就是 1,这类项对总和贡献最大且结构最简单。
a=0\mathbf a=\mathbf 0 等价于两种配对:

  • (A) p=qp=qp=qp'=q'
  • (B) p=pp=p'q=qq=q'

它们一共贡献 2N2N2N^2-N 个四元组,于是得到“冰山项”

Ik(g)=(21N)fkHg2.\mathcal I_k(\mathbf g)=\left(2-\frac{1}{N}\right)|\mathbf f_k^H\mathbf g|^2.

其余非配对项定义为“海平面项” Sk(g)\mathcal S_k(\mathbf g)
进一步你把 (n,m)(n,m) 的求和用 DFT 直接消掉,得到纯四索引形式:

Sk(g)=1N2 ⁣ ⁣(p,q,p,q)P ⁣ ⁣(fk+pqHg)(fkp+qHg)exp ⁣(αS2(p,q,p,q)),\mathcal S_k(\mathbf g) =\frac{1}{N^2}\!\!\sum_{(p,q,p',q')\notin\mathcal P}\!\! (\mathbf f_{\langle k+p-q\rangle}^H\mathbf g)\, (\mathbf f_{\langle -k-p'+q'\rangle}^H\mathbf g)\, \exp\!\big(-\alpha\,S_2(p,q,p',q')\big),

其中 α=h~2/N2\alpha=\tilde h^2/N^2\langle\cdot\rangle 表示模 NN 归约。

这一步非常“TSP”:先抽出可闭式的主导项,再把剩余项写成“结构化矩阵核加权”的形式,为后续数值计算/复杂度分析铺路。


9) 再把海平面项分成 structured layers + bulk:为了低秩/对角/向量化计算

你把非配对集合进一步分成若干“超平面层”(比如 p=qp'=q'p=qp=qq=qq=q'p=pp=p')和剩余 bulk。
在这些超平面上,S2S_2 只依赖一个差分 dd,于是你定义一维权重

κ[d]=exp ⁣(αbTK2(d,0,0,0)b)=exp ⁣(2α(NdΨ(d))),\kappa[d]=\exp\!\left(-\alpha\,\mathbf b^T\mathbf K_2(d,0,0,0)\mathbf b\right) =\exp\!\left(-2\alpha(N_d-\Psi(d))\right),

然后用一次 FFT 得到 s=Fκ~\mathbf s=\mathbf F\tilde{\boldsymbol\kappa},把对应层写成:

  • 两个秩 2\le2 的(对称化后)低秩二次型矩阵
  • 一个对角矩阵层(可用 {s}\Re\{\mathbf s\} 直接构造)
  • bulk 用 vec()T(KPbulk)vec()\operatorname{vec}(\cdot)^T(\mathbf K\odot\mathbf P^{\rm bulk})\operatorname{vec}(\cdot) 的形式写出,方便向量化或近似

最后得到总分解:

Erk2Ik(g)+Sksp(g)+Skbulk(g).\mathbb E|r_k|^2\approx \mathcal I_k(\mathbf g)+\mathcal S_k^{\rm sp}(\mathbf g)+\mathcal S_k^{\rm bulk}(\mathbf g).

这就是你整份推导的“骨架”。


二、IEEE TSP 风格的 LaTeX 推导整理稿(主线清晰、符号统一)

下面这段 LaTeX 我按 TSP 常见写法整理:先定义模型与符号,给关键引理(特征函数近似),再给命题(冰山项闭式、海平面项消元),最后给 structured/bulk 分解。你可以直接替换进 IEEEtran 论文里(只需你自己决定用 F\mathbf F 的指数符号正负号与索引从 0 还是 1,一般 TSP 更喜欢 0..N-1;我这里保持你原式的 1..N + (n1)(n-1) 约定,减少改动成本)。

