A Note on Polynomial Identity Testing for Depth-3 Circuits
Extracted source text ยท arXiv 1805.06692
\documentclass[11pt]{article}
% \usepackage{authblk}
%\documentclass[a4paper,UKenglish]{llncs}
%\documentclass[envcountsame]{llncs}
%\usepackage[normalem]{ulem}
%\usepackage[normalem]{ulem}
\usepackage{amssymb}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{setspace}
\usepackage{cite}
\usepackage{hyperref}
\usepackage{tikz}
\usepackage{verbatim}
\usepackage{graphicx}
\usepackage{subfig}
%\usepackage{algorithm}
%\usepackage[noend]{algpseudocode}
\usetikzlibrary{shapes.geometric}
%\usetikzlibrary{graphs,graphdrawing,arrows.meta}
%\usegdlibrary{trees}
%\usepackage{tikz}
%\usetikzlibrary{calc,positioning,shapes,shadows,arrows,fit}
%\usepackage{showlabels}
\sloppy
%\usepackage[hmargin=2.5cm,vmargin=3cm]{geometry}
%\usepackage{amssymb}
%\usepackage{amsmath}
%\usepackage{setspace}
%\linespread{1.1}
%\usepackage{cite}
\newtheorem{definition}{Definition}
\newtheorem{corollary}{Corollary}
\newtheorem{prop}{Proposition}
\newtheorem{claim}{Claim}
\newtheorem{lemma}{Lemma}
\newtheorem{theorem}{Theorem}
\newtheorem{conj}{Conjecture}
\newtheorem{obs}{Observation}
\newtheorem{example}{Example}
\newtheorem{proposition}{Proposition}
\newtheorem{remark}{Remark}
\sloppy
%\usepackage[hmargin=2.9cm,vmargin=2.9cm]{geometry}
%\newcommand\numberthis{\addtocounter{equation}{1}\tag{\theequation}}
\newcommand{\NW}{\mbox{\small\rm NW}}
\newcommand{\IMM}{\mbox{\small\rm IMM}}
%\newcommand{}{\mbox{\small\rm dim}}
\newcommand{\smallp}{\mbox{\small\rm P}}
\newcommand{\nnl}{\mbox{\small\rm NNL}}
\newcommand{\VP}{\mbox{\small\rm VP}}
\newcommand{\VNP}{\mbox{\small\rm VNP}}
%\renewcommand{\P}{\mbox{\small\rm P}}
\newcommand{\NP}{\mbox{\small\rm NP}}
\newcommand{\LM}{\mbox{\small\rm LM}}
\newcommand{\SL}{\mbox{\small\rm SL}}
\newcommand{\Det}{\mbox{\small\rm Det}}
%\newcommand{\pit}{\mbox{\small\rm PIT}}
\newcommand{\ex}{\mbox{\small\rm X}}
\newcommand{\ey}{\mbox{\small\rm Y}}
\newcommand{\GL}{\mbox{\small\rm GL}}
\newcommand{\el}{L}
\newcommand{\hess}{\mbox{\small\rm H}}
\newcommand{\mathify}[1]{\mbox{\small\rm #1}}
\newcommand{\dc}{\mbox{\small\rm dc}}
\newcommand{\F}{\mathbb{F}}
\renewcommand{\P}{\mathbb{P}}
\newcommand{\V}{\mathbb{V}}
\newcommand{\C}{\mathbb{C}}
\newcommand{\Z}{\mathbb{Z}}
\newcommand{\M}{\mathcal{M}}
\newcommand{\B}{\mathbb{B}}
\newcommand{\fq}{\mathbb{F}_q}
\newcommand{\Q}{\mathbb{Q}}
\newcommand{\R}{\mathbb{R}}
\newcommand{\K}{\mathbb{K}}
\renewcommand{\angle}[1]{\mathopen{\langle} #1\mathclose{\rangle}}
%\newcommand{\M}{\mathcal{M}}
\newcommand{\depththree}{\Sigma\Pi\Sigma}
\newcommand{\depthfour}{\Sigma\Pi\Sigma\Pi}
\newcommand{\depthfoursqrtn}{\Sigma\Pi^{[O(\sqrt{n})]}\Sigma\Pi^{[\sqrt{n}]}}
\newcommand{\depthfourparam}{\Sigma\Pi^{[D]}\Sigma\Pi^{[t]}}
\DeclareMathOperator{\var}{\mbox{\small\rm var}}
%\DeclareMathOperator{\gcd}{\mbox{\small\rm g.c.d}}
\newcommand{\simn}{\mbox{\small\rm sim}}
\newcommand{\sgn}{\mbox{\small\rm sgn}}
\newcommand{\res}{\mbox{\small\rm Res}}
\newcommand{\pit}{\mbox{\small\rm PIT}}
\newcommand{\IP}{\mbox{\small\rm IP}}
\newcommand{\PSPACE}{\mbox{\small\rm PSPACE}}
\newcommand{\corp}{\mbox{\small\rm co-RP}}
\renewcommand{\char}{\mbox{\small\rm Char}}
\DeclareMathOperator{\codim}{\mbox{\small\rm codim}}
\DeclareMathOperator{\rank}{\mbox{\small\rm rank}}
\DeclareMathOperator{\diag}{\mbox{\small\rm diag}}
%\DeclareMathOperator{\perm}{\mbox{\small\rm perm}}
\DeclareMathOperator{\poly}{\mbox{\small\rm poly}}
\DeclareMathOperator{\prob}{\mbox{\small\rm Prob}}
\DeclareMathOperator{\perm}{\mbox{\small\rm Perm}}
\title{A Note on Polynomial Identity Testing for Depth-3 Circuits}
\author{V. Arvind\thanks{Institute of Mathematical Sciences, Chennai,
India, \texttt{email: arvind@imsc.res.in}} \and Abhranil Chatterjee \thanks{Institute of Mathematical Sciences, Chennai,
India, \texttt{email: abhranilc@imsc.res.in}} \and Rajit Datta\thanks{Chennai Mathematical Institute, Chennai, India, \texttt{email: rajit@cmi.ac.in}} \and Partha
Mukhopadhyay\thanks{Chennai Mathematical Institute, Chennai, India,
\texttt{email: partham@cmi.ac.in}}
}
\begin{document}
\maketitle
\begin{abstract}
Let $C$ be a depth-3 arithmetic circuit of size at most $s$, computing a polynomial $f\in\F[x_1,\ldots, x_n]$ (where $\F$ = $\Q$ or $\C$)
and the fan-in of the product gates of $C$ is bounded by $d$. We give a deterministic polynomial identity testing algorithm to check whether $f\equiv 0$ or not in time $2^d \poly(n,s)$.
