From Decision to Random Certificates: Exponential Separation for Edge Estimation with Independent Set Queries
Extracted source text · arXiv 2607.07483
\documentclass[a4paper,11pt]{article}
\usepackage[portrait, margin=1in]{geometry}
\usepackage[utf8]{inputenc}
\usepackage[numbers, sort, compress]{natbib}
\usepackage{nicefrac}
\usepackage{multirow}
\usepackage{xspace}
% \newcommand{\gopi}[1]{\textcolor{purple}{Gopinath: #1 \xspace}}
% \usepackage{amsthm}
% \usepackage{thmtools}
% \usepackage{algorithm2e}
% \usepackage{ulem}
\usepackage[dvipsnames,table]{xcolor}
\input{support}
% \usepackage{lineno}
% \linenumbers
%\usepackage{times}
%\title{From Decision to Random Certificates: Faster Edge Estimation with Independent Set Queries}
\title{From Decision to Random Certificates: Exponential Separation for Edge Estimation with Independent Set Queries}
% \author{
% Debarshi Chanda \\
% Indian Statistical Institute \\
% Kolkata, India\\
% \and
% Buddha Dev Das \\
% Indian Statistical Institute \\
% Kolkata, India\\
% \and
% Arijit Ghosh \\
% Indian Statistical Institute \\
% Kolkata, India\\
% \and
% Gopinath Mishra \\
% Institute of Mathematical Sciences\\
% Chennai, India \\
% }
\usepackage{authblk} % Handles alignment automatically
\author[1]{Debarshi Chanda}
\author[1]{Buddha Dev Das}
\author[1]{Arijit Ghosh}
\author[2]{Gopinath Mishra}
\affil[1]{Indian Statistical Institute, Kolkata, India}
\affil[2]{Institute of Mathematical Sciences, HBNI, Chennai, India }
\date{}
\begin{document}
\maketitle
\thispagestyle{empty}
\begin{abstract}
We study the problem of estimating the number of edges in an undirected,
unweighted graph using sublinear query access. We consider a query model that
preserves the structure of Independent Set (IS) queries, but augments their
output with a random certificate: given a vertex subset, the oracle returns a
uniformly random edge from the induced subgraph if one exists, and returns
null otherwise.
Using this access, we give a randomized algorithm that outputs a
$(1 \pm \varepsilon)$-approximation to the number of edges with constant
success probability using $\widetilde{O}(\log^{2} m)$ queries. This implies
an exponential separation from both standard IS queries and global
random edge-sampling models: estimating the number of edges using
standard IS queries require
$\widetilde{\Theta}\!\left(\min\left\{\sqrt{m},\, \frac{n}{\sqrt{m}}\right\}\right)$
queries, while direct random edge-sample access requires
$\widetilde{\Theta}(\sqrt{m})$ samples. Beyond separation in query complexity, our algorithm is output-sensitive: its query complexity is polylogarithmic in the number of edges in the graph. This aligns with the
classical objective in group testing, where one seeks algorithms that are both
worst-case optimal and instance-adaptive.
Conceptually, our model connects
group testing, the decision--versus--counting dichotomy, graph property
testing, and the ``power of a random certificate'', and can be viewed as a
structured form of conditional sampling of edges in graphs.
\end{abstract}
\newpage
%\tableofcontents
\newpage
\setcounter{page}{1}
\input{Sections/1_Intro}
\input{Sections/3_Notation_And_Overview}
\input{Sections/4_TechOverview}
\input{Sections/5_Technical}
% Acknowledgments---Will not appear in anonymized version
\paragraph{Acknowledgments:}{We thank an anonymous reviewer for their suggestions in improving the query complexity from $\log^3 \edgecount$ (in an earlier version) to $\log^2 \edgecount$.
Arijit Ghosh acknowledges partial support from the Science and Engineering Research Board (SERB), Government of India, through the MATRICS grant MTR/2023/001527, and from the Department of Science and Technology (DST), Government of India, through grant TPN-104427.}
\bibliographystyle{plainnat}
\bibliography{refs}
\appendix
\input{Sections/Appendix}
\end{document}
%Input Basic Packages
\usepackage{float}
%Input Math and Algo Related packages
\usepackage{amsmath, mathtools, amsthm, amssymb}
\usepackage{algorithm, yhmath}
\usepackage[noend]{algpseudocode}
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\usepackage{nicefrac}
\usepackage{thmtools}
\usepackage{thm-restate}
% \allowdisplaybreaks
%Input Packages to Handle Images and Graphs
% \usepackage{wrapfig, epsfig, graphicx, subcaption, pgf, tikz, pgfplots}
\usepackage{tikz}
% Tikz Setup
\usetikzlibrary{calc, shapes.geometric, arrows, automata, decorations.pathreplacing, arrows.meta}
% \pgfplotsset{compat=1.18}
%Include additional various packages
% \usepackage{enumerate, thmtools, soul, setspace}
%Hyperref Setup
\usepackage{hyperref}
\hypersetup{
% colorlinks=true,
% linkcolor=blue,
% filecolor=blue,
% urlcolor=blue,
% citecolor=blue,
pdftitle={Improved Bounds with a Simple Algorithm for Edge Estimation for Graphs of Unknown Size},
pdfpagemode=FullScreen,
}
\usepackage{cleveref}
% Load the parskip package with skip and indent options
% \usepackage[skip=10pt, indent=15pt]{parskip}
\setlength{\marginparwidth }{2cm}
\usepackage[textwidth=4em,]{todonotes}
% Include Theorem related packages
\theoremstyle{plain}
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\newtheorem{notation}[theorem]{Notation}
% \newtheorem{problem}[theorem]{Problem}
\newtheorem{assumption}{Assumption}
\theoremstyle{remark}
\newtheorem{remark}[theorem]{Remark}
% \newcommand\numberthis{\addtocounter{equation}{1}\tag{\theequation}}
%Redefining Command
% \renewcommand{\baselinestretch}{1.25}
% \renewcommand{\chaptermark}[1]{\markboth{#1}{}}
% \renewcommand{\sectionmark}[1]{\markright{\thesection\ #1}}
%Citations Maintenance
% \usepackage[nottoc]{tocbibind}
\input{Support/boxes}
\input{Support/fonts}
\input{Support/macros}
\input{Support/math_symbols}
%Note Taking Snippets
% \newcommand{\todo}[1]{\textcolor{red}{TODO: #1}}
\newcommand{\fixme}[1]{\textcolor{red}{FIXME: #1}}
% \newcommand{\debarshi}[1]{\textcolor{red}{Debarshi: {#1}}}
%Little Boxes
\usepackage[most]{tcolorbox}
\newtcolorbox{idea}[1][]
{
colbacktitle=cyan,
colback=cyan!10,
arc=1pt,
boxrule=1pt,
title=#1 % I would like to make this (one of these in general) assignment optional depending on #1, #2...
}
\newtcolorbox{update}[1][]
{
colbacktitle=gray,
colback=gray!10,
arc=1pt,
boxrule=1pt,
title=#1 % I would like to make this (one of these in general) assignment optional depending on #1, #2...
}
\newtcolorbox{question}[1][]
{
coltitle=black,
colbacktitle=yellow,
colback=yellow!10,
arc=1pt,
boxrule=1pt,
title=#1 % I would like to make this (one of these in general) assignment optional depending on #1, #2...
}
\newtcolorbox{note}[1][]
{
coltitle=black,
colbacktitle=green,
colback=green!10,
arc=1pt,
boxrule=1pt,
title=#1 % I would like to make this (one of these in general) assignment optional depending on #1, #2...
}
\newtcolorbox{problem}[1][]
{
coltitle=black,
colbacktitle=red!60,
colback=red!10,
arc=1pt,
boxrule=1pt,
title=#1 % I would like to make this (one of these in general) assignment optional depending on #1, #2...
}
\newcommand{\colnote}[3]{\textcolor{#1}{{\small $\ll$\textsf{#2}$\gg$\marginpar{\tiny\bf \textcolor[rgb]{1.00,0.00,0.00}{#3}}}}}
%Bold Symbols
\newcommand{\ba}{\boldsymbol{a}}
\newcommand{\bb}{\boldsymbol{b}}
\newcommand{\bc}{\boldsymbol{c}}
\newcommand{\bd}{\boldsymbol{d}}
\newcommand{\be}{\boldsymbol{e}}
\newcommand{\bg}{\boldsymbol{g}}
\newcommand{\bh}{\boldsymbol{h}}
\newcommand{\bi}{\boldsymbol{i}}
\newcommand{\bj}{\boldsymbol{j}}
\newcommand{\bl}{\boldsymbol{l}}
\newcommand{\bm}{\boldsymbol{m}}
\newcommand{\bn}{\boldsymbol{n}}
\newcommand{\bo}{\boldsymbol{o}}
\newcommand{\bp}{\boldsymbol{p}}
\newcommand{\bq}{\boldsymbol{q}}
\newcommand{\br}{\boldsymbol{r}}
\newcommand{\bs}{\boldsymbol{s}}
\newcommand{\bt}{\boldsymbol{t}}
\newcommand{\bu}{\boldsymbol{u}}
\newcommand{\bv}{\boldsymbol{v}}
\newcommand{\bw}{\boldsymbol{w}}
\newcommand{\bx}{\boldsymbol{x}}
\newcommand{\by}{\boldsymbol{y}}
\newcommand{\bz}{\boldsymbol{z}}
\newcommand{\bA}{\boldsymbol{A}}
\newcommand{\bB}{\boldsymbol{B}}
\newcommand{\bC}{\boldsymbol{C}}
\newcommand{\bD}{\boldsymbol{D}}
\newcommand{\bE}{\boldsymbol{E}}
\newcommand{\bF}{\boldsymbol{F}}
\newcommand{\bG}{\boldsymbol{G}}
\newcommand{\bH}{\boldsymbol{H}}
\newcommand{\bI}{\boldsymbol{I}}
\newcommand{\bJ}{\boldsymbol{J}}
\newcommand{\bK}{\boldsymbol{K}}
\newcommand{\bL}{\boldsymbol{L}}
\newcommand{\bM}{\boldsymbol{M}}
\newcommand{\bN}{\boldsymbol{N}}
\newcommand{\bO}{\boldsymbol{O}}
\newcommand{\bP}{\boldsymbol{P}}
\newcommand{\bQ}{\boldsymbol{Q}}
\newcommand{\bR}{\boldsymbol{R}}
\newcommand{\bS}{\boldsymbol{S}}
\newcommand{\bT}{\boldsymbol{T}}
\newcommand{\bU}{\boldsymbol{U}}
\newcommand{\bV}{\boldsymbol{V}}
\newcommand{\bW}{\boldsymbol{W}}
\newcommand{\bX}{\boldsymbol{X}}
\newcommand{\bY}{\boldsymbol{Y}}
\newcommand{\bZ}{\boldsymbol{Z}}
\newcommand{\fA}{\field{A}}
\newcommand{\fB}{\field{B}}
\newcommand{\fC}{\field{C}}
\newcommand{\fD}{\field{D}}
\newcommand{\fE}{\field{E}}
\newcommand{\fF}{\field{F}}
\newcommand{\fG}{\field{G}}
\newcommand{\fH}{\field{H}}
\newcommand{\fI}{\field{I}}
\newcommand{\fJ}{\field{J}}
\newcommand{\fK}{\field{K}}
\newcommand{\fL}{\field{L}}
\newcommand{\fM}{\field{M}}
\newcommand{\fN}{\field{N}}
\newcommand{\fO}{\field{O}}
\newcommand{\fP}{\field{P}}
\newcommand{\fQ}{\field{Q}}
\newcommand{\fR}{\field{R}}
\newcommand{\fS}{\field{S}}
\newcommand{\fT}{\field{T}}
\newcommand{\fU}{\field{U}}
\newcommand{\fV}{\field{V}}
\newcommand{\fW}{\field{W}}
\newcommand{\fX}{\field{X}}
\newcommand{\fY}{\field{Y}}
\newcommand{\fZ}{\field{Z}}
%Mathcal Commands
\newcommand{\sA}{\mathcal{A}}
\newcommand{\sB}{\mathcal{B}}
\newcommand{\sC}{\mathcal{C}}
\newcommand{\sD}{\mathcal{D}}
\newcommand{\sE}{\mathcal{E}}
\newcommand{\sF}{\mathcal{F}}
\newcommand{\sG}{\mathcal{G}}
\newcommand{\sH}{\mathcal{H}}
\newcommand{\sI}{\mathcal{I}}
\newcommand{\sJ}{\mathcal{J}}
\newcommand{\sK}{\mathcal{K}}
\newcommand{\sL}{\mathcal{L}}
\newcommand{\sM}{\mathcal{M}}
\newcommand{\sN}{\mathcal{N}}
\newcommand{\sO}{\mathcal{O}}
\newcommand{\sP}{\mathcal{P}}
\newcommand{\sQ}{\mathcal{Q}}
\newcommand{\sR}{\mathcal{R}}
\newcommand{\sS}{\mathcal{S}}
\newcommand{\sT}{\mathcal{T}}
\newcommand{\sU}{\mathcal{U}}
\newcommand{\sV}{\mathcal{V}}
\newcommand{\sW}{\mathcal{W}}
\newcommand{\sX}{\mathcal{X}}
\newcommand{\sY}{\mathcal{Y}}
\newcommand{\sZ}{\mathcal{Z}}
%Further Specific Fonts (Miscellaneous)
\newcommand{\bhatY}{\boldsymbol{\hat{Y}}}
\newcommand{\bbary}{\boldsymbol{\bar{y}}}
\newcommand{\bhatX}{\boldsymbol{\hat{X}}}
\newcommand{\bbarx}{\boldsymbol{\bar{x}}}
\newcommand{\bbarZ}{\boldsymbol{\bar{Z}}}
\newcommand{\bbarz}{\boldsymbol{\bar{z}}}
\newcommand{\barz}{\bar{z}}
\newcommand{\bbarS}{\boldsymbol{\bar{S}}}
\newcommand{\balpha}{\boldsymbol{\alpha}}
\newcommand{\bbeta}{\boldsymbol{\beta}}
\newcommand{\bmu}{\boldsymbol{\mu}}
\newcommand{\bpi}{\boldsymbol{\pi}}
\newcommand{\bSigma}{\boldsymbol{\Sigma}}
\newcommand{\bsigma}{\boldsymbol{\sigma}}
\newcommand{\btheta}{\boldsymbol{\theta}}
\newcommand{\bTheta}{\boldsymbol{\Theta}}
\newcommand{\bzero}{\boldsymbol{0}}
\newcommand{\bbartheta}{\boldsymbol{\bar{\theta}}}
\newcommand{\bfB}{\mathbf{B}}
\newcommand{\sbarZ}{\bar{\mathcal{Z}}}
\newcommand{\fbag}{\bold{F}}
%Math Operators
\newcommand{\poly}[1]{\mathrm{poly}\fbrac{#1}}
\newcommand{\REQ}{\mathrm{RE}}
\newcommand{\DEG}{\mathrm{DEG}}
\newcommand{\NBR}{\mathrm{NEIGHBOUR}}
\newcommand\Order[1]{ O\left( #1 \right)}
\newcommand\Ordertilde[1]{ \widetilde{O}\left( #1 \right)}
\newcommand{\degreeq}{\texttt{Deg}}
\newcommand{\randedgeq}{\texttt{RandEdge}}
\newcommand{\neighbourq}{\texttt{Nbr}}
\newcommand{\edgeexistsq}{\texttt{Pair}}
\newcommand{\bisq}{\texttt{BIS}}
\newcommand{\isq}{\texttt{IS}}
\newcommand{\iscq}{\texttt{ISC}}
%Graph Macros - To be integrated later
\newcommand{\graph}{G}
\newcommand{\vertexset}{V}
\newcommand{\edgeset}{E}
\newcommand{\vertexcount}{n}
\newcommand{\edgecount}{m}
\newcommand{\vertex}{v}
\newcommand{\altvertex}{u}
\newcommand{\edge}{e}
\newcommand{\neighbour}{{\sf N}}
\newcommand{\degree}[1]{\mathrm{deg}\fbrac{{#1}}}
\newcommand{\arboricity}{\alpha}
\newcommand{\numtriangle}{T}
\newcommand{\esttriangle}{\Bar{T}}
\newcommand{\lighttriangles}{\numtriangle_{light}}
\newcommand{\heavytriangles}{\numtriangle_{heavy}}
\newcommand{\triangleset}{\Delta}
\renewcommand{\triangle}{\tau} %Triangle is taken by latex
\newcommand{\weightfunc}{w}
\newcommand{\weightfamily}{\sW}
\newcommand{\heavyoracle}{\texttt{Heavy}}
\newcommand{\samplededges}{S}
\newcommand{\edgesamplesize}{A}
\newcommand{\vertexsamplesize}{K}
% \newcommand{\samplededges}{S}
\newcommand{\densebucketset}{I_\heavybucket}
\newcommand{\querycount}{r}
\newcommand{\vertexdensity}{\rho}
\newcommand{\deglowerbound}{\ell}
\newcommand{\heavybucket}{\mathrm{H}}
\newcommand{\lightbucket}{\mathrm{L}}
\newcommand{\bucketset}{\mathrm{B}}
\newcommand{\heavydegapprox}{\widetilde{d}_\heavybucket}
\newcommand{\heavydeg}{\degree{\heavybucket}}
\newcommand{\lightdeg}{d_\lightbucket}
\newcommand{\randomvertices}{T}
\newcommand{\heavyprobapprox}{\widetilde{p}_\heavybucket}
\newcommand{\heavyprob}{p_\heavybucket}
\newcommand{\adaptiveestimate}{\bar{X}}
\newcommand{\randomdegreciprocal}{X}
\newcommand{\activevertexcount}{n}
\newcommand{\approxedgecount}{\widehat{\edgecount}}
\newcommand{\nonadapalgo}{\texttt{NonAdaptiveEdge}}
\newcommand{\fullnonadapalgo}{\texttt{FullNonAdaptiveEdge}}
\newcommand{\preapprox}{\gamma}
\newcommand{\collisioncount}{r}
\newcommand{\resamplesize}{s}
\newcommand{\collisionedge}{\widetilde{m}_\text{coll}}
\newcommand{\noncollisionedge}{\widetilde{m}_\text{non-coll}}
\newcommand{\iscollision}{\texttt{IsCollision}}
\newcommand{\collision}{k}
\newcommand{\preresamplesize}{t}
\newcommand{\birthdayestimate}{\texttt{BirthdayEstimate}}
\newcommand{\addestimation}{\texttt{ApproxSubset}}
\newcommand{\mulestimation}{\texttt{MulEstimation}}
\newcommand{\fractionofedges}{p}
\newcommand{\approxfractionofedges}{\widehat{\fractionofedges}}
\newcommand{\edgeestimationISC}{\texttt{EdgeEstimation}}
\newcommand{\threshold}{\tau}
\newcommand{\iscoll}{coll}
\newcommand{\subsequence}{\texttt{ApproxSequence}}
\newcommand{\X}{X}
\newcommand{\Y}{Y}
\newcommand{\birthdayevent}{\mathcal{E}_\mathrm{bir}}
\newcommand{\subseqevent}{\mathcal{E}_\mathrm{sub}}
\newcommand{\mulestevent}{\mathcal{E}_\mathrm{mul}}
%Layerwise approximation
\newcommand{\lightapprox}{\approxerror}
\newcommand{\sampleapprox}{\approxerror}
\newcommand{\constant}{c}
%----------------------------------------------------
\newcommand{\bucket}{B}
% \newcommand{\deg}{d}
\newcommand{\bucketcount}{t}
\newcommand{\edgeestimate}{M}
% \newcommand{\complain}[1]{\textcolor{red}{#1}}
\newcommand{\iterations}{\kappa}
\newcommand{\thresholdconst}{\zeta_\mathrm{sp}}
\newcommand{\trueiter}{t^*}
\newcommand{\aps}{\texttt{APS}}
\newcommand{\apsq}{\texttt{APSQ}}
%Math Fields
\newcommand{\field}[1]{\mathbb{#1}}
\newcommand{\R}{\field{R}}
\newcommand{\Nat}{\field{N}}
\newcommand{\Int}{\field{Z}}
\newcommand{\Rat}{\field{Q}}
%Operations
\newcommand\inn[2]{ \left\langle {#1} \,,\, {#2} \right\rangle }
\newcommand\RE[2]{ D\left({#1} \| {#2}\right) }
\newcommand{\norm}[1]{\left\|{#1}\right\|}
\newcommand{\func}[3]{{#1} : {#2} \rightarrow {#3}}
\DeclarePairedDelimiter{\ceil}{\lceil}{\rceil}
\DeclarePairedDelimiter{\floor}{\lfloor}{\rfloor}
\DeclareMathOperator{\Tr}{Tr}
\newcommand{\defeq}{\stackrel{\rm def}{=}}
\newcommand{\sign}{\mbox{\sc sgn}}
\newcommand{\dt}{\displaystyle}
\newcommand{\wh}{\widehat}
\newcommand{\ve}{\varepsilon}
\newcommand{\hlambda}{\wh{\lambda}}
\newcommand{\yhat}{\wh{y}}
\newcommand{\hDelta}{\wh{\Delta}}
\newcommand{\hdelta}{\wh{\delta}}
%Color Definitions
\newcommand{\blue}[1]{\textcolor{blue}{#1}}
\newcommand{\red}[1]{\textcolor{red}{#1}}
\newcommand{\green}[1]{\textcolor{OliveGreen}{#1}}
\newcommand{\violet}[1]{\textcolor{violet}{#1}}
%Notational Shortcuts
\newcommand{\sequence}[2]{{#1}_1,{#1}_2,...,{#1}_{#2}}
\newcommand{\fbrac}[1]{\left({#1}\right)}
\newcommand{\sbrac}[1]{\left\{{#1}\right\}}
\newcommand{\tbrac}[1]{\left[{#1}\right]}
\newcommand{\abs}[1]{\left|{#1}\right|}
\newcommand{\size}[1]{\left|{#1}\right|}
% \newcommand{\constant}{c}
%Optimization etc.
\newcommand{\argmin}{\mathop{\mathrm{argmin}}}
\newcommand{\argmax}{\mathop{\mathrm{argmax}}}
\newcommand{\conv}{\mathop{\mathrm{conv}}}
\newcommand{\interior}{\mathop{\mathrm{int}}}
\newcommand{\dom}{\mathop{\mathrm{Dom}}}
\newcommand{\range}{\mathop{\mathrm{Range}}}
% \newcommand{\poly}[1]{\mathrm{poly}\fbrac{#1}}
\newcommand{\polylogneps}{\poly{\approxerror^{-1}\log\vertexcount}}
\newcommand{\lipschitzconstant}{\ell}
\DeclareMathOperator{\Proj}{\Pi}
%Online Learning
\DeclareMathOperator{\Regret}{Regret}
\DeclareMathOperator{\Wealth}{Wealth}
\DeclareMathOperator{\Reward}{Reward}
\DeclareMathOperator{\Risk}{Risk}
\DeclareMathOperator{\Prox}{Prox}
% Distributional Distances
\newcommand{\Wasserstein}[3]{\sW_{#1}\left({#2},{#3}\right)}
\newcommand{\TV}[2]{TV\left({#1},{#2}\right)}
\newcommand{\hockeystick}[3]{\operatorname{H_{#1}}\fbrac{{#2}||{#3}}}
\newcommand{\KL}[2]{\operatorname{KL}\left({#1};{#2}\right)}
\newcommand{\hamming}[2]{\operatorname{Ham}\fbrac{{#1},{#2}}}
%Probability
\DeclareMathOperator*{\Prob}{\field{P}}
\DeclareMathOperator*{\E}{\field{E}}
\DeclareMathOperator*{\Var}{\mathrm{Var}}
\newcommand{\sigalg}{\sigma\text{-Algebra}}
%Distributions
\newcommand{\indicator}{\mathbf{1}}
\newcommand{\support}[1]{Supp\fbrac{#1}}
\newcommand{\distribution}{\sD}
\newcommand{\laplace}{\mathrm{Lap}}
\newcommand{\uniform}{\mathrm{Unif}}
\newcommand{\normal}{\sN}
%Random-Approx
\newcommand{\fhat}{\hat{f}}
\newcommand{\ftilde}{\tilde{f}}
\newcommand{\invfhat}{\fhat^{-1}}
\newcommand{\invftilde}{\ftilde^{-1}}
\newcommand{\approxerror}{\varepsilon}
\newcommand{\confidence}{\delta}
\newcommand{\inverse}[1]{{#1}^{-1}}
%Algorithms
\newcommand{\bigo}[1]{O\fbrac{{#1}}}
\newcommand{\bigot}[1]{\widetilde{O}\fbrac{{#1}}}
\newcommand{\bigotln}[1]{\tilde{O}_{\log \vertexcount}\fbrac{{#1}}}
\newcommand{\smallo}[1]{o\fbrac{{#1}}}
\newcommand{\smallot}[1]{\tilde{o}\fbrac{{#1}}}
\newcommand{\bigomega}[1]{\Omega\fbrac{{#1}}}
\newcommand{\bigomegat}[1]{\tilde{\Omega}\fbrac{{#1}}}
\newcommand{\smallomega}[1]{\omega\fbrac{{#1}}}
\newcommand{\smallomegat}[1]{\tilde{\omega}\fbrac{{#1}}}
\newcommand{\bigtheta}[1]{\Theta\fbrac{{#1}}}
\newcommand{\bigthetat}[1]{\tilde{\Theta}\fbrac{{#1}}}
\newcommand{\algo}{\sA}
\newcommand{\polylog}[1]{\poly\log\fbrac{#1}}
%Graph Macros
% \newcommand{\graph}{G}
% \newcommand{\vertexset}{V}
% \newcommand{\edgeset}{E}
% \newcommand{\arboricity}{\alpha}
% \newcommand{\vertexcount}{n}
% \newcommand{\edgecount}{m}
% \newcommand{\vertex}{v}
% \newcommand{\altvertex}{u}
% \newcommand{\edge}{e}
% \newcommand{\neighbour}{Neigh}
% \newcommand{\degree}[1]{d_{#1}}
\newcommand{\hindex}{h}
%Query Macros
% \newcommand{\randedgeq}{\texttt{Random Edge}}
% \newcommand{\neighbourq}{\texttt{Neighbor}}
% \newcommand{\edgeexistsq}{\texttt{EdgeExist}}
% \newcommand{\degreeq}{\texttt{Degree}}
% \newcommand{\isq}{\texttt{IS}}
% \newcommand{\bisq}{\texttt{BIS}}
% \newcommand{\localq}{\texttt{Local Queries}}
%Function Macros
\newcommand{\domain}{\sD}
\newcommand{\ran}{\sR}
% Streaming Macros
\newcommand{\edgearrival}{\text{EA}}
%Lower Bound Macros
\newcommand{\communicationcomplexity}{\textit{CC}}
\newcommand{\ccspace}{Q}
\newcommand{\querycomm}{q}
\newcommand{\disjointness}[1]{\texttt{DISJ}_{#1}}
\newcommand{\promisedisjointness}[1]{\texttt{DISJ}^{prom}_{#1}}
\newcommand{\intersection}{\texttt{INT}}
\newcommand{\stringlength}{M}
\newcommand{\intersectionsize}{r}
\newcommand{\defaultint}{\intersection^\stringlength_\intersectionsize}
\newcommand{\alicestring}{x}
\newcommand{\bobstring}{y}
\newcommand{\ccgraph}{{\graph_{\alicestring,\bobstring}}}
\newcommand{\ccvertexset}{\vertexset_{\alicestring,\bobstring}}
\newcommand{\ccedgeset}{\edgeset_{\alicestring,\bobstring}}
\newcommand{\embedding}{\sE}
%\color{blue}
\section{Introduction}
We consider the problem of estimating the number of edges in an undirected,
unweighted, simple graph using a sublinear number of queries. Let
$\graph = (\vertexset,\edgeset)$ be a graph with
$\vertexset = [\vertexcount] := \{1,2,\dots,\vertexcount\}$ vertices and
$\size{\edgeset} = \edgecount$ edges.