latex
% ============================================================ % IEEE-TSP style derivation (cleaned and linearized) % ============================================================ \subsection{Signal Model and Constellation Assumptions} Let $\mathbf s=[s_1,\ldots,s_{N_d}]^T$ be i.i.d. constellation symbols satisfying \begin{align} \mathbb E\{s_r\}=0,\quad \mathbb E\{s_r^2\}=0,\quad \mathbb E\{|s_r|^2\}=1, \end{align} and define the normalized fourth moment (kurtosis) \begin{align} \mu_4 \triangleq \mathbb E\{|s_r|^4\}. \end{align} To obtain a real-valued OFDM time-domain symbol, map $\mathbf s$ to the Hermitian-symmetric frequency vector $\mathbf v\in\mathbb C^{N}$ such that $\tilde{\mathbf v}=\mathbf F^H\mathbf v\in\mathbb R^N$, where $\mathbf F\in\mathbb C^{N\times N}$ is the unitary DFT matrix with entries \begin{align} [\mathbf F]_{n,d}=\frac{1}{\sqrt N}e^{-j\omega(n-1)d},\qquad \omega\triangleq \frac{2\pi}{N}. \end{align} \subsection{CE-OFDM and the Autocorrelation Sample} Define the constant-envelope OFDM waveform \begin{align} \mathbf x=\tilde A \exp\!\big(j\tilde h\,\mathbf F^H\mathbf v\big)\in\mathbb C^{N}, \end{align} where the exponential is applied elementwise. Let $\mathbf J_k$ denote the $k$-sample cyclic shift matrix. The $k$th autocorrelation sample is \begin{align} r_k \triangleq \mathbf x^H\mathbf J_k\mathbf x. \end{align} Using diagonalization of circulant shifts by the DFT, one can write \begin{align} r_k = \sum_{n=1}^{N} \big|\mathbf f_n^H\mathbf x\big|^2 e^{j\omega k(n-1)}, \label{eq:rk_dft_of_power} \end{align} where $\mathbf f_n^H$ is the $n$th row of $\mathbf F$. For compactness, define the (possibly deterministic) sequence \begin{align} g_n \triangleq \big|\mathbf f_n^H\mathbf x\big|^2,\qquad \mathbf g=[g_1,\ldots,g_N]^T. \end{align} Then \begin{align} |r_k|^2=\sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m e^{j\omega k(n-m)}. \label{eq:rk2_nm} \end{align} \subsection{Four-Index Expansion} Using $\mathbf f_n[d]=\frac{1}{\sqrt N}e^{-j\omega(n-1)(d-1)}$ (indexing consistent with \eqref{eq:rk_dft_of_power}), \begin{align} g_n &=\big|\mathbf f_n^H\mathbf x\big|^2 =\frac{1}{N}\sum_{p=1}^{N}\sum_{q=1}^{N} x[p]x^*[q]\;e^{-j\omega(n-1)(p-q)}. \label{eq:gn_pq} \end{align} Substituting \eqref{eq:gn_pq} into \eqref{eq:rk2_nm} yields the standard four-index form \begin{align} |r_k|^2 = \frac{1}{N^2} \sum_{n,m}\! e^{j\omega k(n-m)} \sum_{p,q,p',q'}\! x[p]x^*[q]x^*[p']x[q']\, e^{-j\omega(n-1)(p-q)}e^{+j\omega(m-1)(p'-q')}. \label{eq:rk2_4idx_raw} \end{align} Under the CE model $x[t]=\tilde A e^{j\tilde h\,\mathbf f_t^H\mathbf v}$, the fourth-order factor becomes \begin{align} x[p]x^*[q]x^*[p']x[q'] =\tilde A^4\exp\!\big(j\mathbf a^H\mathbf v\big), \qquad \mathbf a^H\triangleq \tilde h(\mathbf f_p^H-\mathbf f_q^H-\mathbf f_{p'}^H+\mathbf f_{q'}^H). \label{eq:a_def} \end{align} Hence \begin{align} \mathbb E\{|r_k|^2\} = \frac{\tilde A^4}{N^2} \sum_{n,m}\! e^{j\omega k(n-m)} \sum_{p,q,p',q'}\! e^{-j\omega(n-1)(p-q)}e^{+j\omega(m-1)(p'-q')} \;\mathbb E\!\left[e^{j\mathbf a^H\mathbf v}\right]. \label{eq:Erk2_before_charfun} \end{align} \subsection{Characteristic-Function Approximation via Lindeberg CLT} Due to Hermitian symmetry of $\mathbf v$ and the conjugate pairing of DFT columns, one has \begin{align} a[N-r]=a^*[r]\quad \Rightarrow\quad \mathbf a^H\mathbf v=\sum_{r=1}^{N_d} z_r, \qquad z_r\triangleq a[r]^* s_r + a[r] s_r^* = 2\Re\{a[r]^*s_r\}. \label{eq:aHv_sum} \end{align} The random variables $\{z_r\}$ are independent (not necessarily i.i.d.). \paragraph*{Lemma 1 (Fourth-order cumulant truncation).} Assume the Lindeberg condition \begin{align} \max_r\frac{|a[r]|^2}{\sum_{r=1}^{N_d}|a[r]|^2}\to 0, \qquad \mu_4<\infty. \end{align} Then the characteristic function admits the approximation \begin{align} \mathbb E\!\left[e^{j\mathbf a^H\mathbf v}\right] \approx \exp\!\left( -\sum_{r=1}^{N_d}|a[r]|^2 +\frac{\mu_4-2}{4}\sum_{r=1}^{N_d}|a[r]|^4 \right). \label{eq:charfun_approx} \end{align} \subsection{Two-Difference Parameterization and the Deterministic Sums} After exploiting DFT orthogonality, the sign pattern in \eqref{eq:a_def} induces the modular constraint \begin{align} p-q-p'+q'\equiv 0\ (\mathrm{mod}\ N) \quad \Leftrightarrow\quad p+q'\equiv q+p'\ (\mathrm{mod}\ N). \end{align} A convenient re-parameterization is $(p,q,p',q')\equiv(u+v,v,u,0)\ (\mathrm{mod}\ N)$. Define \begin{align} U_r(u,v) &\triangleq e^{j\omega r(u+v)}-e^{j\omega rv}-e^{j\omega ru}+1 =(1-e^{j\omega ru})(1-e^{j\omega rv}), \end{align} which yields (under the unitary DFT convention) \begin{align} a[r]=\frac{\tilde h}{N}U_r(u,v). \end{align} Introduce the deterministic sums \begin{align} S_2(u,v)\triangleq \sum_{r=1}^{N_d}|U_r(u,v)|^2,\qquad S_4(u,v)\triangleq \sum_{r=1}^{N_d}|U_r(u,v)|^4. \end{align} Then \begin{align} \sum_{r}|a[r]|^2=\frac{\tilde h^2}{N^2}S_2(u,v),\qquad \sum_{r}|a[r]|^4=\frac{\tilde h^4}{N^4}S_4(u,v), \end{align} and \eqref{eq:charfun_approx} becomes \begin{align} \mathbb E\!\left[e^{j\mathbf a^H\mathbf v}\right] \approx \exp\!\left( -\frac{\tilde h^2}{N^2}S_2(u,v) +\frac{\mu_4-2}{4}\frac{\tilde h^4}{N^4}S_4(u,v) \right). \label{eq:charfun_S2S4} \end{align} \subsection{Closed Forms via a Cosine Kernel} Define the cosine-sum kernel \begin{align} \Psi(d)\triangleq \sum_{r=1}^{N_d}\cos(\omega r d),\qquad \Psi(d)=\Psi(-d). \end{align} Its standard closed form is \begin{align} \Psi(d)= \begin{cases} N_d, & d\equiv 0\ (\mathrm{mod}\ N),\\[2mm] \dfrac{\sin\!\left(\frac{N_d\omega d}{2}\right)\cos\!\left(\frac{(N_d+1)\omega d}{2}\right)} {\sin\!\left(\frac{\omega d}{2}\right)}, & \text{otherwise}. \end{cases} \end{align} \paragraph*{Kernel-matrix representation of $S_2$.} Let \begin{align} \mathbf b\triangleq [1,-1,-1,1]^T,\qquad \mathbf n(p,q,p',q')\triangleq [p,q,p',q']^T\ (\mathrm{mod}\ N), \end{align} and define the $4\times4$ kernel matrix \begin{align} [\mathbf K_2(p,q,p',q')]_{ij}\triangleq \Psi(n_i-n_j). \end{align} Then \begin{align} S_2(p,q,p',q')=\mathbf b^T\mathbf K_2(p,q,p',q')\mathbf b. \label{eq:S2_K2} \end{align} (An analogous $16\times16$ kernel yields $S_4$ with $\mathbf w=\mathbf b\otimes \mathbf b$.) \subsection{Second-Order Truncation for $\mathbb E|r_k|^2$} Discarding the $S_4$ correction in \eqref{eq:charfun_S2S4} gives \begin{align} \mathbb E\!\left[e^{j\mathbf a^H\mathbf v}\right]\approx \exp\!\left(-\alpha S_2(p,q,p',q')\right), \qquad \alpha\triangleq \frac{\tilde h^2}{N^2}. \end{align} Substituting into \eqref{eq:Erk2_before_charfun} yields \begin{align} \mathbb E\{|r_k|^2\} \approx \frac{\tilde A^4}{N^3} \sum_{n=1}^{N}\sum_{m=1}^{N} g_n g_m e^{j\omega k(n-m)} \sum_{p,q,p',q'} e^{-j\omega\left((n-1)(q-p)+(m-1)(p'-q')\right)} \exp\!\left(-\alpha S_2(p,q,p',q')\right). \label{eq:Erk2_start} \end{align} \subsection{Pairing Extraction (Iceberg Term)} Let the pairing set be \begin{align} \mathcal P \triangleq \{(p,q,p',q'):\ (p=q\land p'=q')\ \text{or}\ (p=p'\land q=q')\}. \end{align} On $\mathcal P$ one has $\mathbf a=\mathbf 0$ and $S_2=0$, hence $\exp(-\alpha S_2)=1$. The number of pairing quadruples is $2N^2-N$. Extracting them from \eqref{eq:Erk2_start} gives \begin{align} \mathcal I_k(\mathbf g) = \frac{\tilde A^4}{N^3}(2N^2-N)\sum_{n,m} g_n g_m e^{j\omega k(n-m)} =\tilde A^4\left(2-\frac{1}{N}\right)\left|\mathbf f_k^H\mathbf g\right|^2. \label{eq:Ik_closed} \end{align} Define $\mathbb E|r_k|^2\approx \mathcal I_k(\mathbf g)+\mathcal S_k(\mathbf g)$, where $\mathcal S_k$ is \eqref{eq:Erk2_start} restricted to $(p,q,p',q')\notin\mathcal P$. \subsection{Eliminating the $(n,m)$ Sums (Pure 4-Index Sea-Level Form)} Combine the exponentials in \eqref{eq:Erk2_start} to separate $n$ and $m$: \begin{align} e^{j\omega k(n-m)}e^{-j\omega(n-1)(q-p)} &= e^{j\omega(n-1)(k+p-q)},\\ e^{-j\omega k(n-m)}e^{-j\omega(m-1)(p'-q')} &= e^{j\omega(m-1)(-k-p'+q')}. \end{align} Therefore, \begin{align} \sum_{n=1}^{N} g_n e^{j\omega(n-1)(k+p-q)} = \sqrt N\,\mathbf f_{\langle k+p-q\rangle}^H\mathbf g, \quad \sum_{m=1}^{N} g_m e^{j\omega(m-1)(-k-p'+q')} = \sqrt N\,\mathbf f_{\langle -k-p'+q'\rangle}^H\mathbf g, \end{align} where $\langle\cdot\rangle$ denotes modulo-$N$ reduction into $\{0,\ldots,N-1\}$. Substituting back yields \begin{align} \mathcal S_k(\mathbf g) = \frac{\tilde A^4}{N^2} \!\!\sum_{(p,q,p',q')\notin\mathcal P}\!\! \big(\mathbf f_{\langle k+p-q\rangle}^H\mathbf g\big)\, \big(\mathbf f_{\langle -k-p'+q'\rangle}^H\mathbf g\big)\, \exp\!\left(-\alpha\,S_2(p,q,p',q')\right). \label{eq:Sk_4idx} \end{align} \subsection{Structured--Bulk Decomposition (Optional for Fast Computation)} Define the one-dimensional weight (generated by the same $K_2$ kernel) \begin{align} \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{align} and let $\tilde\kappa[0]=0$, $\tilde\kappa[d]=\kappa[d]$ for $d=1,\ldots,N-1$. Compute $\mathbf s=\mathbf F\tilde{\boldsymbol\kappa}$, $\tilde{\boldsymbol\kappa}=[\tilde\kappa[0],\ldots,\tilde\kappa[N-1]]^T$. Then the contributions from the single-pair hyperplanes can be written as a quadratic form \begin{align} \mathcal S_k^{\mathrm{sp}}(\mathbf g)=\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g, \end{align} with $\mathbf Q_k^{\mathrm{sp}}$ consisting of a diagonal layer and two Hermitian rank-$\le2$ layers (all expressible using $\mathbf s$ and Hadamard products; omitted here for brevity, see the construction in the original draft). The remaining bulk can be expressed via vectorization: let $\boldsymbol\gamma=\mathbf F^H\mathbf g$ and define \begin{align} [\mathbf H_k]_{p,q}\triangleq \gamma_{\langle k+p-q\rangle},\qquad [\mathbf G_k]_{p',q'}\triangleq \gamma_{\langle -k+(q'-p')\rangle}, \end{align} and \begin{align} [\mathbf K]_{(p,q),(p',q')}\triangleq \exp\!\left(-\alpha\,S_2(p,q,p',q')\right), \end{align} together with a bulk mask $\mathbf P^{\mathrm{bulk}}$ that removes $\mathcal P$ and the structured hyperplanes. Then \begin{align} \mathcal S_k^{\mathrm{bulk}}(\mathbf g) = \frac{\tilde A^4}{N^2}\, \mathrm{vec}(\mathbf H_k)^T\big(\mathbf K\odot \mathbf P^{\mathrm{bulk}}\big)\mathrm{vec}(\mathbf G_k). \end{align} \subsection{Final Decomposition} Combining \eqref{eq:Ik_closed} and \eqref{eq:Sk_4idx}, one obtains the TSP-style decomposition \begin{align} \boxed{ \mathbb E\{|r_k|^2\} \approx \underbrace{\tilde A^4\left(2-\frac{1}{N}\right)\left|\mathbf f_k^H\mathbf g\right|^2}_{\mathcal I_k(\mathbf g)} + \underbrace{\mathcal S_k(\mathbf g)}_{\text{four-index sea level, \eqref{eq:Sk_4idx}}}. } \end{align}