Over finite fields, for $\char(\F) > d$ we give a deterministic algorithm of running time $2^{\gamma\cdot d}\poly(n,s)$ where $\gamma\leq 5$.
%For $\char(\F)\leq d$, we give deterministic algorithm of running time $2^{\gamma\cdot d \log d}\poly(n,s)$ where $\gamma\leq 5$.
\end{abstract}
\section{Introduction}
Polynomial Identity Testing ($\pit$) is the following problem : Given an arithmetic circuit $C$ computing a polynomial in $\F[x_1, \ldots, x_n]$, determine whether $C$ computes an identically zero polynomial or not.
The problem can be presented either in the \emph{white-box} model or in the \emph{black-box} model. In the white-box model, the arithmetic circuit is given explicitly as the input. In the black-box model, the arithmetic circuit is given black-box access, and the circuit can be evaluated over any point in $\F^n$ (or $F^n$ where $\F \subseteq F$ is an extension field).
Over the years, the problem has played pivotal role in many important results in complexity theory and algorithms: Primality Testing \cite{aks04}, the PCP Theorem \cite{almss98}, $\IP=\PSPACE$ \cite{shamir90}, graph matching algorithms
\cite{lov79,mvv87}. The problem $\pit$ admits a $\corp$ algorithm via the Schwartz-Zippel-Lipton-DeMillo Lemma
\cite{Sch80,Zip79,DL78}, but an efficient deterministic algorithm is known only in some special cases.
An important result of Impagliazzo and Kabanets \cite{KI04} (also, see \cite{HS80, Agr05}) shows a connection between the existence of a subexponential time $\pit$ algorithm and arithmetic circuit lower bounds.
%n particular, it is now known that
%f $\pit$ can be derandomized using a certain type of pseudo-random generator, then the Permanent polynomial can not be computed by a polynomial-size arithmetic circuit \cite{Agr05,KI04}. As a result, it will prove the algebraic analogue of $\smallp$ vs $\NP$ problem: $\VP\neq \VNP$.
We refer the reader to the survey of Shpilka and Yehudayoff \cite{sy10} for the exposition of important results in arithmetic circuit complexity, and polynomial identity testing problem.
%\subsubsection{Depth Reduction and Lower Bounds}\label{depth-lb}
In a surprising result, Agrawal and Vinay \cite{av2008} show that an efficient deterministic $\pit$ algorithm only for depth-4 $\depthfour$ circuits is sufficient for obtaining an efficient deterministic $\pit$ algorithm for the general
arithmetic circuits. The main technical ingredient in their proof is an ingenious depth-reduction technique.
%In nice follow-up works, Koiran and Tavenas further improved their results in terms of certain parameters \cite{koi2012, tav2013}.
Over characteristic zero fields, derandomization of $\pit$ even for depth-3 $\Sigma\Pi\Sigma$ circuits
suffices \cite{gkks2013}.
% will separate $\VP$ from $\VNP$. Currently, there are many impressive partial results in this direction showing depth four lower bounds for explicit polynomials in $\VP$ and in $\VNP$. All these papers use the shifted partial derivative technique (for example, see \cite{gkks2012}).
%, kss2013, flms2013, klss2014, ks2014}.
Motivated by the results of \cite{KI04,Agr05,av2008}, a large body of research consider the polynomial identity testing problem for restricted classes of depth-3 and depth-4 circuits. A particularly popular model in depth three arithmetic circuits is $\depththree(k)$ circuit, where the fan-in of the top $\Sigma$ gate is bounded by $k$. Dvir and Shpilka have shown a \emph{white-box} quasi-polynomial time deterministic $\pit$ algorithm for $\depththree(k)$ circuits \cite{DS07}. Kayal and Saxena have given a deterministic $\poly(d^k,n,s)$ white-box
algorithm for the same problem \cite{KS07}. Following the result of \cite{KS07}(Also see \cite{AM10} for a different analysis).
Karnin and Shpilka have given the first \emph{black-box} quasi-polynomial time algorithm for $\depththree(k)$ circuits \cite{KS11}. Later, Kayal and Saraf \cite{KS09} have shown a polynomial-time deterministic black-box $\pit$ algorithm for the same class of circuits over $\Q$ or $\R$. Finally, Saxena and Sheshadhri have settled the situation completely by giving a deterministic polynomial-time \emph{black-box} algorithm for $\depththree(k)$ circuits \cite{SS12} over any field. Recently, Oliveira et al. have given a sub-exponential $\pit$-algorithm for depth-3 and depth-4 \emph{multilinear} formulas \cite{OSV16}.
For general depth-3 $\Sigma\Pi\Sigma$ circuits with $\times$-gate fan-in bounded by $d$ no deterministic algorithm with running time better than $\min\{ d^n, n^d \} \poly(n,d)$ is known. Our main results are the following.
\begin{theorem}\label{main-thm}
Let $C$ be a depth-3 $\Sigma\Pi\Sigma$ circuit of size at most $s$, computing a polynomial $f \in \F[x_1,\ldots, x_n]$ ( where $\F$ = $\Q$ or $\C$) and the fan-in of the product gates of $C$ is bounded by $d$. We give a \emph{white-box} deterministic polynomial time identity testing algorithm to check whether $f\equiv 0$ or not in time $2^d \poly(n,s)$.
\end{theorem}
As an immediate corollary we get the following.
\begin{corollary}\label{poly-depth3}
Let $C$ be a depth-3 $\Sigma\Pi\Sigma$ circuit of size at most $s$, computing a polynomial $f\in \F[x_1,\ldots, x_n]$ ( where $\F$ = $\Q$ or $\C$) and the fan-in of the product gates of $C$ is bounded by $O(\log n)$. We give a deterministic $\poly(n,s)$ time identity testing algorithm to check whether $f\equiv 0$ or not.
