326 lines
13 KiB
TeX
326 lines
13 KiB
TeX
\chapter{Proximal Decoding}%
|
|
\label{chapter:proximal_decoding}
|
|
|
|
TODO
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\section{Decoding Algorithm}%
|
|
\label{sec:prox:Decoding Algorithm}
|
|
|
|
Proximal decoding was proposed by Wadayama et. al as a novel formulation of
|
|
optimization-based decoding \cite{proximal_paper}.
|
|
With this algorithm, minimization is performed using the proximal gradient
|
|
method.
|
|
In contrast to \ac{LP} decoding, the objective function is based on a
|
|
non-convex optimization formulation of the \ac{MAP} decoding problem.
|
|
|
|
In order to derive the objective function, the authors begin with the
|
|
\ac{MAP} decoding rule, expressed as a continuous maximization problem%
|
|
\footnote{The expansion of the domain to be continuous doesn't constitute a
|
|
material difference in the meaning of the rule.
|
|
The only change is that what previously were \acp{PMF} now have to be expressed
|
|
in terms of \acp{PDF}.}
|
|
over $\boldsymbol{x}$:%
|
|
%
|
|
\begin{align}
|
|
\hat{\boldsymbol{x}} = \argmax_{\tilde{\boldsymbol{x}} \in \mathbb{R}^{n}}
|
|
f_{\tilde{\boldsymbol{X}} \mid \boldsymbol{Y}}
|
|
\left( \tilde{\boldsymbol{x}} \mid \boldsymbol{y} \right)
|
|
= \argmax_{\tilde{\boldsymbol{x}} \in \mathbb{R}^{n}} f_{\boldsymbol{Y}
|
|
\mid \tilde{\boldsymbol{X}}}
|
|
\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)
|
|
f_{\tilde{\boldsymbol{X}}}\left( \tilde{\boldsymbol{x}} \right)%
|
|
\label{eq:prox:vanilla_MAP}
|
|
.\end{align}%
|
|
%
|
|
The likelihood $f_{\boldsymbol{Y} \mid \tilde{\boldsymbol{X}}}
|
|
\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right) $ is a known function
|
|
determined by the channel model.
|
|
The prior \ac{PDF} $f_{\tilde{\boldsymbol{X}}}\left( \tilde{\boldsymbol{x}} \right)$ is also
|
|
known, as the equal probability assumption is made on
|
|
$\mathcal{C}$.
|
|
However, since the considered domain is continuous,
|
|
the prior \ac{PDF} cannot be ignored as a constant during the minimization
|
|
as is often done, and has a rather unwieldy representation:%
|
|
%
|
|
\begin{align}
|
|
f_{\tilde{\boldsymbol{X}}}\left( \tilde{\boldsymbol{x}} \right) =
|
|
\frac{1}{\left| \mathcal{C} \right| }
|
|
\sum_{\boldsymbol{c} \in \mathcal{C} }
|
|
\delta\big( \tilde{\boldsymbol{x}} - \left( -1 \right) ^{\boldsymbol{c}}\big)
|
|
\label{eq:prox:prior_pdf}
|
|
.\end{align}%
|
|
%
|
|
In order to rewrite the prior \ac{PDF}
|
|
$f_{\tilde{\boldsymbol{X}}}\left( \tilde{\boldsymbol{x}} \right)$,
|
|
the so-called \textit{code-constraint polynomial} is introduced as:%
|
|
%
|
|
\begin{align*}
|
|
h\left( \tilde{\boldsymbol{x}} \right) =
|
|
\underbrace{\sum_{i=1}^{n} \left( \tilde{x_i}^2-1 \right) ^2}_{\text{Bipolar constraint}}
|
|
+ \underbrace{\sum_{j=1}^{m} \left[
|
|
\left( \prod_{i\in N_c \left( j \right) } \tilde{x_i} \right)
|
|
-1 \right] ^2}_{\text{Parity constraint}}%
|
|
.\end{align*}%
|
|
%
|
|
The intention of this function is to provide a way to penalize vectors far
|
|
from a codeword and favor those close to one.
|
|
In order to achieve this, the polynomial is composed of two parts: one term
|
|
representing the bipolar constraint, providing for a discrete solution of the
|
|
continuous optimization problem, and one term representing the parity
|
|
constraints, accommodating the role of the parity-check matrix $\boldsymbol{H}$.