三、我帮你“读完后”的几个关键提醒(避免推导里常见踩坑)

  1. fn\mathbf f_n 是“行”还是“列”一定要统一:你原文有时写 fnx\mathbf f_n\mathbf x,有时写 fnHx\mathbf f_n^H\mathbf x。IEEE 风格里一般固定:
  • fnH\mathbf f_n^H = 第 nn
  • fk\mathbf f_k = 第 kk
    否则很容易在共轭、模 NN 索引上出错。
  1. 你把 gngmg_n g_m 拉到期望号外,这隐含了一个“gng_n 当确定量/与 v\mathbf v 独立”的处理。
  • 如果你实际模型里 gn=fnHx2g_n=|\mathbf f_n^H\mathbf x|^2x\mathbf xv\mathbf v 生成,那么严格来说 gng_n 也是随机的。
  • 你现在的推导相当于:随机性的主导来自 exp(jaHv)\exp(j\mathbf a^H\mathbf v),而 gng_n 作为权重被“冻结/近似”。这在 TSP 论文里可以作为近似假设写清楚。
  1. μ42\mu_4-2 的意义非常重要:
  • proper 复高斯时 μ4=2\mu_4=2,四阶修正消失,你的特征函数近似退化成纯二阶(高斯)项;
  • QAM/PSK 等离散星座一般 μ42\mu_4\neq 2,修正项会影响旁瓣/残差项的衰减。

如果你愿意,我还可以把你这份推导进一步“论文化”到可直接投稿的程度:

  • Sksp\mathcal S_k^{\mathrm{sp}} 的三块矩阵(对角 + 两个低秩层)用统一的 Hadamard 表达一次性写成一个命题;
  • 给出复杂度(FFT 一次 + 低秩更新 + bulk 的近似/截断策略);
  • 顺便把符号改成 IEEE 常见的 0-based 索引,减少 (n1)(n-1) 这种易错项。

प्रश्न

我们按照上面的推导策略,把\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)也推导简化到和下面这种结构差不多的样子:
\boxed{
\mathcal S_k^{\mathrm{sp}}(\mathbf g)
=\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g
=\big|(\mathbf Q_k^{\mathrm{sp}})^{1/2}\mathbf g\big|_2^2

\left|
\left(
\Big(\tfrac{1}{N^{3/2}}\mathbf I+\tfrac{\sqrt N}{2N}\mathbf A_k\Big)\odot \tilde{\mathbf S}
\right)^{1/2}\mathbf g
\right|_2^2 .
}

विचार

Simplifying bulk term into quadratic form

The user wants me to simplify a bulk term formulated like vec(H_k)^T (K ⊙ P_bulk) vec(G_k) into something similar to a quadratic form. I'll work through it by relating the vectors to γ (built using circulant matrices), and symmetrizing the bulk term. I'll express it as a bilinear form and look into Kronecker products for simplifications.

Simplifying circulant matrix terms for bulk expression

I’m working with circulant matrices c_k and G_k linked to γ. Both are linear maps, and I need to vectorize them. I’ll use permutation matrices and look for a matrix T_k that relates vec(H_k) and vec(G_k) with γ. Bulk expression will then involve T_k and a structured approach, possibly aligning with Hadamard forms, though it might get complex.

Expressing bulk term in quadratic form

I can express the bulk term as a quadratic form in γ or g, using Kronecker products and permutation matrices. By symmetrizing Q_bulk, we can enforce PSD and represent it as a norm. I’ll break down the bulk term starting with H_k and G_k, using shift and difference matrices, leading to a final form with basis vectors.

Finalizing vectorization and matrix structure

We define matrix B and find that vec(H_k) = B Π_k γ, where Π_k is just a cyclic permutation shifting γ by k. The coefficient for each b_d is γ_{⟨k+d⟩}, and this is consistent with the cyclic shift. Similarly, for G_k, we can express its form using a shift defined by d = p'-q'. This is related to the transpose of a circulant matrix.

Final bulk term simplification and conjugacy check

To express vec(G_k), I use the same matrix B and a transpose permutation P_T, with the overall bulk term involving vec(H_k) and vec(G_k). I need to check if a conjugate is required for the scalar product — particularly since γ is complex. The final expression seems bilinear, but I’ll confirm the conjugate behavior for consistency.

Evaluating bulk expression and PSD behavior

To derive the expectation of |r_k|^2, I need to ensure the result is real, so I should take the real part after computing. G_k might equal the conjugate of H_k, but that’s unsure. The bulk expression involves a Hermitian Q. If Q is PSD, I can norm, but I need to carefully address the mask's impact on PSD.

Deriving the bulk structure with approximations

To approximate the bulk, I need to retain the dependence on difference pairs and simplify K into separable components, potentially using second-order approximations. For the final expression, I aim to use Kronecker products and vector-matrix-vector representations to reach a quadratic form in γ, while addressing the PSD through eigenvalue projection.

Clarifying the approximations for bulk structure

To refine the bulk expression, I'll define QbulkQ_{\text{bulk}} as the symmetrized matrix, ensuring transparency. I’ll check both cases: if QbulkQ_{\text{bulk}} is positive semidefinite, we can use squared norms, or otherwise, apply Jordan decomposition. Next, I might explore a Hadamard product or separable kernels in the frequency domain for approximation.

Deriving the bulk quadratic form

To simplify the expression for SbulkS_{\text{bulk}}, we'll express QbulkQ_{\text{bulk}} in a quadratic form involving FF and PbulkP_{\text{bulk}}, with matrix manipulation. Using the circulant basis, we express BB as a permutation matrix, aiding in understanding the PSD structure. We’ll utilize Kronecker and reversal permutations to finalize the derivation.

Simplifying bulk representation

To express SbulkS_{\text{bulk}}, we use a quadratic form with conjugates for precise complexity. The formula becomes a projection of γ\gamma into coefficients, and by shifting to the gg-space, we can express it as Q+1/2g2|| Q_+^{1/2} g ||^2 for simplicity, simplifying over possible negative components. The user prefers a Hadamard-like structure, so approximations may be needed.