\end{corollary}
Over the fields of positive characteristics, we show the following result.
\begin{theorem}\label{main-thm-2}
Let $C$ be a depth-3 $\Sigma\Pi\Sigma$ circuit of size at most $s$, computing a polynomial $f \in \F[x_1,\ldots, x_n]$ and the fan-in of the product gates of $C$ is bounded by $d$. For $\char(\F) > d$, we give a \emph{white-box} deterministic polynomial time identity testing algorithm to check whether $f\equiv 0$ or not in time $2^{\gamma\cdot d} \poly(n,s)$.
%For $\char(\F)\leq d$, we give a \emph{white-box} deterministic polynomial time identity testing algorithm to check whether $f\equiv 0$ or not in time $2^{\gamma\cdot d\log d} \poly(n,s)$.
The constant $\gamma$ is at most $5$.
\end{theorem}
\section{Orgazination}
The paper is organized as follows. Section \ref{prelim} contains preliminary materials. In Section \ref{main-sec}, we prove Theorem \ref{main-thm}. Theorem \ref{main-thm-2} is proved in Section \ref{main-sec-1}.
%\section{Main Result}
\section{Preliminaries}\label{prelim}
For a monomial $m$ and a polynomial $f$, let $[m]f$ denote the coefficient of the monomial $m$ in $f$.
We denote the field of rational numbers as $\mathbb{Q}$, and the field of complex numbers as $\mathbb{C}$. The depth-3 $\Sigma\Pi\Sigma(s,d)$ circuits compute polynomials of the following form:
\[
C(x_1,\ldots,x_n) = \sum_{i=1}^s \prod_{j=1}^d L_{i,j}(x_1,\ldots,x_n).
\]
where $L_{i,j}$'s are affine linear forms over $\F$. The following observation is well-known and it says that for $\pit$ purpose it is sufficient to consider homogeneous circuits.
\begin{obs}
Let $C(x_1,\ldots,x_n)$ be a $\Sigma\Pi\Sigma(s,d)$ circuit. Then $C\equiv 0$ if and only if
$z^d C(x_1/z, \ldots,x_n/z)\equiv 0$ where $z$ is a new variable.
\end{obs}
We use the notation $\Sigma^{[s]} \Pi^{[d]} \Sigma$ to denote homogeneous depth-3 circuits of top $\Sigma$ gate fan-in $s$, product gates fan-in bounded by $d$.
%computing a polynomial of degree $d$ over $n$ variables.\\
%We use the notation $\Sigma^{[s]} \Pi^{[d]} \Sigma$ to denote homogeneous depth-3 circuits of top $\Sigma$ gate fan-in $s$, product gates fan-in $d$.
%computing a polynomial of degree $d$ over $n$ variables.\\
We recall the definition of Hadamard Product of two polynomials. The concept of Hadamard product is particularly useful in noncommutative computations \cite{AJ09,AS18}.
%which have been useful in noncommutative computation.
\begin{definition}\label{hadprod}
Given two degree $d$ polynomials $f,g \in \F[x_1,x_2,\ldots,x_n]$, the Hadamard Product $f \circ g$ is defined as
$$f \circ g = \sum_{m : \deg(m)\leq d} ([m]f \cdot [m]g) \ m.$$
\end{definition}
For the $\pit$ purpose in the commutative setting, we adapt the notion of Hadamard Product suitably and define a scaled version of Hadamard Product of two polynomials.
%We recall the definition of Inner Product of two polynomials (generally this is defined as a value but here we define it as a polynomial).
%The concept of Inner product is particularly useful in non-commutative computations \cite{AJ09,AS18}.
%which have been useful in non-commutative computation.
\begin{definition}\label{s-hadprod}
Given two degree $d$ polynomials $f,g \in \F[x_1,x_2,\ldots,x_n]$, the scaled version of the Hadamard Product $f \circ^{s} g$ is defined as
$$f \circ^{s} g = \sum_{m : \deg(m)\leq d} (m! \cdot [m]f \cdot [m]g) \ m$$
\end{definition}
where $m=x^{e_1}_{i_1}x^{e_2}_{i_2} \ldots x^{e_r}_{i_r}$ for some $r\leq d$ and by abusing the notation we define $m! = e_1 ! \cdot e_2! \cdot \ldots \cdot e_r!$.
For the purpose of $\pit$ over $\Q$, it is enough to be able to compute $f \circ^{s} f (1,1,\ldots,1)$.
As $f \circ^{s} f$ has only non-negative coefficients, we will see a non-zero value when we compute $f \circ^{s} f (1,1,\ldots,1)$ if and only if $f \not \equiv 0$.
Over $\C$ it is enough to compute $f \circ^{s} \bar{f} (1,1,\ldots,1)$ where $\bar{f}$ denotes the polynomial obtained by conjugating every coefficient of $f$.
We also recall a result of Ryser~\cite{ry63} that gives a $\Sigma^{[2^n]} \Pi^{[n]} \Sigma$ circuit for the Permanent polynomial of $n\times n$ symbolic matrix.
\begin{lemma}[Ryser ~\cite{ry63}]\label{ryser}
For a matrix $X$ with variables $x_{ij} : 1\leq i,j\leq n$ as entries,
$$\perm(X)= (-1)^n \sum_{S \subseteq [n]} (-1)^{|S|} \prod^{n}_{i=1} \left(\sum_{j \in S} x_{ij}\right).$$
\end{lemma}
\begin{lemma}\label{calc}
For a monomial $m = x_{i_1}x_{i_2} \ldots x_{i_d}$ ($i_1 ,\ldots , i_d$ need not be distinct) and a homogeneous $\Pi\Sigma$ circuit $C=\prod^d_{j=1} L_j$ we have:
\[
[m]C = \frac{1}{m!} \sum_{\sigma \in S_d} \prod^d_{j=1}( [x_{i_j}] L_{\sigma(j)}).
