% Created 2023-07-12 Wed 16:48
% Intended LaTeX compiler: pdflatex
\documentclass[serif,professionalfont,colorlinks,aspectratio=169]{beamer}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{graphicx}
\usepackage{grffile}
\usepackage{longtable}
\usepackage{wrapfig}
\usepackage{rotating}
\usepackage[normalem]{ulem}
\usepackage{amsmath}
\usepackage{textcomp}
\usepackage{amssymb}
\usepackage{capt-of}
\usepackage{hyperref}
\usepackage{color}
\usepackage{listings}
\usepackage{pdfpages}
\usepackage{xcolor}
\usepackage{graphicx}
\usepackage{xspace}
\usepackage{libertine}
\usepackage{fontawesome}
\mode<beamer>{\usetheme{Boadilla}}
\setbeamertemplate{navigation symbols}{}
\setbeamerfont{alerted text}{series=\bfseries}
\setbeamercolor{alerted text}{fg=blue}
\usecolortheme{default}
\usefonttheme{serif}
\usefonttheme{professionalfonts}
\lstloadlanguages{[ISO]C++}
\definecolor{codegreen}{rgb}{0,0.6,0}
\definecolor{codegray}{rgb}{0.5,0.5,0.5}
\definecolor{codepurple}{rgb}{0.50,0,0.82}
\definecolor{codeorange}{rgb}{0.95,0.3,0.1}
\definecolor{backcolour}{rgb}{0.95,0.95,0.92}
\lstdefinestyle{mystyle}{
commentstyle=\color{codegreen},
keywordstyle=\color{codeorange},
numberstyle=\tiny\color{codegray},
stringstyle=\color{codepurple},
basicstyle=\footnotesize,
breakatwhitespace=false,
breaklines=true,
captionpos=b,
keepspaces=true,
numbers=left,
numbersep=5pt,
showspaces=false,
showstringspaces=false,
showtabs=false,
tabsize=2
}
\lstset{style=mystyle}
\usetheme{default}
\author{Brett Viren}
\date{\today}
\title{ForWaRD: Fourier-Wavelet Regularized Deconvolution}
\subtitle{Proposed extension to WC signal processing}
\hypersetup{
 pdfauthor={Brett Viren},
 pdftitle={ForWaRD: Fourier-Wavelet Regularized Deconvolution},
 pdfkeywords={},
 pdfsubject={},
 pdfcreator={Emacs 28.2 (Org mode 9.4.6)}, 
 pdflang={English}}
\begin{document}

\maketitle

\begin{frame}[label={sec:org8497278}]{Topics}
\begin{itemize}
\item Problem statement
\item Brief intro to general and Fourier signal basis
\item Naive solution
\item FoRD: Fourier Regularized Deconvolution (currently what WC uses)
\item Problems with FoRD
\item Brief intro to wavelets
\item ForWaRD: Fourier-Wavelet Regularized Deconvolution
\item Application of FoRD and ForWaRD to toy 1D LArTPC sim+sigproc
\item Toward full implementation.
\end{itemize}
\end{frame}

\begin{frame}[label={sec:org06673b3}]{Measure a signal in the presence of noise}
\begin{block}{Linear discrete system}
\Huge

\[ y[n] = \mathcal{H}x[n] + \gamma[n] \]

\normalsize

\begin{itemize}
\item \(x[n]\) \alert{signal} (unknown) as function of discrete time coordinate \(n\in[0,...,N-1]\).