Exploring bulk term simplification

The bulk term can be simplified using shift matrices and circulant structures, reducing the complexity from N2N^2 to N×NN \times N using a difference-domain approach. The matrix BH(KP)BB^H (K \otimes P) B simplifies the bulk kernel, avoiding the need for large matrices. We finalize the expression by defining difference coefficient vectors for HkH_k and GkG_k, ensuring the terms can be treated as bilinear forms for real input gg. This leads to an equivalent structure resembling that of quadratic forms.

Simplifying bulk term in frequency domain

The bulk term can be expressed as a quadratic form using shift-invariant kernels, and we can compute it using a matrix QkbulkQ_k^{bulk}. The structure resembles a Hadamard product after simplifying using the DFT. Specifically, time-domain and frequency-domain relations help express QkbulkQ_k^{bulk} efficiently and lead to simplifications computationally. Key approximations based on the kernel’s shift invariance show that we reduce the complexity without constructing large matrices directly.

Explaining bulk term structure and modulation

The bulk term involves a complex modulation structure that leads to a Hadamard product with a rank-1 matrix and a kernel matrix. We define a modulation vector and explore how it interacts with the kernel in both time and frequency domains. This results in a compact form of the bulk term, with a clearer structure that resembles the user's requested format.

Thought for 2m 39s

उत्तर

可以的。关键思路是:bulk 项虽然看起来是 N2×N2N^2\times N^2 的 vec–matrix–vec 乘积,但 Hk,Gk\mathbf H_k,\mathbf G_k 本身落在“循环矩阵(circulant)子空间”里,所以可以把那个巨大核矩阵先投影到一个 N×NN\times N 的“差分域核”上;然后再把 kk 的依赖写成对一个 kk-无关核矩阵的相位调制(等价于 Hadamard 乘一个秩1矩阵)。这样就会非常像你给的 ()S~\big(\cdot\big)\odot\tilde{\mathbf S} 结构。

下面按你“上面的推导策略”给出一条完整推导链,并给出最终 boxed 形式。


1) 从 bulk 的 vec–matrix–vec 到标准二次型(把 Gk\mathbf G_k 变成 Hk\mathbf H_k^*

你已有的 bulk 定义是(我沿用你原来的记号)

Sk(bulk)(g)=1N2vec(Hk)T(KPbulk)vec(Gk)(1)\boxed{ \mathcal S_k^{(\mathrm{bulk})}(\mathbf g)= \frac{1}{N^2}\, \operatorname{vec}(\mathbf H_k)^{T}\, \big(\mathbf K\odot \mathbf P^{\mathrm{bulk}}\big)\, \operatorname{vec}(\mathbf G_k) } \tag{1}

其中 γ=FHg\boldsymbol\gamma=\mathbf F^H\mathbf g,并且

[Hk]p,q=γk+pq,[Gk]p,q=γk+(qp).[\mathbf H_k]_{p,q}=\gamma_{\langle k+p-q\rangle},\qquad [\mathbf G_k]_{p',q'}=\gamma_{\langle -k+(q'-p')\rangle}.

注意:g\mathbf g 在你这里来自功率/窗函数(例如 gn=fnHx2g_n=|\mathbf f_n^H\mathbf x|^2 或一般的实权重),因此 gRN\mathbf g\in\mathbb R^N,从而

γ=γ.\gamma_{\langle -\ell\rangle}=\gamma_{\langle \ell\rangle}^* .

于是对任意 p,qp',q',

[Gk]p,q=γk(pq)=γk+(pq)=[Hk]p,q,[\mathbf G_k]_{p',q'} =\gamma_{\langle -k-(p'-q')\rangle} =\gamma_{\langle k+(p'-q')\rangle}^* = [\mathbf H_k]_{p',q'}^*,

Gk=Hk,vec(Gk)=vec(Hk).(2)\boxed{\mathbf G_k=\mathbf H_k^*,\qquad \operatorname{vec}(\mathbf G_k)=\operatorname{vec}(\mathbf H_k)^*.} \tag{2}

另一方面,你的核 KPbulk\mathbf K\odot\mathbf P^{\mathrm{bulk}}exp(αS2())\exp(-\alpha S_2(\cdot))(全为实正)和 0–1 mask 组成,因此它是实矩阵;并且在自然索引对称下可视为(或取其 Hermitian part)

MbulkHerm{KPbulk}.\mathbf M_{\mathrm{bulk}}\triangleq \operatorname{Herm}\{\mathbf K\odot\mathbf P^{\mathrm{bulk}}\}.

把 (2) 代回 (1),并把 “T()^T(\cdot)^*” 改写成标准的 sesquilinear 形式,就得到

Sk(bulk)(g)=1N2vec(Hk)HMbulkvec(Hk)(3)\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) } \tag{3}

(严格等号需要 Mbulk\mathbf M_{\mathrm{bulk}} 取到合适对称;在 TSP 写法里通常直接写 Herm{}\operatorname{Herm}\{\cdot\} 以确保结果为实。)


2) 用“循环矩阵子空间投影”把 N2N^2 维降到 NN

定义循环移位基矩阵(与你前面的 Jd\mathbf J_d 一致)

[Jd]p,q1{pqd (mod N)},dZN.[\mathbf J_d]_{p,q}\triangleq \mathbb 1\{p-q\equiv d\ (\mathrm{mod}\ N)\},\qquad d\in\mathbb Z_N.

任何循环矩阵都可写成

circ(c)=d=0N1c[d]Jd.\operatorname{circ}(\mathbf c)=\sum_{d=0}^{N-1} c[d]\mathbf J_d.

Hk\mathbf H_k 写成循环矩阵形式:令

ck[d]γk+d,ck=[ck[0],,ck[N1]]T,c_k[d]\triangleq \gamma_{\langle k+d\rangle},\qquad \mathbf c_k=[c_k[0],\ldots,c_k[N-1]]^T,

Hk=circ(ck).(4)\boxed{\mathbf H_k=\operatorname{circ}(\mathbf c_k).} \tag{4}

接着定义“循环合成矩阵”

B[vec(J0),,vec(JN1)]RN2×N.\mathbf B \triangleq \big[\operatorname{vec}(\mathbf J_0),\ldots,\operatorname{vec}(\mathbf J_{N-1})\big]\in\mathbb R^{N^2\times N}.

则由线性展开立刻有

vec(circ(c))=Bc.(5)\boxed{\operatorname{vec}(\operatorname{circ}(\mathbf c))=\mathbf B\,\mathbf c.} \tag{5}

因此

vec(Hk)=Bck.\operatorname{vec}(\mathbf H_k)=\mathbf B\,\mathbf c_k.

把 (5) 代回 (3),得到一个NN的二次型:

Sk(bulk)(g)=1N2ckH(BHMbulkB)RbulkCN×Nck(6)\boxed{ \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 } \tag{6}

这里

RbulkBHMbulkB(7)\boxed{\mathbf R_{\mathrm{bulk}}\triangleq \mathbf B^{H}\mathbf M_{\mathrm{bulk}}\mathbf B} \tag{7}

可以理解为:把原来作用在 (p,q)(p,q) 平面上的巨大核 Mbulk\mathbf M_{\mathrm{bulk}} 投影到“差分索引 d=pqd=p-q”张成的循环子空间里。


3) 把 ck\mathbf c_k 写成 γ\boldsymbol\gamma(再写成 g\mathbf g),并把 kk 依赖做成 Hadamard 调制

先把 ck\mathbf c_k 写成对 γ\boldsymbol\gamma 的循环置换。定义置换矩阵 Πk\mathbf\Pi_k 使得

(Πkγ)[d]=γk+d.(\mathbf\Pi_k\boldsymbol\gamma)[d]=\gamma_{\langle k+d\rangle}.

ck=Πkγ,γ=FHg.(8)\boxed{\mathbf c_k=\mathbf\Pi_k\boldsymbol\gamma,\qquad \boldsymbol\gamma=\mathbf F^H\mathbf g.} \tag{8}

将 (8) 代入 (6):

Sk(bulk)(g)=1N2γHΠkHRbulkΠkγ=gH(1N2FΠkHRbulkΠkFH)Qkbulkg.\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.