Given query access to $\graph$ and an approximation parameter
$\approxerror \in (0,1)$, the objective is to output an estimate
$\approxedgecount$ such that
\[
(1-\approxerror)\edgecount
\;\le\;
\approxedgecount
\;\le\;
(1+\approxerror)\edgecount
\]
with success probability at least $2/3$.
\paragraph{Query model and our result.}
We work in a query model built on \emph{Independent Set} (\isq{}) queries,
introduced by \citet{harpeled18}. The \isq{} query takes a subset of vertices
$U \subseteq \vertexset$ and returns whether $U$ is an independent set.
We consider the \emph{Independent Set with Certificate} query, denoted by
$\iscq{}$, which takes as input a subset of vertices $U \subseteq \vertexset$
and returns
\[
\iscq(U) =
\begin{cases}
\edge \sim \uniform(\edgeset(U)), & \text{if } \edgeset(U) \neq \emptyset,\\[4pt]
0, & \text{otherwise},
\end{cases}
\]
where $\edgeset(U)$ denotes the set of edges in the subgraph of $\graph$
induced by $U$, and $\uniform(S)$ denotes the uniform distribution over a
(nonempty) set $S$.
Importantly, the $\iscq{}$ query preserves the structural constraint of a
standard $\isq{}$ query: the algorithm specifies exactly the same vertex
subsets as in the IS model, and the only difference lies in the response,
which returns a uniformly random edge as a certificate whenever the induced
subgraph is nonempty.
\citet{harpeled18} showed that the number of edges in a graph can be estimated using
\[
\widetilde{O}\!\left( \min\!\left\{ \frac{n^{2}}{m}, \sqrt{m} \right\} \right)
\;=\;
\widetilde{O}(n^{2/3})
\]
$\isq{}$ queries, establishing a polynomial separation from local query based algorithms that require $\Theta(n)$ queries in the worst case~\citep {eden_lower_bound18}. In a subsequent work,
\citet{Xichen19} nearly resolved the query complexity of edge estimation using
IS queries by showing that
\[
\widetilde{\Theta}\!\left( \min\!\left\{ \frac{n}{\sqrt{m}}, \sqrt{m} \right\}
\right)
\]
queries are both necessary and sufficient. In particular, this implies a
worst-case complexity of $\widetilde{\Theta}(\sqrt{n})$ $\isq{}$ queries.
%Our main result shows that this natural enrichment of the $\isq{}$ query output
%suffices to obtain a polylogarithmic-query algorithm for edge estimation.
Our main result shows that the natural enrichment of $\isq{}$ queries with random certificates yields an exponential separation from standard $\isq{}$ queries for edge estimation, achieving polylogarithmic query complexity in $m$.
\begin{restatable}[Main result]{theorem}{MainUB}\label{Thm: Main Upper Bound}
There exists a randomized algorithm that, given access to a graph $\graph$
through $\iscq{}$ queries and an approximation parameter
$\approxerror \in (0,1/3)$, outputs an estimate $\approxedgecount$ of the number
of edges $\edgecount$ in $\graph$ such that, with probability at least
$\nicefrac{2}{3}$,
\[
(1-\approxerror)\edgecount
\;\le\;
\approxedgecount
\;\le\;
(1+\approxerror)\edgecount .
\]
The algorithm uses
$O(\approxerror^{-2}\log^{2}\edgecount \cdot \log\log \edgecount)$
$\iscq{}$ queries in expectation.
\end{restatable}
The query complexity of the algorithm is \emph{strictly output-sensitive} in
the sense that it depends only on the number of edges, with no dependence on
$n$, even in lower-order terms.
\Cref{Thm: Main Upper Bound} not only demonstrates the power gained by
augmenting $\isq{}$ access with random certificates over standard $\isq{}$
access, but also underscores the strength of sampling within induced
subgraphs. When edges can only be sampled uniformly at random from the entire
graph, estimating the number of edges requires
$\widetilde{\Theta}(\sqrt{m})$ samples~\citep{Mitzenmacher_Upfal_2005}. In contrast, allowing the algorithm to sample a uniformly random edge from a queried induced subgraph yields an exponential separation from global random edge-sampling models, achieving query complexity $\widetilde{O}(\log^2 m)$.
%In contrast, enabling the ability to
%sample a uniformly random edge from a queried induced subgraph reduces the
%query complexity to polylogarithmic in $m$.
In what follows, we highlight conceptual connections between our results and
query model with several well-studied themes, including group testing, graph
property testing, the power of random certificates, decision versus counting
problems, and conditional sampling.
\paragraph{Group testing perspective.}
The introduction of $\isq{}$ queries in graph algorithms was originally
motivated by ideas from \emph{group testing}. In classical group testing,
introduced by \citet{Dorfman/AnnalsofMathStat/1943/GroupTestingFirst} and
studied extensively since then, the goal is to infer the size of an unknown
set of defective items by querying subsets of a known universe
\citep{AldridgeJS/Book/2019/GroupTestingIT}. A standard group test returns only
a binary result indicating whether the queried subset contains at least one
defective item.
This viewpoint applies naturally to graph edge estimation. Here, edges play
the role of defective items, and the allowed vertex pairs for queries impose
structural constraints on the possible subsets of edges that can be queried.
\citet{harpeled18} introduced both $\isq{}$ and Bipartite Independent Set
($\bisq{}$) queries through this lens. A $\bisq{}$ query takes two disjoint
vertex sets as input and returns whether there exists an edge with one
endpoint in each set.
In the same vein, \citet{BishnuGM23} and \citet{AssadiCK21} introduced the
Edge Emptiness query $\mathrm{EE}$ to study triangle estimation and
connectivity in graphs, respectively. An $\mathrm{EE}$ query specifies a set
of candidate edges and reports whether at least one of them exists in the
graph.
\paragraph{Decision versus counting.}
In a series of seminal works,
\citet{Stockmeyer/stoc/83/ComplxtyApprxCnt,
Stockmeyer/siamcomp/85/ApprxAlgoforSharpP} studied the relationship between
approximate counting and decision access for Boolean formulas. In particular,
they showed that given decision access to satisfiability, one can
approximately count the number of satisfying assignments of any Boolean
formula on $n$ variables using only $O(\log n)$ such queries.
Our work establishes an analogous phenomenon in the setting of graph edge
estimation: augmenting decision-style $\isq{}$ queries with a random
certificate enables efficient approximate counting of edges in a graph.
\paragraph{Conditional sampling.}
Our model also bears a close conceptual relationship to work on property
testing and estimation with conditional sampling access. In distribution
testing, a conditional sampling oracle~\citep{ChakrabortyFischerGoldhirshMatsliah/ITCS/2013/CondSamp,DBLP:journals/siamcomp/CanonneRS15/Conditional} allows the algorithm to specify a
subset of the domain and receive a sample drawn from the distribution
restricted to that subset. Such structured sampling can lead to exponential
savings over standard sampling in tasks such as uniformity testing, support
size estimation, and testing of joint distributions
\citep{CanonneRS14,BhattacharyyaC18,CanonneCKLW21,Bhattacharyya0P24}.
In our setting, the $\iscq{}$ oracle plays an analogous role: the algorithm
specifies a vertex subset, and if the induced subgraph is nonempty, it
receives a uniformly random edge from that subgraph. Thus, $\iscq{}$ can be
viewed as a structured conditional sampling primitive over the edge set of
the graph, where the conditioning is constrained to induced subgraphs. Our
results demonstrate that, as in distribution testing, such conditional
access can fundamentally change the complexity of estimation problems in
graphs.
\paragraph{Power of a random certificate.}
Studying the power of random certificates for approximate counting is an
important question across a variety of query models
\citep{RonTsur/TOCT/2016/PowerofAnExample,
ChakrabortyCKM/ICALP/2023/SATvsNPOracle}. These works show that augmenting
group queries with access to random certificates can lead to significant
reductions in query complexity.
In the graph setting, several queries have been introduced from the viewpoint
of group queries while retaining the structural properties of the graph. Two
notable examples are $\isq{}$ and $\bisq{}$, introduced by
\citet{harpeled18} for estimating the number of edges. Our $\iscq{}$ query fits
naturally within this framework: it preserves the structure of $\isq{}$ while
adding a random certificate. In particular, our algorithm achieves
substantially lower query complexity compared to algorithms using only
$\isq{}$ access, and also improves over the best known
$\widetilde{O}(\log^{5} n)$-query algorithm based on $\bisq{}$ access due to
\citet{addanki_et_al:LIPIcs.ESA.2022.2}.
The main difference between the power of random certificates studied in
\citep{RonTsur/TOCT/2016/PowerofAnExample,
ChakrabortyCKM/ICALP/2023/SATvsNPOracle} and our setting is that our query
oracles are group graph queries that retain the structural properties of the
underlying graph.
\paragraph{Graph property testing.}
Estimating the number of edges in a graph with sublinear queries has been
studied in several access models. Early works focused on local queries such as
degree, neighbour, and pair queries
\citep{feige04,avg_deg_danaron08,eden_lower_bound18,
TetekThorup/STOC/2022/EdgeSamplingFullNbrhood}, which provide fine-grained
access but face inherent limits for edge estimation.
This led to models that incorporate global information. In particular,
augmenting local access with random edge sampling
\citep{Aliakbarpour/Algorithmica/2018/StarSubgraphEdgeSampling,
EdenRS19,BerettaCS/SODA/2026/FasterEdgeEstimation} enables more efficient
estimation. More recently, a finer-grained analysis of edge estimation under
local queries enriched with random edge sampling have provided an understanding of the tradeoffs
between local and global access
\citep{BishnuCDM2025,BishnuCM_degree_distribution_2025,
Chanda_edge_estimation_unknown,BerettaCS/SODA/2026/FasterEdgeEstimation}.
Another line of work studies structured global queries, notably $\isq{}$ and
$\bisq{}$ \citep{harpeled18,Xichen19,addanki_et_al:LIPIcs.ESA.2022.2}. These
queries provide decision-only access over vertex subsets while preserving
graph structure, and have been extended to hypergraphs
\citep{DellLM22,BhattacharyaBGM24,DellLM24}. We discussed their known upper and
lower bounds earlier in the Introduction.
Finally, \citet{AdarHotamLevi/Arxiv/2601.21457} showed that augmenting
$\isq{}$ queries with additional local queries yields a quadratic separation
from using $\isq{}$ queries alone, further illustrating the power of combining
different query primitives. Other than the \bisq{} query, all query models discussed above requires polynomial number of queries in the worst case.
\subsection*{Notations}
Throughout, the notation
$\widetilde{O}(f)$, suppress factors polynomial in
$1/ \approxerror$ and $\log f$. For an integer $k$, we define $[k] = \sbrac{1,\dots,k}$.
We write $a = (1 \pm \approxerror) b$ to indicate that $a$ is a $(1 \pm \approxerror)$-approximation of $b$.
Throughout, we use $G = (\vertexset, \edgeset) $ to denote the input graph with $ \edgecount = \abs{\edgeset} $ number of edges and $ \vertexset = \abs{\vertexset} $ vertices. For a subset $ U \subseteq \vertexset $, let $ \edgecount( U ) $ be the number of edges in the graph induced by the subset $ U $.
% For any pair of subsets $ U' \subseteq U $, we define $ \fractionofedges_{U} ( U' ) = \edgecount ( U' ) / \edgecount( U ) $ to be the fraction of edges in $ U' $ relative to $ U $. We omit the subscript and write $p(U')$ when the corresponding $U$ is clear from the context.
% \gopi{Notation $\fractionofedges_{U} ( U' )$ may either be removed or used consistently with no omission of subscript. I think we can remove it those are not many?}
% Throughout, for all constants $\alpha>0$, the notations
% $\widetilde{O}(n^{\alpha})$, $\widetilde{\Theta}(n^{\alpha})$, and
% $\widetilde{\Omega}(n^{\alpha})$ suppress factors polynomial in $\log n$ and
% $1/\approxerror$; similarly,
% $\widetilde{O}(\log^{\alpha} n)$,
% $\widetilde{\Theta}(\log^{\alpha} n)$, and
% $\widetilde{\Omega}(\log^{\alpha} n)$ suppress factors polynomial in
% $1/\approxerror$ and $\log\log n$.
% For an integer $k$, we define $[k] = \{1,\dots,k\}$. We write
% $a = (1 \pm \approxerror) b$ to indicate that $a$ is a
% $(1 \pm \approxerror)$-approximation of $b$.
% We use $G = (\vertexset,\edgeset)$ to denote the input graph with
% $\edgecount = |\edgeset|$ edges and $|\vertexset|$ vertices. For a subset
% $U \subseteq \vertexset$, let $\edgecount(U)$ be the number of edges in the
% subgraph induced by $U$.
\color{black}
% \section{Introduction}
% We consider the problem of estimating the number of edges in an undirected,
% unweighted, simple graph using a sublinear number of queries. Let
% $\graph = (\vertexset,\edgeset)$ be a graph with
% $\vertexset = [\vertexcount] := \{1,2,\dots,\vertexcount\}$ vertices and
% $\size{\edgeset} = \edgecount$ edges.
% Given query access to $\graph$ and an approximation parameter
% $\approxerror \in (0,1)$, the objective is to output an estimate
% $\approxedgecount$ such that
% \[
% (1-\approxerror)\edgecount
% \;\le\;
% \approxedgecount
% \;\le\;
% (1+\approxerror)\edgecount
% \]
% with a success probability at least $2/3$.
% \paragraph{Query model and our result.}
% We work in a query model built on \emph{Independent Set} (\isq{}) queries,
% introduced by \citet{harpeled18}. The \isq{} query works on a subset of vertices $U \subseteq \vertexset$, and outputs whether $U$ is an independent set. We consider the \emph{Independent Set with Certificate} query that works on a subset of vertices $U \subseteq \vertexset$, denoted by $\iscq{}$.
% The response to such a query is
% \[
% \iscq(U) =
% \begin{cases}
% \edge \sim \uniform(\edgeset(U)), & \text{if } \edgeset(U) \neq \emptyset,\\[4pt]
% 0, & \text{otherwise},
% \end{cases}
% \]
% where $\edgeset(U)$ denotes the set of edges in the subgraph of $\graph$
% induced by $U$, and $\uniform(S)$ denotes the uniform distribution over a
% (nonempty) set $S$.
% Importantly, the $\iscq{}$ query preserves the structural constraint of a standard $\isq{}$ query: it provides access to induced subgraphs as in the IS
% model, and the only difference lies in the response, which returns a uniformly random edge as a certificate whenever the induced subgraph is non-empty.
% \citet{harpeled18} showed that the number of edges in a graph can be estimated using
% \[
% \widetilde{O}\!\left( \min\!\left\{ \frac{n^{2}}{m}, \sqrt{m} \right\} \right)
% \;=\;
% \widetilde{O}(n^{2/3})
% \]
% $\isq{}$ queries, yielding a worst-case improvement over algorithms based on local queries~\citep{eden_lower_bound18}. In subsequent work, \citet{Xichen19} nearly resolved the query complexity of edge estimation using IS queries by showing that
% \[
% \widetilde{\Theta}\!\left( \min\!\left\{ \frac{n}{\sqrt{m}}, \sqrt{m} \right\}
% \right)
% \]
% queries are both necessary and sufficient. In particular, this implies a
% worst-case complexity of $\widetilde{\Theta}(\sqrt{n})$ $\isq{}$ queries.
% Our main result shows that this natural enrichment of the $\isq{}$ query output is
% sufficient to achieve a significant improvement in the query complexity of
% edge estimation.
% \begin{restatable}[Main result]{theorem}{MainUB}\label{Thm: Main Upper Bound}
% There exists a randomized algorithm that, given access to a graph $\graph$
% through $\iscq{}$ queries and an approximation parameter
% $\approxerror \in (0,1)$, outputs an estimate $\approxedgecount$ of the number
% of edges $\edgecount$ in $\graph$ such that, with probability at least
% $\nicefrac{3}{4}$,
% \[
% (1-\approxerror)\edgecount
% \;\le\;
% \approxedgecount
% \;\le\;
% (1+\approxerror)\edgecount .
% \]
% The algorithm uses
% $O(\approxerror^{-2}\log^{3}\edgecount \cdot \log\log \edgecount)$
% $\iscq{}$ queries in expectation.
% \end{restatable}
% \blue{The query complexity of the algorithm is \emph{strictly} output-sensitive in the sense that it depends only on the number of edges, with no dependence on
% $n$, even in lower-order terms.}
% % \blue{The above result highlights the significance of our model from two
% % complementary viewpoints. First, it demonstrates the power gained by augmenting \isq{} access with
% % random certificates. Equipped with \iscq{} queries, one can estimate the
% % number of edges in a graph using significantly fewer queries than what is
% % possible under standard \isq{} access. Second,
% \blue{ \Cref{Thm: Main Upper Bound} not only demonstrates the power gained by augmenting
% \isq{} access with random certificates over standard \isq{} access, but
% also underscores the strength of sampling within induced subgraphs.
% When edges can only be sampled uniformly at random from the entire graph,
% estimating the number of edges requires
% $\widetilde{\Theta}(\sqrt{m})$ samples. In contrast, enabling the ability
% to sample a uniformly random edge from a queried induced subgraph reduces
% the query complexity to polylogarithmic in $m$.
% }
% In what follows, we highlight conceptual connections between our results and query model and several well-studied themes, including group testing, graph property testing, the power of random certificates, decision versus counting problems, and conditional sampling.
% \paragraph{Group testing perspective.}
% The introduction of $\isq{}$ queries in graph algorithms was originally motivated
% by ideas from \emph{group testing}.
% In classical group testing, introduced by \citet{Dorfman/AnnalsofMathStat/1943/GroupTestingFirst}
% and studied extensively since then, the goal is to infer the size of an unknown set of defective items by querying subsets of a known universe, see \cite{AldridgeJS/Book/2019/GroupTestingIT}.
% A standard group test returns only a binary result that indicates whether the queried subset contains at least one defective item.
% This viewpoint from group testing applies naturally to graph edge estimation. Here, edges play the role of defective items, and the allowed vertex pairs for queries impose structural constraints on the possible subset of edges that can be queried. \citet{harpeled18} introduced both $\isq{}$ and Bipartite Independent Set ($\bisq{}$) queries through this lens and studied edge estimation problem. Note that the \bisq{} query takes two disjoint vertex sets as input and returns whether there exists an edge with one endpoint in the first set and the other endpoint in the second set. In the same vein, \citet{BishnuGM23} and \citet{AssadiCK21} introduced the Edge Emptiness\footnote{\citet{AssadiCK/ESA/2021/EmptinessForConnectivity} called it $\mathrm{OR}$ query.} query $\mathrm{EE}$ to study the triangle estimation and connectivity in graphs, respectively. An $\mathrm{EE}$ query specifies a set of candidate edges and reports whether at least one of them exists in the graph.
% % This viewpoint applies naturally to graph edge estimation. Here, edges play the role of defective items, and querying a subset of vertices amounts to testing whether the induced subgraph contains at least
% % one edge. \blue{However, the choice of the vertices that can be queried adds a structural constraints to the subsets of edges that can be queried}. An $\isq{}$ query corresponds precisely to such a \emph{decision-only} group test: it reports whether the queried vertex set induces at least one edge, but provides no further information. \citet{harpeled18} introduced both $\isq{}$ and Bipartite Independent Set ($\bisq{}$) queries through this lens, demonstrating that such global queries
% % can outperform purely local access models.\gopi{BIS may be defined here.}
% % \blue{The \bisq{} query is more powerful than the \isq{} query as $\isq{(U)}$ can be simulated as $\bisq{(U,U)}$. Their work established an exponential separation between the \isq{} and the \bisq{} queries. They showed that $\bisq{}$ queries are
% % exponentially more powerful than $\isq{}$ queries for edge estimation; in
% % particular, the number of edges can be estimated using only a
% % polylogarithmic number of $\bisq{}$ queries. The best known such algorithm,
% % due to \citet{addanki_et_al:LIPIcs.ESA.2022.2}, uses
% % $\widetilde{O}(\log^{5} n)$ $\bisq{}$ queries.}
% % \red{At the same time, subsequent work has revealed inherent limitations of
% % decision-only access.
% % Despite their global nature, $\isq{}$ queries admit an
% % $\Omega(\sqrt{n})$ lower bound for approximate edge counting, implying that
% % they are insufficient for estimating the number of edges using
% % polylogarithmically many queries in general~\citep{Xichen19}.} \gopi{We have already stated this one. We may just conclude that our query follow the same suit as of group testing.}
% %These limitations closely parallel classical phenomena in group testing, where detecting the presence of a defective item is often significantly easier than accurately estimating the total number of defectives.
% \paragraph{Decision versus counting.}
% %These limitations are consistent with the classical gap between decision and counting problems.
% In a series of seminal works,
% \citet{Stockmeyer/stoc/83/ComplxtyApprxCnt,
% Stockmeyer/siamcomp/85/ApprxAlgoforSharpP} studied the relationship between approximate counting and decision access for Boolean formulas. In particular, they showed that given decision access to the satisfiability of Boolean formulas, one can approximately count the number of satisfying assignments of any Boolean formula on $n$ variables using only $O(\log n)$ such queries. Our work establishes an analogous phenomenon in the setting of graph edge estimation: we show that augmenting decision-style $\isq{}$ with a random certificate enables efficient approximate counting of edges in a graph.
% %\blue{From this perspective, the difficulty of edge estimation using $\isq{}$ queriesdoes not stem from the lack of global structure in the queries, but from theabsence of certifying information in the query responses.IS queries can detect the presence of an edge in a queried subset, but theyreveal nothing further that would enable efficient estimation of the total number of edges.} \gopi{The second paragraph may be edited a bit.}
% \paragraph{Conditional sampling.}
% Our model also bears a close conceptual relationship to work on property testing and estimation with conditional sampling access. In distribution testing, a conditional sampling oracle allows the algorithm to specify a subset of the domain and receive a sample drawn from the distribution restricted to that subset. A growing body of work has shown that such structured sampling can lead to exponential savings over standard sampling in tasks such as uniformity testing, support size estimation, and testing of joint distributions~\citep{CanonneRS14,BhattacharyyaC18,CanonneCKLW21,Bhattacharyya0P24}.
% In our setting, the $\iscq{}$ oracle plays an analogous role: the algorithm specifies a vertex subset, and if the induced subgraph is nonempty, it receives a uniformly random edge from that subgraph. Thus, $\iscq{}$ can be viewed as a structured conditional sampling primitive over the edge set of the graph, where the conditioning is constrained to induced subgraphs. Our results demonstrate that, as in distribution testing, such conditional access can fundamentally change the complexity of estimation problems in graphs.
% \paragraph{Power of a random certificate.}
% \color{blue}
% %\gopi{We may start with the power of random certificate in approximate counting and then go to graph setting.}% We should also somewhere say that the power of random certificate has been useful in graph setting when we have local queries augmented with random edge query.}
% Studying the power of random certificates for approximate counting is an
% important question across a variety of query models
% \citep{RonTsur/TOCT/2016/PowerofAnExample,ChakrabortyCKM/ICALP/2023/SATvsNPOracle}.
% These works show that augmenting group queries with access to random
% certificates can lead to significant reductions in query complexity.
% In the graph setting, several queries have been introduced from the
% viewpoint of group queries while retaining the structural properties of
% the graph. Two notable examples are \isq{} and \bisq{}, introduced by
% \citet{harpeled18} for estimating the number of edges. Our \iscq{} query
% fits naturally within this framework: it can be viewed as a graph group
% query augmented with the power of random certificates. In particular,
% our algorithm achieves significantly lower query complexity compared to
% algorithms using only \isq{} access, and also improves over the (best known)
% $\widetilde{O}(\log^{5} n)$-query algorithm based on \bisq{} access due
% to \citet{addanki_et_al:LIPIcs.ESA.2022.2}.
% The main difference between the power of random certificates studied in \citep{RonTsur/TOCT/2016/PowerofAnExample, ChakrabortyCKM/ICALP/2023/SATvsNPOracle} and our setting is that our query oracles are group graph queries that retain the structural properties of the underlying graph.
% \color{black}
% % However
% % %Our work shows that enriching an $\isq{}$ query with a single random
% % certificate, without changing the structure of the queried subsets,
% % fundamentally changes the complexity of edge estimation.
% % The $\iscq{}$ query returns a uniformly random edge whenever the queried
% % induced subgraph is non-empty, which, from a group testing perspective,
% % corresponds to a test that not only detects the presence of a defective item
% % but also reveals a random example.