|
|
The prior \ac{PDF} is then approximated using the code-constraint polynomial as:%
|
|
%
|
|
\begin{align}
|
|
f_{\tilde{\boldsymbol{X}}}\left( \tilde{\boldsymbol{x}} \right)
|
|
\approx \frac{1}{Z}\mathrm{e}^{-\gamma h\left( \tilde{\boldsymbol{x}} \right) }%
|
|
\label{eq:prox:prior_pdf_approx}
|
|
.\end{align}%
|
|
%
|
|
The authors justify this approximation by arguing, that for
|
|
$\gamma \rightarrow \infty$, the approximation in equation
|
|
(\ref{eq:prox:prior_pdf_approx}) approaches the original function in equation
|
|
(\ref{eq:prox:prior_pdf}).
|
|
This approximation can then be plugged into equation (\ref{eq:prox:vanilla_MAP})
|
|
and the likelihood can be rewritten using the negative log-likelihood
|
|
$L \left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right) = -\ln\left(
|
|
f_{\boldsymbol{Y} \mid \tilde{\boldsymbol{X}}}\left(
|
|
\boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right) \right) $:%
|
|
%
|
|
\begin{align*}
|
|
\hat{\boldsymbol{x}} &= \argmax_{\tilde{\boldsymbol{x}} \in \mathbb{R}^{n}}
|
|
\mathrm{e}^{- L\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right) }
|
|
\mathrm{e}^{-\gamma h\left( \tilde{\boldsymbol{x}} \right) } \\
|
|
&= \argmin_{\tilde{\boldsymbol{x}} \in \mathbb{R}^n} \big(
|
|
L\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)
|
|
+ \gamma h\left( \tilde{\boldsymbol{x}} \right)
|
|
\big)%
|
|
.\end{align*}%
|
|
%
|
|
Thus, with proximal decoding, the objective function
|
|
$g\left( \tilde{\boldsymbol{x}} \right)$ considered is%
|
|
%
|
|
\begin{align}
|
|
g\left( \tilde{\boldsymbol{x}} \right) = L\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}}
|
|
\right)
|
|
+ \gamma h\left( \tilde{\boldsymbol{x}} \right)%
|
|
\label{eq:prox:objective_function}
|
|
\end{align}%
|
|
%
|
|
and the decoding problem is reformulated to%
|
|
%
|
|
\begin{align*}
|
|
\text{minimize}\hspace{2mm} &L\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)
|
|
+ \gamma h\left( \tilde{\boldsymbol{x}} \right)\\
|
|
\text{subject to}\hspace{2mm} &\tilde{\boldsymbol{x}} \in \mathbb{R}^n
|
|
.\end{align*}
|
|
%
|
|
|
|
For the solution of the approximate \ac{MAP} decoding problem, the two parts
|
|
of equation (\ref{eq:prox:objective_function}) are considered separately:
|
|
the minimization of the objective function occurs in an alternating
|
|
fashion, switching between the negative log-likelihood
|
|
$L\left( \boldsymbol{y} \mid \boldsymbol{x} \right) $ and the scaled
|
|
code-constraint polynomial $\gamma h\left( \boldsymbol{x} \right) $.
|
|
Two helper variables, $\boldsymbol{r}$ and $\boldsymbol{s}$, are introduced,
|
|
describing the result of each of the two steps.
|
|
The first step, minimizing the log-likelihood, is performed using gradient
|
|
descent:%
|
|
%
|
|
\begin{align}
|
|
\boldsymbol{r} \leftarrow \boldsymbol{s} - \omega \nabla
|
|
L\left( \boldsymbol{y} \mid \boldsymbol{s} \right),
|
|
\hspace{5mm}\omega > 0
|
|
\label{eq:prox:step_log_likelihood}
|
|
.\end{align}%
|
|
%
|
|
For the second step, minimizing the scaled code-constraint polynomial, the
|
|
proximal gradient method is used and the \textit{proximal operator} of
|
|
$\gamma h\left( \tilde{\boldsymbol{x}} \right) $ has to be computed.
|
|
It is then immediately approximated with gradient-descent:%
|
|
%
|
|
\begin{align*}
|
|
\textbf{prox}_{\gamma h} \left( \tilde{\boldsymbol{x}} \right) &\equiv
|
|
\argmin_{\boldsymbol{t} \in \mathbb{R}^n}
|
|
\left( \gamma h\left( \boldsymbol{t} \right) +
|
|
\frac{1}{2} \lVert \boldsymbol{t} - \tilde{\boldsymbol{x}} \rVert \right)\\
|
|
&\approx \tilde{\boldsymbol{x}} - \gamma \nabla h \left( \tilde{\boldsymbol{x}} \right),
|
|
\hspace{5mm} \gamma > 0, \text{ small}
|
|
.\end{align*}%
|
|
%
|
|
The second step thus becomes%
|
|
%
|
|
\begin{align*}
|
|
\boldsymbol{s} \leftarrow \boldsymbol{r} - \gamma \nabla h\left( \boldsymbol{r} \right),
|
|
\hspace{5mm}\gamma > 0,\text{ small}
|
|
.\end{align*}
|
|
%
|
|
While the approximation of the prior \ac{PDF} made in equation (\ref{eq:prox:prior_pdf_approx})
|
|
theoretically becomes better
|
|
with larger $\gamma$, the constraint that $\gamma$ be small is important,
|
|
as it keeps the effect of $h\left( \tilde{\boldsymbol{x}} \right) $ on the landscape
|
|
of the objective function small.