\]
\end{lemma}
\begin{proof}
The monomial $m$ can be obtained from $C$ by first fixing a bijection $\sigma : [d] \mapsto [d]$ and considering the coefficient $[m]C_{\sigma}= \prod^{d}_{j=1} [x_{i_{\sigma(j)}}] L_j = \prod^{d}_{j=1} [x_{i_{j}}] L_{\sigma^{-1}(j)}$.
This is one way of generating this monomial and this monomial $m$ can be generated in many different orders. The final $[m]C$ is the sum of all coefficients $[m]C_{\sigma}$ generated in all distinct orders.
Now if $m=x^{e_1}_{i_1}x^{e_2}_{i_2} \ldots x^{e_r}_{i_r}$ for some $r\leq d$ then for a fixed $\sigma$ one can obtain $m!$ different bijections that do not change the string $x_{i_{\sigma(1)}}x_{i_{\sigma(2)}} \ldots x_{i_{\sigma(d)}}$ and these will generate
the same coefficient $\prod^{d}_{j=1} [x_{i_{\sigma(j)}}] L_j$ . Thus only the bijections that produce a different string from $x_{i_{\sigma(1)}}x_{i_{\sigma(2)}} \ldots x_{i_{\sigma(d)}}$ are relevant
($[m]C_{\sigma} = [m]C_{\pi}$ if the strings $x_{i_{\sigma(1)}}x_{i_{\sigma(2)}} \ldots x_{i_{\sigma(d)}}$ and $x_{i_{\pi(1)}}x_{i_{\pi(2)}} \ldots x_{i_{\pi(d)}}$ are identical). To account for the coefficients produced by
the extra bijections we divide by $m!$
% We shall prove the formula using induction on the degree.
% Let $m=m' x_{i_d}$ since $C$ is a product of linear forms we have,
% $$[m]C = \sum^d_{j=1} ([x_{i_d}]L_j) \cdot [m']\prod^d_{i=1: \ i \neq j} L_i$$
% Now using induction hypothesis we get,
% $$[m]C = \sum^d_{j=1} ([x_{i_d}]L_j) \cdot \frac{1}{m' !} \sum_{\sigma \in B(d-1,[d] \setminus \{ j \})} \prod^{d-1}_{t=1} ([x_{i_t}]L_{\sigma(t)})$$
% Where $B(d-1,[d] \setminus \{ j \})$ is the set of all bijections from $[d-1]$ to $[d] \setminus \{ j \})$.
% Now taking the product gives us a complete permutation of $[d]$.
%We note that $[m]C = m! \partial_{x_{i_1}} \partial_{x_{i_2}}\ldots \partial_{x_{i_d}} C |_{x_i=0}$.
%From the product rule of derivative we get
%$$m! \partial_{x_{i_1}} \partial_{x_{i_2}}\ldots \partial_{x_{i_d}} C |_{x_i=0} = m! \partial_{x_{i_1}} \partial_{x_{i_2}}\ldots \partial_{x_{i_{d-1}}}( \sum^d_{j=1} (\partial_{x_{i_d}}L_j) \prod^d_{i=1: \ i \neq j} L_i )$$
%Now using linearity of derivative we get,
%$$ m! \sum^d_{j=1} (\partial_{x_{i_d}}L_j) \partial_{x_{i_1}} \partial_{x_{i_2}}\ldots \partial_{x_{i_{d-1}}} \prod^d_{i=1: \ i \neq j} L_i ) $$
%Now using induction hypothesis we get,
%$$m! \sum^d_{j=1} (\partial_{x_{i_d}}L_j) \partial_{x_{i_1}} \partial_{x_{i_2}}\ldots \partial_{x_{i_{d-1}}} \prod^d_{i=1: \ i \neq j} L_i ) $$
\end{proof}
Now we are ready to prove the main theorems.
\section{The results over zero characteristics}\label{main-sec}
To prove Theorem \ref{main-thm}, the following theorem is sufficient.
\begin{theorem}\label{main}
Given a homogeneous $\Sigma^{[s]} \Pi^{[d]} \Sigma$ circuit $C$ computing a degree $d$ polynomial in $\F[x_1 ,x_2 ,\ldots , x_n]$
(where $\F$ = $\Q$ or $\C$), we can test whether $C \equiv 0$ or not deterministically in $2^{d} \poly(s,n)$ time.
\end{theorem}
\begin{proof}
For simplicity, we present the proof only over $\Q$. Over $\C$, we need a minor modification as explained in Remark \ref{rmk1}.
Given the circuit $C$ we compute $C\circ^{s} C$ and evaluate at $(1,1,\ldots, 1)$ point. Notice that over rationals, $C \circ^{s} C$ has non-negative coefficients. This also implies that $C\equiv 0$
if and only if $C\circ^{s} C(1,1,\ldots,1) = 0$. So it is sufficient to show that $C\circ^{s} C(1,\ldots,1)$ can be computed deterministically in time $2^d \poly(s,n)$.
%Here we show that $C \circ C$ has a $\sum^{[s^2 2^d]} \prod^{[d]} \sum$ size circuit which can be computed in $2^{d} s^2 n^{O(1)}$.
Since the scaled Hadamard Product distributes over addition, we only need to show that the scaled Hadamard Product of
two $\Pi\Sigma$ circuits can be computed efficiently.
\begin{lemma}\label{thetrick}
Given two homogeneous $\Pi^{[d]} \Sigma$ circuits $C_1 =
\prod^d_{i=1} L_i$ and $C_2 = \prod^d_{i=1} {L'}_i$ we have:
\[
C_1 \circ^{s} C_2 = \sum_{\sigma \in S_d} \prod^d_{i=1} (L_i \circ^{s} {L'}_{\sigma(i)}).
\]
\end{lemma}
\begin{proof}
We prove the formula monomial by monomial. Let $m= x_{i_1} x_{i_2}
\ldots x_{i_d}$ be a monomial in $C_1$ (Note that $i_1 ,
i_2,\ldots,i_d$ need not be distinct).
Now let $m$ be a monomial that appears in both $C_1$ and $C_2$. From Lemma~\ref{calc} the coefficients are $$[m]C_1 = \alpha_1 = \frac{1}{m!} \left(\sum_{\sigma \in S_d} \prod^d_{j=1}
[x_{i_j}] {L}_{\sigma(j)}\right)$$ and $$[m]C_2 = \alpha_2 = \frac{1}{m!} \left(\sum_{\pi \in S_d} \prod^d_{j=1}
[x_{i_j}] {L'}_{\pi(j)}\right)$$ respectively.