% % \red{Notably, \citet{harpeled18} introduced $\isq{}$ and, in the same
% % paper, also introduced \emph{Bipartite Independent Set} queries, denoted by
% % $\bisq{}$: a $\bisq{}$ query takes two (possibly overlapping) vertex sets
% % $U,V\subseteq\vertexset$ and returns whether there exists an edge with one
% % endpoint in $U$ and the other in $V$. They showed that $\bisq{}$ queries are
% % exponentially more powerful than $\isq{}$ queries for edge estimation; in
% % particular, the number of edges can be estimated using only a
% % polylogarithmic number of $\bisq{}$ queries. The best known such algorithm,
% % due to \citet{addanki_et_al:LIPIcs.ESA.2022.2}, uses
% % $\widetilde{O}(\log^{5} n)$ $\bisq{}$ queries.}
% % %\gopi{In a foot note we can argue that each IS query can be simulated using $O(\log n)$ IS query with high probability.}
% % \debarshi{Include \cite{ChakrabortyCKM/ICALP/2023/SATvsNPOracle}?}
% % We show that augmenting $\isq{}$ queries with random certificates suffices to
% % \emph{surpass the best known algorihtms using \bisq{} queries} while additionally achieving
% % output-sensitive query complexity. This additional information overcomes the
% % decision-level barriers inherent in both $\isq{}$ and $\bisq{}$ queries,
% % yielding substantially faster algorithms that adapt to the number of edges in
% % the graph. From a group testing viewpoint, this matches the gold standard of
% % algorithms that are both worst-case optimal and instance-adaptive.\gopi{We may use the last line in the group testing paragraph.}
% % Conceptually, this phenomenon is closely related to the ``power of a random
% % example'' studied by \citet{RonTsur/TOCT/2016/PowerofAnExample}.
% \paragraph{Graph property testing}
% \color{blue}
% Estimating the number of edges in a graph with sublinear queries has been
% studied in several access models. Early work focused on local queries such
% as degree, neighbour, and pair queries
% \citep{feige04,avg_deg_danaron08,eden_lower_bound18,
% TetekThorup/STOC/2022/EdgeSamplingFullNbrhood}. These queries provide
% fine-grained access but face inherent limits for edge estimation. This led
% to models that incorporate global information. In particular, augmenting
% local access with random edge sampling
% \citep{Aliakbarpour/Algorithmica/2018/StarSubgraphEdgeSampling,
% EdenRS19,BerettaCS/SODA/2026/FasterEdgeEstimation} enables more efficient
% estimation. Another line of work studies structured global queries, notably \isq{} and
% \bisq{} \citep{harpeled18,Xichen19,addanki_et_al:LIPIcs.ESA.2022.2}, which
% retain graph structure while providing decision access over vertex
% subsets, and have been extended to hypergraphs
% \citep{DellLM22,BhattacharyaBGM24,DellLM24}.
% Our work introduces \iscq{} queries, which augment \isq{} access with random
% certificates. This model combines structured group queries with random
% sampling from induced subgraphs, enabling more efficient edge estimation.
% \color{black}
% \gopi{May be we expand the above paragraph and have no related work section.}
% %\%gopi{We may add a paragraph in the context of graph property testing.}
% %\gopi{Also, we may add a paragraph on support estimation using conditional sampling from a distribution? I will help us to high highlight that having power of sampling from any induced subgraph over the power of sampling over the entire graph. The improvement is from $O(\sqrt{m})$ to $poly(\log m)$.}
% % \paragraph{Conceptual takeaway.}
% % Taken together, our results identify \emph{random certification} as the key
% % ingredient separating hard and easy regimes of global graph queries for edge
% % estimation.
% % While prior work on IS and BIS queries carefully exploits global structure
% % yet remains constrained by decision-only access, our work shows that access
% % to a single random witness per positive IS query is sufficient to enable
% % efficient and output-sensitive approximate counting. \gopi{We may also say that the result and query model are conceptual connection with group testing, graph property testing, power of random certificate, decision vs counting, and conditional sampling. We may add a similar sentence in abstract too.}
% \subsection*{Notations}
% %\debarshi{Do we actually need these suppressions?}
% % Throughout, for all constants $\alpha>0$, the notations
% % $\widetilde{O}(n^{\alpha})$, $\widetilde{\Theta}(n^{\alpha})$, and
% % $\widetilde{\Omega}(n^{\alpha})$ suppress factors polynomial in $\log n$ and
% % $1/ \approxerror$; similarly,
% % $\widetilde{O}(\log^{\alpha} n)$,
% % $\widetilde{\Theta}(\log^{\alpha} n)$, and
% % $\widetilde{\Omega}(\log^{\alpha} n)$ suppress factors polynomial in
% % $1/ \approxerror$ and $\log\log n$.
% Throughout, for all constants $\alpha>0$, the notation
% $\widetilde{O}(\log^{\alpha} n)$, suppress factors polynomial in
% $1/ \approxerror$ and $\log\log n$. For an integer $k$, we define $[k] = \sbrac{1,\dots,k}$.
% We write $a = (1 \pm \approxerror) b$ to indicate that $a$ is a $(1 \pm \approxerror)$-approximation of $b$.
% Throughout, we use $G = (\vertexset, \edgeset) $ to denote the input graph with $ \edgecount = \abs{\edgeset} $ number of edges and $ \vertexset = \abs{\vertexset} $ vertices. For a subset $ U \subseteq \vertexset $, let $ \edgecount( U ) $ be the number of edges in graph induced by $ U $. For any pair of subsets $ U' \subseteq U $, we define $ \fractionofedges_{U} ( U' ) = \edgecount ( U' ) / \edgecount( U ) $ to be the fraction of edges in $ U' $ relative to $ U $. We omit the subscript and write $p(U')$ when the corresponding $U$ is clear from the context.
% \gopi{Notation $\fractionofedges_{U} ( U' )$ may either be removed or used consistently with no omission of subscript. I think we can remove it those are not many?}
% \section{Related Work}
% Estimating the number of edges in a graph using sublinear query access is a
% fundamental problem in graph property testing and sublinear-time algorithms.
% Given oracle access to a graph, the goal is to approximate the number of edges
% while minimizing the number of queries.
% % \red{Do we give a small overview of queries?}
% Early work focused on \emph{local query models}, including degree, neighbour,
% and pair queries. A degree query returns the degree of a vertex, a neighbour query returns the $i$-th neighbour of a vertex, a pair query returns whether two given a vertices have an edge between them. Using degree queries, \citet{feige04} gave a
% $\widetilde{O}(n/\sqrt{m})$-query algorithm that achieves a
% $(2+\epsilon)$-approximation to the number of edges, and showed that
% $\Omega(n^2/m)$ queries are necessary to obtain an approximation factor better
% than $2$.
% Subsequently, \citet{avg_deg_danaron08} showed that augmenting degree access
% with neighbour queries enables a $(1\pm\epsilon)$-approximation using
% $\widetilde{O}(n/\sqrt{m})$ queries, and that this bound is essentially optimal.
% Lower bounds were later extended to richer local access models, including pair
% queries, by \citet{eden_lower_bound18}, and refinements of these bounds were
% studied by \citet{TetekThorup/STOC/2022/EdgeSamplingFullNbrhood}.
% Taken together, these results delineate the inherent limitations of purely
% local access for edge estimation.
% Motivated by these limitations, several works have explored \emph{global}
% query models.
% \citet{Aliakbarpour/Algorithmica/2018/StarSubgraphEdgeSampling} introduced the
% \emph{random edge query}, which returns an edge sampled uniformly at random from
% the graph.
% This model is closely related to weighted sampling frameworks
% \citep{motwani07,BerettaTetek/TALG/2024/BetterSumEstimationViaWeightedSampling}.
% Using degree and random edge queries, \citet{EdenRS19} obtained a
% $\widetilde{O}(\min\{m/n,\,n/\sqrt{m}\})$-query algorithm for edge estimation,
% and \citet{BerettaCS/SODA/2026/FasterEdgeEstimation} showed that this bound is
% essentially optimal. \gopi{Here, random edge query is augmented with local query.}
% More recently, a finer-grained analysis of edge estimation under local queries
% augmented with random edge queries has been developed
% \citep{
% BishnuCDM2025,BishnuCM_degree_distribution_2025,
% Chanda_edge_estimation_unknown,
% BerettaCS/SODA/2026/FasterEdgeEstimation}.
% These works sharpen our understanding of the interplay between local and
% global access in sublinear edge estimation.
% A closely related line of work studies edge estimation under
% \emph{Independent Set} (\isq{}) and \emph{Bipartite Independent Set} (\bisq{})
% queries \citep{harpeled18,Xichen19,addanki_et_al:LIPIcs.ESA.2022.2}.
% These queries provide structured, global, but decision-only access to the
% graph, and have been generalized to hypergraph settings in subsequent work \citep{DellLM22,BhattacharyaBGM24,DellLM24}.
% We refer the reader to the Introduction for a detailed discussion of this literature, including known upper and lower bounds for $\isq{}$ and $\bisq{}$ queries and their relation to our model.
% Finally, \citet{AdarHotamLevi/Arxiv/2601.21457} showed that augmenting \isq{}
% queries with additional local queries yield a quadratic improvement over
% using \isq{} queries alone, highlighting the benefit of combining different
% query primitives.
% \paragraph{Group testing, decision versus counting, and random certificates.}
% Independent Set and Bipartite Independent Set queries are closely related to
% classical problems in \emph{group testing}
% \citep{Dorfman/AnnalsofMathStat/1943/GroupTestingFirst,
% AldridgeJS/Book/2019/GroupTestingIT}.
% In the group testing framework, one seeks to infer properties of an unknown
% set of defective items by querying subsets of a universe, where each query
% returns only a binary outcome indicating whether the subset contains at least
% one defective item.
% From this perspective, edges correspond to defective items, and IS or BIS
% queries correspond to highly structured group tests over the edge set of the
% graph.
% The limitations of IS and BIS queries for edge estimation reflect a broader
% gap between \emph{decision} and \emph{counting}.
% In seminal work,
% \citet{Stockmeyer/stoc/83/ComplxtyApprxCnt,
% Stockmeyer/siamcomp/85/ApprxAlgoforSharpP} showed that approximate counting is
% strictly more powerful than decision access unless the algorithm is augmented
% with additional randomness or sampling capabilities.
% This separation has since appeared in a variety of oracle-based models.
% Ron and Tsur~\cite{RonTsur/TOCT/2016/PowerofAnExample} systematically studied
% the effect of receiving a random example upon a positive query, demonstrating
% that such random certificates can lead to exponential improvements in query
% complexity, even when the allowed query sets are highly structured.
% Related manifestations of this phenomenon have appeared in approximate model
% counting~\citep{ChakrabortyCKM/ICALP/2023/SATvsNPOracle} and in conditional
% sampling models for distribution testing
% \citep{BhattacharyyaC/TOCT/18/SubCONDforJoint,
% CanonneCKLW/SODA/2021/UnifTestSubCond,
% BhattacharyyaCPSS/AAAI/2024/TestingSelfReducibleSamplers}.
% \subsection{Organization of the paper}
% \section{Introduction}
% % \begin{note}[Points to Mention]
% % Contextualize w.r.t.
% % \begin{itemize}
% % \item Group Testing - Defective edges in a network?
% % \item Conditional distribution testing - in particular support size estimation.
% % \item IS and BIS query standard results.
% % \end{itemize}
% % \end{note}
% We consider the problem of estimating the number of edges in an undirected,
% unweighted, simple graph using a sublinear number of queries. Let
% $\graph = (\vertexset,\edgeset)$ be a graph with
% $\vertexset = [\vertexcount] := \left\{1, 2, \dots, n \right\}$ vertices and
% $\size{\edgeset} = \edgecount$ edges.
% Given query access to $\graph$ and an approximation parameter $\approxerror \in (0,1)$, the objective is to output an estimate
% $\approxedgecount$ such that
% $(1-\approxerror)\edgecount \le \approxedgecount \le (1+\approxerror)\edgecount$
% with probability at least $\nicefrac{3}{4}$.
% We work in the following query model.
% For any subset of vertices $U \subseteq \vertexset$, the algorithm may make
% an $\iscq{}$ query, which returns
% \[
% \iscq{(U)} =
% \begin{cases}
% \edge \sim \uniform(\edgeset(U)), & \text{if } \edgeset(U) \neq \emptyset, \\[4pt]
% 0, & \text{otherwise},
% \end{cases}
% \]
% where $\edgeset(U)$ denotes the set of edges in the subgraph of $\graph$
% induced by $U$, and $\uniform(S)$ denotes the uniform distribution over a
% (nonempty) set $S$.
% Our main result in this paper is a randomized algorithm that estimates the
% number of edges in any graph $G$ using
% $O(\approxerror^{-2}\log^{3} m \cdot \log \log m)$ many $\iscq{}$ queries to $G$.
% \begin{restatable}[Main result]{theorem}{MainUB}\label{Thm: Main Upper Bound}
% There exists a randomized algorithm that, given access to a graph $\graph$
% through $\iscq{}$ queries and an approximation parameter
% $\approxerror \in (0,1)$, outputs an estimate $\approxedgecount$ of the number
% of edges $\edgecount$ in $\graph$ such that, with probability at least
% $\nicefrac{2}{3}$,
% \[
% (1-\approxerror)\edgecount \le \approxedgecount \le (1+\approxerror)\edgecount .
% \]
% \end{restatable}
%section{Preliminaries}
% \begin{definition}[\iscq{}]\label{Def: Independent Set Certificate Query}
% Given a subset of vertices $U \subseteq \vertexset$, $\iscq{}$ returns:
% \begin{align*}
% \iscq{(U)}= \begin{cases}
% \edge \sim \uniform{(\edgeset(U))} &\text{If } \edgeset(U) \neq \emptyset \\
% 0 &\text{Otherwise}
% \end{cases}
% \end{align*}
%\end{definition}
% \subsection{Important Results}
% \begin{restatable}[Graph Sparsification for 2 colors]{lemma}{}\label{lemma: graph sparsification for k = 2}
% For a given $ \confidence \in (0, 1)$, there exists an absolute constant $ \constant $ such that the following holds. Let $ \graph = ( \vertexset, \edgeset ) $ be a graph with $ \vertexcount $ vertices and $ \edgecount $ edges. Let $ U_1 $ and $ U_2 $ be a uniformly random partition of $ \vertexset $. Then,
% \begin{equation*}
% \Pr\tbrac{ \abs{ \frac{ \edgecount }{ 2 } - \sum_{ i = 1 }^{ 2 } \edgecount( U_i ) } \geq \constant \sqrt{ \edgecount } \log \fbrac{\frac{ \edgecount }{ \confidence }} } \leq \confidence
% \end{equation*}
% \end{restatable}
\section{Technical overview}
We give a high-level description of our algorithm for estimating the number of edges in an undirected graph using Independent Set with Certificate (\iscq{}) queries. Conceptually, our approach converts approximate counting into a sequence of controlled decisions and sampling steps that gradually reduce the graph while preserving a constant fraction of its edges.
A central conceptual contribution is the use of \emph{approximation preserving subsets} and \emph{approximation preserving sequences}, which formalise how much of the graph’s edge mass is retained at each stage and guide both the algorithm and its analysis. We leverage graph sparsification techniques to enable iterative construction of approximation-preserving sequences on graphs using \iscq{} queries.
\paragraph{Outline of the Algorithm.} At a high level, our algorithm maintains a nested sequence of vertex subsets \[ \vertexset = U_0 \supseteq U_1 \supseteq \cdots \supseteq U_\kappa, \] where each $U_j$ contains a constant fraction of the edges of $U_{j-1}$. The algorithm alternates between two modes:
\begin{itemize}
\item If the current induced subgraph is ``small'' (that is, has fewer than $\threshold$ edges), we estimate its edge count directly.
\item Otherwise, we construct a smaller approximation preserving subset of the graph by randomly partitioning the vertex set and selecting a part that provably retains a constant fraction of the edges.
\end{itemize}
This iterative reduction of the graph continues until we reach a subset $U_\kappa$ whose edge count is small enough to estimate accurately; we set the threshold for $U_\kappa$ to be a sufficiently large constant. The final estimate of $m$ is then reconstructed by multiplying the relative edge losses across all steps.
\paragraph{Where the $\widetilde{O}(\log^2 m)$ query complexity comes from.} A key novelty of our approach is that all complexity bounds are driven by $\log m$ rather than $\log n$. This is achieved through three ideas that are not present in prior $\isq{}$ or $\bisq{}$ based algorithms:
\begin{itemize}
\item \emph{Random certificates enable output-sensitive sparsification:} Unlike decision-only $\isq{}$ or $\bisq{}$ queries, ISC queries allow us to estimate how many edges of a subset lie in a randomly chosen part, which lets us iteratively construct smaller induced subgraphs to work on while tracking edge mass locally.
\item \emph{Approximation Preserving Sequence:} Instead of estimating $m$ directly, we construct a chain of subsets $U_0 \supseteq U_1 \supseteq \cdots \supseteq U_\kappa$ such that each $U_j$ retains a constant fraction of edges of $U_{j-1}$.
Crucially, we establish $\kappa$ to be $O(\log m)$ in expectation, independent of $n$. We also bound the number of queries required to construct this sequence as $O(\kappa^2)$ and bound its expectation.
\item \emph{Local multiplicative reconstruction:} We only need to estimate the ratios $p_j = \frac{m(U_j)}{m(U_{j-1})}$, each of which is $\Omega(1)$ due to the $U_0,\ldots,U_\kappa$ being an approximation preserving subsequence. This keeps the per-level query complexity small and makes the overall cost in expectation $\widetilde{O}(\kappa^2) = \widetilde{O}(\log^2 m)$.
\end{itemize}
\paragraph{Estimating Small Edge Counts.}
Fix a threshold $\threshold$. One can design an algorithm that makes $O(\sqrt{\tau})$ \iscq{} queries and guarantees the following with desired probability about a given subset $U \subseteq V$. If $m(U)\leq \tau$, then the algorithm reports an approximation of $m(U)$. Moreover, if $m(U) \geq 8\tau$, the algorithm reports that $m(U)$ is large. The proof is based on the birthday paradox. The details can be found in \Cref{sec: threshold checking} and Appendix \ref{appendix: threshold checking}.
\paragraph{Graph Sparsification.}
The idea here is to leverage graph sparsification to construct a smaller subset of vertices to work on at each step. We partition $U$ uniformly at random into two subsets $U^1$ and $U^2$. A graph sparsification lemma, along the result of \citet{harpeled18}, shows that with desired probability,
\[
\left| \frac{\edgecount(U)}{2} - \sum_{i=1}^{2} \edgecount(U^i) \right| \leq c \sqrt{\edgecount(U)} \log \edgecount(U).
\]
We provide the proof in \Cref{sec: sparsification}. Intuitively, this means that the total number of edges inside $U^1$ and $U^2$ is close to half of $\edgecount(U)$, up to a lower-order error term. As long as $\edgecount(U)$ is sufficiently large, this implies that at least one of $U^1$ or $U^2$ must contain a constant fraction of the edges of $U$. More precisely, there exists an absolute constant $\thresholdconst$ such that if $\edgecount(U)\ge \thresholdconst$, then with probability at least $1-1/\edgecount(U)$, at least one of $m(U^1)/m(U)$ and $m(U^2)/m(U)$ is at least $1/8$, and both $m(U^1)/m(U)$ and $m(U^2)/m(U)$ are at most $3/4$.
\paragraph{Choosing the Good Part.} After partitioning $U$ into $U^1$ and $U^2$, we must identify which part retains an appropriate fraction of the edges. Note that this step is fundamentally enabled by the $\iscq{}$ queries. The idea is to take a random sample of the edges in $U$, and check whether it is preserved in the induced subgraph of either $U^1$ or $U^2$.
To do this, we draw random edges from $U$ via \iscq{} queries. Let $\widehat{p}(U^i)$ be the fraction of sampled edges that fall inside $U^i$, for $i \in [2]$. If $\widehat{p}(U^i) \ge 1/16$, we select $U^i$ as the next subset; otherwise, we pick an arbitrary part. A Chernoff bound guarantees that, with desired probability, we correctly identify a subset satisfying
\[
\frac{1}{32}\leq \frac{\edgecount(U^i)}{\edgecount(U)} \le \frac{3}{4}.
\]
Such subsets are called \emph{approximation preserving subsets}. The upper bound of $3/4$ is useful to bound the final query complexity and the lower bound of $1/32$ is useful for guaranteeing the quality of the estimator.
\paragraph{Constructing an Approximation Preserving Sequence.} Starting with $U_0=\vertexset$, we iterate:
\begin{enumerate}
\item Run threshold checking on $U_j$.
\item If $\edgecount(U_j) < \threshold$, stop and set $U_\kappa = U_j$.
\item Otherwise, sparsify $U_j$ and select an approximation preserving subset $U_{j+1}$.
\end{enumerate}
This yields a sequence \[ \vertexset = U_0 \supseteq U_1 \supseteq \cdots \supseteq U_\kappa, \] where, with probability at least $4/5$, each $U_j$ satisfies
\[
\frac{1}{32}\leq \frac{\edgecount(U_j)}{\edgecount(U_{j-1})} \le \frac{3}{4}.
\]
\paragraph{Final Estimation.} For each $j\in [\iterations]$, define
\[
\fractionofedges_j = \frac{\edgecount(U_j)}{\edgecount(U_{j-1})}.
\]
Then we have the identity
\[
\edgecount = \frac{\edgecount(U_\kappa)}{\prod_{j=1}^{\kappa} \fractionofedges_j}.
\]
Instead of estimating each $m(U_j)$ from scratch, we estimate only the ratios $p_j$. Because each $p_j = \Omega(1)$ by construction of the approximation preserving sequence, we can estimate each $p_j$ using small number of queries, independent of $n$ or $m(U_j)$. This yields an estimate $\widehat{\fractionofedges}_j$ satisfying the following with desired probability.
\[
\widehat{\fractionofedges}_j = (1\pm \approxerror/O( \sqrt{\kappa}))\,\fractionofedges_j
\]
Note that the entire approximation preserving subsequence is constructed a priori. Hence, we know $\kappa$ during the final estimation phase that we use to fix the approximation parameters appropriately for the multiplicative estimation for each $p_j, j \in [\kappa]$.
Let $\approxedgecount(U_\kappa)$ be the estimate obtained via the birthday-based estimator. Our final output is \[ \widehat{\edgecount} = \frac{\approxedgecount(U_\kappa)} {\prod_{j=1}^{\kappa} \widehat{\fractionofedges}_j}. \] A careful multiplicative error analysis shows that this is a $(1\pm \approxerror)$-approximation to $m$.
\paragraph{Proof Roadmap.} The correctness proof proceeds in the following stages:
\begin{itemize}
\item \emph{Threshold correctness:} We show that the collision-based test reliably distinguishes between $\edgecount(U)<\threshold$ and $\edgecount(U)\ge 8\threshold$, and that the birthday-paradox estimator is accurate when $\edgecount(U)<\threshold$.
\item \emph{Constructing approximation preserving subsequence:} {While we decrease the failure probability of choosing a good subset at each iteration, the probability of failure of the sparsification step depends on the number of edges remaining. This necessitates a careful analysis of the failure probability of the sparsification step and of our threshold-checking mechanisms to establish the correctness of our algorithm and query complexity bounds.}
\item \emph{Multiplicative reconstruction:} Given our approximation preserving subsequence, we estimate each $p_j$ with appropriately scaled approximation error. Then, we bound the cumulative error from estimating all $\fractionofedges_j$ and show that the product estimator preserves a global $(1\pm \approxerror)$ approximation.
\item \emph{Complexity analysis:} We show that $\mathbb{E}[\kappa] = O(\log m)$ and that the total query complexity is dominated by estimating the $p_j$'s, yielding an overall expected complexity of $\widetilde{O}(\log^2 m/\varepsilon^2)$.
\end{itemize}
\paragraph{Outline}
We begin in \Cref{sec: threshold checking}, where we present the threshold-checking algorithm and show how to estimate the edge count when the number of edges is below a given threshold. In \Cref{sec: sparsification}, we describe the algorithm for constructing an approximation-preserving subsequence. In \Cref{section: approx subset}, we establish how graph sparsification yields an approximation-preserving subset, and use this result in \Cref{section: approximate sequence} to construct an approximation-preserving subsequence. Finally, in \Cref{section: full algorithm}, we present the complete algorithm for edge estimation.
% \section{Technical Section}
% In this section, we establish an algorithm for approximating the number of edges $ \edgecount $ of an undirected graph using only \iscq{} queries. We can construct a sublinear approximation algorithm that outperforms the existing algorithms in the \isq{} and \bisq{} regimes.
\section{Threshold checking}\label{sec: threshold checking}
Let a fixed threshold ($ \threshold \geq 1$) be given. We consider the following threshold-checking problem: given a subset of vertices $ U \subseteq \vertexset $, determine whether $ \edgecount( U ) < \threshold $. If this condition holds, the goal is to estimate the $ (1 \pm \approxerror ) $ approximation of $ \edgecount( U )$.
To address this task, we present two algorithms \iscollision{} and \birthdayestimate{} based on the birthday paradox techniques \citep[Chapter 5]{Mitzenmacher_Upfal_2005}. \iscollision{} distinguishes between the cases $ \edgecount( U ) < \threshold $ or $ \edgecount( U ) \geq 8\threshold $ (\Cref{lemma: collision detection}). \birthdayestimate{} outputs an $ (1 \pm \approxerror) $ approximation of $ \edgecount( U ) $, if $ \edgecount( U ) < \threshold $ (\Cref{lemma: birthdayestimate}). The corresponding guarantees are stated below. For details, refer to Appendix~\ref{appendix: threshold checking}.
% \begin{algorithm}
% \caption{\iscollision($U, \threshold, \confidence$)}\label{alg:collision}
% \begin{algorithmic}[1]
% \Require \iscq{} access to the subset $U \subseteq \vertexset$ and a threshold $\threshold$
% \Ensure $0$ or $1$
% \For{$i \gets 1$ to $O(\log(\confidence^{-1}))$} \label{alg: collision line 1}
% \State Uniformly and independently sample $\preresamplesize = \sqrt{2\threshold}$ random edges from $U$
% \label{alg: collision line 2}
% \If{there is a collision}
% \State $\collision_i \gets 1$
% \Else
% \State $\collision_i \gets 0$
% \EndIf
% \EndFor
% \If{more than half of $\collision_i$ are $1$}
% \State \Return $1$
% \Else
% \State \Return $0$
% \EndIf
% \end{algorithmic}
% \end{algorithm}
\begin{restatable}[Correctness of \iscollision{}]{lemma}{isollison}
For a subset $ U \subseteq \vertexset $ and a threshold $ \threshold \geq 1 $, the following statements hold for \iscollision{} with probability at least $ 1 - \confidence$:\label{lemma: collision detection}
\begin{enumerate}
\item[(i)] If $ \edgecount( U ) \leq \threshold $, then $\iscollision{}$ outputs $1$.
\item[(ii)] If $ \edgecount( U ) \geq 8 \threshold $, then $\iscollision{}$ outputs $0$.
\end{enumerate}
The number of \iscq{} queries made is $ O( \sqrt{ \threshold } \log ( \confidence^{ -1} ))$
% The number of \iscq{} queries made is $ O( \sqrt{ \threshold } \log ( \nicefrac{ 1 }{ \confidence } ) ) $.
\end{restatable}
% \begin{proof}
% Fix an iteration $i$ of \iscollision{} (Line~\ref{alg: collision line 1}).