于是得到你想要的“和 sp 一样的二次型结构”:

Sk(bulk)(g)=gHQkbulkg,Qkbulk1N2FΠkHRbulkΠkFH.(9)\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}. } \tag{9}

进一步把 kk 依赖写成 Hadamard 乘一个秩1相位矩阵(和你的 sp 形式更像)

循环置换在频域等价于对角调制:存在

Πk=FHDkF,Dkdiag(ejωk(n1))n=1N.\mathbf\Pi_k=\mathbf F^{H}\mathbf D_k\mathbf F,\qquad \mathbf D_k\triangleq \operatorname{diag}\big(e^{-j\omega k(n-1)}\big)_{n=1}^N.

把它代入 Qkbulk\mathbf Q_k^{\mathrm{bulk}},并令

KbulkFRbulkFH,\mathbf K_{\mathrm{bulk}}\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^{H},

Qkbulk=1N2DkKbulkDk.\mathbf Q_k^{\mathrm{bulk}} = \frac{1}{N^2}\mathbf D_{-k}\mathbf K_{\mathrm{bulk}}\mathbf D_{k}.

而对任意矩阵 X\mathbf X,有恒等式

DkXDk=(dkdkH)X,dk[ejωk(n1)]n=1N.\mathbf D_{-k}\mathbf X\mathbf D_k = (\mathbf d_k\mathbf d_k^{H})\odot \mathbf X,\qquad \mathbf d_k\triangleq [e^{j\omega k(n-1)}]_{n=1}^N.

因此

Qkbulk=1N2(dkdkH)Kbulk,Kbulk=FRbulkFH,Rbulk=BHMbulkB.(10)\boxed{ \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}, \qquad \mathbf R_{\mathrm{bulk}}=\mathbf B^{H}\mathbf M_{\mathrm{bulk}}\mathbf B. } \tag{10}

你会发现这和你的 sp 结构非常一致:

  • sp:(简单的 k-依赖矩阵)S~\big(\text{简单的 }k\text{-依赖矩阵}\big)\odot \tilde{\mathbf S}
  • bulk:(秩1的 k-依赖相位矩阵 dkdkH)Kbulk\big(\text{秩1的 }k\text{-依赖相位矩阵 }\mathbf d_k\mathbf d_k^H\big)\odot \mathbf K_{\mathrm{bulk}}

区别只是:sp 的 S~\tilde{\mathbf S} 可以用一维 κ[d]\kappa[d] + FFT 构出来,bulk 的 Kbulk\mathbf K_{\mathrm{bulk}} 是更“满”的核。


4) 最终写成你想要的 “()1/2g22\|(\cdot)^{1/2}\mathbf g\|_2^2” 形式

只要 Qkbulk0\mathbf Q_k^{\mathrm{bulk}}\succeq \mathbf 0(半正定),就可以完全平行于你给的 sp boxed 式写:

Sk(bulk)(g)=gHQkbulkg=(Qkbulk)1/2g22=(1N2(dkdkH)Kbulk)1/2g22.(11)\boxed{ \mathcal S_k^{(\mathrm{bulk})}(\mathbf g) =\mathbf g^{H}\mathbf Q_k^{\mathrm{bulk}}\mathbf g =\big\|(\mathbf Q_k^{\mathrm{bulk}})^{1/2}\mathbf g\big\|_2^2 = \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 . } \tag{11}

实务上:Mbulk=KPbulk\mathbf M_{\mathrm{bulk}}=\mathbf K\odot \mathbf P^{\mathrm{bulk}} 的 mask 可能破坏 PSD(Schur 乘积要 PSD 需要 mask 也 PSD)。如果不想在论文里卡死,可以用 TSP 常见写法:

Qk,+bulkΠ0(Herm{Qkbulk}),Sk(bulk)(g)(Qk,+bulk)1/2g22,\mathbf Q_{k,+}^{\mathrm{bulk}}\triangleq \Pi_{\succeq 0}\big(\operatorname{Herm}\{\mathbf Q_k^{\mathrm{bulk}}\}\big), \quad \mathcal S_k^{(\mathrm{bulk})}(\mathbf g)\approx \|(\mathbf Q_{k,+}^{\mathrm{bulk}})^{1/2}\mathbf g\|_2^2,

其中 Π0\Pi_{\succeq 0} 表示把负特征值截断为 0(PSD 投影)。这样“结构”完全保留,而且数值上稳定。


如果你希望我把 Rbulk=BHMbulkB\mathbf R_{\mathrm{bulk}}=\mathbf B^H\mathbf M_{\mathrm{bulk}}\mathbf B 进一步写成显式的差分求和公式(也就是把 N2×N2N^2\times N^2 的投影彻底展开成 N×NN\times N 的双和,明确每个 [d,d][d,d'] 元素如何由 exp(αS2)\exp(-\alpha S_2) 聚合得到),我也可以继续往下推,这样 bulk 的“核生成方式”会更像你 sp 里由 κ[d]\kappa[d] 生成 S~\tilde{\mathbf S} 的那种“可实现表达”。

प्रश्न

继续往下写,然后把这个推导过程整理成latex代码

विचार

Deriving R_bulk formula

I need to continue developing the formula for R_bulk as sums over p, q, etc. I'll craft it in Chinese and provide the necessary LaTeX code. The formula will involve a 4-index weight W(p,q,p',q'), considering shift invariance, and an aggregated kernel to handle differences and sums. I'll adjust for the bulk mask and its impact on hyperplanes, too. This will reduce from N^4 to N^2 for each (d,d') step.

Examining S2 with pairwise differences

S2 depends on differences between pairs like p-q, p-p', and others. When we substitute with d and d', it still depends on q-q' (let's call it t). We can reframe the variables for shift invariance. The bulk mask conditions involve inequalities, based on d, d', and t. Conditions for non-zero d, d', and t are important for defining the bulk mask.

Defining weight dependence on differences and shift invariance

So, now I’ve realized that if p≠p', then t ≠ d' - d; if q≠q', then t ≠ 0, and so on with other conditions on pairings/hyperplanes.

This suggests the weight depends only on t and the differences. Shift invariance allows us to simplify the formula. The total number of quadruples is N, so R_bulk can be written as a sum over t with a weight function w that involves the differences. This leads to a more compact representation.

Reducing R_bulk using kernel κ_bulk

We can write R_bulk as a sum over t, with a kernel function κ_bulk[d,d',t]. But we’re looking to simplify to 2D. The bulk mask restricts certain conditions, so ι=1 when d≠0, d'≠0, and t is not in {0, -d, d', d'-d}. Then we compute S2 using the difference terms for p, q, p', and q'.

Crafting explicit expressions for bulk model

We derive the formula for R_bulk and simplify using shift invariance. The S2 function is expressed in terms of Ψ, and ι_bulk filters certain conditions for the bulk mask, excluding hyperplanes. Then, we express S_bulk as a norm and finalize by representing it in LaTeX, including step-by-step derivations.