|
|
Otherwise, unwanted stationary points, including local minima, are introduced.
|
|
The authors say that ``in practice, the value of $\gamma$ should be adjusted
|
|
according to the decoding performance.'' \cite[Sec. 3.1]{proximal_paper}.
|
|
|
|
%The components of the gradient of the code-constraint polynomial can be computed as follows:%
|
|
%%
|
|
%\begin{align*}
|
|
% \frac{\partial}{\partial x_k} h\left( \boldsymbol{x} \right) =
|
|
% 4\left( x_k^2 - 1 \right) x_k + \frac{2}{x_k}
|
|
% \sum_{i\in \mathcal{B}\left( k \right) } \left(
|
|
% \left( \prod_{j\in\mathcal{A}\left( i \right)} x_j\right)^2
|
|
% - \prod_{j\in\mathcal{A}\left( i \right) }x_j \right)
|
|
%.\end{align*}%
|
|
%\todo{Only multiplication?}%
|
|
%\todo{$x_k$: $k$ or some other indexing variable?}%
|
|
%%
|
|
In the case of \ac{AWGN}, the likelihood
|
|
$f_{\boldsymbol{Y} \mid \tilde{\boldsymbol{X}}}
|
|
\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)$
|
|
is%
|
|
%
|
|
\begin{align*}
|
|
f_{\boldsymbol{Y} \mid \tilde{\boldsymbol{X}}}
|
|
\left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)
|
|
= \frac{1}{\sqrt{2\pi\sigma^2}}\mathrm{e}^{
|
|
-\frac{\lVert \boldsymbol{y}-\tilde{\boldsymbol{x}}
|
|
\rVert^2 }
|
|
{2\sigma^2}}
|
|
.\end{align*}
|
|
%
|
|
Thus, the gradient of the negative log-likelihood becomes%
|
|
\footnote{For the minimization, constants can be disregarded. For this reason,
|
|
it suffices to consider only proportionality instead of equality.}%
|
|
%
|
|
\begin{align*}
|
|
\nabla L \left( \boldsymbol{y} \mid \tilde{\boldsymbol{x}} \right)
|
|
&\propto -\nabla \lVert \boldsymbol{y} - \tilde{\boldsymbol{x}} \rVert^2\\
|
|
&\propto \tilde{\boldsymbol{x}} - \boldsymbol{y}
|
|
,\end{align*}%
|
|
%
|
|
allowing equation (\ref{eq:prox:step_log_likelihood}) to be rewritten as%
|
|
%
|
|
\begin{align*}
|
|
\boldsymbol{r} \leftarrow \boldsymbol{s}
|
|
- \omega \left( \boldsymbol{s} - \boldsymbol{y} \right)
|
|
.\end{align*}
|
|
%
|
|
|
|
One thing to consider during the actual decoding process, is that the gradient
|
|
of the code-constraint polynomial can take on extremely large values.
|
|
To avoid numerical instability, an additional step is added, where all
|
|
components of the current estimate are clipped to $\left[-\eta, \eta \right]$,
|
|
where $\eta$ is a positive constant slightly larger than one:%
|
|
%
|
|
\begin{align*}
|
|
\boldsymbol{s} \leftarrow \Pi_{\eta} \left( \boldsymbol{r}
|
|
- \gamma \nabla h\left( \boldsymbol{r} \right) \right)
|
|
,\end{align*}
|
|
%
|
|
$\Pi_{\eta}\left( \cdot \right) $ expressing the projection onto
|
|
$\left[ -\eta, \eta \right]^n$.
|
|
|
|
The iterative decoding process resulting from these considerations is shown in
|
|
figure \ref{fig:prox:alg}.