% The probability that no collision occurs in the $i$-th iteration is
% \begin{align*}
% \Pr[\collision_i = 0]
% = \prod_{j=0}^{\preresamplesize - 1}
% \left(1 - \frac{j}{\edgecount(U)}\right).
% \end{align*}
% \noindent If $\edgecount(U) \leq \threshold$,
% We upper bound the probability of no collision as follows:
% \begin{align*}
% \Pr[\collision_i = 0]
% &\leq \exp\left(
% - \sum_{j=0}^{\preresamplesize - 1}
% \frac{j}{\edgecount(U)}
% \right)
% && \text{($1 + x \leq e^{x}$ for all $x \in \R$)} \\
% &= \exp\left(
% - \frac{\preresamplesize(\preresamplesize - 1)}{2 \edgecount(U)}
% \right) \\
% &\leq \exp\left(
% - \frac{\preresamplesize(\preresamplesize - 1)}{2 \threshold}
% \right)
% && \text{(since $\edgecount(U) \leq \threshold$)} \\
% &\leq \frac{1}{e}
% && \text{(using $\preresamplesize = \sqrt{2\threshold}$)}.
% \end{align*}
% \noindent If $\edgecount(U) \geq 8\threshold$, we lower bound the probability of no collision:
% \begin{align*}
% \Pr[\collision_i = 0]
% &\geq \exp\left(
% - \sum_{j=0}^{\preresamplesize - 1}
% \left(
% \frac{j}{\edgecount(U)} +
% \frac{j^2}{\edgecount(U)^2}
% \right)
% \right)
% && \text{($1 - x \geq e^{-x - x^2}$ for $x \in [0, 2/3]$)} \\
% &\geq \exp\left(
% - \frac{\preresamplesize(\preresamplesize - 1)}{16 \threshold}
% + \Theta\!\left(\frac{1}{\sqrt{\threshold}}\right)
% \right) \\
% &\geq \frac{1}{2}
% && \text{(using $\preresamplesize = \sqrt{2\threshold}$)}.
% \end{align*}
% \noindent If $ \edgecount( U ) \leq \threshold $, $ \Pr[ \collision_i = 0 ] \leq \frac{ 1 }{ e } $. \iscollision{} outputs $0$ when more than half of $ \collision_i = 0 $. Over $ \log ( \confidence^{-1} ) $ parallel runs, using the upper tail estimate of the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), the probability that \iscollision{} outputs $0$ is at most $ \exp ( - \log ( \confidence^{-1} ) ) = \confidence $. Similarly, if $ \edgecount( U ) \geq 8\threshold $, the probability that \iscollision{} outputs $1$ is at most $ \confidence $.
% \end{proof}
Suppose that a collision is detected, i.e., \iscollision{} outputs $1$. We can obtain a
$ (1 \pm \approxerror)$-approximation of $\edgecount(U)$ using additional
\iscq{} queries. This estimation procedure is implemented in \birthdayestimate{}.
% \begin{algorithm}
% \caption{\birthdayestimate{}($ U, \threshold, \approxerror $)}\label{alg: birthdayestimate}
% \begin{algorithmic}[1]
% \Require \iscq{} access to the subset $ U \subseteq \vertexset $ and a threshold $ \threshold
% \geq 1 $.
% \Ensure An estimate of $ \edgecount( U ) $
% \State Uniformly and independently sample $ \Theta \fbrac{ \sqrt { \threshold } / \approxerror } $ random edges from $ U $.
% \State Let $ \resamplesize = \Theta\fbrac{ \nicefrac{ \sqrt{\threshold} }{ \approxerror } } $ be the number of samples drawn and $ \collisioncount $ is the number of collisions.
% \State \Return $\binom{\resamplesize}{2}/\collisioncount $.
% \end{algorithmic}
% \end{algorithm}
\begin{restatable}[Correctness of \birthdayestimate{}]{lemma}{birth}\label{lemma: birthdayestimate}
For a subset $ U \subseteq \vertexset $, threshold $ \threshold \geq 1 $ and approximation error $ \approxerror > 0 $. If $ \edgecount( U ) \leq \threshold $, then \birthdayestimate{} outputs an $ (1 \pm \approxerror) $ estimate of $ \edgecount( U ) $ with probability at least $\nicefrac{ 15 }{ 16 }$. The number of \iscq{} query made is $ \Theta\fbrac{ \approxerror^{-1} \sqrt{\threshold} } $.
\end{restatable}
% \begin{proof}
% Let $\sbrac{e_1, e_2, \ldots, e_{\resamplesize}}$ be the sampled edges. or $1 \le i < j \le \resamplesize$, define the indicator random variable $X_{ij}$ to be $1$ if $e_i = e_j$ and $0$ otherwise . Then,
% $$\Pr[X_{ij}=1] = \frac{1}{ \edgecount( U ) } \text{ and } \E[X_{ij}] = \frac{1}{ \edgecount( U ) }$$
% Define the total number of collisions by $ X = \sum_{1 \le i < j \le \resamplesize} X_{ij} $. By the linearity of expectation
% \begin{align*}
% \E[X] = \binom{\resamplesize}{2} \frac{1}{ \edgecount( U ) }
% %= O\left(\frac{\resamplesize^{2}}{\edgecount}\right).
% \end{align*}
% Since the random variables $ X_{ij} $s are not independent. We will bound the variance using
% \begin{align}\label{eq: variance equation}
% \Var(X) &= \sum_{i < j } \Var(X_{ij}) + 2\sum_{i < j, k < \ell} \mathrm{Cov}(X_{ij},X_{k\ell})
% \end{align}
% Let us bound the first term $ \Var(X_{ij}) $ in~\eqref{eq: variance equation}
% \begin{align*}
% \sum_{i < j } \Var(X_{ij}) \leq \sum_{ i < j } \E[X_{ij}^2] \leq \binom{\resamplesize}{2} \frac{1}{ \edgecount( U ) } = \E[X]
% \end{align*}
% Next, consider the covariance terms in~\ref{eq: variance equation}. If $X_{ij}$ and $X_{k\ell}$ correspond to disjoint pairs, then they are independent and the covariance is zero. If they overlap in exactly one index (share exactly one index), then
% \begin{align*}
% \mathrm{Cov}(X_{ij},X_{j\ell}) &= \E[X_{ij}X_{j\ell}] - \E[X_{ij}]\E[X_{j\ell}]\\
% &= \Pr[X_{ij} = X_{j\ell} = 1] - \frac{1}{ \edgecount( U )^2}\\
% &= \Pr[e_i = e_j = e_\ell] - \frac{1}{ \edgecount( U )^2 } \\
% &= 0
% \end{align*}
% So, the bound on the variance is $ \Var(X) = \sum_{i < j } \Var(X_{ij}) \leq \E[X] $. Using Chebyshev’s Inequality (Lemma~\ref{lemma: chebyshev}), we have
% \begin{align}
% \nonumber\Pr[\abs{X - \E[X]} > \frac{\approxerror}{3} \E[X]] &\leq \frac{9\Var(X)}{\approxerror^2 (\E[X])^2}\\
% \nonumber&\leq \frac{9}{\approxerror^2\E[X]} &\text{$\fbrac{\Var(X) \leq \E[X]}$}\\
% \nonumber&= \frac{ 18\edgecount( U ) }{ \approxerror^2\resamplesize(\resamplesize - 1) } &\text{$\fbrac{ \E[X] = \binom{\resamplesize}{2} \frac{ 1 }{ \edgecount( U ) } }$}\\
% &\leq \frac{1}{ 16 } &\text{($ \edgecount_U \leq \threshold $ and $\resamplesize = \Theta ( \sqrt{ \threshold } /{ \approxerror } )$)}\label{ineq: collision bound}
% \end{align}
% To complete the proof, we need to show with probability at least $ 15/ 16 $
% \begin{align*}
% \fbrac{1 - \approxerror} \edgecount( U ) \leq \binom{\resamplesize}{2} \frac{1}{X} \leq \edgecount( U ) \fbrac{1 + \approxerror}
% \end{align*}
% From~\eqref{ineq: collision bound}, with probability at least $ 15/ 16 $
% \begin{align*}
% \fbrac{1 - \frac{\approxerror}{3}}\E\tbrac{X} \leq X \leq \fbrac{1 + \frac{\approxerror}{3}} \E\tbrac{X}
% \end{align*}
% This implies
% \begin{align*}
% \Rightarrow \frac{ \edgecount( U ) }{\fbrac{1 + \frac{\approxerror}{3}}}\leq \binom{\resamplesize}{2} \frac{1}{X} \leq \frac{\edgecount( U ) }{\fbrac{1 - \frac{\approxerror}{3}}}
% \end{align*}
% For any $ \approxerror \in (0, 1) $
% \begin{equation*}
% 1 - \approxerror \leq \fbrac{1 + \frac{\approxerror}{3}}^{-1} \text{ and } \fbrac{1 - \frac{\approxerror}{3}}^{-1} \leq 1 + \approxerror
% \end{equation*}
% Since $X$ is the number of collisions. we conclude that
% \begin{align*}
% \fbrac{1 - \approxerror} \edgecount( U ) \leq \binom{\resamplesize}{2} \frac{1}{\collisioncount} \leq \edgecount( U ) \fbrac{1 + \approxerror}
% \end{align*}
% which completes the proof.
% \end{proof}
\section{Sparsification}
\label{sec: sparsification}
In this section, we will describe a procedure to generate an approximation preserving sequence (formally defined below) $U_0, \ldots, U_\kappa$ with probability at least $7/8$. Moreover, the expected query complexity of the procedure is $O(\log^2 m)$.
\begin{definition}[Approximation preserving subset]
For $U,U' \subset V$ such that $U' \subseteq U$, we say that $U'$ is an \emph{approximation preserving subset} of $U$ if
\[
\frac{1}{32} \le \frac{m(U')}{m(U)}\le \frac{3}{4}.\]
\end{definition}
\begin{definition}[Approximation preserving sequence]
For a sequence $U_0, U_1, \ldots, U_\kappa \subseteq V$ with
$U_0 \supseteq U_1 \supseteq \ldots \supseteq U_\iterations$,
We say that $U_0, U_1, \ldots, U_\kappa$ is an \emph{approximation preserving sequence} if $U_{i-1}$ is an approximation preserving subset of $U_i$
for all $i \in [\kappa]$.%~\footnote{Note that a singleton set $U$ is trivially approximation preserving sequence and any subsequence of an approximation preserving sequence is also an approximation preserving sequence.}
\end{definition}
Formally, we will prove the following lemma in this section.
\begin{lemma}[Approximate sequence lemma] \label{lem: sequence}%label{lemma: subsequence}
There exists an algorithm that outputs an approximation preserving sequence $ U_0, U_1, \ldots, U_\iterations $ with probability at least $ \nicefrac{4}{5}$. The expected value of $\kappa$ is $O(\log \edgecount)$, and the expected number of \iscq{} queries made is $ O( \log^2 m)$.
\end{lemma}
% \gopi{The statement may be a bit tight.}
To prove the above lemma, we start with $U_0 = V$, and find $ U_i $ iteratively. Informally speaking, we stop with $U_\kappa$ when $m(U_\kappa)$ is \emph{small enough}. In \Cref{section: approx subset}, we discuss a procedure (\addestimation{} corresponding to \Cref{Lemma: AddEst + Sparsifier}) that generates a $U'$ which is an approximation preserving subset of $ U $, with specified probability, for any given $U$ as long as $m(U)$ is \emph{large}. In \Cref{section: approximate sequence}, we describe the procedure (\subsequence{} corresponding to \Cref{lemma: subsequence}) that uses \iscollision{} and \addestimation{} iteratively to generate an approximation preserving subsequence $ U_0, \ldots, U_\iterations $. Essentially, \iscollision{} is used to determine whether the number of edges in the current subset $U_i$ is small or large and \addestimation{} is used to generate a good subset as long as $m(U_i)$ is large.
\subsection{Procedure to find approximate preserving subset}\label{section: approx subset}
In this section, we describe the procedure \addestimation{} (see \Cref{alg: addestimation}) for finding the approximation preserving subset $ U' \subseteq U $, for any given $ U $ and prove its guarantee.
\begin{algorithm}
\caption{\addestimation($ U, \confidence $)}\label{alg: addestimation}
\begin{algorithmic}[1]
\Require \iscq{} query access to $G$ and a confidence parameter $\delta \in (0,1)$.
\Ensure A subset $U' \subseteq U$.
\State Uniformly and independently partition $ U $ into $ U^1 $ and $ U^2 $.\label{addestimation: line 1}
\State Uniformly and independently sample $ O( \log ( \confidence^{-1} )) $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges. \label{addestimation: line 2}
\State Compute $ \approxfractionofedges ( U^i ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U^i } $ for $ i \in \{1, 2\} $. \label{addestimation: line 3}
\If{There exists $ i \in \{1, 2\} $ such that $ \approxfractionofedges( U^i ) \geq \frac{ 1 }{ 16 } $} \label{addestimation: line 4}
\State \Return $U^i$ as $U'$ \label{addestimation: line 5}
\Else
\State \Return $U^1$ or $U^2$ arbitrarily as $U'$. \label{addestimation: line 7}
\EndIf
\end{algorithmic}
\end{algorithm}
Formally, we will prove the following.
\begin{lemma}[Approximate subset lemma] \label{Lemma: AddEst + Sparsifier}
Let $\thresholdconst$ be a suitably large constant. Consider \addestimation{} that takes $U \subseteq V$ as input and produces $U' \subseteq U$ as the output.
\begin{itemize}
\item[(i)] $\E[\edgecount(U')] \leq \edgecount(U)/2$;
\item[(ii)] %If $m(U) \geq \zeta$, then $\frac{m(U')}{ m(U)}\geq \frac 1 {32}$ with probability $1-\delta-1/m(U)$.
If $m(U) \geq \thresholdconst$, with probability $1-\delta-1/m(U)$, $U'$ is an approximation preserving subset of $U$.%, that is, $ \frac{1}{32} \leq \frac{m(U')}{ m(U)}\leq \frac 3 4$.
%\item[(iii)] If $m(U) \geq \zeta$, with probability $1-1/m(U)$, $\frac{m(U')}{ m(U)}\leq \frac 3 {4}$.
\end{itemize}
\end{lemma}
The construction of an approximation-preserving subset $ U' $ proceeds in two steps. First, we sparsify the graph by randomly partitioning the set $ U $ into $U^1$ and $U^2$ (see \Cref{lemma: graph sparsification for k = 2}). When $\edgecount( U )$ is large enough, this partition guarantees the existence of $U^i$ such that $\edgecount( U^i )$ is a constant fraction of $\edgecount( U )$ (\Cref{lemma: spasification k = 2}). In the second step, we identify such an approximation-preserving subset from $U^1$ and $U^2$ (\Cref{lemma: addestimation}). Combining
these results yields the proof of \Cref{Lemma: AddEst + Sparsifier}.
We prove the following variant of the graph sparsification result of Beame et al.~\citep{harpeled18}.
\begin{restatable}[Graph Sparsification]{lemma}{graphsparsification}\label{lemma: graph sparsification for k = 2}
For a given $ \confidence \in (0, 1)$, there exists an absolute constant $ \constant $ such that the following holds. Let $ \graph = ( \vertexset, \edgeset ) $ be a graph with $ \vertexcount $ vertices and $ \edgecount $ edges. Let $ U^1 $ and $ U^2 $ be a uniformly random partition of $ \vertexset $. Then,
\begin{itemize}
\item[(i)] $\E[ \edgecount(U^1) + \edgecount(U^2)] = \edgecount(U)/2$;
\item[(ii)]
$\Pr\tbrac{ \abs{ \edgecount/2 - \sum_{ i = 1 }^{ 2 } \edgecount( U^i ) } \geq \constant \sqrt{ \edgecount } \log \fbrac{ \edgecount / \confidence} } \leq \confidence$.
\end{itemize}
\end{restatable}
We will need the following lemma to prove the above graph sparsification lemma.
\begin{lemma}\label{lemma: support for sparsification}
Let $ A $ be a set of $ \ell $ elements. For every element $ t \in A $, one assigns a color independently and uniformly from $ \sbrac{1, 2} $. Let $ C^{(i)} $ be the number of vertices in $ A $ with color $i$, for any $ i \in \sbrac{ 1, 2 }$. Then for $ \confidence \in (0, 1) $, we have
\begin{equation*}
\Pr \tbrac{ \abs{ C^{(1)} - C^{(2)} } \geq \sqrt{ { 6\ell\log (4\confidence^{-1}) } }} \leq \confidence \text{ and } \E\tbrac{ \abs{ C^{(1)} - C^{(2)} } } \leq \sqrt{ \ell }
\end{equation*}
\end{lemma}
\begin{proof}
For $ t \in A$, let $ X_t $ be the indicator random variable that is $ 1 $ if $ t $ is assigned color $ 1 $ and $ 0 $ otherwise. Then, $ C^{(1)} = \sum_{t \in A} X_t $ and $ \E \tbrac{ C^{(1)} } = \ell/2$. Using Chernoff's inequality (Lemma~\ref{lemma: Multiplicative Chernoff Bound}) with $ \approxerror = \sqrt{ (6 / \ell) \log (4 / \confidence) }$, we have
\begin{equation}\label{ineq: color 1}
\Pr \tbrac{ \abs{ C^{(1)} - \frac{ \ell }{ 2 } } > \frac{\approxerror \ell }{ 2 } } \leq 2 \exp \fbrac{ \frac{ 6 \log (4\confidence^{-1})}{ \ell } \cdot \frac{ \ell }{ 6 } } \leq \frac{ \confidence }{ 2 }
\end{equation}
Similarly, we can have the same bound for color $2$.
\begin{equation}\label{ineq: color 2}
\Pr \tbrac{ \abs{ C^{(2)} - \frac{ \ell }{ 2 } } > \frac{\approxerror \ell }{ 2 } } \leq \frac{ \confidence }{ 2 }
\end{equation}
Observe that $ \abs{C^{(1)} - C^{(2)}} \leq \abs{ C^{(1)} - (\approxerror\ell/2) } + \abs{ C^{(2)} - (\approxerror\ell/2)} $. Therefore
\begin{equation*}
\Pr \tbrac{ \abs{C^{(1)} - C^{(2)}} > \approxerror \ell } \leq \confidence
\end{equation*}
For $ t \in A$, let $ Y_t $ be a random variable that is $ 1 $ if $ t $ is assigned the color $ 1 $ and $ -1 $ otherwise. Then $ C^{(1)} - C^{(2)} = \sum_{ t \in A} Y_t = Y $. We have that $ \Pr \tbrac{ Y_t = 1 } = \Pr \tbrac{ Y_t = 1 } = 1/2$, then
\begin{equation*}
\E[ Y_t ] = 0 \text{ and } \E[ Y_t^2] = 1
\end{equation*}
By the independence of $ Y_t $, we get
\begin{equation*}
\E[ Y^2 ] = \sum_{ t \in A } \E[ Y_t^2 ] + \sum_{ u \neq v} \E[ Y_u Y_v ] = \ell
\end{equation*}
Finally, we have $ \E[ \abs{Y} ] \leq \sqrt{ \E [ \abs{ Y }^2 ] } = \sqrt{ \E [ { Y }^2 ] } \leq \sqrt{ \ell }$.
\end{proof}
We now prove \Cref{lemma: graph sparsification for k = 2}.
\begin{proof}[Proof of Lemma~\ref{lemma: graph sparsification for k = 2}] Let each vertex $t \in \vertexset$ be associated with a random variable
$Y_t \in \sbrac{1,2}$ such that $\Pr\tbrac{Y_t = 1} = \nicefrac{1}{2}$. For $i \in \sbrac{1,2}$, define $U^i = \sbrac{ t \in \vertexset \mid Y_t = i }$ and
\begin{equation*}
f\fbrac{Y_1, \ldots, Y_{\vertexcount}} = \edgecount(U^1) + \edgecount(U^2).
\end{equation*}
% For $i \in \sbrac{1,2}$, define
% $U_i = \sbrac{ t \in \vertexset \mid Y_t = i }$ and \[
% f\fbrac{Y_1, \ldots, Y_{\vertexcount}}
% = \edgecount(U_1) + \edgecount(U_2).
% \]
A fixed edge is counted by $f$ if and only if both its endpoints are assigned the same label,
which occurs with probability $\nicefrac{1}{2}$.
Therefore, $\E[f] = \edgecount / 2$.
Consider the Doob martingale $ X_0, X_1, \ldots, X_\vertexcount $, where $ X_t = \E[f(Y_1, \ldots, Y_\vertexcount)|Y_1, \ldots, Y_t]$. For a fix values of $ Y_1, \ldots, Y_{t-1}$, we have,
\begin{align*}
X_{t-1} = \frac{ g(1) + g(2) }{ 2 }, \text{ where } g(i) = \E[f(Y_1, \ldots, Y_\vertexcount)| (Y_1, \ldots, Y_{t-1}) \cap (Y_t = i)]
\end{align*}
Since, $X_t$ is either $g(1)$ or $g(2)$, we have $ \abs{X_t - X_{t-1}} \leq \abs{g(1) - g(2)}$. Let $ N(t) $ be the set of neighbours of the vertex $ t $ and $ \deg(t) = \abs{ N(t) }$. Let $ N_{ < t} = N(t) \cap [t-1] $ and $ N_{ > t} = N(t) \cap \sbrac{(t + 1), \ldots, \vertexcount}$. Let $C_{<t}^{(i)}$ and $C_{>t}^{(i)}$ be the number of vertices in $ N_{ < t} $ and $ N_{ > t} $ respectively. So,
\begin{equation*}
\Delta_t = \abs{g(1) - g(2)} = \abs{ C_{<t}^{(1)} + \E\tbrac{ C_{>t}^{(1)} \mid Y_t = 1} - C_{<t}^{(2)} - \E\tbrac{ C_{>t}^{(2)} \mid Y_t = 2} }
\end{equation*}
The above equation is true because any edge with vertices in $ [t-1] $ will have the same contribution to $ g(1) $ and $ g(2) $. Similarly, any edge which have one vertex in $ [t-1] $ and another in $ \sbrac{(t + 1), \ldots, \vertexcount} $ or both of its vertices in $ \sbrac{(t + 1), \ldots, \vertexcount} $ will have the same contribution.
Using Lemma~\ref{lemma: support for sparsification}, we have with probability at least $ 1 - \confidence'$,
\begin{equation*}
\Delta_t \leq \sqrt{ 6 { \deg(t) } \log \frac{ 4 }{ \confidence' } } + \sqrt{ \deg(t) } \leq 2\constant \sqrt{ \deg(t) \log \frac{ 1 }{ \confidence' } } = \constant_t
\end{equation*}
Let $ \mathcal{B} $ be the event that there exits a $ t $ such that $ \Delta_t > \constant_t $. Let $ \vertexcount_a $ be the set of active vertices (vertices with degree at least $ 1 $). Therefore, using union bound $ \Pr[\mathcal{B}] \leq 4 \vertexcount_a \confidence' \leq 4 \edgecount \confidence' $.
Let $ S = \sum_{ t = 1}^\vertexcount \constant_t^2 = \sum_{ t = 1}^\vertexcount 4 \constant^2 \deg(t) \log( 4 / \confidence') = O( \edgecount \log( 1 / \confidence' ) ) $ and $ s = \constant \sqrt{\edgecount} \log( 1 / \confidence' ) $. Using \Cref{lemma: martingle inequality}, we have
\begin{equation*}
\Pr \tbrac{ \abs{ f - \frac{ \edgecount }{ 2 }} > s } \leq 2 \exp \fbrac{ - \frac{ s^2 }{ 2S } } + \Pr \tbrac{ \mathcal{ B } } \leq \confidence' + 4 \edgecount \confidence'
\end{equation*}
Replacing $ \confidence' $ with $ \confidence / \constant \edgecount $ for a sufficiently large constant, we obtain
\begin{equation*}
\Pr\tbrac{ \abs{ \frac{ \edgecount }{ 2 } - \sum_{ i = 1 }^{ 2 } \edgecount( U^i ) } \geq \constant \sqrt{ \edgecount } \log \fbrac{\frac{ \edgecount }{ \confidence }} } \leq \confidence
\end{equation*}
which completes the proof.
\end{proof}
\begin{restatable}[]{corollary}{}\label{lemma: spasification k = 2}
For a subset $U \subseteq \vertexset$, suppose there exists a sufficiently large
absolute constant $\thresholdconst$ such that $\edgecount(U) \ge \thresholdconst$.
Let $U^1$ and $U^2$ be a uniformly random partition of $U$. Then, with probability at least $1 - \frac{1}{\edgecount(U)}$, the following hold.
\begin{enumerate}
\item[(i)] $ \frac{ \edgecount( U^1 ) }{ \edgecount( U ) } \geq 1/8$ or $\frac{ \edgecount( U^2 ) }{ \edgecount( U )} \geq 1/8$;
\item [(ii)] $\frac{ \edgecount( U^1 ) }{ \edgecount( U ) } \leq 3/4$ and $\frac{ \edgecount( U^2 ) }{ \edgecount( U )} \leq 3/4$.
\end{enumerate}
% For a subset $ U \subseteq \vertexset $, if there exists a large enough absolute constant $ \thresholdconst $, such that $ \edgecount( U ) \geq \thresholdconst $. Let $ U_1 $ and $ U_2 $ be a uniform random partition of $ U $. %(Line~\ref{addestimation: line 1} of \addestimation{}).
% Then, there exists $ i \in \sbrac{1, 2}$ such that $ \fractionofedges_U( U_i ) \geq \frac{ 1 }{ 8 } $ with probability at least $ 1 - \frac{1 }{\edgecount( U ) }$. Furthermore, for all $i \in \sbrac{1,2}$, $\fractionofedges_U( U_i ) \leq \frac{3}{4}$ with probability at least $ 1 - \frac{1 }{ \edgecount( U ) }$.
\end{restatable}
\begin{proof}
Taking $\thresholdconst$ to be sufficiently large, we have $c \sqrt{m(U)} \log m(U) \leq \frac{m(U)}{4}$. So, by taking $\delta=1/m(U)$ in Lemma~\ref{lemma: graph sparsification for k = 2}, we
% Let the subset $ U $ be partitioned in $ U_1 $ and $ U_2 $ as mentioned. Therefore, by Lemma~\ref{lemma: graph sparsification for k = 2} we have,
% \begin{align*}
% \Pr\tbrac{
% \abs{
% \frac{ \edgecount( U ) }{ 2 }
% - \sum_{ i = 1 }^{ 2 } \edgecount( U_i )
% }
% \geq
% \constant \sqrt{ \edgecount( U ) } \log \edgecount( U )
% }
% \leq
% \frac{ 1 }{ \edgecount( U ) }.