\item \(\gamma[n]\) \alert{noise}
\item \(y[n]\) \alert{measurement}
\item \(\mathcal{H}\) \alert{detector response}:
\begin{itemize}
\item Assume linear convolution model: \(\mathcal{H}x[n] \equiv (h * x)[n]\), \(h[n]\) is \alert{impulse response} of \(H\).
\end{itemize}
\end{itemize}
\end{block}

\begin{block}{Problem: find good estimate of the signal}
\Huge

\[\tilde{x}[n] \approx x[n]\]
\end{block}
\end{frame}

\begin{frame}[label={sec:org1b2e8a3}]{Orthonormal vector basis: \(\{\psi_k\}\) on \(\mathbb{R}^N\) and its \alert{dual set} \(\{\tilde\psi_k\}\).}
\begin{block}{Decomposition of \alert{sampled signal vector} \(x[n],\ n \in [0,N-1]\) onto basis vectors}
\[ x[n]  = \sum_k X[k]\psi_k[n] \]
\end{block}

\begin{block}{Representation of signal as vector of coefficients (dual-basis dot product)}
\[ X[k] \equiv \sum_{n=0}^{N-1} \tilde\psi^*_k[n]x[n] \equiv\ <\tilde\psi_k, x>\]
\end{block}

\begin{block}{Reconstruction of vector in original basis}
\[ x[n] = \sum_{k=0}^{N-1} X[k]\psi_k[n] \]
\end{block}
\end{frame}


\begin{frame}[label={sec:orgfce7193}]{The Fourier basis}
\begin{block}{Fourier vector basis and its dual in terms of ``\(N^{th}\) root of unity'' \(z_N\)}
\[ \psi_k[n] = \frac{1}{N} (z_N)^{kn},\ \tilde\psi_k[n] \equiv (z_N)^{kn},\ z_N \equiv e^{i2\pi/N} \]
\end{block}

\begin{block}{Convolution}
\[ (a*b)[n] \leftrightarrow A[k]B[k] \]
\end{block}

\begin{block}{Parseval energy equality}
\[ \sum_n|x[n]|^2 = \frac{1}{N}\sum_k|X[k]|^2 \]
\end{block}
\end{frame}

\begin{frame}[label={sec:org7025f94}]{Aside: generating a basis from a ``mother''}
\begin{block}{Euclidean \(\mathbb{R}^N\) (\emph{eg}, time domain) basis from Dirac-\(\delta\)}
\[ \psi \equiv \delta_{0,n},\ shift(\psi_0, k) \to \delta_{k,n},\ \{\psi_k\} = \{\delta_{k,n}\} \]
\end{block}

\begin{block}{Fourier basis from ``\(N^{th}\) root of unity'' \(z_N\)}
\[ \psi_k[n] = \frac{1}{N} (z_N)^{kn},\ \tilde\psi_k[n] \equiv (z_N)^{kn},\ z_N \equiv e^{i2\pi/N} \]
\end{block}

\begin{block}{Wavelet basis from mother wavelet + scaling functions}
We will get to this.
\end{block}
\end{frame}


\begin{frame}[label={sec:orgf97d9b7}]{The ``naive'' estimate}
Back to solving for \(x\) in \(y[n] = \mathcal{H}x[n] + \gamma[n]\)

\begin{block}{Apply inverse detector response \(\mathcal{H}^-1\) in Fourier basis for estimate \(\tilde x\)}
\[ \tilde{x}[n] \equiv \mathcal{H}^{-1} y[n] = x[n] + \mathcal{H}^{-1}\gamma[n] \]


\[ \tilde{X}[k] = \frac{Y[k]}{H[k]} = X[k] + \frac{\Gamma[k]}{H[k]} \]
\end{block}
\begin{block}{Why is it called ``naive''?}
\begin{itemize}
\item Influence of noise \(\Gamma\) remains and it will be \alert{amplified} when \(H\) is small (or even zero).
\end{itemize}

Next, look at \(\mathcal{H}\) and \(\gamma\) for real detectors\ldots{}.
\end{block}
\end{frame}


\begin{frame}[label={sec:orge6e9e08}]{Impulse response \(\mathcal{H}\) for MicroBooNE}
\begin{columns}
\begin{column}{0.6\columnwidth}
\begin{center}
\includegraphics[width=\textwidth,page=1]{fwd-uboone-responses.pdf}
\end{center}
\end{column}

\begin{column}{0.4\columnwidth}
Waveform:
\begin{itemize}
\item \footnotesize U-plane, bipolar, asymmetric.
\end{itemize}
Spectrum:
\begin{itemize}
\item \footnotesize Strongly peaked at low frequency
\item Sharp dive at \(k=0\) frequency
\begin{itemize}
\item \scriptsize Should go to zero but calculation domain was too small.
\end{itemize}
\item Strong attenuation above \(\approx 200\) kHz
\end{itemize}
\end{column}
\end{columns}
\end{frame}

\begin{frame}[label={sec:orgccc03e8}]{Noise spectrum for MicroBooNE}
Mean spectrum model \(\to\) generate waveforms \(\to\) reconstruct mean spectrum.