|
|
|
|
\begin{figure}[H]
|
|
\centering
|
|
|
|
\begin{genericAlgorithm}[caption={}, label={}]
|
|
$\boldsymbol{s} \leftarrow \boldsymbol{0}$
|
|
for $K$ iterations do
|
|
$\boldsymbol{r} \leftarrow \boldsymbol{s} - \omega \left( \boldsymbol{s} - \boldsymbol{y} \right) $
|
|
$\boldsymbol{s} \leftarrow \Pi_\eta \left(\boldsymbol{r} - \gamma \nabla h\left( \boldsymbol{r} \right) \right)$
|
|
$\boldsymbol{\hat{x}} \leftarrow \text{sign}\left( \boldsymbol{s} \right) $
|
|
if $\boldsymbol{H}\boldsymbol{\hat{c}} = \boldsymbol{0}$ do
|
|
return $\boldsymbol{\hat{c}}$
|
|
end if
|
|
end for
|
|
return $\boldsymbol{\hat{c}}$
|
|
\end{genericAlgorithm}
|
|
|
|
|
|
\caption{Proximal decoding algorithm for an \ac{AWGN} channel}
|
|
\label{fig:prox:alg}
|
|
\end{figure}
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\section{Implementation Details}%
|
|
\label{sec:prox:Implementation Details}
|
|
|
|
The algorithm as first implemented in Python because of the fast development
|
|
process and straightforward debugging.
|
|
They have subsequently been reimplemented in C++ using the Eigen%
|
|
\footnote{\url{https://eigen.tuxfamily.org}}
|
|
linear algebra library to achieve higher performance.
|
|
The focus has been set on a fast implementation, sometimes at the expense of
|
|
memory usage.
|
|
The evaluation of the simulation results has been wholly realized in Python.
|
|
|
|
The gradient of the code-constraint polynomial is given by \cite[Sec. 2.3]{proximal_paper}%
|
|
%
|
|
\begin{align*}
|
|
\nabla h\left( \boldsymbol{x} \right) &= \begin{bmatrix}
|
|
\frac{\partial}{\partial x_1}h\left( \boldsymbol{x} \right) &
|
|
\ldots &
|
|
\frac{\partial}{\partial x_n}h\left( \boldsymbol{x} \right) &
|
|
\end{bmatrix}^\text{T}, \\[1em]
|
|
\frac{\partial}{\partial x_k}h\left( \boldsymbol{x} \right) &= 4\left( x_k^2 - 1 \right) x_k
|
|
+ \frac{2}{x_k} \sum_{j\in N_v\left( k \right) }\left(
|
|
\left( \prod_{i \in N_c\left( j \right)} x_i \right)^2
|
|
- \prod_{i\in N_c\left( j \right) } x_i \right)
|
|
.\end{align*}
|
|
%
|
|
Evidently, the products
|
|
$\prod_{i\in N_c\left( j \right) } x_i,\hspace{1mm}\forall i\in \mathcal{J}$
|
|
can be precomputed, as they are the same for all components $x_k$ of $\boldsymbol{x}$.
|
|
Defining%
|
|
%
|
|
\begin{align*}
|
|
\boldsymbol{p} := \begin{bmatrix}
|
|
\prod_{i\in N_c\left( 1 \right) }x_i \\
|
|
\vdots \\
|
|
\prod_{i\in N_c\left( m \right) }x_i \\
|
|
\end{bmatrix}
|
|
\hspace{5mm}
|
|
\text{and}
|
|
\hspace{5mm}
|
|
\boldsymbol{v} := \boldsymbol{p}^{\circ 2} - \boldsymbol{p}
|
|
,\end{align*}
|
|
%
|
|
the gradient can be written as%
|
|
%
|
|
\begin{align*}
|
|
\nabla h\left( \boldsymbol{x} \right) =
|
|
4\left( \boldsymbol{x}^{\circ 3} - \boldsymbol{x} \right)
|
|
+ 2\boldsymbol{x}^{\circ -1} \circ \boldsymbol{H}^\text{T}
|
|
\boldsymbol{v}
|
|
,\end{align*}
|
|
%
|
|
enabling the computation of the gradient primarily with element-wise
|
|
operations and matrix-vector multiplication.
|
|
This is beneficial, as the libraries used for the implementation are
|
|
heavily optimized for such calculations (e.g., through vectorization of the
|
|
operations).
|
|
\todo{Note about how the equation with which the gradient is calculated is
|
|
itself similar to a message-passing rule}
|
|
|
|
The projection $\prod_{\eta}\left( . \right)$ also proves straightforward to
|
|
compute, as it amounts to simply clipping each component of the vector onto
|
|
$[-\eta, \eta]$ individually.
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\section{Results}%
|
|
\label{sec:prox:Results}
|
|
|
|
|
|
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
|
\section{Improved Implementation}%
|
|
\label{sec:prox:Improved Implementation}
|
|
|