% \end{align*}
% If $\edgecount(U) > \thresholdconst$, where $\thresholdconst$ is sufficiently large such that
% $c \sqrt{m(U)} \log m(U) \leq \frac{m(U)}{4}$.
% % \begin{equation*}
% % \frac{ \sqrt{ \edgecount( U ) } }{ 4 \log \edgecount( U ) }
% % \geq
% % \constant .
% % \end{equation*}
% Hence,
\[
\Pr\!\tbrac{
\frac{\edgecount(U)}{4}
\le
\sum_{i=1}^{2} \edgecount(U^i)
\le
\frac{3\,\edgecount(U)}{4}
}
\ge
1 - \frac{1}{\edgecount(U)} \, .
\]
% % \begin{align*}
% % \Pr\!\tbrac{
% % \abs{
% % \frac{\edgecount(U)}{2}
% % - \sum_{i=1}^{2} \edgecount(U_i)
% % }
% % \le
% % \frac{\edgecount(U)}{4}
% % }
% % &\ge
% % 1 - \frac{1}{\edgecount(U)} \\[6pt]
% % \Rightarrow\quad
% % \Pr\!\tbrac{
% % \frac{\edgecount(U)}{4}
% % \le
% % \sum_{i=1}^{2} \edgecount(U_i)
% % \le
% % \frac{3\,\edgecount(U)}{4}
% % }
% % &\ge
% % 1 - \frac{1}{\edgecount(U)} .
% % \end{align*}
% Since,
% $ \fractionofedges_U( U_i ) = \frac{ \edgecount( U_i ) }{ \edgecount( U ) } $.
% It follows that,
% \begin{align*}
% \Pr\tbrac{
% \frac{ 1 }{ 4 }
% \leq
% \sum_{ i = 1 }^{ 2 } \fractionofedges_U( U_i )
% \leq
% \frac{ 3 }{ 4 }
% }
% \geq
% 1 - \frac{ 1 }{ \edgecount( U ) } .
% \end{align*}
This implies that, with probability at least $1 - 1/m(U)$,
at least one of $ \edgecount( U^1 ) / \edgecount( U ) $ and $\edgecount( U^2 ) / \edgecount( U )$ is at least $1/8$,
and both $\edgecount( U^1 ) / \edgecount( U )$ and $\edgecount( U^2 ) / \edgecount( U )$ are at most $3/4$.
\end{proof}
% \begin{remark}
% The above corollary applies only when the number of edges induced by \( U \) exceeds the threshold $ \thresholdconst $. Consequently, before the sparsification step, the algorithm must perform a threshold check. This threshold test is performed using \iscollision{}, as described in
% \Cref{sec: threshold checking}, by setting the threshold $ \threshold = \thresholdconst $.
%Now, we describe our procedure \addestimation{} and the corresponding guarantee is given in \Cref{Lemma: AddEst + Sparsifier}.
% \end{remark}
% \begin{algorithm}
% \caption{\addestimation($ U, \confidence $)}\label{alg: addestimation}
% \begin{algorithmic}[1]
% \Require \iscq{} query access to $G$ and a confidence parameter $\delta \in (0,1)$.
% \Ensure A subset $U' \subseteq U$.
% \State Uniformly and independently partition $ U $ into $ U_1 $ and $ U_2 $.\label{addestimation: line 1}
% \State Uniformly and independently sample $ O( \log ( \confidence^{-1} )) $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges. \label{addestimation: line 2}
% \State Compute $ \approxfractionofedges ( U_i ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U_i } $ for $ i \in \{1, 2\} $. \label{addestimation: line 3}
% \If{There exists $ i \in \{1, 2\} $ such that $ \approxfractionofedges( U_i ) \geq \frac{ 1 }{ 16 } $} \label{addestimation: line 4}
% \State \Return $U_i$ as $U'$ \label{addestimation: line 5}
% \Else
% \State \Return $U_1$ or $U_2$ arbitrarily as $U'$. \label{addestimation: line 7}
% \EndIf
% \end{algorithmic}
% \end{algorithm}
%\gopi{Chenge the name of $\addestimation{}$.}
%\gopi{few connecting lines for the lemma.}
To prove the guarantees of \addestimation{} described in \Cref{alg: addestimation}, we need \Cref{lemma: addestimation}.
\begin{restatable}[]{lemma}{addest}\label{lemma: addestimation}
With probability at least $ 1 - \confidence$, the following statements hold for \addestimation{}
\begin{enumerate}
\item[(i)] If $ \frac{ \edgecount( U^i ) }{ \edgecount( U ) } \geq \frac{ 1 }{ 8 } $, then $ \approxfractionofedges( U^i ) \geq \frac{ 1 }{ 16 } $.%
\item[(ii)] If $ \frac{ \edgecount( U^i ) }{ \edgecount( U ) } < \frac{ 1 }{ 32 } $, then $ \approxfractionofedges( U^i ) < \frac{ 1 }{ 16 } $.% and \addestimation{} fail.
\end{enumerate}
The number of $\iscq$ queries made is $ O\fbrac{ \log ( \confidence^{-1} ) } $.
\end{restatable}
\begin{proof}
Consider a subset $ U^i \subseteq U $ for any $ i \in \{1,2\} $.
For each sampled edge, define the indicator random variable
\[
X_j =
\begin{cases}
1 & \text{if the $ j^{\text{th}} $ sampled edge lies in $ U^i $,} \\
0 & \text{otherwise.}
\end{cases}
\]
Observe that $\E[X_j] = \frac{m(U^i)}{m(U)}$.
Define the fraction of sampled edges that belong to $ U^i $ by
$
X = \frac{1}{\size{\samplededges}}
\sum_{j \in [\size{\samplededges}]} X_j.
$
where $ \samplededges $ is the multiset of sampled edges (Line~\ref{addestimation: line 2}) obtained via \iscq{} queries on $ U $.
Note that $ X = \approxfractionofedges(U^i) $ and hence
\[
\E[\approxfractionofedges(U^i)]
= \frac{1}{\size{\samplededges}}
\sum_{j \in [\size{\samplededges}]}\E[X_j]
= \frac{\edgecount(U^i)}{\edgecount(U)}
= \frac{m(U^i)}{m(U)}.
\]
Applying Hoeffding's bound~(\Cref{lemma: hoeffding}), we get
\[
\Pr\!\left( \left| \approxfractionofedges(U^i) -
\frac{m(U^i)}{m(U)}
\right|
\ge \tfrac{1}{32}
\right)
\le
\exp\!\left(
-\frac{\size{S}}{2}\cdot (1/32)^2
\right) \le \delta/2.
\]
The last inequation holds since $\size{\samplededges}=O(\log 1/\confidence)$. Using the union bound over $i\in \{1,2\}$, we get the desired result.
\end{proof}
% \begin{proof} Refer to Appendix~\ref{appendix: concentration proofs}
% Consider a subset $ U_i \subseteq U $ for any $ i \in \{1,2\} $.
% For each sampled edge, define the indicator random variable
% \[
% X_j =
% \begin{cases}
% 1 & \text{if the $ j^{\text{th}} $ sampled edge lies in $ U_i $,} \\
% 0 & \text{otherwise.}
% \end{cases}
% \]
% Observe that $\E[X_j] = \frac{m(U_i)}{m(U)}$.
% Define the fraction of sampled edges that belong to $ U_i $ by
% $
% X = \frac{1}{\size{\samplededges}}
% \sum_{j \in [\size{\samplededges}]} X_j.
% $
% where $ \samplededges $ is the multiset of sampled edges (Line~\ref{addestimation: line 2}) obtained via \iscq{} queries on $ U $.
% Note that $ X = \approxfractionofedges(U_i) $ and hence
% \[
% \E[\approxfractionofedges(U_i)]
% = \frac{1}{\size{\samplededges}}
% \sum_{j \in [\size{\samplededges}]}\E[X_j]
% = \frac{\edgecount(U_i)}{\edgecount(U)}
% = \fractionofedges(U_i).
% \]
% Applying Hoeffding's bound~(\Cref{lemma: hoeffding}),
% \[
% \Pr\!\left(
% \left|
% \approxfractionofedges(U_i)
% -
% \fractionofedges(U_i)
% \right|
% \ge \tfrac{1}{32}
% \right)
% \le
% \exp\!\left(
% -\frac{\size{S}}{2}\cdot (1/32)^2
% \right)
% \le \delta/2.
% \]
% The last inequation holds since $\size{\samplededges}=O(\log 1/\confidence)$.
% Using the union bound over $i\in \{1,2\}$, we get the desired result.
% \end{proof}
%\gopi{few connecting lines for the lemma.}
We now give the proof of \Cref{Lemma: AddEst + Sparsifier}, using \Cref{lemma: graph sparsification for k = 2}, \Cref{lemma: spasification k = 2} and \Cref{lemma: addestimation}
\begin{proof}[Proof of \Cref{Lemma: AddEst + Sparsifier}]
In \addestimation{}, the set $U$ is partitioned uniformly at random into $U^1$ and $U^2$.
From \Cref{lemma: graph sparsification for k = 2}(i),
\[
\E[\edgecount(U^1)+\edgecount(U^2)] = \edgecount(U)/2.
\]
Since $U'$ is either $U^1$ or $U^2$, it follows that
$\E[\edgecount(U')] \le \edgecount(U)/2$, proving part~(i).
\smallskip
We now prove part~(ii). Let $\mathcal{E}_{\text{apx}}$ denote the event that $U'$ is an approximation-preserving subset of $U$, that is, $ 1/32 \leq m(U')/m(U)\leq 3/4$.
Let $\mathcal{E}$ be the event that at least one of $\edgecount( U^1 ) / \edgecount( U )$ and
$\edgecount( U^2 ) / \edgecount( U )$ is at least $1/8$, and both $\edgecount( U^1 ) / \edgecount( U )$ and $\edgecount( U^2 ) / \edgecount( U )$ are at
most $3/4$. By \Cref{lemma: spasification k = 2},
\[
\Pr[\mathcal{E}] \ge 1 - \frac{1}{m(U)}.
\]
Let us condition on the event $\mathcal{E}$. It suffices to bound $\Pr[\mathcal{E}_{\text{apx}} \mid \mathcal{E}]$.
Then both $\edgecount( U^1 ) / \edgecount( U )$ and $\edgecount( U^2 ) / \edgecount( U )$ are at most $3/4$, and at least one of
$\edgecount( U^1 ) / \edgecount( U )$ and $\edgecount( U^2 ) / \edgecount( U )$ is at least $1/8$; w.l.o.g.\ assume $\edgecount( U^1 ) / \edgecount( U ) \ge 1/8$. If $\edgecount( U^2) / \edgecount( U ) \ge 1/32$, then since $U'\in\{U^1,U^2\}$ we immediately have $1/32 \le m(U')/m(U) \le 3/4$, that is, $\Pr[\mathcal{E}_{\text{apx}}\mid\mathcal{E}]=1$. Otherwise $\edgecount( U^2) / \edgecount( U )<1/32$. By \Cref{lemma: addestimation}, with probability at least $1-\delta$, we obtain $\widehat{p}(U^1)\ge 1/16$ and $\widehat{p}(U^2)<1/16$, implying $U'=U^1$. Consequently $1/32 \le m(U')/m(U) \le 3/4$ holds with probability at least $1-\delta$ conditioned on $\mathcal{E}$. Thus in all cases $$\Pr[\mathcal{E}_{\text{apx}}\mid\mathcal{E}] \ge 1-\delta.$$
Therefore,
\[
\Pr[\mathcal{E}_{\text{apx}}]
\ge
\Pr[\mathcal{E}] \cdot \Pr[\mathcal{E}_{\text{apx}}\mid\mathcal{E}]
\ge
1 - \frac{1}{m(U)} - \delta.
\]
% Then $\edgecount( U_1 ) / \edgecount( U )$, \edgecount( U_2) / \edgecount( U )$ \leq 3/4$
% Since $U'$ is either $U_1$ or $U_2$, the bound
% $\edgecount(U')/\edgecount(U) \le 3/4$ always holds whenever
% $\mathcal{E}$ occurs. It remains to argue that
% \[
% \Pr\!\left(
% \frac{\edgecount(U')}{\edgecount(U)} \ge \frac{1}{32}
% \;\middle|\;
% \mathcal{E}
% \right)
% \ge 1 - \delta.
% \]
% Otherwise, $\edgecount( U_1 ) / \edgecount( U )$\geq 1/8$ and $\edgecount( U_2) / \edgecount( U )$< 1/32$.
% there exists $i$, such that $U_i$ satisfies $\approxfractionofedges(U_i) \geq \frac{1}{16}$ with probability at least $1 - \confidence/2$. This implies \addestimation{} does not execute line~\ref{addestimation: line 7} with probability $1-\confidence/2 - 1/\edgecount(U)$. So, the algorithm reports some $U_i$ as $U'$ with $\fractionofedges(U_i)\geq 1/16$.
% Furthermore, the probability that an $U_i$ such that $\fractionofedges(U_i) \leq \frac{1}{32}$ satisfies $\approxfractionofedges(U_i) < \frac{1}{16}$ is at most $\confidence/2$, by \Cref{lemma: addestimation} (ii). Putting everything together, we are done.
% Part~(iii) follows directly from \Cref{lemma: spasification k = 2}.
%\red{A better rewrite?}
\end{proof}
\subsection{Procedure to find approximate preserving sequence}\label{section: approximate sequence}
We now describe the algorithm \subsequence{} (see \Cref{alg: subsequence}), which constructs the sparsification sequence used by the main estimator.
\begin{algorithm}
\caption{\subsequence{}($ \vertexset $)}\label{alg: subsequence}
\begin{algorithmic}[1]
\Require \iscq{} access to graph $G$.
\Ensure A nested sequence $U_1, U_2, \cdots, U_\iterations$.
\State Initialize $ U_0 \gets V $; $ \iscoll \gets 1 $; $ \iterations \gets 0 $ \label{subsequence: Line 1}
\While{$ true $}\label{subsequence: Line 2}
\State $\iscoll \gets \iscollision{}( U_\iterations, \thresholdconst, \frac{ 1 }{ 2^{ \iterations + 5 } } )$
\If{$\iscoll$ equals $ 1 $}
\State \textbf{break}
\Else
\State $ U_{ \iterations + 1 } \gets \addestimation( U_\iterations, \frac{ 1 }{ 2^{ \iterations + 5 } } ) $ \label{subsequence: Line 7}
\State $ \iterations \gets \iterations + 1 $
\EndIf
\EndWhile
\State \Return $U_1, U_2, \cdots, U_\iterations$
\end{algorithmic}
\end{algorithm}
% \begin{algorithm}
% \setcounter{AlgoLine}{0}
% \SetAlgoNoLine
% \LinesNumbered
% \SetAlgoNoEnd
% \caption{ \subsequence{}($ \vertexset $) }
% \label{alg: subsequence}
% \KwIn{\iscq{} access to graph $G$.}
% \KwOut{A sequence $U_1, U_2, \cdots, U_\iterations$.}
% Initialize $ U_0 \gets V $; $ \iscoll \gets 1 $; $ \iterations \gets 0 $ \label{subsequence: Line 1}
% \While{$ true $ }{\label{subsequence: Line 2}
% $\iscoll \gets \iscollision{}( U_\iterations, \thresholdconst, \frac{ 1 }{ \cdot 2^{ \iterations + 5 } } )$
% \eIf{ $\iscoll$ equals $ 1 $ } {
% \textbf{break}
% } {
% $ U_{ \iterations + 1 } \gets \addestimation( U_\iterations, \frac{ 1 }{ \cdot 2^{ \iterations + 5 } } ) $ {\label{subsequence: Line 7}
% $ \iterations \gets \iterations + 1 $
% }
% }
% }
% \KwRet $U_1, U_2, \cdots, U_\iterations$
% \end{algorithm}
\begin{restatable}[Correctness of \subsequence{}]{lemma}{}\label{lemma: subsequence}
Consider \subsequence{} as described in \Cref{alg: subsequence}. With probability at least $ \nicefrac{4}{5}$, it outputs an approximation preserving sequence $ U_1, U_2, \ldots, U_\iterations $ such that $\edgecount (U_{\kappa})\leq 8\thresholdconst$. Here $\thresholdconst$ is given large constant. The number of $\iscq$ queries made is $ O( \kappa^2 )$.
\end{restatable}
%\gopi{Needs editing.}
\begin{proof}
\noindent\textbf{Correctness: }
Let us consider the event $\mathcal{E}_{\text{col}}$ that denotes
\begin{itemize}
\item[(i)] There does not exist a $j$ such that the algorithm goes to the next iteration if $m(U_j) \leq \thresholdconst.$
\item[(ii)] For any $j$ such that the algorithm goes to the next iteration if $m(U_j) \geq 8\thresholdconst.$
\end{itemize}
The probability of $\mathcal{E}_{\text{col}}$ is lower bounded by \iscollision{} succeeds in all iterations during the run of the algorithm. Using the union bound, we have
\[
\Pr[\mathcal{E}_{\text{col}}]
\ge 1 - \sum_{i = 1}^{\infty} \frac{1}{2^{i+5}}
\ge 1 - \frac{1}{2^5} \sum_{i = 1}^{\infty} \frac{1}{2^{i}}
\ge 1 - \frac{1}{16}
= \frac{15}{16}.
\]
% \debarshi{Fix how to state the assumption on $\edgecount(U_0)$.}
Let us continue the argument under the conditional space that $\mathcal{E}_{\text{col}}$ has occurred.
We assume $m(U_0) = \edgecount$ is sufficiently large; hence, the algorithm does not terminate in the first iteration, given that $\mathcal{E}_{\mathrm{col}}$ has occurred. Let $U_0,\ldots, U_\kappa$ be the output of the algorithm, where $\kappa$ is a positive integer.
Conditioned on the event $\mathcal{E}_{\text{col}}$, we have $m(U_j) \geq \thresholdconst$ for every $j \leq \kappa - 1$. %We now proceed with the argument under the assumption that $\mathcal{E}_{\text{col}}$ occurs and that $U_0, U_1, \ldots, U_\kappa$ is the output of the algorithm for some non-negative integer $\kappa$.
We prove the claim by induction on $\kappa$. For each $j \ge 0$, let $\mathcal{E}_j$ denote the event which is that the algorithm executes iteration $j$ and either terminates with, $\edgecount(U_j) \leq 8\thresholdconst$, or it produces an approximation preserving subset $U_{j+1}$ such that $\edgecount(U_{j+1}) \leq \frac{3\edgecount(U_j)}{4}$.
Let $\mathcal{E}_{\le j}$ denote the event that $\mathcal{E}_i$ occurs for all $i \in [j]$, that is,
\[
\mathcal{E}_{\le j} \;=\; \bigcap_{i=1}^{j} \mathcal{E}_i .
\]
Equivalently, $\mathcal{E}_{\le j}$ is the event that the algorithm runs for $j$ iterations, such that
$U_0, \ldots, U_{j}$ is an approximation preserving sequence, and at iteration $j$ either it produces $U_{j+1}$, an approximation preserving subset of $U_{j}$ such that $\edgecount(U_{j+1}) \leq \frac{3\edgecount(U_j)}{4}$, or the algorithm terminates with
$\edgecount(U_j) \leq 8\thresholdconst$.
The induction hypothesis is that
\[
\Pr[\mathcal{E}_{\le j}] \;\ge\; 1 - \delta_{\leq j},
\]
where
\[
\delta_i
\;=\;
\frac{1}{\edgecount(U_i)}
+
\frac{1}{2^{i+5}}
,
\qquad
\text{and}
\qquad
\delta_{\le j}
\;=\;
\sum_{i=0}^{j} \delta_i .
\]
We first prove the base case. That is, $\Pr[\mathcal{E}_{0}]\geq 1-\delta_0$. Recall that we assume that $m(U_0) = m$ is sufficiently large. Moreover, we assume that event $\mathcal{E}_{\text{col}}$ has occurred and algorithm produced a sequence $U_0, \ldots,U_\iterations$ with $\iterations \geq 1$. So, $\mathcal{E}_0$ occurs if \addestimation{} reports an approximate preserving subset $U_{j+1}$. By \Cref{Lemma: AddEst + Sparsifier}, this occurs with probability at least $1-\confidence_0$.
For the inductive step, we mainly show that $\Pr[\mathcal{E}_j~|~\mathcal{E}_{\leq j-1}] \geq 1-\delta_j$ for $j\geq 1$. Let us argue that it is sufficient.
Note that
$\delta_{\le j} = \delta_{\le j-1} + \delta_j$ for $j\geq 1$ and
\[
\Pr[\mathcal{E}_{\le j}]
=
\Pr[\mathcal{E}_{\le j-1} \cap \mathcal{E}_j]
=
\Pr[\mathcal{E}_{\le j-1}]\cdot
\Pr[\mathcal{E}_j \mid \mathcal{E}_{\le j-1}] .
\]
By the induction hypothesis, $
\Pr[\mathcal{E}_{\le j-1}] \ge 1 - \delta_{\le j-1}.$
We will show that
\begin{equation*}\label{eqn:ind}
\Pr[\mathcal{E}_j \mid \mathcal{E}_{\le j-1}] \ge 1 - \delta_j .
\end{equation*}
Therefore,
\begin{align*}
\Pr \left[ \mathcal{E}_{\le j} \right]
&=
\Pr \left[ \mathcal{E}_{\le j-1} \right] \cdot \Pr \left[ \mathcal{E}_j \mid \mathcal{E}_{\le j-1} \right] \\
&\ge
(1-\delta_{\le j-1})(1-\delta_j) \\
&\ge
1-\delta_{\le j-1}-\delta_j \\
&=
1-\delta_{\le j}.
\end{align*}
%\red{Base case: $\kappa = 0$:}
Consider any $j$ with $1\leq j \leq \kappa-1$. Under the conditional space $\mathcal{E}_{\text{col}}$ and along with the fact that we assume $m(U_0) = m$ is sufficiently large, $m(U_j) \geq \thresholdconst$ for each $j \leq \kappa-1$.
We consider two cases: $m(U_j) \geq 8\thresholdconst$, and $\thresholdconst \leq m(U_j) \leq 8\thresholdconst$, and analyse them individually.
\paragraph{Case $1$: $(\edgecount(U_j) > 8\thresholdconst)$.}
Recall that we are assuming that the event $\mathcal{E}_{\text{col}}$ has occurred. So,
\subsequence{} does not terminate after executing \iscollision{} and execute \addestimation{}. Moreover it produces an approximate preserving subset $U_{j+1}$ with $\edgecount(U_{j+1}) \leq \frac{3\edgecount(U_j)}{4}$, that is event $\mathcal{E}_j$ occurs, when \addestimation{} reports an approximate preserving subset $U_{j+1}$ with $\edgecount(U_{j+1}) \leq \frac{3\edgecount(U_j)}{4}$. By \Cref{Lemma: AddEst + Sparsifier} and Line 7 of \subsequence{}, this occurs with probability at least $1-\confidence_j$ where,
\[
\delta_j
=
\frac{1}{\edgecount(U_j)} + \frac{1}{2^{j+5}}
% +
% \frac{1}{2^{j+5}} .
\]
%\subsequence{} outputs an approximate preserving subset $U_{j+1}$ with probability at least $1-\confidence^{(1)}_j$, where $$\delta^{(1)}_j=1-1m(U_j)- \frac{1}{2^{j+5}}.$$
% From the definition of the event $\mathcal{E}_j$, it occurs if\addestimation{} returns an approximate preserving subset $U_{j+1}$ of $U_i$. Hence,
Hence,
\[
\Pr[\mathcal{E}_j \mid \mathcal{E}_{\le j-1}] \ge 1-\delta_j^{(1)},
\]
% Hence, it satisfies our induction step
% \[
% \Pr(\mathcal{E}_j \mid \mathcal{E}_{\le j-1}) \ge 1 - \delta_j .
% \]
\paragraph{Case $2$: $( \thresholdconst \leq \edgecount(U_j) \leq 8\thresholdconst )$.}
In this case, we show that
\[
\Pr\!\left[\mathcal{E}_j \mid \mathcal{E}_{\le j-1} \right] \ge 1-\delta_j
\]
regardless of whether \subsequence{} terminates after executing \iscollision{} or proceeds to execute \addestimation{}.
In the former case, the claim follows directly from the definition of $\mathcal{E}_j$ together with the fact that $\edgecount(U_j) \le 8\thresholdconst$. In the latter case, \subsequence{} produces an approximate preserving subset $U_{j+1}$ of $U_j$ provided that \addestimation{} returns an approximate preserving subset $U_{j+1}$. By \Cref{Lemma: AddEst + Sparsifier}, this occurs with probability at least $1-\delta_j$, where
\[
\delta_j
\le
\frac{1}{\edgecount(U_j)} + \frac{1}{2^{j+5}
}.
\]
Hence,
\[
\Pr\!\left[ \mathcal{E}_j \mid \mathcal{E}_{\le j-1} \right]
\ge
1 - \delta_j .
\]
We are done with the proof of the inductive step. We summarize the above discussion as follows. $\Pr[\mathcal{E}_{\text{col}}]\geq 15/16$. When the event $\mathcal{E}_{\text{col}}$ occurs, $\Pr[\mathcal{E}_{j}~|~\mathcal{E}_{j-1}] \geq 1-\delta_j$ for all $j \geq 0$. Moreover, $\Pr[\mathcal{E}_{\leq j}] \geq 1-\delta_{\leq j}$ for all $j \geq 0$. Recall that the algorithm outputs a sequence
$U_0, \ldots, U_{\kappa-1}$.
The events $\mathcal{E}_{\text{col}}$ and
$\mathcal{E}_{\leq \iterations-1}$ occur with probability at least
\[
\Pr[\mathcal{E}_{\text{col}}]
- \confidence_{\leq \iterations-1}.
\]
Hence, the following facts hold with the same probability:
\begin{itemize}
\item[1.] $U_0, \ldots, U_\kappa$ is an approximate preserving sequence, that is, $$ \frac{1}{32} \leq \frac{m(U_{i})}{m(U_{i-1})} \leq \frac{3}{4},~\mbox{where}~i \in [\kappa].$$
\item [2.] $\edgecount(U_j) \geq \thresholdconst$ for each $j \leq \iterations - 1$,
and $\edgecount(U_\iterations) \leq 8\thresholdconst$.
\end{itemize}
Note that \subsequence{} produces an approximation-preserving sequence
$U_1, U_2, \cdots, U_\iterations$ such that
$\edgecount(U_{\kappa}) \leq 8\thresholdconst$ whenever both of the above
conditions hold. This occurs with probability at least $1/5$, provided
$\delta_{\iterations-1} \leq 1/10$, since
$\Pr[\mathcal{E}_{\text{col}}] \geq 15/16$.