\begin{columns}
\begin{column}{0.6\columnwidth}
\includegraphics[width=\textwidth,page=1]{fwd-uboone-noise.pdf}
\end{column}

\begin{column}{0.4\columnwidth}
\footnotesize
Noise is substantial across a broad band.
\begin{itemize}
\item \scriptsize Particularly a \(k\approx 0\) where \(H\) is small
\end{itemize}

\vfill

\scriptsize

Note: need care in \alert{mean} \emph{vs} \alert{mode} when generating waveforms and applying Parseval energy.
\end{column}
\end{columns}
\end{frame}

\begin{frame}[label={sec:org60f8e29}]{Toy signal model}
\begin{block}{``Detector''}
\begin{itemize}
\item quasi-1D field response integrating 2D response over central wire
\item simple Gaussian smearing for diffusion.
\item look at voltage level, no ADC
\end{itemize}
\end{block}

\begin{block}{``Kinematics''}
\begin{itemize}
\item Ideal ``track'' as square wave in time with MIP-level electron density.
\item Slightly less ideal ``blips'' uniformly random Gaussians filling a square wave.
\end{itemize}
\end{block}
\end{frame}


\begin{frame}[label={sec:org01ce8d2}]{}
\begin{center}
\includegraphics[height=\textheight,page=1]{fwd-uboone-square.pdf}
\end{center}
\end{frame}

\begin{frame}[label={sec:orgd4b996c}]{}
\begin{center}
\includegraphics[height=\textheight,page=5]{fwd-uboone-square.pdf}
\end{center}
\end{frame}

\begin{frame}[label={sec:org3d3e6c5}]{}
\begin{center}
\includegraphics[height=\textheight,page=4]{fwd-uboone-square.pdf}
\end{center}
\end{frame}

\begin{frame}[label={sec:orgc25e83d}]{}
\begin{center}
\includegraphics[height=\textheight,page=8]{fwd-uboone-square.pdf}
\end{center}
\end{frame}


\begin{frame}[label={sec:org74cbbf3}]{}
\begin{center}
\includegraphics[height=\textheight,page=8]{fwd-uboone-blips.pdf}
\end{center}
\end{frame}

\begin{frame}[label={sec:org1bead1f}]{}
\begin{center}
\includegraphics[height=\textheight,page=14]{fwd-uboone-blips.pdf}
\end{center}
\end{frame}




\begin{frame}[label={sec:org25139b0}]{Protect against divide-by-zero}
\begin{block}{Modify the naive estimate in obvious way}
\[\tilde{X}[k] \equiv \begin{cases}
X[k] + \frac{\Gamma[k]}{H[k]} & \text{if}\ H[k] > 0 \\
0                             & \text{otherwise}\end{cases}\]

That is, we \alert{destroy} all noise at any frequency where response is zero.
\begin{itemize}
\item This idea is extended next to provide a \alert{shink} some frequencies instead of destroy them.
\end{itemize}
\end{block}
\end{frame}

\begin{frame}[label={sec:orge1e3c71}]{FoRD: Fourier Regularized Deconvolution (WC's ``SP filtering'')}
\begin{block}{Introduce \alert{Fourier shrinkage} coefficients \(\lambda^f[k]\) with \alert{regularization terms} \(\Lambda[k]\)}
\[ \lambda^f[k] = \frac{|H[k]|^2}{|H[k]|^2 + \Lambda[k]},\ \Lambda[k] \ge 0 \]
\end{block}

