Back to Arithmetic Circuits
Back to notes

# Multilinear circuits and formulas

20%

Let $f(\bx)\in\mathbb{F}[\bx]$, we say $f$ is multilinear if $f$ is linear in each of the variable $x_i$. That is, $\deg_i(f)\leq1$ for all $i\in[n]$. One can immediately see that most of the common polynomials are multilinear. For example, determinant, permanent, elementary symmetric polynomials etc.

What about computational models such as arithmetic circuits and formulas, can we have similar notion of multilinearity? We say a circuit (resp. formula) is multilinear if all its gates compute a multilinear polynomial.

An arithmetic circuit (resp. formula) is multilinear if all of its gates computed a multilinear polynomial. Further, we say it is syntactically multilinear if the children of every product gates contain disjoint set of variables.

Note that multilinearity is a restriction for arithmetic circuits/formulas since in general cancellation could happen. Especially, there is exponential lower bound for multilinear formula while the state-of-the-art formula lower bound is only quadratic.

An even more restricted definition is set-multilinear defined as follows.

Let $f\in\mathbb{F}[\bx]$. We say $f$ is $r$-set-multilinear if there exists a partition for $\bx$ of size $r$ $(\bx_1,\bx_2\dots,\bx_r)$ such that $f$ is multilinear w.r.t. each variable set $\bx_1$. Similarly, one can define set\multilinear circuits and formulas.

## Multilinear versus syntactically multilinear

It turns out that for formula, multilinearity is not far away from syntactically multilinearity.

Suppose $f\in\mathbb{F}[\bx]$ is computed by a size $s$ multilinear formula, then there exists a size $s$ syntactically multilinear formula computes $f$.

The proof is constructive. Given a multilinear formula of size $s$ for $f$. Suppose there is a product gate $g$ which is not syntactically multilinear. That is, there are two children $v_1,v_2$ of $g$ having the same variable $x_i$ in their leafs. As $g$ is multilinear, this means that one of the $x_i$ in $v_1,v_2$ does not contribute to the $x_i$ in $g$. Wlog, assume its the $x_i$ in $v_1$. Then, define a new formula by replacing the $x_i$ in the leafs of $v_1$ with 0. Note that this does not change the polynomial computed by the formula and now $g$ is syntactically multilinear in $x_i$. Furthermore, the size of the formula does not increase. By repeating this process, we yield a syntactically multilinear formula of the same size for the same polynomial.

Note that the above proof does not hold for multilinear circuits.

## Multilinearization

Given a circuit (resp. formula) for a multilinear polynomial, a natural question is asking whether there is a circuit (resp. formula) of similar size. A naive idea by duplicating each gates for $2^n$ times gives the following.

Let $f\in\mathbb{F}[\bx]$ be a multilinear polynomial. Suppose there exists a size $s$ circuit (resp. formula) computes $f$, then there exists a multilinear circuit (resp. formula) of size $s\cdot 2^n$ for $f$.

The idea is as trivial as it sounds. For each gate $g$ in the original circuit (resp. formula), duplicating it for $2^n$ where $g_{\ba}$ computes the coefficient of the monomial $x_1^{\ba_1}x_2^{\ba_2}\cdots x_n^{\ba_n}$ in $g$ for each $\ba\in\{0,1\}^n$. The construction can be done easily with induction. For the base case where $g=x_i$ for some $i$, simply let $g_{e_i}=1$ and $g_{\ba}=0$ for all $\ba\neq e_i$. For a sum gate $g=u+v$, let $g_{\ba}=u_{\ba}+v_{\ba}$ for each $\ba\in\{0,1\}^n$. For a product gate $g=u\cdot v$, $g_{\ba} = \sum_{\ba'\leq\ba}u_{\ba'}\cdot v_{\ba-\ba'}$ where the inequality and substraction are coordinate-wise.

The $2^n$ blow-up is definitely not desirable, it turns out that one can do much better for $r$-set-multilinear formula for small $r$ as shown by Raz.

Let $f\in\mathbb{F}[\bx]$ be a multilinear polynomial of degree $r$. Suppose $f$ is computed by a formula of size $s$ of depth $d$, then there exists a multilinear formula of size $O(s\cdot(d+2)^r)$ for $f$.

The proof mainly uses the fact that each gate in a formula has an unique path to the root. The idea is also based on duplicating gates, but in a more clever way. Some notations before going into the details. Let $C$ be a formula computing a multilinear polynomial. For each gate $g$ of $C$, let $path(g)$ denote the set of gates in the path from $g$ to the output gate of $C$. Next, let us consider the following observation. For a leaf $g$ in a formula computing a multilinear polynomials. Look at $path(g)$, and note that the *contribution* of $g$ along this path is *monotone* in the sense that the degree of $g$ would not decrease. (Here we treat two different leafs of same variable as different) This is due to the fact that formula would not reuse a gate. This observation is very good to us because once the degree of $g$ exceeds 1 at some point in $path(g)$, we know it would not contribute in the final output since the formula computes a multilinear polynomal. Thus, we don't need to keep track of $g$ beyond this point. Formally, this observation can be turned into the following construction. **TODO**

Some remarks on the theorem. First, by the depth reduction for formula, one can wlog assume $d=O(\log s)$. Thus, when $r=O(\frac{\log n}{\log\log n})$, the blow-up of the above multilinearization is only of a polynomial factor! Finally, this multilinearization is also a key step of the connection between formula lower bound and tensor rank in which a $\geq n^{(1-o(1))r}$ lower bound for order $r$ tensor implies super-polynomial formula lower bound.