Thought for 34s

उत्तर

下面我继续把 bulk 项从你现在的 vec ⁣ ⁣matrix ⁣ ⁣vec\operatorname{vec}\!-\!\text{matrix}\!-\!\operatorname{vec} 形式,一步步压缩到和你给的 structured(sp) 项同型的

Sk(bulk)(g)=gHQkbulkg=(Qkbulk)1/2g22\mathcal S_k^{(\mathrm{bulk})}(\mathbf g) =\mathbf g^H\mathbf Q_k^{\mathrm{bulk}}\mathbf g =\big\|(\mathbf Q_k^{\mathrm{bulk}})^{1/2}\mathbf g\big\|_2^2

并且进一步把“生成 Qkbulk\mathbf Q_k^{\mathrm{bulk}} 的核”写成显式的差分域双重求和(只剩 (d,d,t)(d,d',t) 三个模差变量),这样就和你 sp 里用 κ[d]\kappa[d] 的思路一致:** 先把四索引聚合成差分核,再用 FFT/调制表达 kk 依赖**。


推导要点(先用中文概括)

  1. 由于 gRN\mathbf g\in\mathbb R^N(你这个 gng_n 是功率/窗权重,通常取实),令 γ=FHg\boldsymbol\gamma=\mathbf F^H\mathbf g,则 γ=γ\gamma_{\langle-\ell\rangle}=\gamma_{\langle\ell\rangle}^*。这会导出 Gk=Hk\mathbf G_k=\mathbf H_k^*,从而 bulk 变成标准二次型(Hermitian 形式)。

  2. Hk\mathbf H_k循环矩阵[Hk]p,q=γk+pqHk=circ(ck)[\mathbf H_k]_{p,q}=\gamma_{\langle k+p-q\rangle}\Rightarrow \mathbf H_k=\mathrm{circ}(\mathbf c_k),其中 (ck)[d]=γk+d(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}

  3. 用循环基 {Jd}\{\mathbf J_d\} 展开并向量化:vec(circ(ck))=Bck\mathrm{vec}(\mathrm{circ}(\mathbf c_k))=\mathbf B\mathbf c_k。把 N2N^2 维的大核投影到循环子空间得到 Rbulk=BHMbulkBCN×N\mathbf R_{\mathrm{bulk}}=\mathbf B^H\mathbf M_{\mathrm{bulk}}\mathbf B\in\mathbb C^{N\times N},bulk 就降到 NN 维二次型。

  4. 进一步把 [Rbulk]d,d[\mathbf R_{\mathrm{bulk}}]_{d,d'} 展开成显式求和,并利用核 S2(p,q,p,q)S_2(p,q,p',q') 只依赖“差分”,可把 (p,q,p,q)(p,q,p',q') 的四维求和压成 (d,d,t)(d,d',t)(其中 d=pqd=p-q, d=pqd'=p'-q', t=qqt=q-q'),得到一个差分域聚合核

[Rbulk]d,d=NtZNιbulk(d,d,t)exp ⁣(αS2(t+d,t,d,0)).[\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).

其中 ιbulk\iota_{\mathrm{bulk}} 是把 pairings + 单对超平面剔除后的 bulk 指示函数。

  1. 最后用“循环置换 Πk\Pi_k”的“频域对角调制”等价,把 kk 依赖写成 Hadamard 调制:
Qkbulk=1N2(dkdkH)Kbulk,Kbulk=FRbulkFH.\mathbf Q_k^{\mathrm{bulk}}=\frac{1}{N^2}\,(\mathbf d_k\mathbf d_k^H)\odot \mathbf K_{\mathrm{bulk}}, \quad \mathbf K_{\mathrm{bulk}}=\mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^H.

这就和你 sp 的 ()S~(\cdot)\odot \tilde{\mathbf S} 完全同味。


按 IEEE TSP 风格整理后的 LaTeX 推导(可直接贴进论文)

说明:下面默认 \langle\cdot\rangle 为模 NN 归约到 {0,,N1}\{0,\dots,N-1\}ω=2π/N\omega=2\pi/NF\mathbf F 为 unitary DFT。
mask 可能破坏 PSD,我在最后给了“取 Hermitian / PSD 投影”的标准 TSP 写法。

latex
% ============================================================ % Bulk-term reduction to an sp-like quadratic / norm form % ============================================================ \subsection{Bulk Term: From vec--matrix--vec to an $N\times N$ Difference-Domain Kernel} Let $\boldsymbol\gamma\triangleq \mathbf F^H\mathbf g$ so that $\gamma_\ell=\mathbf f_\ell^H\mathbf g$. Define the two $N\times N$ 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 $\mathbf g\in\mathbb R^N$, hence $\gamma_{\langle-\ell\rangle}=\gamma_{\langle\ell\rangle}^*$. Then for all $(p',q')$, \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 $c_k[d]\triangleq \gamma_{\langle k+d\rangle}$ and $\mathbf c_k=[c_k[0],\ldots,c_k[N-1]]^T$. 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 $N^2\times N^2$ 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 $\mathbf R_{\mathrm{bulk}}$} Let $[\mathbf R_{\mathrm{bulk}}]_{d,d'}=\operatorname{vec}(\mathbf J_d)^H\mathbf M_{\mathrm{bulk}}\operatorname{vec}(\mathbf J_{d'})$. 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 $p=q+d$ and $p'=q'+d'$ (all indices modulo $N$), 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 $t=q-q'$.} The term $S_2(p,q,p',q')$ and the bulk mask $\mathbf P^{\mathrm{bulk}}$ depend only on modular differences among $(p,q,p',q')$, hence the summand in \eqref{eq:Rbulk_qqprime} depends on $(q,q')$ only through $t\triangleq q-q'$. For each fixed $t\in\mathbb Z_N$, there are exactly $N$ pairs $(q,q')$ such that $q-q'=t$. 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 $\iota_{\mathrm{bulk}}(d,d',t)\in\{0,1\}$ is the indicator that removes pairings and single-pair hyperplanes. \paragraph*{Bulk indicator in $(d,d',t)$.} 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 $S_2(t+d,t,d',0)$.} Recall the cosine kernel $\Psi(\Delta)\triangleq\sum_{r=1}^{N_d}\cos(\omega r\Delta)$ 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 $(p,q,p',q')=(t+d,\ t,\ d',\ 0)$ 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 $\mathbf R_{\mathrm{bulk}}$. \subsection{Back to $\mathbf g$: sp-like Hadamard-modulated quadratic / norm form} Let $\boldsymbol\gamma=\mathbf F^H\mathbf g$ and define the circulant shift (permutation) matrix $\mathbf\Pi_k$ such that $\mathbf c_k=\mathbf\Pi_k\boldsymbol\gamma$ (i.e., $(\mathbf c_k)[d]=\gamma_{\langle k+d\rangle}$). 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 $k$.} Using $\mathbf\Pi_k=\mathbf F^H\mathbf D_k\mathbf F$ with $\mathbf D_k\triangleq\operatorname{diag}(e^{-j\omega k(n-1)})_{n=1}^N$, let $\mathbf K_{\mathrm{bulk}}\triangleq \mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^H$ and $\mathbf d_k\triangleq[e^{j\omega k(n-1)}]_{n=1}^N$. 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 $\mathbf R_{\mathrm{bulk}}$ is given explicitly by \eqref{eq:Rbulk_ddprime_tsum}--\eqref{eq:S2_bulk_ddprime_t}. \paragraph*{Norm form (PSD case).} If $\mathbf Q_k^{\mathrm{bulk}}\succeq \mathbf 0$, \begin{align} \boxed{ \mathcal S_k^{(\mathrm{bulk})}(\mathbf g) =\big\|(\mathbf Q_k^{\mathrm{bulk}})^{1/2}\mathbf g\big\|_2^2 = \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 $\mathbf Q_{k,+}^{\mathrm{bulk}}\triangleq \Pi_{\succeq 0}(\operatorname{Herm}\{\mathbf Q_k^{\mathrm{bulk}}\})$, yielding $\mathcal S_k^{(\mathrm{bulk})}(\mathbf g)\approx\|(\mathbf Q_{k,+}^{\mathrm{bulk}})^{1/2}\mathbf g\|_2^2$.

你现在得到的“最终结构”长什么样(和 sp 对齐)

  • sp:
Sksp(g)=gHQkspg=(()S~)1/2g22.\mathcal S_k^{\mathrm{sp}}(\mathbf g)=\mathbf g^H\mathbf Q_k^{\mathrm{sp}}\mathbf g =\left\|\Big(\big(\cdots\big)\odot \tilde{\mathbf S}\Big)^{1/2}\mathbf g\right\|_2^2.
  • bulk(我们刚推完):
Sk(bulk)(g)=(1N2(dkdkH)Kbulk)1/2g22,\boxed{ \mathcal S_k^{(\mathrm{bulk})}(\mathbf g) =\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, }

其中 Kbulk=FRbulkFH\mathbf K_{\mathrm{bulk}}=\mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^H,而

[Rbulk]d,d=Ntιbulk(d,d,t)exp ⁣(αS2(t+d,t,d,0)),[\mathbf R_{\mathrm{bulk}}]_{d,d'} = N\sum_{t}\iota_{\mathrm{bulk}}(d,d',t)\exp\!\big(-\alpha S_2(t+d,t,d',0)\big), S2(t+d,t,d,0)=(4Nd+2)[Ψ(t+d)+Ψ(td)Ψ(d)Ψ(d)Ψ(t)Ψ(t+dd)].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].