\begin{block}{DFT of \alert{FoRD estimate} \(\tilde{X}_{\lambda^f}\)}
\[ \tilde{X}_{\lambda^f}[k] \equiv \tilde{X}[k]\lambda^f[k] \equiv X_{\lambda^f}[k] + \frac{\Gamma_{\lambda^f}[k]}{H[k]}\]
\begin{itemize}
\item Attenuate (``shink'') the amplified noise in naive estimate \(\tilde{X}\) when \(\Lambda > 0,\ \lambda_f < 1\)
\begin{itemize}
\item And may \alert{distort the signal}
\end{itemize}
\item Define terms: DFT of \alert{retained signal} \(X_{\lambda^f}\) and DFT of \alert{leaked noise} \(\frac{\Gamma_{\lambda^f}[k]}{H[k]}\)
\end{itemize}
\end{block}
\end{frame}

\begin{frame}[label={sec:orgd529316}]{FoRD: defining shrinkage coefficients}
\begin{block}{Wiener regularization term}
\[ \Lambda[k] \equiv \alpha \frac{|\Gamma[k]|^2}{|X[k]|^2} \to \lambda^f[k] = \frac{|Y[k]|^2}{|Y[k]|^2 + \alpha |\Gamma[k]|^2}\]

Optimize a single parameter \(\alpha\) across all frequencies (and events).
\begin{itemize}
\item \(\alpha = 0\) naive deconvolution.
\item \(\alpha \approx 1\) more strongly shrink noisy frequencies.
\end{itemize}
\end{block}

\begin{block}{Other forms}
\begin{itemize}
\item Tikhonov: \(\Lambda[k] \equiv \tau\).
\item Allpass: \(\lambda[k] = |H[k]|\)
\end{itemize}
\end{block}
\end{frame}


\begin{frame}[label={sec:orgad62e1c}]{FoRD shrinkage vs WCT filtering}
\begin{block}{FoRD}
A single Wiener filter with \alert{fixed noise} model.
\begin{itemize}
\item Ends up being naturally peaked in frequency (\alert{band pass}).
\item May be \alert{adaptive}, based on each \alert{measured} waveform or \alert{fixed} based on some average signal.
\end{itemize}
\end{block}

\begin{block}{WCT}
Pair of \alert{fixed} filters based on average \alert{simulated signal} and \alert{fixed noise model}.
\begin{itemize}
\item ``Wiener-inspired'' \alert{low-pass} filter provides ``half'' Wiener filter.
\item Add one of two ad-hoc \alert{high-pass} (``low-frequency'') filters to form \alert{band pass}.
\begin{itemize}
\item ``tight'' low noise, attenuates long signals.
\item ``loose'' higher noise, retains long signals.
\end{itemize}
\end{itemize}
\end{block}
\end{frame}


\begin{frame}[label={sec:org85df5a8}]{FoRD: signal and noise}
\end{frame}


\begin{frame}[label={sec:org11e73b5}]{FoRD: pros and cons}
\end{frame}


\begin{frame}[label={sec:orgd8133c9}]{Brief intro to wavelets}
A wavelet basis is \alert{generated} from:
\begin{itemize}
\item A ``mother'' wavelet function \(\psi[n]\)
\item A scaling function \(\phi[n]\)
\item A time-shift operator \(\tau_l\)
\end{itemize}



\begin{block}{Similar to ``normal'' basis}
\begin{itemize}
\item Euclidean basis fully localized in time, generated by a shift operator
\end{itemize}
\end{block}
\end{frame}

\begin{frame}[label={sec:org4e44f4d}]{ForWaRD: Fourier-Wavelet Regularized Deconvolution}
\end{frame}
\begin{frame}[label={sec:org0c6990f}]{Application of FoRD and ForWaRD to toy 1D LArTPC sim+sigproc}
\end{frame}

\begin{frame}[label={sec:org73901bd}]{Toward full implementation in Wire-Cell toolkit}
\end{frame}
\end{document}