Read-Once Formulas in Complexity

• Read-once formulas are computation trees where each variable appears only once, producing multilinear polynomials in both Boolean and arithmetic contexts.
• Key decompositions, such as monomial pairing and inductive methods, establish that every multilinear polynomial can be expressed as a bounded sum of ROFs, underpinning sharp hierarchy theorems.
• The study of ROFs drives advances in efficient property testing, learning algorithms, and circuit lower bounds, while revealing open challenges in explicit lower bounds within arithmetic complexity.

A read-once formula is a central construct in algebraic and Boolean complexity theory, characterized by the requirement that each input variable appears at most once within the computation tree. In the arithmetic setting, read-once formulas (ROFs) generate only multilinear polynomials. The study of read-once formulas provides structural insight into the representation of both Boolean and arithmetic functions, supports sharp hierarchy theorems, enables efficient property testing and learning algorithms, and underpins strong circuit lower bounds on sums of read-once formulas.

1. Formal Definition and Structural Properties

Let $\mathbb{F}$ be a field and $X = \{x_1, ..., x_n\}$ a set of variables. An arithmetic read-once formula (ROF) is a binary computation tree with internal nodes labeled by $+$ or $\times$, and leaves labeled by variables or field constants, where each variable labels at most one leaf, and all fan-out is 1. Each node $v$ carries coefficients $(\alpha_v, \beta_v) \in \mathbb{F}^2$; a leaf labeled $x_i$ with $(\alpha, \beta)$ outputs $\alpha x_i + \beta$, while an internal node with children $f_1, f_2$ and operation $X = \{x_1, ..., x_n\}$0 computes $X = \{x_1, ..., x_n\}$1.

Every polynomial computed by such a formula is multilinear. The class of read-once polynomials (ROPs) consists of all polynomials computable by some ROF. The key property is that any ROF computes a polynomial of individual degree at most one in each variable.

A combinatorial analog exists in the Boolean case: a Boolean read-once formula is a tree over $X = \{x_1, ..., x_n\}$2 in which every variable (possibly negated) appears once, supporting, e.g., monotone read-once CNF and DNF.

2. Representation Power: Sum-of-ROFs

A fundamental result is that every multilinear polynomial can be expressed as a sum of a bounded number of ROPs. Two constructive upper bounds are established:

• Monomial pairing: Group $X = \{x_1, ..., x_n\}$3 nonzero multilinear monomials in $X = \{x_1, ..., x_n\}$4 into pairs. For each pair $X = \{x_1, ..., x_n\}$5, the polynomial $X = \{x_1, ..., x_n\}$6 is a ROP. Thus $X = \{x_1, ..., x_n\}$7 is the sum of at most $X = \{x_1, ..., x_n\}$8 ROPs.
• Inductive decomposition: For $X = \{x_1, ..., x_n\}$9, every $+$0-variate multilinear polynomial can be written as the sum of at most $+$1 ROPs by setting $+$2 to peel off variables recursively (Mahajan et al., 2016).

This guarantees that the sum-of-ROPs model is universal for multilinear polynomials.

3. Hierarchy Theorem: Tight Lower and Upper Bounds

The main technical achievement is a hierarchy theorem quantifying the number of ROP summands required for specific explicit families. For the symmetric polynomials

$+$3

and more generally for

$+$4

it is shown that:

• For every field $+$5, every $+$6, and every $+$7, the polynomial $+$8 cannot be written as a sum of fewer than $+$9 ROPs, but can be written as a sum of exactly $\times$0 ROPs (Mahajan et al., 2016, Mahajan et al., 2015).
• For example, the degree-$\times$1 symmetric polynomial $\times$2 requires $\times$3 ROPs.

The decomposition exploits explicit variable-pairing: for $\times$4 even, each summand is a degree-$\times$5 monomial with two variables "opened," the rest multiplied together; the construction generalizes to all $\times$6.

4. Proof Techniques and Structural Insights

The hierarchy theorem leverages several technical tools:

• Partial derivative elimination: If $\times$7 for some ROP summand $\times$8, then $\times$9 can be eliminated from the sum by differentiating, reducing to the lower-$v$0 case and supporting induction.
• Multiplicative ROFs and variable-splitting property: Any ROF with nonzero mixed second derivatives on all variable pairs is multiplicative (no $v$1 gates). A key lemma shows that for multiplicative ROPs, some variable can be set to a constant to lower the degree structure, allowing an induction step.
• Commutator analysis: The commutator operator, $v$2, is used to manage small-variable (notably $v$3) cases by detecting linear dependencies or structural factorizations in the sum of two ROPs.

5. Complexity Implications and Open Problems

The sum-of-ROPs model sits strictly between small-depth multilinear circuits and general formulas:

• The separation result $v$4 holds for $v$5 up to about $v$6.
• Any superpolynomial lower bound for sums of ROPs computing explicit polynomials (e.g., determinant or permanent) would sharply separate complexity classes relevant for polynomial identity testing (PIT) (Mahajan et al., 2016).
• The current best-known upper bound for expressing a general multilinear polynomial is $v$7 ROP summands, but the best explicit lower bound is only $v$8.
• Characterizing fields over which certain low-variable separations (e.g., 4 variables, $v$9) exist remains open.

A counting argument shows that a random multilinear polynomial requires $(\alpha_v, \beta_v) \in \mathbb{F}^2$0 ROPs, but no explicit construction is known to match this lower bound (Mahajan et al., 2016).

6. Connections to Determinantal and Circuit Complexity

Read-once structures play a significant role in algebraic complexity. Read-once determinants (ROD$(\alpha_v, \beta_v) \in \mathbb{F}^2$1) are a strict subclass of low-determinantal complexity polynomials; the best known unconditional limitations for the permanent and certain symmetric polynomials in this model use similar variable occurrence constraints. Lower bounds for the sum-of-ROPs model would yield significant complexity class separations and have implications for PIT and hardness-vs-randomness paradigms in arithmetic circuit complexity (Aravind et al., 2015, Ramya et al., 2015).

7. Summary Table: Key Quantitative Results

Polynomial Family	Lower Bound (min ROP summands)	Matching Upper Bound
General multilinear, $(\alpha_v, \beta_v) \in \mathbb{F}^2$2 vars	--	$(\alpha_v, \beta_v) \in \mathbb{F}^2$3
$(\alpha_v, \beta_v) \in \mathbb{F}^2$4 (degree $(\alpha_v, \beta_v) \in \mathbb{F}^2$5 symmetric)	$(\alpha_v, \beta_v) \in \mathbb{F}^2$6	$(\alpha_v, \beta_v) \in \mathbb{F}^2$7
Random multilinear polynomial	$(\alpha_v, \beta_v) \in \mathbb{F}^2$8 (non-constructive)	$(\alpha_v, \beta_v) \in \mathbb{F}^2$9

Tightness is achieved for the explicit $x_i$0 family; the exponential gap for general polynomials highlights a major open area.

---

References:

Mahajan, Tawari, "Sums of read-once formulas: How many summands suffice?" (Mahajan et al., 2016, Mahajan et al., 2015). Shpilka, Volkovich, CCC 2014. Raz, STOC 2004. See also (Ramya et al., 2015) and (Aravind et al., 2015) for lower bounds and implications for related arithmetic models.