We complete the proof of correctness of \subsequence{} by showing that $\delta_{\iterations-1} \leq 1/10$.
\begin{align*}
\confidence_{\leq \iterations - 1}
= \sum_{i=0}^{\iterations - 1} \frac{1}{\edgecount(U_j)}
+ \frac{1}{2^{j+5}}
\leq \sum_{i=0}^{\infty} \frac{1}{2^{j+4}}
+ \sum_{i=0}^{\iterations - 1} \frac{1}{\edgecount(U_j)}
\leq \frac{1}{16}
+ \sum_{i=0}^{\iterations - 1}
\frac{3^{\iterations - 1 - i}}
{4^{\iterations - 1 - i}\thresholdconst}.
\end{align*}
The last inequation holds since $\edgecount(U_{\iterations - 1}) \geq \thresholdconst$ and
$\edgecount(U_i) \leq \frac{3\edgecount(U_{i-1})}{4}$ for all
$i \in [\iterations - 1]$. We obtain the bound of $\confidence_{\leq \iterations - 1}$ as below:
\begin{align*}
\confidence_{\leq \iterations - 1}
= \frac{1}{16}
+ \frac{1}{\thresholdconst}
\sum_{i=0}^{\iterations - 1}
\frac{3^{\iterations - 1 - i}}
{4^{\iterations - 1 - i}}
= \frac{1}{16}
+ \frac{1}{\thresholdconst}
\sum_{i=0}^{\iterations - 1}
\frac{3^{i}}{4^{i}}
\leq \frac{1}{16} + \frac{2}{\thresholdconst}\leq \frac{1}{10}.
\end{align*}
Here we have used the fact that $\thresholdconst$ is a large constant.
% \begin{align*}
% \confidence_{\leq \iterations} &= \sum_{i=1}^\iterations \frac{1}{\edgecount(U_j)} + \frac{1}{2^{j+5}} + \frac{1}{2^{j+5}} \\
% &\leq \sum_{i=1}^\infty \frac{1}{2^{j+4}} + \sum_{i=1}^\iterations \frac{1}{\edgecount(U_j)}\\
% &\leq \frac{1}{16} + \sum_{i=1}^\iterations \frac{3^{\iterations-i}}{4^{\iterations-i}\thresholdconst} &\text{As $\edgecount_\iterations \geq \thresholdconst$, and $\edgecount(U_i) \leq \frac{3\edgecount(U_{i-1})}{4}$, for all $i \in [\iterations]$}\\
% &= \frac{1}{16} + \frac{1}{\thresholdconst}\sum_{i=1}^\iterations \frac{3^{\iterations-i}}{4^{\iterations-i}}\\
% &= \frac{1}{16} + \frac{1}{\thresholdconst}\sum_{i=1}^\iterations \frac{3^{i}}{4^{i}}\\
% &\leq \frac{1}{16} + \frac{3}{\thresholdconst}
% \end{align*}
% \begin{align*}
% \confidence_{\leq \iterations-1}
% &= \sum_{i=1}^{\iterations-1} \frac{1}{\edgecount(U_j)}
% + \frac{1}{2^{j+5}}
% + \frac{1}{2^{j+5}} \\
% &\leq \sum_{i=1}^{\infty} \frac{1}{2^{j+4}}
% + \sum_{i=1}^{\iterations-1} \frac{1}{\edgecount(U_j)} \\
% &\leq \frac{1}{16}
% + \sum_{i=1}^{\iterations-1}
% \frac{3^{(\iterations-1)-i}}
% {4^{(\iterations-1)-i}\thresholdconst}
% \qquad
% \text{as $\edgecount ({U_{\iterations-1}}) \ge \thresholdconst$, and
% $\edgecount(U_i) \le \frac{3\edgecount(U_{i-1})}{4}$} \\
% &= \frac{1}{16}
% + \frac{1}{\thresholdconst}
% \sum_{i=1}^{\iterations-1}
% \frac{3^{(\iterations-1)-i}}
% {4^{(\iterations-1)-i}} \\
% &= \frac{1}{16}
% + \frac{1}{\thresholdconst}
% \sum_{i=1}^{\iterations-1}
% \frac{3^{i}}{4^{i}} \\
% &\leq \frac{1}{16}
% + \frac{3}{\thresholdconst}.
% \end{align*}
%gopi{Here we bound $\delta_{\kappa-1}$. }
%This establishes the result as $\thresholdconst$ is sufficiently large.
\paragraph{Query Complexity:} From Lemma~\ref{lemma: collision detection} and~\ref{lemma: addestimation}, the number of \iscq{} queries made by \iscollision{} (for a constant threshold $ \threshold = \thresholdconst $) and \addestimation{} over $\iterations$ iterations is,
\begin{align*}
O\fbrac{\sum_{ j = 1 }^{\iterations} \log( 2^{ \iterations + 5 } ) } = O\fbrac{\iterations^2}
\end{align*}
\end{proof}
Note that the number of iterations and the query complexity of \subsequence{} (as mentioned in the above lemma) are random variables. We will argue their expectation in the following lemma. Moreover,
when we discuss the full algorithm in \Cref{section: full algorithm} (that uses \subsequence{} as a subroutine), we show that the query complexity of the full algorithm is $O(\kappa^2 \cdot \log \kappa)$, where $\kappa$ is the number of iterations of \subsequence{}. The following lemmas will be useful in establishing the final query complexity in \Cref{section: full algorithm}.
\begin{lemma}\label{lemma: kappa}
$\E[\iterations]=O(\log \edgecount)$, $\E[\iterations^2]=O(\log^2 \edgecount)$, and $\E[\iterations^2 \log \iterations]=O(\log^2 \edgecount \cdot \log \log \edgecount)$.
\end{lemma}
From \Cref{lemma: subsequence} and \Cref{lemma: kappa}, \Cref{lem: sequence} follows. In the rest of the section, we prove \Cref{lemma: subsequence} and \Cref{lem: sequence}.
\begin{proof} Here, we prove that
$$\E\tbrac{ \iterations^2 \log \iterations} = O( \log^2 \edgecount \log \log \edgecount ) $$ The remaining bounds follow similarly.
We want to analyse $\E[\iterations^2 \log \iterations]$ where the expectation is over the distribution of $\iterations$ for a run of the algorithm. To analyse this, we consider two different cases. For the first case, let $\trueiter$ denote the time when the algorithm generates a subgraph with less than or equal to $\thresholdconst$ edges. Formally, we define $\trueiter$ as the time $t$ such that $\edgecount(U_{t}) \leq \thresholdconst$, and $\edgecount(U_{t-1}) > \thresholdconst$. If this is not the case, i.e. when $m(U_t) > \thresholdconst$ holds throughout the algorithm, we show that $ \iterations = O( \log \edgecount )$.
First, we bound the expectation of $\edgecount(U_t)$ for some $t \in [\iterations]$. By \Cref{Lemma: AddEst + Sparsifier}, we have
\begin{align*}
\E\left[\edgecount(U_t)~|~m(U_{t-1})\right] \leq \frac{\edgecount(U_{t-1})}{2}.
\end{align*}
By a inductive argument, and the fact that $\edgecount(U_0) = \edgecount$, we have
\begin{align}
\E[\edgecount(U_t)]=\E[\E[\edgecount(U_{t})~\mid~\edgecount(U_{t-1})]] \leq \frac{\E[\edgecount(U_{t-1})]}{2} \leq \ldots \leq \frac{\edgecount}{2^t} \label{Eq: Edge bound at step t}
\end{align}
Now, when $m(U_j) > \thresholdconst$ holds throughout the algorithm, we have
\begin{equation*}
\thresholdconst < \E\tbrac{ \edgecount( U_\iterations ) } \leq \frac{ \edgecount}{ 2^{\iterations} } \implies \iterations = O ( \log \edgecount )
\end{equation*}
Hence, we now focus on the case when there exists a $\trueiter \leq \iterations$. Note that $\trueiter$ is unique for a run of the algorithm. Once, the algorithm reaches $\trueiter$, the probability of the algorithm terminating depends on the calls to \iscollision{} succeeding. We start with analysing the conditional expectation of $\iterations^2 \log \iterations$ given $\trueiter$.
\begin{align}
\nonumber\E \tbrac{ \iterations^2 \log \iterations \mid \trueiter } &= \sum_{ j = 1 }^\infty j^2 \log j \Pr \tbrac{ \iterations = j \mid \trueiter}\\\nonumber
&= \sum_{ j = 1}^{\trueiter -1} j^2 \log j \Pr \tbrac{ \iterations = j \mid \trueiter} + \sum_{ j = \trueiter}^\infty j^2\log j \Pr \tbrac{ \iterations = j \mid \trueiter} \\\nonumber
&= \sum_{ j = \trueiter}^\infty j^2 \log j \Pr \tbrac{ \iterations = j \mid \trueiter} &\text{(As $\trueiter \leq \iterations$)}\\\nonumber
& \leq\sum_{ j = \trueiter}^\infty j^2 \log j \Pr \tbrac{ \iterations \geq j \mid \trueiter}\\\nonumber
& \leq \sum_{ j = \trueiter}^\infty \frac{ j^2 \log j }{2^{j + 5}}\\
& = \bigo{{\trueiter}^2 \log \trueiter}\label{Eq: Conditional iteration bound} %\text{\red{Do we elaborate?}}
\end{align}
Now, we analyse the distribution of $\trueiter$. We use~\eqref{Eq: Edge bound at step t} to obtain a bound on the probabilities of different values of $\trueiter$.
\begin{align}
\nonumber\Pr \tbrac{ \trueiter = t } \leq \Pr \tbrac{ \trueiter \geq t } &\leq \Pr \tbrac{ \edgecount( U_{ t } ) > \thresholdconst }\\\nonumber
&\leq \frac{ \E[ \edgecount( U_{ t } ) ] }{ \thresholdconst } &\text{Markov's Inequality (Lemma~\ref{Markov's Inequality})}\\
&\leq \frac{ \edgecount }{2^t \thresholdconst } &\text{\Cref{Eq: Edge bound at step t}}\label{Eq: Markov Edge bound}
\end{align}
Now, we bound the expectation of $\iterations^2 \log \iterations$.
\begin{align*}
\E[ \iterations^2 \log \iterations ] &= \E \tbrac{\E \tbrac{ \iterations^2 \log \iterations \mid \trueiter } }\\
&= \E[ {\trueiter}^2 \log \trueiter ] &\text{(\Cref{Eq: Conditional iteration bound})}\\
&= \sum_{j = 1}^\infty j^2 \log j \Pr[\trueiter = j]\\
&= \sum_{j = 1}^{2\log \edgecount} j^2 \log j \Pr[\trueiter = j] + \sum_{j = 2\log\edgecount + 1}^\infty j^2 \log j \Pr[\trueiter = j]\\
&\leq 8\log^2\edgecount \log \log\edgecount\sum_{j = 1}^{2\log \edgecount} \Pr[\trueiter = j] + \sum_{2\log\edgecount}^\infty \frac{\edgecount j^2 \log j}{2^j\thresholdconst} \\
&= \bigo{\log^2 \edgecount \log \log \edgecount} %&\text{\red{Do we elaborate?}}
\end{align*}
which proves the desired result.
\end{proof}
\Cref{lem: sequence} follows from the algorithm \subsequence{} (\Cref{alg: subsequence}) along with \Cref{lemma: subsequence} and \Cref{lemma: kappa}.
% Now we finally prove \Cref{lem: sequence}.
% \begin{proof}[Proof of \Cref{lem: sequence}]
% This follows
% \end{proof}
% \section{Full algorithm}\label{section: full algorithm}
% In this section, we present the full algorithm for estimating the number of edges using \iscq{} queries and prove our main result stated below.
% \MainUB*
% \edgeestimationISC{} (\Cref{alg: edgeestimationISC}) begins by calling \subsequence{} which outputs an approximation preserving sequence $U_1 , \ldots, U_\iterations $, where each $ \fractionofedges_i = \edgecount( U_i ) / \edgecount( U_{i-1} ) $ is bounded below by a constant. This ensures that each ratio can be estimated efficiently. To estimate each $ \fractionofedges_i $, we use the subroutine \mulestimation{} which returns an $ (1 \pm \approxerror)$-approximation of $ \fractionofedges_i$ . The procedure \mulestimation{} is described in \Cref{alg: mulestimation} and its guarantee is established in \Cref{lemma: mulestimation}.
% \begin{algorithm}[H]
% \caption{\mulestimation($ U', U, \approxerror, \confidence $)}\label{alg: mulestimation}
% \begin{algorithmic}[1]
% \Require $ U' \subseteq U$ and \iscq{} access to the subset $ U \subseteq \vertexset $.
% \Ensure A subset $U' \subseteq U$.
% \State Uniformly and Independently sample $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges.
% \State \Return $ \approxfractionofedges ( U' ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U' } $.
% \end{algorithmic}
% \end{algorithm}
% % \begin{algorithm}
% % \setcounter{AlgoLine}{0}
% % \SetAlgoNoLine
% % \LinesNumbered
% % \SetAlgoNoEnd
% % \caption{ \mulestimation($ U', U, \approxerror, \confidence $)}
% % \label{alg: mulestimation}
% % \KwIn{$ U' \subseteq U$ and \iscq{} access to the subset $ U \subseteq \vertexset $.}
% % \KwOut{A subset $U' \subseteq U$.}
% % Uniformly and Independently sample $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges.
% % \KwRet $ \approxfractionofedges ( U' ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U' } $.
% % \end{algorithm}
% % Now,
% % \begin{equation*}
% % \edgecount = \frac{ \edgecount( U_\iterations ) }{ \prod_{i = 1}^\iterations \fractionofedges_i }
% % \end{equation*}
% \begin{algorithm}[H]
% \caption{\edgeestimationISC($ \vertexset, \approxerror $)}\label{alg: edgeestimationISC}
% \begin{algorithmic}[1]
% \Require \iscq{} access to any subset $ U \subseteq \vertexset $.
% \Ensure An estimate of $ \approxedgecount $
% \State $ U_1, U_2, \cdots, U_\iterations \gets \subsequence{}(\vertexset)$
% \State $ \approxedgecount_b \gets \birthdayestimate( U_k, 8\thresholdconst, \frac{ 3\approxerror }{ \iterations + 1 } )$\label{edgeestimation: line 2}
% \State Initialize $ \approxfractionofedges_i \gets 1 $, for all $ i \in [\iterations]$
% \For{$ i = 1 $ to $ \iterations $}
% \State $ \approxfractionofedges_i \gets \mulestimation( U_i, U_{i - 1}, \frac{ \approxerror }{ 12(\iterations + 1) }, \frac{ 1 }{ 16 \iterations })$ \label{edgeestimation: line 5}
% \EndFor
% \State $ \approxedgecount \gets \approxedgecount_b / { \prod_{ i = 1 }^{ k } \approxfractionofedges_i }$
% \State \Return $ \approxedgecount $
% \end{algorithmic}
% \end{algorithm}
% %Before proving the correctness of \edgeestimationISC{}, let us first prove the correctness of \mulestimation{} using the Chernoff bound. We provide the proof in \Cref{appendix: concentration proofs}.
% \begin{restatable}[Correctness of \mulestimation{}]{lemma}{mulest}\label{lemma: mulestimation}
% For subsets $ U' \subseteq U \subseteq \vertexset $ and $ \approxerror > 0$ , If $ \fractionofedges( U' ) = \frac{ \edgecount( U' ) }{ \edgecount( U ) } \geq \frac{ 1 }{ 32 } $, then \mulestimation{} outputs $ \approxfractionofedges ( U' ) \in ( 1 \pm \approxerror ) \fractionofedges( U' ) $ with probability at least $ 1 - \confidence $. The number of \iscq{} queries made is $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $.
% \end{restatable}
% \begin{proof}
% Let $ X_j$ be the indicator random variable such that,
% \begin{equation*}
% X_j =
% \begin{cases}
% 1 & \text{If the $ j^{th} $ sampled edge is in $ U' $} \\
% 0 & \text{Otherwise }
% \end{cases}
% \end{equation*}
% Define number of edges sampled that belong to $ U' $ is $ X = \sum_{ j \in [\size{ \samplededges }] } X_j$, where $ \samplededges $ is the multi-set of sampled edges (Line~\ref{addestimation: line 2}) and the expected size of $ X $ is
% \begin{equation*}
% \E\tbrac{ X } = \sum_{ j \in [\size{S}] } \E\tbrac{ X_j } = \size{ S } \cdot \frac{ \edgecount{ (U') }}{ \edgecount{ (U) } } = \size{ S } \cdot \fractionofedges( U' ) \geq \frac{ \size{ \samplededges } }{ 32 }
% \end{equation*}
% By the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), we have
% % \begin{align*}
% % \Pr\tbrac{ \approxfractionofedges ( U' ) \in ( 1 \pm \approxerror ) \fractionofedges( U ) } &= \Pr\tbrac{ X \in ( 1 \pm \approxerror ) \E[X] } \\
% % % &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \E[X] }{ 3 } }\\
% % &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \size{ \samplededges } }{ 96 } }
% % &\text{$\fbrac{ \E\tbrac{ X } \geq \size{ \samplededges } / 32 }$}\\
% % &\geq 1 - \confidence &\text{$\fbrac{ \size{ \samplededges } = O\fbrac{ \frac{ \log ( \confidence^{-1} ) }{ \approxerror^2 }} }$}
% % \end{align*}
% \[
% \Pr\tbrac{ \approxfractionofedges(U') \in (1\pm\approxerror)\fractionofedges(U') } = \Pr\tbrac{ X \in (1\pm\approxerror)\E[X] } \ge 1 - 2\exp\fbrac{ - \frac{\approxerror^2 \size{\samplededges}}{96} } \ge 1 - \confidence,
% \]
% where the last step uses $\size{\samplededges}=O\left(\log (\delta^{-1})/\varepsilon^2\right)$.
% \end{proof}
% % \textcolor{purple}{Gopinath: It should be 1/32?}
% % \begin{proof} Refer to Appendix~\ref{appendix: concentration proofs}
% % Let $ X_j$ be the indicator random variable such that,
% % \begin{equation*}
% % X_j =
% % \begin{cases}
% % 1 & \text{If the $ j^{th} $ sampled edge is in $ U' $} \\
% % 0 & \text{Otherwise }
% % \end{cases}
% % \end{equation*}
% % Define number of edges sampled that belong to $ U' $ is $ X = \sum_{ j \in [\size{ \samplededges }] } X_j$, where $ \samplededges $ is the multi-set of sampled edges (Line~\ref{addestimation: line 2}) and the expected size of $ X $ is
% % \begin{equation*}
% % \E\tbrac{ X } = \sum_{ j \in [\size{S}] } \E\tbrac{ X_j } = \size{ S } \cdot \frac{ \edgecount{ (U') }}{ \edgecount{ (U) } } = \size{ S } \cdot \fractionofedges( U' ) \geq \frac{ \size{ \samplededges } }{ 32 }
% % \end{equation*}
% % By the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), we have
% % \[
% % \Pr\tbrac{ \approxfractionofedges(U')
% % \in (1\pm\approxerror)\fractionofedges(U') }
% % = \Pr\tbrac{ X \in (1\pm\approxerror)\E[X] }
% % \ge 1 - 2\exp\fbrac{ - \frac{\approxerror^2 \size{\samplededges}}{96} }
% % \ge 1 - \confidence,
% % \]
% % where the last step uses $\size{\samplededges}=O\left(\log^{\delta^-1}/\varepsilon^2\right)$.
% % \end{proof}
% % \begin{restatable}[]{lemma}{}
% % For any $ \approxerror > 0$, \edgeestimationISC{} outputs $ \approxedgecount \in (1 \pm \approxerror) \edgecount$ with probability at least $ \nicefrac{ 3 }{ 4 } $. The number of \iscq{} queries made is $ O\fbrac{ \approxerror^{-2} \log ( \log \edgecount ) \log^3 \edgecount }$ in expectation.
% % \end{restatable}
% Now, we prove the guarantees of \edgeestimationISC{}
% In particular, we prove the correctness of \edgeestimationISC{} and its query complexity in \Cref{lem:final-correct} and \Cref{lem:final:query}, respectively.
% \begin{lemma}\label{lem:final:query}
% The expected query complexity of \edgeestimationISC{} is $O(\log^3 m \cdot \log \log m)$.
% \end{lemma}
% \begin{proof}
% From Lemma~\ref{lemma: mulestimation}, the number of \iscq{} queries made by \mulestimation{} in $j$-th iteration is
% \begin{align*}
% O\fbrac{ \frac{\log(8 \iterations )}{ \approxerror'^2 } } \text{where } \approxerror' = \frac{ \approxerror }{ 12( \iterations + 1 )}
% \end{align*}
% Summing over $ \iterations $ iterations, we get
% \begin{align*}
% O\fbrac{ \sum_{ j = 1 }^{ \iterations } \frac{ \log( \iterations ) }{ \approxerror'^2 } } = O\fbrac{\iterations^3 \frac{ \log \iterations }{ \approxerror^2 } }
% \end{align*}
% From Lemma~\ref{lemma: subsequence}, the number of \iscq{} queries made by \subsequence{} is $ O( \iterations^2 ) $. Finally, from Lemma~\ref{lemma: birthdayestimate}, the number of \iscq{} queries made by \birthdayestimate{} is $O\fbrac{ { \sqrt{8\thresholdconst} \cdot \iterations }/{ \approxerror}}= O(\iterations/\approxerror)$. Combining all contributions, the total number of \iscq{} queries made by \edgeestimationISC{} is dominated by \mulestimation{}, and hence is, $ O( \iterations^3 \log \iterations / \approxerror^2)$. From \Cref{lemma: kappa},
% $$
% \E[\iterations^3 \log \iterations] = O(\log^3 \edgecount \log \log \edgecount)
% $$
% which bounds the expected query complexity.
% \end{proof}
% \begin{lemma}\label{lem:final-correct}
% In \edgeestimationISC{}, the estimate $\widehat{m}$ is an
% $(1 \pm \approxerror)$-approximation to $m$ with probability at least $3/4$.
% \end{lemma}
% \begin{proof}
% Let $\mathcal{E}_1 = \subseqevent \cap \birthdayevent$, where $\subseqevent$ be the event that \subsequence{} outputs an approximation-preserving sequence $U_1, U_2, \ldots, U_\iterations$ with $ \edgecount( U_\iterations ) \leq 8 \thresholdconst$. From \Cref{lemma: subsequence} we have $\Pr[\subseqevent] \geq 4/5$.
% $ \birthdayevent $ be the event that \birthdayestimate{} outputs an $(1 \pm \approxerror')$ approximation of $ \edgecount( U_\iterations ) $ where $ \approxerror' = \approxerror / 12 ( \iterations + 1 )$, from \Cref{lemma: birthdayestimate}. From \Cref{lemma: birthdayestimate}, we have $\Pr[\birthdayevent \mid \subseqevent] \geq 15/16$. Therefore,
% \begin{equation*}
% \Pr[\mathcal{E}_1] = \Pr[\subseqevent]\Pr[\birthdayevent \mid \subseqevent] \geq 1 - \fbrac{ \frac{1}{5} + \frac{1}{16} } \geq \frac{59}{80}
% \end{equation*}
% Let $\mathcal{E}_2$ be the event that \mulestimation{} outputs an $(1 \pm \approxerror')$ approximation of $\fractionofedges_i$ for all $i \in [\iterations]$ . Therefore
% \begin{equation*}
% \Pr[\mathcal{E}_2 \mid \mathcal{E}_1] = 1 - \sum_{ j = 1 }^{ \iterations } \frac{ 1 }{ 16 \iterations } \geq \frac{15}{16}.
% \end{equation*}
% %
% Hence,
% \begin{align*}
% \Pr[ \mathcal{E}_1 \cap \mathcal{E}_2] &= \Pr[\mathcal{E}_1] \Pr[\mathcal{E}_2 \mid \mathcal{E}_1] \geq 1 - \fbrac{ \frac{21}{80} + \frac{ 1 }{ 16 }} \geq \frac{2}{3} .
% \end{align*}
% In what follows, we assume that when both $\mathcal{E}_1$ and $\mathcal{E}_2$ hold. Since, $\Pr[ \mathcal{E}_1 \cap \mathcal{E}_2] \geq 2/3$, it is enough to argue that \edgeestimationISC{} returns an $(1 \pm \approxerror)$-approximation to $m$ when both $\mathcal{E}_1$ and $\mathcal{E}_2$ hold.