这就是你要的“同型简化”:bulk 也变成了“一个 kk-依赖的简单 Hadamard 调制 \odot 一个 kk-无关核矩阵”的二次型/范数形式;并且这个核矩阵可以由 (d,d,t)(d,d',t) 的显式求和生成。


如果你接下来想把 bulk 也做得更“像 sp 那样快”,我建议下一步是:研究 Rbulk\mathbf R_{\mathrm{bulk}}(d,d)(d,d') 上是否近似Toeplitz/circulant(很多差分核在大 NN 或忽略边界时会接近),那样 Kbulk=FRbulkFH\mathbf K_{\mathrm{bulk}}=\mathbf F\mathbf R_{\mathrm{bulk}}\mathbf F^H 会变得近似对角或低秩,bulk 就能进一步 FFT 化/低秩化。你要我也可以顺着这个方向继续推一个“可计算复杂度更低”的近似版本。

प्रश्न

把下面的公式中的exp都换成e和指数的形式:
\begin{align}
|r_k|^2 &= \sum_{n=1}^N\big|\mathbf{f}n\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-1)}{N}\right)\sum{m=1}^N\big|\mathbf{f}m\mathbf{x}\big|^2\exp!\left(-j\frac{2\pi k (m-1)}{N}\right)\notag\
& = \sum
{n=1}^N\sum_{m=1}^N \big|\mathbf{f}_n\mathbf{x}\big|^2\big|\mathbf{f}_m\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-m)}{N}\right)
\end{align}

\begin{align}
|r_k|^2 & = \sum_{n=1}^N\sum_{m=1}^N \big|\mathbf{f}n\mathbf{x}\big|^2\big|\mathbf{f}m\mathbf{x}\big|^2\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\notag\
& =\frac{1}{N^2}\sum
{n=1}^N\sum
{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\exp(j\mathbf{a}^H\mathbf{v})
\end{align}

\begin{align}
\mathbb{E}{|r_k|^2}& =\frac{1}{N^2}\mathbb{E}\left{\sum_{n=1}^N\sum_{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\exp(j\mathbf{a}^H\mathbf{v})\right}\notag\
& =\frac{1}{N^2}\sum_{n=1}^N\sum_{m=1}^N\exp!\left(j\frac{2\pi k (n-m)}{N}\right)\sum_{p,q,p'q'}\exp\left(-j\frac{2\pi(n-1)(p-q)}{N}\right)\exp\left(j\frac{2\pi(m-1)(p'-q')}{N}\right)\notag\
& \times \mathbb{E}\left{\exp(j\mathbf{a}^H\mathbf{v})\right}\notag\
\end{align}

उत्तर

下面我仅做符号层面的等价替换:把所有的 \exp(\cdot) 统一改写成 e()e^{(\cdot)},其余内容(求和结构、指数里的符号与系数)保持完全一致,方便你直接替换进论文或推导中。


第一组公式

latex
\begin{align} |r_k|^2 &= \sum_{n=1}^N\big|\mathbf{f}_n\mathbf{x}\big|^2 e^{\,j\frac{2\pi k (n-1)}{N}} \sum_{m=1}^N\big|\mathbf{f}_m\mathbf{x}\big|^2 e^{-j\frac{2\pi k (m-1)}{N}} \notag\\ &= \sum_{n=1}^N\sum_{m=1}^N \big|\mathbf{f}_n\mathbf{x}\big|^2 \big|\mathbf{f}_m\mathbf{x}\big|^2 e^{\,j\frac{2\pi k (n-m)}{N}} . \end{align}

第二组公式

latex
\begin{align} |r_k|^2 &= \sum_{n=1}^N\sum_{m=1}^N \big|\mathbf{f}_n\mathbf{x}\big|^2 \big|\mathbf{f}_m\mathbf{x}\big|^2 e^{\,j\frac{2\pi k (n-m)}{N}} \notag\\ &= \frac{1}{N^2} \sum_{n=1}^N\sum_{m=1}^N e^{\,j\frac{2\pi k (n-m)}{N}} \sum_{p,q,p',q'} e^{-j\frac{2\pi (n-1)(p-q)}{N}} e^{\,j\frac{2\pi (m-1)(p'-q')}{N}} e^{\,j\mathbf{a}^H\mathbf{v}} . \end{align}

第三组公式(期望)

latex
\begin{align} \mathbb{E}\{|r_k|^2\} &= \frac{1}{N^2} \mathbb{E}\Bigg\{ \sum_{n=1}^N\sum_{m=1}^N e^{\,j\frac{2\pi k (n-m)}{N}} \sum_{p,q,p',q'} e^{-j\frac{2\pi (n-1)(p-q)}{N}} e^{\,j\frac{2\pi (m-1)(p'-q')}{N}} e^{\,j\mathbf{a}^H\mathbf{v}} \Bigg\} \notag\\ &= \frac{1}{N^2} \sum_{n=1}^N\sum_{m=1}^N e^{\,j\frac{2\pi k (n-m)}{N}} \sum_{p,q,p',q'} e^{-j\frac{2\pi (n-1)(p-q)}{N}} e^{\,j\frac{2\pi (m-1)(p'-q')}{N}} \,\mathbb{E}\!\left\{e^{\,j\mathbf{a}^H\mathbf{v}}\right\}. \end{align}

小提示(IEEE TSP 写作习惯)

  • 数学上完全等价\exp(j\theta)e^{j\theta} 没有任何区别;
  • TSP / 信号处理论文中,在长指数链或多重指数并列时,e^{(\cdot)} 通常更常见、更紧凑;
  • 若后面还要做 指数合并(如 eAeB=eA+Be^{A}e^{B}=e^{A+B}),用 e^{} 的形式视觉上也更清晰。

如果你愿意,我也可以帮你统一全文指数写法风格(例如:全部改为 e()e^{(\cdot)},并把 ω=2π/N\omega=2\pi/N 抽出来,进一步简化符号密度)。

प्रश्न

分析一下下面的代码错在哪里,并修改:
\begin{align}
r_ &= \mathbf{\tilde{x}}^H \mathbf{J}k \mathbf{\tilde{x}} = \sqrt{LN} \mathbf{\tilde{x}}^H \mathbf{F}{LN}^H \operatorname{Diag}(\tilde{\mathbf{f}}{LN-k+1}) \mathbf{F}{LN} \mathbf{\tilde{x}} \notag\
&= \mathbf{x}{\text{up}}^H \mathbf{P}^H \mathbf{F}{LN}^H \operatorname{Diag}(\tilde{\mathbf{f}}{LN-k+1}) \mathbf{F}{LN} \mathbf{P}\mathbf{x}{\text{up}}\notag \
&= (LN)^{\frac{3}{2}}\mathbf{x}
{\text{up}}^H \mathbf{F}{LN}^H \operatorname{Diag} (\mathbf{F}{LN}\mathbf{P}\odot\tilde{\mathbf{f}}{LN-k+1}\mathbf{F}{LN}^\mathbf{P}^) \notag \
& \cdot \mathbf{F}{LN}\mathbf{x}{\text{up}}.
\end{align}

उत्तर

你这段 LaTeX/推导里主要有 3 类错误:** (i) 记号/下标写错导致语法错误;(ii) 维度不匹配;(iii) 第 3 行把“对角化/逐元素乘”的恒等式用错了(还夹杂了不该出现的 P\mathbf P^*)**。我按行指出问题并给出一份可用的改写版本。


1) 第一行的明显语法错误:r_\tilde{x}

错在哪里

  • r_ 在 LaTeX 里是“下标未给出”,会直接报错。你想写的是 rkr_krr_{\ell} 之类。
  • \mathbf{\tilde{x}} 这种写法不规范:\tilde 应该作用在符号上,粗体用 \tilde{\mathbf x} 更标准(IEEE 风格也更常用)。

改法

把第一行开头改成:

latex
r_k = \tilde{\mathbf x}^H \mathbf J_k \tilde{\mathbf x}

2) 第二行:漏了 LN\sqrt{LN}(或你应当明确它去哪里了)

你第一行写

rk=LNx~HFLNHDiag(f~LNk+1)FLNx~,r_k=\sqrt{LN}\,\tilde{\mathbf x}^H \mathbf F_{LN}^H \mathrm{Diag}(\tilde{\mathbf f}_{LN-k+1})\mathbf F_{LN}\tilde{\mathbf x},

第二行代入 x~=Pxup\tilde{\mathbf x}=\mathbf P\mathbf x_{\rm up} 后,按代数应当是

rk=LNxupHPHFLNHDiag(f~LNk+1)FLNPxup.r_k=\sqrt{LN}\,\mathbf x_{\rm up}^H \mathbf P^H \mathbf F_{LN}^H \mathrm{Diag}(\tilde{\mathbf f}_{LN-k+1})\mathbf F_{LN}\mathbf P\mathbf x_{\rm up}.