From the definition~\ref{s-hadprod} we have $$[m] (C_1 \circ^s C_2)=m! \cdot \alpha_1 \cdot \alpha_2.$$
Now let us consider the matrix $T$ where $T_{ij} = L_i \circ^s {L'}_j :
1\leq i,j\leq d$ and $\perm(T)= \sum_{\sigma \in S_d} \prod^d_{i=1}
L_i \circ^s {L'}_{\sigma(i)}$. The coefficient of $m$ in $\perm(T)$ is
$$ [m]\perm(T) = \sum_{\sigma \in S_d} [m] \left(\prod^d_{j=1} L_j \circ^s
{L'}_{\sigma(j)}\right).$$
Similar to Lemma~\ref{calc}, we notice the following.
$$[m]\perm(T)=\sum_{\sigma \in S_d} \frac{1}{m!} \sum_{\pi \in S_d} \prod^d_{j=1} [x_{i_j}] (L_{\pi(j)} \circ^s {L'}_{\sigma(\pi(j))})$$
$$= \frac{1}{m!} \sum_{\sigma \in S_d} \sum_{\pi \in S_d} \prod^d_{j=1} ([x_{i_j}] L_{\pi(j)}) \cdot ([x_{i_j}] {L'}_{\sigma(\pi(j))})$$
$$=\frac{1}{m!}\sum_{\sigma \in S_d} \sum_{\pi \in S_d} \prod^d_{j=1} ([x_{i_j}] L_{\pi(j)}) \cdot \prod^d_{j=1} ([x_{i_j}] {L'}_{\sigma(\pi(j))})$$
$$=\sum_{\pi \in S_d} \left(\prod^d_{j=1} ([x_{i_j}] L_{\pi(j)}) \cdot \frac{1}{m!} \sum_{\sigma \in S_d} \prod^d_{j=1} ([x_{i_j}] {L'}_{\sigma(\pi(j))})\right)$$
$$= m! \cdot \frac{1}{m!} \sum_{\pi \in S_d} \left(\prod^d_{j=1} ([x_{i_j}] L_{\pi(j)}) \cdot \frac{1}{m!} \sum_{\sigma \in S_d} \prod^d_{j=1}( [x_{i_j}] {L'}_{\sigma(\pi(j))})\right).$$
% $$=(\sum_{\pi \in S_d} \prod^d_{j=1} ([x_{i_j}] L_{\pi(j)})) \cdot (\sum_{\sigma \in S_d} \prod^d_{j=1} [x_{i_j}] {L'}_{\sigma(j)})$$
%$$=\alpha_1 . \alpha_2.$$
Clearly, for any fixed $\pi\in S_d$, we have that $\sum_{\sigma \in
S_d} \prod^d_{j=1} [x_{i_j}] {L'}_{\sigma(\pi(j))}= m! \alpha_2$. Hence,
$[m]\perm(T) = m! \cdot \alpha_1 \cdot \alpha_2$ and the lemma follows.
%Which shows that the coefficient of $m$ is same in $C_1 \circ C_2$ and $perm(T)$. Since this works for every monomial
%the claimed formula is true.
\end{proof}
\begin{lemma}\label{d2}
Given two $\Pi^{[d]} \Sigma$ circuits $C_1$ and $C_2$ we can compute
a $\Sigma^{[2^d]} \Pi^{[d]} \Sigma$ for $C_1 \circ^s C_2$ in time $2^d
\poly(n,d)$.
\end{lemma}
\begin{proof}
From Lemma~\ref{thetrick} we observe that $\perm(T)$ gives a circuit
for $C_1 \circ^s C_2$. A $\Sigma^{[2^d]} \Pi^{[d]} \Sigma$ circuit for
$\perm(T)$ can be computed in $2^d \poly(n,d)$ time using
Lemma~\ref{ryser}.
\end{proof}
Now we show how to take the scaled Hadamard Product of two $\Sigma \Pi \Sigma$ circuits.
\begin{lemma}\label{d3}
Given two $\Sigma \Pi^{[d]} \Sigma$ circuits $C = \sum^{s}_{i=1} P_i$
and $\widetilde{C} = \sum^{\tilde{s}}_{i=1} \widetilde{P_i}$ We can
compute a $\Sigma^{[2^d s \tilde{s}]} \Pi^{[d]} \Sigma$ circuit for
$C \circ^s \widetilde{C}$ in time $2^d \poly(s,\tilde{s},d,n)$.
\end{lemma}
\begin{proof}
We first note that by distributivity,
$$C \circ^s \widetilde{C} = \sum^{s}_{i=1} \sum^{\tilde{s}}_{j=1} P_i
\circ^s \widetilde{P_j}.$$
Using Lemma~\ref{d2} for each pair $P_i \circ^s \widetilde{P_j}$ we get
a $\Sigma^{[2^d]} \Pi^{[d]} \Sigma$ circuit $P_{ij}$. Now the formula
$\sum^{s}_{i=1} \sum^{\tilde{s}}_{j=1} P_{ij}$ is a $\Sigma^{[2^d s
\tilde{s}]} \Pi^{[d]} \Sigma$ formula which can be computed in
$2^d \poly(s,\tilde{s},d,n)$ time.
\end{proof}
Now given a $\Sigma^{[s]} \Pi^{[d]} \Sigma$ circuit $C$ we can compute
$C \circ^s C$ using Lemma~\ref{d3} and finally evaluating $C \circ^s C
(1,1,\ldots,1)$ completes the $\pit$ algorithm. Clearly all the
computation can be done in $2^{d} \poly(s,n)$ time. This completes the
proof of Theorem~\ref{main}.
\end{proof}
\begin{remark}\label{rmk1}
To adapt the algorithm over $\mathbb{C}$, we need to just compute $C
\circ^s \bar{C}$ where $\bar{C}$ is the polynomial obtained from $C$ by
conjugating each coefficient. Note that a circuit computing $\bar{C}$
can be obtained from $C$ by just conjugating the scalars that appear
in the linear forms of $C$. This follows from the fact that the
conjugation operation distributes over addition and multiplication.