% In line~\ref{edgeestimation: line 5} of \edgeestimationISC{}, $ \approxfractionofedges_i $ is the output of $\mulestimation( U_i, U_{ i - 1 }, \approxerror', 1/16\iterations ) $ where $ \approxerror' = \frac{ \approxerror }{ 12 (\iterations + 1) } $. From Lemma~\ref{lemma: mulestimation}. We have,
% \begin{align*}
% \fbrac{ 1 - \approxerror' } \cdot \fractionofedges_i \leq \approxfractionofedges_i \leq \fbrac{ 1 + \approxerror'} \cdot \fractionofedges_i
% \end{align*}
% where, $ \fractionofedges_i = \frac{ \edgecount( U_i ) }{ \edgecount( U_{ i - 1 } ) } $. Using the fact that for any $\approxerror \in (0, 1/3)$, we have
% \begin{equation*}
% 1 - 3\approxerror \leq \frac{1}{\fbrac{1 + \approxerror}} \;\text{ and }\; \frac{1}{\fbrac{1 - \approxerror}} \leq 1 + 3\approxerror
% \end{equation*}
% We get,
% \begin{align*}
% \fbrac{ 1 - 3\approxerror' } \cdot \frac{ 1 }{\fractionofedges_i} \leq \frac{ 1 }{\approxfractionofedges_i} \leq \fbrac{ 1 + 3\approxerror'} \cdot \frac{ 1 }{\fractionofedges_i}
% \end{align*}
% Combining the output of \mulestimation{} over the $ \iterations $ iterations, we get,
% \begin{align} \label{ineq: mulplicateest}
% \fbrac{ 1 - 3\approxerror' }^{ \iterations } \cdot \frac{ 1 }{ \prod_{i = 1 }^{ \iterations } \fractionofedges_i} \leq \frac{ 1 }{ \prod_{i = 1}^{ \iterations } \approxfractionofedges_i} \leq \fbrac{ 1 + 3\approxerror'}^{ \iterations } \cdot \frac{ 1 }{ \prod_{i = 1}^{ \iterations } \fractionofedges_i}
% \end{align}
% In line~\ref{edgeestimation: line 2} of \edgeestimationISC{}, $ \approxedgecount_b $ is the output of $ \birthdayestimate( U_\iterations, 8\thresholdconst, 3 \approxerror / ( \iterations + 1 ) ) $, Therefore, from Lemma~\ref{lemma: birthdayestimate} we get,
% \begin{align} \label{ineq: birthdayest}
% \fbrac{ 1 - 3\approxerror' } \cdot \edgecount( U_k ) \leq \approxedgecount_b \leq \fbrac{ 1 + 3\approxerror' } \cdot \edgecount( U_k )
% \end{align}
% Combining~\eqref{ineq: mulplicateest} and \eqref{ineq: birthdayest} we get,
% \begin{align}\label{ineq: intermediate edge est}
% \fbrac{ 1 - 3\approxerror' }^{ k + 1 } \cdot \frac{ \edgecount( U_k ) }{ \prod_{i = 1 }^{ \iterations } \fractionofedges_i} \leq \frac{ \approxedgecount_b }{ \prod_{i = 1}^{ \iterations } \approxfractionofedges_i} \leq \fbrac{ 1 + 3\approxerror'}^{ k + 1 } \cdot \frac{ \edgecount( U_k ) }{ \prod_{i = 1}^{ \iterations } \fractionofedges_i}
% \end{align}
% Since, $ \fractionofedges_i = \frac{ \edgecount( U_i) }{ \edgecount( U_{i - 1} )}$ and $ U_0 = \vertexset $. Therefore, $ \frac{ \edgecount( U_k ) }{ \prod_{i = 1 }^{ \iterations } \fractionofedges_i} = \edgecount( U_0 ) = \edgecount $ and~\eqref{ineq: intermediate edge est} becomes,
% \begin{align}\label{ineq: intermediate edge est-2}
% \fbrac{ 1 - 3\approxerror' }^{ k + 1 } \cdot \edgecount \leq \approxedgecount \leq \fbrac{ 1 + 3\approxerror'}^{ k + 1 } \cdot \edgecount
% \end{align}
% % \debarshi{Use Weirstrass Inequalities. See~\cite{Doerr_2019}, 1.4.10, 1.4.11.}
% For the upper estimation, we need to show that $ \fbrac{ 1 + 3\approxerror'}^{ \iterations + 1 } \leq ( 1 + \approxerror) $
% Observe that,
% \begin{align*}
% \fbrac{1 + 3\approxerror'}^{ \iterations + 1 }
% &\leq \exp\fbrac{ 3 \approxerror' ( \iterations + 1 ) } &\text{$\fbrac{1 + x \leq \exp(x), \text{ for any } x > 0}$}\\
% &= \exp\fbrac{ \frac{ \approxerror }{ 4 }} &\text{$\fbrac{ \approxerror' = \frac{ \approxerror }{ 12 ( \iterations + 1 )} }$}\\
% &\leq ( 1 + \approxerror) &\text{$\fbrac{ \frac{ x }{ 4 } \leq \log \fbrac{1 + x}, \text{ for any } x \in (0, 1) } $}
% \end{align*}
% For the lower estimation, we need to show that $ \fbrac{1 - 3\approxerror'}^{ \iterations + 1 } \geq ( 1 - \approxerror) $. Observe that,
% \begin{align*}
% \fbrac{1 - 3\approxerror'}^{ \iterations + 1 } &\geq \fbrac{ 1 - 3\approxerror' ( \iterations + 1 ) } &\text{$\fbrac{ (1 + x)^{n} \geq 1 + nx \text{ for any } x > -1 }$}\\
% &= 1 - \frac{ \approxerror }{ 4 } &\text{$\fbrac{ \approxerror' = \frac{ \approxerror }{12 (\iterations + 1)} }$}\\
% &\geq \fbrac{1 - \approxerror}
% \end{align*}
% % Therefore, from~\eqref{ineq: intermediate edge est-2}, we have
% % \begin{align*}
% % \fbrac{ 1 - \approxerror } \edgecount \leq \approxedgecount \leq \fbrac{ 1 + \approxerror'} \edgecount
% % \end{align*}
% % \
% % Now we finally conclude the proof of \Cref{Thm: Main Upper Bound}.
% After simplification using the value of $\approxerror'=\approxerror/12(\iterations +1)$,
% \begin{align*}
% \fbrac{ 1 - \approxerror } \edgecount \leq \approxedgecount \leq \fbrac{ 1 + \approxerror'} \edgecount
% \end{align*}
% \end{proof}
% % \begin{proof}[Proof of \Cref{Thm: Main Upper Bound}]
% \Cref{Thm: Main Upper Bound} follows from the algorithm \edgeestimationISC{} (\Cref{alg: edgeestimationISC}) along with \Cref{lem:final:query} an \Cref{lem:final-correct}.
% \end{proof}
% \noindent\textbf{Query Complexity:}
%\textcolor{purple}{Gopinath: I guess here also we need to talk about the expectation of $\kappa^3 \log \kappa$.}
\section{Full algorithm}\label{section: full algorithm}
In this section, we present the full algorithm for estimating the number of edges using \iscq{} queries and prove our main result stated below.
\MainUB*
\edgeestimationISC{} (\Cref{alg: edgeestimationISC}) begins by calling \subsequence{} which outputs an approximation preserving sequence $U_1 , \ldots, U_\iterations $, where each $ \fractionofedges_i = \edgecount( U_i ) / \edgecount( U_{i-1} ) $ is bounded below by a constant. This ensures that each ratio can be estimated efficiently. To estimate each $ \fractionofedges_i $, we use the subroutine \mulestimation{} which returns an $ (1 \pm \approxerror)$-approximation of $ \fractionofedges_i$ . The procedure \mulestimation{} is described in \Cref{alg: mulestimation} and its guarantee is established in \Cref{lemma: mulestimation}.
\begin{algorithm}[H]
\caption{\mulestimation($ U', U, \approxerror, \confidence $)}\label{alg: mulestimation}
\begin{algorithmic}[1]
\Require $ U' \subseteq U$ and \iscq{} access to the subset $ U \subseteq \vertexset $.
\Ensure $\approxfractionofedges ( U' )$
\State Uniformly and Independently sample $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges.
\State \Return $ \approxfractionofedges ( U' ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U' } $.
\end{algorithmic}
\end{algorithm}
% \begin{algorithm}
% \setcounter{AlgoLine}{0}
% \SetAlgoNoLine
% \LinesNumbered
% \SetAlgoNoEnd
% \caption{ \mulestimation($ U', U, \approxerror, \confidence $)}
% \label{alg: mulestimation}
% \KwIn{$ U' \subseteq U$ and \iscq{} access to the subset $ U \subseteq \vertexset $.}
% \KwOut{A subset $U' \subseteq U$.}
% Uniformly and Independently sample $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $ random edges from $ U $ and let $ \samplededges $ be the multi-set of sampled edges.
% \KwRet $ \approxfractionofedges ( U' ) = \frac{ 1 }{ \size{ \samplededges } } \sum_{ \edge \in \samplededges} \indicator_{ \edge \in U' } $.
% \end{algorithm}
% Now,
% \begin{equation*}
% \edgecount = \frac{ \edgecount( U_\iterations ) }{ \prod_{i = 1}^\iterations \fractionofedges_i }
% \end{equation*}
\begin{algorithm}[H]
\caption{\edgeestimationISC($ \vertexset, \approxerror $)}\label{alg: edgeestimationISC}
\begin{algorithmic}[1]
\Require \iscq{} access to any subset $ U \subseteq \vertexset $.
\Ensure An estimate of $ \approxedgecount $
\State $ U_1, U_2, \cdots, U_\iterations \gets \subsequence{}(\vertexset)$
\State $ \approxedgecount_b \gets \birthdayestimate( U_\iterations, 8\thresholdconst, \frac{\approxerror }{ 3 } )$\label{edgeestimation: line 2}
\State Initialize $ \approxfractionofedges_i \gets 1 $, for all $ i \in [\iterations]$
\For{$ i = 1 $ to $ \iterations $}
\State $ \approxfractionofedges_i \gets \mulestimation( U_i, U_{i - 1}, \frac{ \approxerror }{ 200 \sqrt{\iterations} }, \frac{ 1 }{ 16 \iterations })$ \label{edgeestimation: line 5}
\EndFor
\State $ \approxedgecount \gets \approxedgecount_b / { \prod_{ i = 1 }^{ \iterations } \approxfractionofedges_i }$
\State \Return $ \approxedgecount $
\end{algorithmic}
\end{algorithm}
%Before proving the correctness of \edgeestimationISC{}, let us first prove the correctness of \mulestimation{} using the Chernoff bound. We provide the proof in \Cref{appendix: concentration proofs}.
% \begin{restatable}[Correctness of \mulestimation{}]{lemma}{mulest}\label{lemma: mulestimation}
% For subsets $ U' \subseteq U \subseteq \vertexset $ and $ \approxerror > 0$ , If $ \fractionofedges( U' ) = \frac{ \edgecount( U' ) }{ \edgecount( U ) } \geq \frac{ 1 }{ 32 } $, then \mulestimation{} outputs $ \approxfractionofedges ( U' ) \in ( 1 \pm \approxerror ) \fractionofedges( U' ) $ with probability at least $ 1 - \confidence $. The number of \iscq{} queries made is $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $.
% \end{restatable}
\begin{restatable}[Correctness of \mulestimation{}]{lemma}{mulest}\label{lemma: mulestimation}
For subsets $ U' \subseteq U \subseteq \vertexset $ and $ \approxerror > 0$ , If $ \fractionofedges( U' ) = \frac{ \edgecount( U' ) }{ \edgecount( U ) } \geq \frac{ 1 }{ 32 } $, then \mulestimation{} outputs $ \approxfractionofedges ( U' ) = ( 1 + x ) \fractionofedges( U' ) $, where $\E[x] = 0 $ and $x \in (-\approxerror, \approxerror)$ with probability at least $ 1 - \confidence $. The number of \iscq{} queries made is $ O\fbrac{ \log ( \confidence^{-1} ) / \approxerror^2 } $.
\end{restatable}
\begin{proof}
Let $ X_j$ be the indicator random variable such that,
\begin{equation*}
X_j =
\begin{cases}
1 & \text{If the $ j^{th} $ sampled edge is in $ U' $} \\
0 & \text{Otherwise }
\end{cases}
\end{equation*}
Define number of edges sampled that belong to $ U' $ is $ X = \sum_{ j \in [\size{ \samplededges }] } X_j$, where $ \samplededges $ is the multi-set of sampled edges (Line~\ref{addestimation: line 2}) and the expected size of $ X $ is
\begin{equation*}
\E\tbrac{ X } = \sum_{ j \in [\size{S}] } \E\tbrac{ X_j } = \size{ S } \cdot \frac{ \edgecount{ (U') }}{ \edgecount{ (U) } } = \size{ S } \cdot \fractionofedges( U' ) \geq \frac{ \size{ \samplededges } }{ 32 }
\end{equation*}
Therefore,
\begin{align*}
\E[x] = \E\tbrac{ \frac{\approxfractionofedges ( U' )}{\fractionofedges( U' )}} -1 = \frac{\E[X]}{\size{ S } \fractionofedges( U' )} -1 = 0
\end{align*}
By the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), we have
% \begin{align*}
% \Pr\tbrac{ \approxfractionofedges ( U' ) \in ( 1 \pm \approxerror ) \fractionofedges( U ) } &= \Pr\tbrac{ X \in ( 1 \pm \approxerror ) \E[X] } \\
% % &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \E[X] }{ 3 } }\\
% &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \size{ \samplededges } }{ 96 } }
% &\text{$\fbrac{ \E\tbrac{ X } \geq \size{ \samplededges } / 32 }$}\\
% &\geq 1 - \confidence &\text{$\fbrac{ \size{ \samplededges } = O\fbrac{ \frac{ \log ( \confidence^{-1} ) }{ \approxerror^2 }} }$}
% \end{align*}
\[
\Pr[x \in (-\approxerror, \approxerror)] = \Pr\tbrac{ \approxfractionofedges(U') \in (1\pm\approxerror)\fractionofedges(U') } = \Pr\tbrac{ X \in (1\pm\approxerror)\E[X] } \ge 1 - 2\exp\fbrac{ - \frac{\approxerror^2 \size{\samplededges}}{96} } \ge 1 - \confidence,
\]
where the last step uses $\size{\samplededges}=O\left(\log (\delta^{-1})/\varepsilon^2\right)$.
\end{proof}
% \textcolor{purple}{Gopinath: It should be 1/32?}
% \begin{proof} Refer to Appendix~\ref{appendix: concentration proofs}
% Let $ X_j$ be the indicator random variable such that,
% \begin{equation*}
% X_j =
% \begin{cases}
% 1 & \text{If the $ j^{th} $ sampled edge is in $ U' $} \\
% 0 & \text{Otherwise }
% \end{cases}
% \end{equation*}
% Define number of edges sampled that belong to $ U' $ is $ X = \sum_{ j \in [\size{ \samplededges }] } X_j$, where $ \samplededges $ is the multi-set of sampled edges (Line~\ref{addestimation: line 2}) and the expected size of $ X $ is
% \begin{equation*}
% \E\tbrac{ X } = \sum_{ j \in [\size{S}] } \E\tbrac{ X_j } = \size{ S } \cdot \frac{ \edgecount{ (U') }}{ \edgecount{ (U) } } = \size{ S } \cdot \fractionofedges( U' ) \geq \frac{ \size{ \samplededges } }{ 32 }
% \end{equation*}
% By the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), we have
% \[
% \Pr\tbrac{ \approxfractionofedges(U')
% \in (1\pm\approxerror)\fractionofedges(U') }
% = \Pr\tbrac{ X \in (1\pm\approxerror)\E[X] }
% \ge 1 - 2\exp\fbrac{ - \frac{\approxerror^2 \size{\samplededges}}{96} }
% \ge 1 - \confidence,
% \]
% where the last step uses $\size{\samplededges}=O\left(\log^{\delta^-1}/\varepsilon^2\right)$.
% \end{proof}
% \begin{restatable}[]{lemma}{}
% For any $ \approxerror > 0$, \edgeestimationISC{} outputs $ \approxedgecount \in (1 \pm \approxerror) \edgecount$ with probability at least $ \nicefrac{ 3 }{ 4 } $. The number of \iscq{} queries made is $ O\fbrac{ \approxerror^{-2} \log ( \log \edgecount ) \log^3 \edgecount }$ in expectation.
% \end{restatable}
Now, we prove the guarantees of \edgeestimationISC{}
In particular, we prove the correctness of \edgeestimationISC{} and its query complexity in \Cref{lem:final-correct} and \Cref{lem:final:query}, respectively.
\begin{lemma}\label{lem:final:query}
The expected query complexity of \edgeestimationISC{} is $O(\log^2 m \cdot \log \log m)$.
\end{lemma}
\begin{proof}
From Lemma~\ref{lemma: mulestimation}, the number of \iscq{} queries made by \mulestimation{} in $j$-th iteration is
\begin{align*}
O\fbrac{ \frac{\log(16 \iterations )}{ \approxerror'^2 } } \text{where } \approxerror' = \frac{ \approxerror }{ 200 \sqrt{\iterations} }
\end{align*}
Summing over $ \iterations $ iterations, we get
\begin{align*}
O\fbrac{ \sum_{ j = 1 }^{ \iterations } \frac{ \log( \iterations ) }{ \approxerror'^2 } } = O\fbrac{\iterations^2 \frac{ \log \iterations }{ \approxerror^2 } }
\end{align*}
From Lemma~\ref{lemma: subsequence}, the number of \iscq{} queries made by \subsequence{} is $ O( \iterations^2 ) $. Finally, from Lemma~\ref{lemma: birthdayestimate}, the number of \iscq{} queries made by \birthdayestimate{} is $O\fbrac{ { \sqrt{8\thresholdconst} }/{ \approxerror}}= O(1/\approxerror)$. Combining all contributions, the total number of \iscq{} queries made by \edgeestimationISC{} is dominated by \mulestimation{}, and hence is, $ O( \iterations^2 \log \iterations / \approxerror^2)$. From \Cref{lemma: kappa},
$$
\E[\iterations^2 \log \iterations] = O(\log^2 \edgecount \log \log \edgecount)
$$
which bounds the expected query complexity.
\end{proof}
\begin{lemma}\label{lem:final-correct}
In \edgeestimationISC{}, the estimate $\widehat{m}$ is an
$(1 \pm \approxerror)$-approximation to $m$ with probability at least $3/4$.
\end{lemma}
\begin{proof}
Let $\mathcal{E}_1 = \subseqevent \cap \birthdayevent$, where $\subseqevent$ be the event that \subsequence{} outputs an approximation-preserving sequence $U_1, U_2, \ldots, U_\iterations$ with $ \edgecount( U_\iterations ) \leq 8 \thresholdconst$. From \Cref{lemma: subsequence} we have $\Pr[\subseqevent] \geq 4/5$.
$ \birthdayevent $ be the event that \birthdayestimate{} outputs an $(1 \pm \approxerror/3)$ approximation of $ \edgecount( U_\iterations ) $ from \Cref{lemma: birthdayestimate}. Also, from \Cref{lemma: birthdayestimate}, we have $\Pr[\birthdayevent \mid \subseqevent] \geq 15/16$. Therefore,
\begin{equation*}
\Pr[\mathcal{E}_1] = \Pr[\subseqevent]\Pr[\birthdayevent \mid \subseqevent] \geq 1 - \fbrac{ \frac{1}{5} + \frac{1}{16} } \geq \frac{59}{80}
\end{equation*}
Let $\mathcal{E}_2$ be the event that \mulestimation$\left( U_i, U_{i - 1}, \varepsilon', { 1 }/{ 16 \iterations}\right)$ outputs $\widehat{p_i}=(1 \pm x_i/4)p_i$ such that $\E[x_i]=0$ and $x_i \in \fbrac{ -4{\approxerror'}, 4\approxerror'}$ for all $i \in [\iterations]$, where $ \approxerror' = \frac{ \approxerror }{ 200 \sqrt{ \iterations }}$ . Therefore, using Lemma~\ref{lemma: mulestimation} and applying the union bound over all $i \in [\iterations]$,
\begin{equation*}
\Pr[\mathcal{E}_2 \mid \mathcal{E}_1] = 1 - \sum_{ j = 1 }^{ \iterations } \frac{ 1 }{ 16 \iterations } \geq \frac{15}{16}.
\end{equation*}
%
% Hence,
% \begin{align*}
% \Pr[ \mathcal{E}_1 \cap \mathcal{E}_2] &= \Pr[\mathcal{E}_1] \Pr[\mathcal{E}_2 \mid \mathcal{E}_1] \geq 1 - \fbrac{ \frac{21}{80} + \frac{ 1 }{ 16 }}
% \end{align*}
% In Line~\ref{edgeestimation: line 5} of \edgeestimationISC{}, $ \approxfractionofedges_i $ is the output of $\mulestimation( U_i, U_{ i - 1 }, \approxerror', 1/16\iterations ) $. From Lemma~\ref{lemma: mulestimation}. We have,
% \begin{align*}
% \approxfractionofedges_i = \fbrac{1 + \frac{x_i}{4}} \cdot \fractionofedges_i
% \end{align*}
% where, $ \E \tbrac{x_i} = 0 $ and $ x_i \in \fbrac{ -\frac{\approxerror'}{50}, \frac{\approxerror'}{50}}$ for each $i \in [\iterations]$.
Let $\mathcal{E}_3$ be the event that $\abs{\sum_{i = 1 }^\iterations x_i} \leq \approxerror/9$.
By Hoeffding's bound (Lemma~\ref{lemma: hoeffding}),
\begin{align}
\Pr\!\left[
\mathcal{E}_3 \mid \mathcal{E}_1 \cap \mathcal{E}_2
\right]
&=
1-\Pr\!\left[
\left|
\sum_{i=1}^{\iterations} x_i
\right|
>
\frac{\approxerror}{9}
\,\middle|\,
\mathcal{E}_1 \cap \mathcal{E}_2
\right]
\nonumber \\
&\ge 1-
2\exp\!\left(
-
\frac{
2\approxerror^2
}{
81 \sum_{i=1}^{\iterations}
\left(
\frac{2\approxerror}{50\sqrt{\iterations}}
\right)^2
}
\right)
\ge 1-\frac{1}{e^{12}}. \label{ineq: hoeff}
\end{align}
% Therefore $\Pr\tbrac{\mathcal{E}_3 \mid \mathcal{E}_1 \cap \mathcal{E}_2} \geq 1 - e^{-12}$.
Hence,
\begin{align*}
\Pr[ \mathcal{E}_1 \cap \mathcal{E}_2 \cap \mathcal{E}_3] &= \Pr[\mathcal{E}_1] \cdot \Pr[\mathcal{E}_2 | \mathcal{E}_1] \cdot \Pr[\mathcal{E}_3 \mid \mathcal{E}_1 \cap \mathcal{E}_2] \geq 1 - \fbrac{ \frac{21}{80} + \frac{ 1 }{ 16 } + \frac{ 1 }{ e^{12} } } \geq \frac{2}{3}
\end{align*}
In what follows, we assume that all of the events $\mathcal{E}_1$, $\mathcal{E}_2$ and $\mathcal{E}_3$ occur. Since, $\Pr[ \mathcal{E}_1 \cap \mathcal{E}_2 \cap \mathcal{E}_3] \geq 2/3$, it is enough to argue that \edgeestimationISC{} returns an $(1 \pm \approxerror)$-approximation to $m$ when all $\mathcal{E}_1$, $\mathcal{E}_2$ and $\mathcal{E}_3$ holds.
Combining the output of \mulestimation{} over the $ \iterations $ iterations, we get,
\begin{align*} %\label{ineq: mulplicateest}
\prod_{i = 1}^{ \iterations } \approxfractionofedges_i = \prod_{i = 1 }^{ \iterations } \fractionofedges_i \cdot \prod_{i = 1}^{ \iterations } \fbrac{ 1 + \frac{x_i}{4} }
\end{align*}
Using the fact that for each $x_i \in (0,1)$, it can be shown that
\begin{align*}%\label{ineq: binomial bound}
1 - \frac{ \sum_{i = 1}^{ \iterations } x_i }{ 4 } \leq \prod_{i = 1}^{ \iterations } \fbrac{ 1 + \frac{x_i}{4} } \leq 1 + \sum_{i = 1}^{ \iterations } x_i
\end{align*}
\noindent Since event $\mathcal{E}_3$ occurs,
\begin{align*}
\prod_{i = 1 }^{ \iterations } \fractionofedges_i \cdot \fbrac{1 - \frac{\approxerror}{36}} \leq \prod_{i = 1}^{ \iterations } \approxfractionofedges_i \leq \prod_{i = 1 }^{ \iterations } \fractionofedges_i \cdot \fbrac{1 + \frac{\approxerror}{9}}
\end{align*}
Using the fact that for any $\approxerror \in (0, 1/3)$, we have $1 - 3\approxerror \leq (1 + \approxerror)^{-1}$ and $ (1 - \approxerror)^{-1} \leq 1 + 3\approxerror $. We get,
% \begin{equation*}
% 1 - 3\approxerror \leq \frac{1}{\fbrac{1 + \approxerror}} \;\text{ and }\; \frac{1}{\fbrac{1 - \approxerror}} \leq 1 + 3\approxerror
% \end{equation*}
\begin{align}\label{ineq: xyz}
\fbrac{1 - \frac{\approxerror}{3}} \cdot \frac{1}{\prod_{i = 1 }^{ \iterations } \fractionofedges_i}
\leq \frac{1}{\prod_{i = 1}^{ \iterations } \approxfractionofedges_i} \leq \frac{1}{\prod_{i = 1 }^{ \iterations } \fractionofedges_i} \cdot \fbrac{1 + \frac{\approxerror}{12}}
\end{align}
Since $\mathcal{E}_1$ occurs, i.e., $\birthdayevent$ occurs, $ \approxedgecount_b$ satisfies the following:
\begin{align} \label{ineq: birthdayest}
\fbrac{ 1 - \frac{\approxerror}{3} } \cdot \edgecount( U_\iterations ) \leq \approxedgecount_b \leq \fbrac{ 1 + \frac{\approxerror}{3} } \cdot \edgecount( U_\iterations )
\end{align}
Combining~\eqref{ineq: xyz} and \eqref{ineq: birthdayest}, we get,
\begin{align*}%\label{ineq: intermediate edge est}
\fbrac{ 1 - \frac{\approxerror}{3} }^2 \cdot \frac{ \edgecount( U_\iterations ) }{ \prod_{i = 1 }^{ \iterations } \fractionofedges_i}
\leq
\widehat{m}=\frac{ \approxedgecount_b }{ \prod_{i = 1}^{ \iterations } \approxfractionofedges_i}
\leq
\fbrac{ 1 + \frac{\approxerror}{12}}\fbrac{ 1 + \frac{\approxerror}{3}} \cdot \frac{ \edgecount( U_\iterations ) }{ \prod_{i = 1}^{ \iterations } \fractionofedges_i}
\end{align*}
Since, $ \fractionofedges_i = \frac{ \edgecount( U_i) }{ \edgecount( U_{i - 1} )}$ for each $i \in [\kappa]$ and $ U_0 = \vertexset $, $ \frac{ \edgecount( U_\iterations ) }{ \prod_{i = 1 }^{ \iterations } \fractionofedges_i} = \edgecount( U_0 ) = \edgecount $. So, simplifying the above expression, we have
\begin{align*}
\fbrac{ 1 - \approxerror } \edgecount \leq \approxedgecount \leq \fbrac{ 1 + \approxerror} \edgecount
\end{align*}
\end{proof}
\section{Concentration Inequalities}
We use the following concentration bounds.
\begin{lemma}[Markkov's inequlaity, see~\citep{Mitzenmacher_Upfal_2005}]\label{Markov's Inequality}
Let $X$ be a random variable that assumes only nonnegative values. Then, for all $ a > 0 $
\begin{equation*}
\Pr \tbrac{ X \geq a} \leq \frac{ \E [ X ] }{ a }.
\end{equation*}
\end{lemma}
\begin{lemma}[Chebyshev’s inequality, see~\citep{Mitzenmacher_Upfal_2005}]\label{lemma: chebyshev}
Let $ X $ be a random variable and $ a > 0 $, Then
\begin{equation*}
\Pr[ \abs{X - \E[X]} > a] \leq \frac{ \Var( X ) }{ a^2 }.