但你第二行把前面的 LN\sqrt{LN} 丢了,这在推导上是不一致的(除非你把它吸收到 P\mathbf Pxup\mathbf x_{\rm up} 的定义里,但文中没说)。


3) 第三行:核心恒等式用错 + 维度和对象都不对

你写成了

(LN)3/2xupHFLNHDiag(FLNPf~LNk+1FLNP)FLNxup.(LN)^{3/2}\mathbf x_{\rm up}^H \mathbf F_{LN}^H \mathrm{Diag}(\mathbf F_{LN}\mathbf P\odot \tilde{\mathbf f}_{LN-k+1}\mathbf F_{LN}^*\mathbf P^*)\cdot \mathbf F_{LN}\mathbf x_{\rm up}.

这里错得比较多,关键点如下

(a) 系数 (LN)3/2(LN)^{3/2} 来历不对。
从前两行最多只会有一个 LN\sqrt{LN}。除非你又引入了 FLN\mathbf F_{LN} 非酉归一化(例如没有 1/LN1/\sqrt{LN}),或者 P\mathbf P 带有 LN\sqrt{LN} 级别的缩放,否则不可能凭空变成 (LN)3/2(LN)^{3/2}

(b) Diag()\operatorname{Diag}(\cdot) 里面必须是长度 LNLN 的向量。
但你塞进去的是类似“矩阵 \odot 矩阵”的东西:

  • FLNP\mathbf F_{LN}\mathbf P 是矩阵
  • f~LNk+1FLNP\tilde{\mathbf f}_{LN-k+1}\mathbf F_{LN}^*\mathbf P^* 也不像是合法的向量表达(而且 f~LNk+1\tilde{\mathbf f}_{LN-k+1} 如果是“列向量”,不能左乘去乘矩阵;如果是“对角矩阵”,那你又写错了符号)

(c) 你想用的正确恒等式应是: 对任意矩阵 A\mathbf A 和向量 d\mathbf d,有

AHDiag(d)A=(Ad1T)HA\mathbf A^H\mathrm{Diag}(\mathbf d)\mathbf A = (\mathbf A\odot \mathbf d\mathbf 1^T)^H \mathbf A

或更常用的逐元素写法:

AHDiag(d)A=ndnananH,\mathbf A^H\mathrm{Diag}(\mathbf d)\mathbf A =\sum_{n} d_n\,\mathbf a_n \mathbf a_n^H,

其中 an\mathbf a_nA\mathbf A 的第 nn 行/列(取决于你定义)。

你现在那种把 FLNP\mathbf F_{LN}^*\mathbf P^* 硬塞进来是不对的:在 PHFHDiag()FP\mathbf P^H\mathbf F^H\mathrm{Diag}(\cdot)\mathbf F\mathbf P 这种结构里,Diag()\mathrm{Diag}(\cdot) 只作用在中间,不会凭空生成一个 P\mathbf P^*


4) 一份“维度正确、推导正确、写法干净”的修改版本

下面给两种等价的改写,你选一种风格即可。


版本 A:最直接、最不容易出错(推荐)

latex
\begin{align} r_k &= \tilde{\mathbf x}^H \mathbf J_k \tilde{\mathbf x} \notag\\ &= \sqrt{LN}\,\tilde{\mathbf x}^H \mathbf F_{LN}^H \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)\, \mathbf F_{LN}\tilde{\mathbf x} \notag\\ &= \sqrt{LN}\,\mathbf x_{\mathrm{up}}^H \mathbf P^H \mathbf F_{LN}^H \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)\, \mathbf F_{LN}\mathbf P\,\mathbf x_{\mathrm{up}}. \end{align}

这版修复了:

  • r_r_k
  • \mathbf{\tilde{x}}\tilde{\mathbf x}
  • 第二行漏掉的 LN\sqrt{LN}
  • 不引入错误的第三行变形(你那行本来就不对)

版本 B:如果你确实想把 P\mathbf PF\mathbf F 合并成一个“矩阵-对角-矩阵”的标准二次型

AFLNPCLN×N.\mathbf A \triangleq \mathbf F_{LN}\mathbf P\in\mathbb C^{LN\times N}.

rk=LNxupHAHDiag(f~LNk+1)Axup.r_k=\sqrt{LN}\,\mathbf x_{\mathrm{up}}^H \mathbf A^H \mathrm{Diag}(\tilde{\mathbf f}_{LN-k+1}) \mathbf A\,\mathbf x_{\mathrm{up}}.

LaTeX:

latex
\begin{align} r_k &= \sqrt{LN}\,\mathbf x_{\mathrm{up}}^H \mathbf P^H \mathbf F_{LN}^H \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)\, \mathbf F_{LN}\mathbf P\,\mathbf x_{\mathrm{up}} \notag\\ &= \sqrt{LN}\,\mathbf x_{\mathrm{up}}^H \mathbf A^H \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)\, \mathbf A\,\mathbf x_{\mathrm{up}}, \qquad \mathbf A \triangleq \mathbf F_{LN}\mathbf P. \end{align}

如果你后面要做快速计算(FFT/逐元素乘),从这一步再推更安全。


5) 你第三行如果“本意”是把中间对角矩阵换成 Hadamard 形式,该怎么写才对?

一个正确且常用的恒等式是:对任意 ACM×N\mathbf A\in\mathbb C^{M\times N}dCM\mathbf d\in\mathbb C^{M}

AHDiag(d)A=(Diag(d)1/2A)H(Diag(d)1/2A).\mathbf A^H\mathrm{Diag}(\mathbf d)\mathbf A = (\mathrm{Diag}(\mathbf d)^{1/2}\mathbf A)^H(\mathrm{Diag}(\mathbf d)^{1/2}\mathbf A).

所以你可以写成“范数平方”:

rk=LNDiag(f~LNk+1)1/2FLNPxup22r_k=\sqrt{LN}\,\big\|\mathrm{Diag}(\tilde{\mathbf f}_{LN-k+1})^{1/2}\mathbf F_{LN}\mathbf P\,\mathbf x_{\rm up}\big\|_2^2

(如果 f~LNk+1\tilde{\mathbf f}_{LN-k+1} 非负实才有“平方根对角”这一步;否则只能写 Diag()\mathrm{Diag}(\cdot) 不开根。)

对应 LaTeX:

latex
\begin{align} r_k &= \sqrt{LN}\,\mathbf x_{\mathrm{up}}^H \mathbf A^H \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)\, \mathbf A\,\mathbf x_{\mathrm{up}} \notag\\ &= \sqrt{LN}\, \big\| \operatorname{Diag}\!\big(\tilde{\mathbf f}_{LN-k+1}\big)^{1/2}\, \mathbf A\,\mathbf x_{\mathrm{up}} \big\|_2^2, \qquad \mathbf A=\mathbf F_{LN}\mathbf P, \end{align}

如果你把 P\mathbf P 的具体定义(上采样插零矩阵?重复矩阵?含不含 L\sqrt{L} 归一化?)和 f~LNk+1\tilde{\mathbf f}_{LN-k+1} 到底是 DFT 的列向量还是 对角的特征值向量 贴一下,我还能把上面“范数/Hadamard/FFT 友好”的版本继续推到你想要的最终实现式(例如把它写成一次 FFT + 点乘 + 一次 IFFT 的形式)。

इस प्रश्नोत्तर को साझा करें