Now we have $[m] (C \circ^s \bar{C}) ={|[m]C|}^2$, so the coefficients
are all positive and thus evaluating $C \circ^s \bar{C}(1,1,\ldots,1)$
is sufficient for the PIT algorithm.
\end{remark}
%\section{The results over positive characteristics}\label{main-sec-1}
\section{The results over finite fields}\label{main-sec-1}
In this section we extend the $\pit$ results over the finite fields.
%First, we note a simple adaptation of Lemma \ref{calc}.
%\begin{lemma}\label{calc-1}
% For a monomial $m = x_{i_1}x_{i_2} \ldots x_{i_d}$ ($i_1 ,\ldots , i_d$ need not be distinct) and a homogeneous $\Pi^{[d]}\Sigma$ circuit $C=\prod^r_{j=1} L_j^{e_j}$ we have:
%\[
%[m]C = \sum_{\sigma \in S_C} \prod^d_{j=1}( [x_{i_j}] L_{\sigma(j)})
%\]
%where $S_C$ is the set of all permutations of $d$ elements where the type $k$ element is repeated
%$e_k$ times for $1\leq k\leq r$.
%\end{lemma}
%\begin{proof}
%By abusing the notation, identify the linear forms as $L_{1}, L_2, \ldots, L_d$ where $L_1, \ldots,L_{e_1}$ are the same, $L_{e_1 + 1}, \ldots, L_{e_1 + e_2}$ are the same and so on.
%The monomial $m$ can be obtained from $C$ by first fixing a bijection $\sigma : [d] \mapsto [d]$ and considering the coefficient $[m]C_{\sigma}= \prod^{d}_{j=1} [x_{i_j}] L_{\sigma(j)}$.
%This is one way of generating this monomial and this monomial $m$ can be generated in many different orders. The final $[m]C$ is the sum of all coefficients $[m]C_{\sigma}$ generated in all distinct orders and the lemma follows.
%\end{proof}
Now we state the main theorem of the section.
\begin{theorem}\label{main-thm-2}
Let $C$ be a depth-3 $\Sigma\Pi\Sigma$ circuit of size at most $s$, computing a polynomial $f \in \F[x_1,\ldots, x_n]$ and the fan-in of the product gates of $C$ is bounded by $d$. For $\char(\F) > d$, we give a \emph{white-box} deterministic polynomial time identity testing algorithm to check whether $f\equiv 0$ or not in time $2^{\gamma\cdot d} \poly(n,s)$.
%For $\char(\F)\leq d$, we give a \emph{white-box} deterministic polynomial time identity testing algorithm to check whether $f\equiv 0$ or not in time $2^{\gamma\cdot d\log d} \poly(n,s)$.
The constant $\gamma$ is at most $5$.
\end{theorem}
\begin{proof}
Consider first the case when $p=\char(\F) > d$.
From Lemma \ref{calc}, notice that for any $\Pi^{[d]}\Sigma$ circuit $P$,
\[
[m]P = \frac{1}{m!} \sum_{\sigma \in S_d} \prod^d_{j=1}( [x_{i_j}] L_{\sigma(j)}).
\]
and $m!\neq 0 \mod p$. Now define the $d\times d$ matrix $T_P$ such that each row of $T_P$ is just the linear forms
$L_1 L_2 \ldots L_d$ appearing in $P$ \footnote{Again, we identify the linear forms as $L_{1}, L_2, \ldots, L_d$ where $L_1, \ldots,L_{e_1}$ are the same, $L_{e_1 + 1}, \ldots, L_{e_1 + e_2}$ are the same and so on.}. Clearly the following is true.
\[
\perm(T_P)=\sum_{\sigma \in S_d} \prod^d_{j=1} L_{\sigma(j)}.
\]
Use Ryser's formula given by Lemma \ref{ryser}, to express $\perm(T_P)$ as a depth-3 $\Sigma^{[2^d]}\Pi^{[d]}\Sigma$ circuit. If $C=P_1 + \ldots + P_s$, consider the polynomial $f_C = \sum_{i=1}^s \perm(T_{P_i})$. Notice that $f_C$ can be expressed as $\Sigma^{[2^d \cdot s]}\Pi^{[d]}\Sigma$ circuit. Consider the \emph{noncommutative} version of the polynomial $f_C$ which we denote as $f_C^{nc}$. Clearly we have a noncommutative ABP for $f_C^{nc}$ of width $w=2^d \cdot s$ and $d$ many layers.
%the number of layers is $d$.
%Now we make an important observation from the proof of Lemma \ref{calc}. Suppose $m$ be a monomial
%in $f$ and $[m] f = \alpha_m$. The monomial $m$ can be present in $f_C^{nc}$ in different orders: $m_1, \ldots, m_k$. Now we claim that $f\equiv 0$ if and only if $f_C^{nc}\equiv 0$. So see the claim, the following simple lemma suffices.
%
Now we make an important observation from the proof of Lemma \ref{calc}. Suppose $\mathcal{M}$ be the set of all monomials of degree $d$ over $x_1,\ldots,x_n$. For a fixed monomial $m\in\mathcal{M}$ of form $x_{i_1}x_{i_2}\ldots x_{i_d}$ where $i_1\leq i_2\leq\ldots \leq i_d$ and $\sigma\in S_d$, define $m^\sigma = x_{i_{\sigma(1)}}x_{i_{\sigma(2)}}\ldots x_{i_{\sigma(d)}}$. The monomial $m$ can be present in $f_C^{nc}$ in different orders $m^\sigma$. We claim that $f\equiv 0$ if and only if $f_C^{nc}\equiv 0$. To see the claim, the following simple lemma suffices.
%\begin{lemma}\label{important}
%Let $f = \sum_{j=1} ^K \alpha_{m_j} \cdot {m_j}$ where $\alpha_{m_j} \in\F$ %for all monomials $m_j$ in f. Then $f_C^{nc} = \sum_{j=1}^K \sum_{\ell=1}%^{s_{j}} (m_j)! \cdot \alpha_{m_j} \cdot m_{\ell}^{j}$ where
%$m_1^{j}, m_2^{j}, \ldots, m_{s_j}^{j}$ are the different ordering of $m_j$ %appearing in $f_C^{nc}$.