\end{equation*}
\end{lemma}
\begin{lemma}[Multiplicative Chernoff bound, see~\citep{Mitzenmacher_Upfal_2005}]\label{lemma: Multiplicative Chernoff Bound}
Let $X_1,X_2,...,X_t$ be i.i.d. random variables
where $\Pr[X_i = 1] = p$ and $\Pr[X_i = 0] = 1-p$ for all $i \in [t]$, and $X = \sum_{i \in [t]} X_i$. Then, we have
\begin{align*}
% \Pr[X \geq (1+\approxerror) \Exp\tbrac{X}] &\leq \exp{\fbrac{-\frac{\Exp\tbrac{X}\approxerror^2}{3}}} & 0 \leq \approxerror <1\\
\Pr \left[ X \leq (1-\approxerror) \E\tbrac{X} \right] &\leq \exp{\fbrac{-\frac{\approxerror^2\E\tbrac{X}}{3}}} & 0 \leq \approxerror <1\\
% \Pr[\abs{X - \Exp\tbrac{X}} \geq \approxerror \Exp\tbrac{X}] &\leq 2\exp{\fbrac{-\frac{\Exp\tbrac{X}\approxerror^2}{3}}} & 0 \leq \approxerror <1\\
\Pr \left[ X \geq (1+\approxerror) \E\tbrac{X} \right] &\leq \exp{\fbrac{-\frac{\approxerror^2\E\tbrac{X}}{2}}} &
\approxerror \geq 0
\end{align*}
\end{lemma}
\begin{lemma}[Hoeffding's bound, see~\citep{Mitzenmacher_Upfal_2005}]\label{lemma: hoeffding}
Let $\X_1, \X_2, \ldots, \X_t$ be independent random variables such that for all $i \in [t]$, $\E[\X] = \mu $ and $\Pr[a \leq \X_i \leq b] = 1$. Then
\begin{equation*}
\Pr \tbrac{ \abs{ \frac{ 1 }{ t } \sum_{i = 1}^t \X_i - \mu } \geq \approxerror } \leq 2 \exp{ \fbrac{ - \frac{ 2t\approxerror^2 }{ (b - a)^2 } } }.
\end{equation*}
\end{lemma}
\begin{lemma}[See~\citep{martingle06}]\label{lemma: martingle inequality}
Let f be any function of $t$ independent random variables $Y_1, \ldots, Y_t$ and let $X_i = \E[f(Y_1, \ldots, Y_t)|Y_1, \ldots, Y_i]$ for $i \in [t]$ and $ X_0 = \E[f(Y_1, \ldots, Y_t)]$. Let $ \mathcal{B} $ be the event that there exists $ i \in [t]$ such that $ \abs{X_i - X_{i-1}} > c_i $ where $c_1, \ldots, c_t$ are some non negative numbers. Let $S = \sum_{i = 1}^t c_i^2$, then
\begin{equation*}
\Pr \tbrac{\abs{X_t - X_0} > s } \leq 2\exp{ \fbrac{-\frac{s^2}{2S}}} + \Pr[ \mathcal{B}].
\end{equation*}
\end{lemma}
% \section{Proof of graph sparsification}\label{appendix: spasification}
% We will need the following lemma to prove the graph sparsification lemma.
% \begin{lemma}\label{lemma: support for sparsification}
% Let $ A $ be a set of $ \ell $ elements. For every element $ t \in A $, one assigns a color independently and uniformly from $ \sbrac{1, 2} $. Let $ C^{(i)} $ be the number of vertices in $ A $ with color $i$, for any $ i \in \sbrac{ 1, 2 }$. Then for $ \confidence \in (0, 1) $, we have
% \begin{equation*}
% \Pr \tbrac{ \abs{ C^{(1)} - C^{(2)} } \geq \sqrt{ { 6\ell\log (4\confidence^{-1}) } }} \leq \confidence \text{ and } \E\tbrac{ \abs{ C^{(1)} - C^{(2)} } } \leq \sqrt{ \ell }
% \end{equation*}
% \end{lemma}
% \begin{proof}
% For $ t \in A$, let $ X_t $ be the indicator random variable that is $ 1 $ if $ t $ is assigned color $ 1 $ and $ 0 $ otherwise. Then, $ C^{(1)} = \sum_{t \in A} X_t $ and $ \E \tbrac{ C^{(1)} } = \ell/2$. Using Chernoff's inequality (Lemma~\ref{lemma: Multiplicative Chernoff Bound}) with $ \approxerror = \sqrt{ (6 / \ell) \log (4 / \confidence) }$, we have
% \begin{equation}\label{ineq: color 1}
% \Pr \tbrac{ \abs{ C^{(1)} - \frac{ \ell }{ 2 } } > \frac{\approxerror \ell }{ 2 } } \leq 2 \exp \fbrac{ \frac{ 6 \log (4\confidence^{-1})}{ \ell } \cdot \frac{ \ell }{ 6 } } \leq \frac{ \confidence }{ 2 }
% \end{equation}
% Similarly, we can have the same bound for color $2$.
% \begin{equation}\label{ineq: color 2}
% \Pr \tbrac{ \abs{ C^{(2)} - \frac{ \ell }{ 2 } } > \frac{\approxerror \ell }{ 2 } } \leq \frac{ \confidence }{ 2 }
% \end{equation}
% Observe that $ \abs{C^{(1)} - C^{(2)}} \leq \abs{ C^{(1)} - (\approxerror\ell/2) } + \abs{ C^{(2)} - (\approxerror\ell/2)} $. Therefore
% \begin{equation*}
% \Pr \tbrac{ \abs{C^{(1)} - C^{(2)}} > \approxerror \ell } \leq \confidence
% \end{equation*}
% For $ t \in A$, let $ Y_t $ be a random variable that is $ 1 $ if $ t $ is assigned the color $ 1 $ and $ -1 $ otherwise. Then $ C^{(1)} - C^{(2)} = \sum_{ t \in A} Y_t = Y $. We have that $ \Pr \tbrac{ Y_t = 1 } = \Pr \tbrac{ Y_t = 1 } = 1/2$, then
% \begin{equation*}
% \E[ Y_t ] = 0 \text{ and } \E[ Y_t^2] = 1
% \end{equation*}
% By the independence of $ Y_t $, we get
% \begin{equation*}
% \E[ Y^2 ] = \sum_{ t \in A } \E[ Y_t^2 ] + \sum_{ u \neq v} \E[ Y_u Y_v ] = \ell
% \end{equation*}
% Finally, we have $ \E[ \abs{Y} ] \leq \sqrt{ \E [ \abs{ Y }^2 ] } = \sqrt{ \E [ { Y }^2 ] } \leq \sqrt{ \ell }$.
% \end{proof}
% We now prove \Cref{lemma: graph sparsification for k = 2}. The argument follows the
% analysis of \citet{harpeled18}, which we restate here for completeness.
% \graphsparsification*
% \begin{proof}[Proof of Lemma~\ref{lemma: graph sparsification for k = 2}] Let each vertex $t \in \vertexset$ be associated with a random variable
% $Y_t \in \sbrac{1,2}$ such that $\Pr\tbrac{Y_t = 1} = \nicefrac{1}{2}$. For $i \in \sbrac{1,2}$, define $U_i = \sbrac{ t \in \vertexset \mid Y_t = i }$ and
% \begin{equation*}
% f\fbrac{Y_1, \ldots, Y_{\vertexcount}} = \edgecount(U_1) + \edgecount(U_2).
% \end{equation*}
% % For $i \in \sbrac{1,2}$, define
% % $U_i = \sbrac{ t \in \vertexset \mid Y_t = i }$ and \[
% % f\fbrac{Y_1, \ldots, Y_{\vertexcount}}
% % = \edgecount(U_1) + \edgecount(U_2).
% % \]
% A fixed edge is counted by $f$ if and only if both its endpoints are assigned the same label,
% which occurs with probability $\nicefrac{1}{2}$.
% Therefore, $\E[f] = \edgecount / 2$.
% Consider the doob martingale $ X_0, X_1, \ldots, X_\vertexcount $, where $ X_t = \E[f(Y_1, \ldots, Y_\vertexcount)|Y_1, \ldots, Y_t]$. For a fix values of $ Y_1, \ldots, Y_{t-1}$, we have,
% \begin{align*}
% X_{t-1} = \frac{ g(1) + g(2) }{ 2 }, \text{ where } g(i) = \E[f(Y_1, \ldots, Y_\vertexcount)| (Y_1, \ldots, Y_{t-1}) \cap (Y_t = i)]
% \end{align*}
% Since, $X_t$ is either $g(1)$ or $g(2)$, we have $ \abs{X_t - X_{t-1}} \leq \abs{g(1) - g(2)}$. Let $ N(t) $ be the set of neighbours of the vertex $ t $ and $ \deg(t) = \abs{ N(t) }$. Let $ N_{ < t} = N(t) \cap [t-1] $ and $ N_{ > t} = N(t) \cap \sbrac{(t + 1), \ldots, \vertexcount}$. Let $C_{<t}^{(i)}$ and $C_{>t}^{(i)}$ be the number of vertices in $ N_{ < t} $ and $ N_{ > t} $ respectively. So,
% \begin{equation*}
% \Delta_t = \abs{g(1) - g(2)} = \abs{ C_{<t}^{(1)} + \E\tbrac{ C_{>t}^{(1)} \mid Y_t = 1} - C_{<t}^{(2)} - \E\tbrac{ C_{>t}^{(2)} \mid Y_t = 2} }
% \end{equation*}
% The above equation is true because any edge with vertices in $ [t-1] $ will have the same contribution to $ g(1) $ and $ g(2) $. Similarly, any edge which have one vertex in $ [t-1] $ and another in $ \sbrac{(t + 1), \ldots, \vertexcount} $ or both of its vertices in $ \sbrac{(t + 1), \ldots, \vertexcount} $ will have the same contribution.
% Using Lemma~\ref{lemma: support for sparsification}, we have with probability at least $ 1 - \confidence'$,
% \begin{equation*}
% \Delta_t \leq \sqrt{ 6 { \deg(t) } \log \frac{ 4 }{ \confidence' } } + \sqrt{ \deg(t) } \leq 2\constant \sqrt{ \deg(t) \log \frac{ 1 }{ \confidence' } } = \constant_t
% \end{equation*}
% Let $ \mathcal{B} $ be the event that there exits a $ t $ such that $ \Delta_t > \constant_t $. Let $ \vertexcount_a $ be the set of active vertices (vertices with degree at least $ 1 $). Therefore, using union bound $ \Pr[\mathcal{B}] \leq 4 \vertexcount_a \confidence' \leq 4 \edgecount \confidence' $.
% Let $ S = \sum_{ t = 1}^\vertexcount \constant_t^2 = \sum_{ t = 1}^\vertexcount 4 \constant^2 \deg(t) \log( 4 / \confidence') = O( \edgecount \log( 1 / \confidence' ) ) $ and $ s = \constant \sqrt{\edgecount} \log( 1 / \confidence' ) $. Using \Cref{lemma: martingle inequality}, we have
% \begin{equation*}
% \Pr \tbrac{ \abs{ f - \frac{ \edgecount }{ 2 }} > s } \leq 2 \exp \fbrac{ - \frac{ s^2 }{ 2S } } + \Pr \tbrac{ \mathcal{ B } } \leq \confidence' + 4 \edgecount \confidence'
% \end{equation*}
% Replacing $ \confidence' $ with $ \confidence / \constant \edgecount $ for a sufficiently large constant, we obtain
% \begin{equation*}
% \Pr\tbrac{ \abs{ \frac{ \edgecount }{ 2 } - \sum_{ i = 1 }^{ 2 } \edgecount( U_i ) } \geq \constant \sqrt{ \edgecount } \log \fbrac{\frac{ \edgecount }{ \confidence }} } \leq \confidence
% \end{equation*}
% which completes the proof.
% \end{proof}
%
%
%
%
%
%
\section{Threshold checking}\label{appendix: threshold checking}
In this section, we present the algorithms \iscollision{}, \birthdayestimate{} and the proofs of \Cref{lemma: collision detection} and \Cref{lemma: birthdayestimate}.
\begin{algorithm}
\caption{\iscollision($U, \threshold, \confidence$)}\label{alg:collision}
\begin{algorithmic}[1]
\Require \iscq{} access to the subset $U \subseteq \vertexset$ and a threshold $\threshold$
\Ensure $0$ or $1$
\For{$i \gets 1$ to $O(\log(\confidence^{-1}))$} \label{alg: collision line 1}
\State Uniformly and independently sample $\preresamplesize = \sqrt{2\threshold}$ random edges from $U$
\label{alg: collision line 2}
\If{there is a collision}
\State $\collision_i \gets 1$
\Else
\State $\collision_i \gets 0$
\EndIf
\EndFor
\If{more than half of $\collision_i$ are $1$}
\State \Return $1$
\Else
\State \Return $0$
\EndIf
\end{algorithmic}
\end{algorithm}
% \begin{algorithm}
% \setcounter{AlgoLine}{0}
% \SetAlgoNoLine
% \LinesNumbered
% \SetAlgoNoEnd
% \caption{\iscollision{}($U, \threshold, \confidence$)}
% \KwIn{\iscq{} access to the subset $U \subseteq \vertexset$ and a threshold $\threshold$}
% \KwOut{ $0$ or $1$}
% \For{$i \gets 1$ \KwTo $O(\log(\confidence^{-1}))$}{ \label{alg: collision line 1}
% Uniformly and independently sample
% $\preresamplesize = \sqrt{2\threshold}$ random edges from $U$
% \label{alg: collision line 2}
% \eIf{there is a collision}{
% $\collision_i \gets 1$
% }{
% $\collision_i \gets 0$
% }
% \eIf{more than half of $\collision_i$ are $1$}{
% \KwRet $1$
% }{
% \KwRet $0$
% }
% \end{algorithm}
\isollison*
\begin{proof}
Fix an iteration $i$ of \iscollision{} (Line~\ref{alg: collision line 1}).
The probability that no collision occurs in the $i$-th iteration is
\begin{align*}
\Pr[\collision_i = 0]
= \prod_{j=0}^{\preresamplesize - 1}
\left(1 - \frac{j}{\edgecount(U)}\right).
\end{align*}
\noindent If $\edgecount(U) \leq \threshold$,
We upper bound the probability of no collision as follows:
\begin{align*}
\Pr[\collision_i = 0]
&\leq \exp\left(
- \sum_{j=0}^{\preresamplesize - 1}
\frac{j}{\edgecount(U)}
\right)
&& \text{($1 + x \leq e^{x}$ for all $x \in \R$)} \\
&= \exp\left(
- \frac{\preresamplesize(\preresamplesize - 1)}{2 \edgecount(U)}
\right) \\
&\leq \exp\left(
- \frac{\preresamplesize(\preresamplesize - 1)}{2 \threshold}
\right)
&& \text{(since $\edgecount(U) \leq \threshold$)} \\
&\leq \frac{1}{e}
&& \text{(using $\preresamplesize = \sqrt{2\threshold}$)}.
\end{align*}
\noindent If $\edgecount(U) \geq 8\threshold$, we lower bound the probability of no collision:
\begin{align*}
\Pr[\collision_i = 0]
&\geq \exp\left(
- \sum_{j=0}^{\preresamplesize - 1}
\left(
\frac{j}{\edgecount(U)} +
\frac{j^2}{\edgecount(U)^2}
\right)
\right)
&& \text{($1 - x \geq e^{-x - x^2}$ for $x \in [0, 2/3]$)} \\
&\geq \exp\left(
- \frac{\preresamplesize(\preresamplesize - 1)}{16 \threshold}
+ \Theta\!\left(\frac{1}{\sqrt{\threshold}}\right)
\right) \\
&\geq \frac{1}{2}
&& \text{(using $\preresamplesize = \sqrt{2\threshold}$)}.
\end{align*}
\noindent If $ \edgecount( U ) \leq \threshold $, $ \Pr[ \collision_i = 0 ] \leq \frac{ 1 }{ e } $. \iscollision{} outputs $0$ when more than half of $ \collision_i = 0 $. Over $ \log ( \confidence^{-1} ) $ parallel runs, using the upper tail estimate of the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), the probability that \iscollision{} outputs $0$ is at most $ \exp ( - \log ( \confidence^{-1} ) ) = \confidence $. Similarly, if $ \edgecount( U ) \geq 8\threshold $, the probability that \iscollision{} outputs $1$ is at most $ \confidence $.
\end{proof}
\begin{algorithm}
\caption{\birthdayestimate{}($ U, \threshold, \approxerror $)}\label{alg: birthdayestimate}
\begin{algorithmic}[1]
\Require \iscq{} access to the subset $ U \subseteq \vertexset $ and a threshold $ \threshold
\geq 1 $.
\Ensure An estimate of $ \edgecount( U ) $
\State Uniformly and independently sample $ \Theta \fbrac{ \sqrt { \threshold } / \approxerror } $ random edges from $ U $.
\State Let $ \resamplesize = \Theta\fbrac{ \nicefrac{ \sqrt{\threshold} }{ \approxerror } } $ be the number of samples drawn and $ \collisioncount $ is the number of collisions.
\State \Return $\binom{\resamplesize}{2}/\collisioncount $.
\end{algorithmic}
\end{algorithm}
% \begin{algorithm}
% \setcounter{AlgoLine}{0}
% \SetAlgoNoLine
% \LinesNumbered
% \SetAlgoNoEnd
% \caption{\birthdayestimate{}($ U, \threshold, \approxerror $)}
% \label{alg: birthdayestimate}
% \KwIn{\iscq{} access to the subset $ U \subseteq \vertexset $ and a threshold $ \threshold \geq 1 $.}
% \KwOut{An estimate of $ \edgecount( U ) $}
% Uniformly and independently sample $ \Theta \fbrac{ \sqrt { \threshold } / \approxerror } $ random edges from $ U $.
% Let $ \resamplesize = \Theta\fbrac{ \nicefrac{ \sqrt{\threshold} }{ \approxerror } } $ be the number of samples drawn and $ \collisioncount $ is the number of collisions.
% \KwRet $\binom{\resamplesize}{2}/\collisioncount $.
% \end{algorithm}
\birth*
\begin{proof}
Let $\sbrac{e_1, e_2, \ldots, e_{\resamplesize}}$ be the sampled edges. or $1 \le i < j \le \resamplesize$, define the indicator random variable $X_{ij}$ to be $1$ if $e_i = e_j$ and $0$ otherwise . Then,
$$\Pr[X_{ij}=1] = \frac{1}{ \edgecount( U ) } \text{ and } \E[X_{ij}] = \frac{1}{ \edgecount( U ) }$$
Define the total number of collisions by $ X = \sum_{1 \le i < j \le \resamplesize} X_{ij} $. By the linearity of expectation
\begin{align*}
\E[X] = \binom{\resamplesize}{2} \frac{1}{ \edgecount( U ) }
%= O\left(\frac{\resamplesize^{2}}{\edgecount}\right).
\end{align*}
Since the random variables $ X_{ij} $ are not independent. We will bound the variance using
\begin{align}\label{eq: variance equation}
\Var(X) &= \sum_{i < j } \Var(X_{ij}) + 2\sum_{i < j, k < \ell} \mathrm{Cov}(X_{ij},X_{k\ell})
\end{align}
Let us bound the first term $ \Var(X_{ij}) $ in~\eqref{eq: variance equation}
\begin{align*}
\sum_{i < j } \Var(X_{ij}) \leq \sum_{ i < j } \E[X_{ij}^2] \leq \binom{\resamplesize}{2} \frac{1}{ \edgecount( U ) } = \E[X]
\end{align*}
Next, consider the covariance terms in~\eqref{eq: variance equation}. If $X_{ij}$ and $X_{k\ell}$ correspond to disjoint pairs, then they are independent and the covariance is zero. If they overlap in exactly one index (share exactly one index), then
\begin{align*}
\mathrm{Cov}(X_{ij},X_{j\ell}) &= \E[X_{ij}X_{j\ell}] - \E[X_{ij}]\E[X_{j\ell}]\\
&= \Pr[X_{ij} = X_{j\ell} = 1] - \frac{1}{ \edgecount( U )^2}\\
&= \Pr[e_i = e_j = e_\ell] - \frac{1}{ \edgecount( U )^2 } \\
&= 0
\end{align*}
So, the bound on the variance is $ \Var(X) = \sum_{i < j } \Var(X_{ij}) \leq \E[X] $. Using Chebyshev’s Inequality (Lemma~\ref{lemma: chebyshev}), we have
\begin{align}
\nonumber\Pr[\abs{X - \E[X]} > \frac{\approxerror}{3} \E[X]] &\leq \frac{9\Var(X)}{\approxerror^2 (\E[X])^2}\\
\nonumber&\leq \frac{9}{\approxerror^2\E[X]} &\text{$\fbrac{\Var(X) \leq \E[X]}$}\\
\nonumber&= \frac{ 18\edgecount( U ) }{ \approxerror^2\resamplesize(\resamplesize - 1) } &\text{$\fbrac{ \E[X] = \binom{\resamplesize}{2} \frac{ 1 }{ \edgecount( U ) } }$}\\
&\leq \frac{1}{ 16 } &\text{($ \edgecount_U \leq \threshold $ and $\resamplesize = \Theta ( \sqrt{ \threshold } /{ \approxerror } )$)}\label{ineq: collision bound}
\end{align}
To complete the proof, we need to show with probability at least $ 15/ 16 $
\begin{align*}
\fbrac{1 - \approxerror} \edgecount( U ) \leq \binom{\resamplesize}{2} \frac{1}{X} \leq \edgecount( U ) \fbrac{1 + \approxerror}
\end{align*}
From~\eqref{ineq: collision bound}, with probability at least $ 15/ 16 $
\begin{align*}
\fbrac{1 - \frac{\approxerror}{3}}\E\tbrac{X} \leq X \leq \fbrac{1 + \frac{\approxerror}{3}} \E\tbrac{X}
\end{align*}
This implies
\begin{align*}
\Rightarrow \frac{ \edgecount( U ) }{\fbrac{1 + \frac{\approxerror}{3}}}\leq \binom{\resamplesize}{2} \frac{1}{X} \leq \frac{\edgecount( U ) }{\fbrac{1 - \frac{\approxerror}{3}}}
\end{align*}
For any $ \approxerror \in (0, 1) $
\begin{equation*}
1 - \approxerror \leq \fbrac{1 + \frac{\approxerror}{3}}^{-1} \text{ and } \fbrac{1 - \frac{\approxerror}{3}}^{-1} \leq 1 + \approxerror
\end{equation*}
Since $X$ is the number of collisions. we conclude that
\begin{align*}
\fbrac{1 - \approxerror} \edgecount( U ) \leq \binom{\resamplesize}{2} \frac{1}{\collisioncount} \leq \edgecount( U ) \fbrac{1 + \approxerror}
\end{align*}
which completes the proof.
\end{proof}
% \section{Proof of \Cref{lemma: addestimation} and \Cref{lemma: mulestimation}}\label{appendix: concentration proofs}
% In this section, we present the proofs of Proof of \Cref{lemma: addestimation} and \Cref{lemma: mulestimation} using concentration bounds.
% \addest*
% \begin{proof}
% Consider a subset $ U_i \subseteq U $ for any $ i \in \{1,2\} $.
% For each sampled edge, define the indicator random variable
% \[
% X_j =
% \begin{cases}
% 1 & \text{if the $ j^{\text{th}} $ sampled edge lies in $ U_i $,} \\
% 0 & \text{otherwise.}
% \end{cases}
% \]
% Observe that $\E[X_j] = \frac{m(U_i)}{m(U)}$.
% Define the fraction of sampled edges that belong to $ U_i $ by
% $
% X = \frac{1}{\size{\samplededges}}
% \sum_{j \in [\size{\samplededges}]} X_j.
% $
% where $ \samplededges $ is the multiset of sampled edges (Line~\ref{addestimation: line 2}) obtained via \iscq{} queries on $ U $.
% Note that $ X = \approxfractionofedges(U_i) $ and hence
% \[
% \E[\approxfractionofedges(U_i)]
% = \frac{1}{\size{\samplededges}}
% \sum_{j \in [\size{\samplededges}]}\E[X_j]
% = \frac{\edgecount(U_i)}{\edgecount(U)}
% = \frac{m(U_i)}{m(U)}.
% \]
% Applying Hoeffding's bound~(\Cref{lemma: hoeffding}),
% \[
% \Pr\!\left( \left| \approxfractionofedges(U_i) -
% \frac{m(U_i)}{m(U)}
% \right|
% \ge \tfrac{1}{32}
% \right)
% \le
% \exp\!\left(
% -\frac{\size{S}}{2}\cdot (1/32)^2
% \right) \le \delta/2.
% \]
% The last inequation holds since $\size{\samplededges}=O(\log 1/\confidence)$. Using the union bound over $i\in \{1,2\}$, we get the desired result.
% \end{proof}
% \mulest*
% \begin{proof}
% Let $ X_j$ be the indicator random variable such that,
% \begin{equation*}
% X_j =
% \begin{cases}
% 1 & \text{If the $ j^{th} $ sampled edge is in $ U' $} \\
% 0 & \text{Otherwise }
% \end{cases}
% \end{equation*}
% Define number of edges sampled that belong to $ U' $ is $ X = \sum_{ j \in [\size{ \samplededges }] } X_j$, where $ \samplededges $ is the multi-set of sampled edges (Line~\ref{addestimation: line 2}) and the expected size of $ X $ is
% \begin{equation*}
% \E\tbrac{ X } = \sum_{ j \in [\size{S}] } \E\tbrac{ X_j } = \size{ S } \cdot \frac{ \edgecount{ (U') }}{ \edgecount{ (U) } } = \size{ S } \cdot \fractionofedges( U' ) \geq \frac{ \size{ \samplededges } }{ 32 }
% \end{equation*}
% By the Chernoff Bound (see Lemma~\ref{lemma: Multiplicative Chernoff Bound}), we have
% % \begin{align*}
% % \Pr\tbrac{ \approxfractionofedges ( U' ) \in ( 1 \pm \approxerror ) \fractionofedges( U ) } &= \Pr\tbrac{ X \in ( 1 \pm \approxerror ) \E[X] } \\
% % % &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \E[X] }{ 3 } }\\
% % &\geq 1 - 2\exp\fbrac{ - \frac{ \approxerror^2 \size{ \samplededges } }{ 96 } }
% % &\text{$\fbrac{ \E\tbrac{ X } \geq \size{ \samplededges } / 32 }$}\\
% % &\geq 1 - \confidence &\text{$\fbrac{ \size{ \samplededges } = O\fbrac{ \frac{ \log ( \confidence^{-1} ) }{ \approxerror^2 }} }$}
% % \end{align*}
% \[
% \Pr\tbrac{ \approxfractionofedges(U') \in (1\pm\approxerror)\fractionofedges(U') } = \Pr\tbrac{ X \in (1\pm\approxerror)\E[X] } \ge 1 - 2\exp\fbrac{ - \frac{\approxerror^2 \size{\samplededges}}{96} } \ge 1 - \confidence,
% \]
% where the last step uses $\size{\samplededges}=O\left(\log (\delta^{-1})/\varepsilon^2\right)$.
% \end{proof}