%\end{lemma}
\begin{lemma}\label{important}
Let $f = \sum_{m\in\mathcal{M}} [m]f \cdot {m}$ where $[m]f \in\F$ for all monomials $m\in\mathcal{M}$. Then
\[f_C^{nc} = \sum_{m\in\mathcal{M}} \sum_{\sigma\in S_d} m! \cdot [m]f \cdot m^{\sigma}.\]
\end{lemma}
\begin{proof}
Let $x_{i_1}\ldots x_{i_d}$ be a fixed ordering of a monomial $m$ appearing in $f_{C}^{nc}$. The coefficient of $x_{i_1}\ldots x_{i_d}$ in $\perm(T_P)=\sum_{\sigma \in S_d} \prod^d_{j=1} L_{\sigma(j)}$ is simply
%$\sum_{\sigma\in S_d} \prod_{j=1}^d a_{\sigma(j), i_j}$
$\sum_{\sigma\in S_d} \prod_{j=1}^d [x_{i_j}]L_{\sigma(j)}.$
%where each $L_i = \sum_{j=1}^n a_{i,j} x_j$.
But from Lemma \ref{calc},
%$\sum_{\sigma\in S_d} \prod_{j=1}^d a_{\sigma(j), i_j}$
$\sum_{\sigma\in S_d} \prod_{j=1}^d [x_{i_j}]L_{\sigma(j)}$ is exactly
%$m! \cdot \alpha_{m}^P$ where $\alpha_m^{P}$ is the coefficient of $m$ in the $\Pi^{[d]}\Sigma$ circuit $P$.
$m! \cdot [m]P$.
Since $[m]f = \sum_{i=1}^s [m]{P_i}$, the lemma follows.
\end{proof}
%But $[m_i] f_C^{nc} = m! \cdot \alpha_i$ where $\alpha_m = \sum_{i=1}^k \alpha_i$. Hence, we have that $f\equiv 0$ if and only if $f_{C}^{nc}\equiv 0$.
Now we apply the identity testing algorithm of
Raz and Shpilka for noncommutative ABPs on the ABP of $f_C^{nc}$ to get the desired result \cite{RS05}. The bound on $\gamma$ comes from Theorem 4 of their paper \cite{RS05}.
%Now we consider the case when $\char(\F)\leq d$. From Lemma \ref{calc-1}, notice that for any $\Pi^{[d]}\Sigma$ gate $P$
%\[
%[m]P = \sum_{\sigma \in S_P} \prod^d_{j=1}( [x_{i_j}] L_{\sigma(j)})
%\]
%where $S_P$ is the set of all permutations of $d$ elements where the type $k$ element is repeated
%$e_k$ times for $1\leq k\leq r$.
%
%A depth-3 $\Sigma^{2^{N}}\Pi^{[d]}\Sigma$ circuit can be easily constructed for the polynomial
%$T_P=\sum_{\sigma \in S_P} \prod^d_{j=1} L_{\sigma(j)}$ where $N\leq {d\log d}$. Similar to the case above,
%construct a depth-3 $\Sigma^{[2^{N}\cdot s]}\Pi^{[d]}\Sigma$ circuit for the polynomial $f_C=\sum_{i=1}^{s}T_{P_i}$. Again we consider the noncommutative version of the polynomial $f_C$ which we call as $f_C^{nc}$. Now, using an identical argument like the large characteristics case, we observe that
%$f\equiv 0$ if and only if $f_C^{nc}\equiv 0$ \footnote{For this, we can easily prove a lemma similar to Lemma \ref{important}. The only difference will be that the scaling terms $(m_j)!$ will not appear as we are summing only over distinct permutations. }. Clearly we have a noncommutative ABP for $f_C^{nc}$ of width $w=2^N \cdot s$ and the height is $d$. Now, we apply the algorithm of Raz and Shpilka to get the desired result.
%\begin{lemma}\label{calc}
%For a monomial $m = x_{i_1}x_{i_2} \ldots x_{i_d}$ ($i_1 ,\ldots , i_d$ need not be distinct) and a homogeneous $\Pi\Sigma$ circuit $C=\prod^d_{j=1} L_j$ we have:
%\[
%[m]C = \frac{1}{m!} \sum_{\sigma \in S_d} \prod^d_{j=1}( [x_{i_j}] L_{\sigma(j)}).
% \]
%\end{lemma}
%\begin{proof}
\end{proof}
As an immediate application of Theorem \ref{main-thm-2}, we state the following corollary.
\begin{corollary}\label{poly-depth3-finite}
Let $C$ be a depth-3 $\Sigma\Pi\Sigma$ circuit of size at most $s$, computing a polynomial $f\in \F[x_1,\ldots, x_n]$ and the fan-in of the product gates of $C$ is bounded by $d$. Suppose
that $\char(\F) > d$. For $d=O(\log n)$, we give a deterministic $\poly(n,s)$ time identity testing algorithm to check whether $f\equiv 0$ or not.
%Suppose that $\char(\F)\leq d$. For $d=O(\log n/\log\log n)$, we give a deterministic $\poly(n,s)$ time identity testing algorithm to check whether $f\equiv 0$ or not.
\end{corollary}
%\bibliographystyle{alpha}
%\bibliography{ref1}
\newcommand{\etalchar}[1]{$^{#1}$}
\begin{thebibliography}{ALM{\etalchar{+}}98}
\bibitem[Agr05]{Agr05}
Manindra Agrawal.
\newblock Proving lower bounds via pseudo-random generators.
\newblock In {\em {FSTTCS} 2005: Foundations of Software Technology and
Theoretical Computer Science, 25th International Conference}, pages 92--105,
2005.
\bibitem[AJ09]{AJ09}
Vikraman Arvind and Pushkar~S. Joglekar.
\newblock Arithmetic circuits, monomial algebras and finite automata.
\newblock In {\em Mathematical Foundations of Computer Science 2009, 34th
International Symposium, {MFCS} 2009, Novy Smokovec, High Tatras, Slovakia,
August 24-28, 2009. Proceedings}, pages 78--89, 2009.
\bibitem[AKS04]{aks04}
Manindra Agrawal, Neeraj Kayal, and Nitin Saxena.
\newblock {PRIMES} is in {P}.
\newblock {\em Ann. of Math}, 160(2):781--793, 2004.
\bibitem[ALM{\etalchar{+}}98]{almss98}
Sanjeev Arora, Carsten Lund, Rajeev Motwani, Madhu Sudan, and Mario Szegedy.
\newblock Proof verification and the hardness of approximation problems.
\newblock {\em J. {ACM}}, 45(3):501--555, 1998.
\bibitem[AM10]{AM10}
Vikraman Arvind and Partha Mukhopadhyay.
\newblock The ideal membership problem and polynomial identity testing.
\newblock {\em Inf. Comput.}, 208(4):351--363, 2010.
\bibitem[AS18]{AS18}
Vikraman Arvind and Srikanth Srinivasan.
\newblock On the hardness of the noncommutative determinant.
\newblock {\em Computational Complexity}, 27(1):1--29, 2018.
\bibitem[AV08]{av2008}
Manindra Agrawal and V~Vinay.
\newblock Arithmetic circuits: A chasm at depth four.
\newblock In {\em Proceedings-Annual Symposium on Foundations of Computer
Science}, pages 67--75. IEEE, 2008.
\bibitem[DL78]{DL78}
Richard~A. DeMillo and Richard~J. Lipton.
\newblock A probabilistic remark on algebraic program testing.
\newblock {\em Inf. Process. Lett.}, 7:193--195, 1978.
\bibitem[dOSlV16]{OSV16}
Rafael~Mendes de~Oliveira, Amir Shpilka, and Ben lee Volk.
\newblock Subexponential size hitting sets for bounded depth multilinear
formulas.
\newblock {\em Computational Complexity}, 25(2):455--505, 2016.
\bibitem[DS07]{DS07}
Zeev Dvir and Amir Shpilka.
\newblock Locally decodable codes with two queries and polynomial identity
testing for depth 3 circuits.
\newblock {\em {SIAM} J. Comput.}, 36(5):1404--1434, 2007.
\bibitem[GKKS13]{gkks2013}
Ankit Gupta, Pritish Kamath, Neeraj Kayal, and Ramprasad Saptharishi.
\newblock Arithmetic circuits: A chasm at depth three.
\newblock In {\em FOCS}, pages 578--587, 2013.
\bibitem[HS80]{HS80}
Joos Heintz and Claus{-}Peter Schnorr.
\newblock Testing polynomials which are easy to compute (extended abstract).
\newblock In {\em Proceedings of the 12th Annual {ACM} Symposium on Theory of
Computing, 1980}, pages 262--272, 1980.
\bibitem[KI04]{KI04}
Valentine Kabanets and Russell Impagliazzo.
\newblock Derandomizing polynomial identity tests means proving circuit lower
bounds.
\newblock {\em Computational Complexity}, 13(1-2):1--46, 2004.
\bibitem[KS07]{KS07}
Neeraj Kayal and Nitin Saxena.
\newblock Polynomial identity testing for depth 3 circuits.
\newblock {\em Computational Complexity}, 16(2):115--138, 2007.
\bibitem[KS09]{KS09}
Neeraj Kayal and Shubhangi Saraf.
\newblock Blackbox polynomial identity testing for depth 3 circuits.
\newblock In {\em 50th Annual {IEEE} Symposium on Foundations of Computer
Science, {FOCS} 2009}, pages 198--207, 2009.
\bibitem[KS11]{KS11}
Zohar~Shay Karnin and Amir Shpilka.
\newblock Black box polynomial identity testing of generalized depth-3
arithmetic circuits with bounded top fan-in.
\newblock {\em Combinatorica}, 31(3):333--364, 2011.
\bibitem[Lov79]{lov79}
L{\'{a}}szl{\'{o}} Lov{\'{a}}sz.
\newblock On determinants, matchings, and random algorithms.
\newblock In {\em {FCT}}, pages 565--574, 1979.
\bibitem[MVV87]{mvv87}
Ketan Mulmuley, Umesh~V. Vazirani, and Vijay~V. Vazirani.
\newblock Matching is as easy as matrix inversion.
\newblock {\em Combinatorica}, 7(1):105--113, 1987.
\bibitem[RS05]{RS05}
Ran Raz and Amir Shpilka.
\newblock Deterministic polynomial identity testing in non-commutative models.
\newblock {\em Computational Complexity}, 14(1):1--19, 2005.
\bibitem[Rys63]{ry63}
H.J. Ryser.
\newblock {\em Combinatorial Mathematics}.
\newblock Carus mathematical monographs. Mathematical Association of America,
1963.
\bibitem[Sch80]{Sch80}
Jacob~T. Schwartz.
\newblock Fast probabilistic algorithms for verification of polynomial
identities.
\newblock {\em J. {ACM}}, 27(4):701--717, 1980.
\bibitem[Sha90]{shamir90}
Adi Shamir.
\newblock {I}{P}={P}{S}{P}{A}{C}{E}.
\newblock In {\em 31st Annual Symposium on Foundations of Computer Science, St.
Louis, Missouri, USA, October 22-24, 1990, Volume {I}}, pages 11--15, 1990.
\bibitem[SS12]{SS12}
Nitin Saxena and C.~Seshadhri.
\newblock Blackbox identity testing for bounded top-fanin depth-3 circuits: The
field doesn't matter.
\newblock {\em {SIAM} J. Comput.}, 41(5):1285--1298, 2012.
\bibitem[SY10]{sy10}
Amir Shpilka and Amir Yehudayoff.
\newblock Arithmetic circuits: A survey of recent results and open questions.
\newblock {\em Foundations and Trends in Theoretical Computer Science},
5(3-4):207--388, 2010.
\bibitem[Zip79]{Zip79}
Richard Zippel.
\newblock Probabilistic algorithms for sparse polynomials.
\newblock In {\em Symbolic and Algebraic Computation, {EUROSAM} '79, An
International Symposiumon Symbolic and Algebraic Computation}, pages
216--226, 1979.
\end{thebibliography}
\end{document}
\end{document}