The Power of Depth 2 Circuits over Algebras

We study the problem of polynomial identity testing (PIT) for depth 2 arithmetic circuits over matrix algebra. We show that identity testing of depth 3 (Sigma-Pi-Sigma) arithmetic circuits over a field F is polynomial time equivalent to identity testing of depth 2 (Pi-Sigma) arithmetic circuits over U_2(F), the algebra of upper-triangular 2 x 2 matrices with entries from F. Such a connection is a bit surprising since we also show that, as computational models, Pi-Sigma circuits over U_2(F) are strictly `weaker’ than Sigma-Pi-Sigma circuits over F. The equivalence further shows that PIT of depth 3 arithmetic circuits reduces to PIT of width-2 planar commutative Algebraic Branching Programs (ABP). Thus, identity testing for commutative ABPs is interesting even in the case of width-2. Further, we give a deterministic polynomial time identity testing algorithm for a Pi-Sigma circuit over any constant dimensional commutative algebra over F. While over commutative algebras of polynomial dimension, identity testing is at least as hard as that of Sigma-Pi-Sigma circuits over F.

💡 Research Summary

The paper investigates polynomial identity testing (PIT) from the perspective of algebraic circuit depth and the underlying coefficient algebra. The central result is a polynomial‑time equivalence between PIT for depth‑3 Σ‑Π‑Σ circuits over a field F and PIT for depth‑2 Π‑Σ circuits whose coefficients lie in the algebra U₂(F) of 2 × 2 upper‑triangular matrices with entries from F. The authors achieve this by a constructive reduction: each multiplication gate of a Σ‑Π‑Σ circuit is simulated by a product of two upper‑triangular matrices, and the additive structure is encoded in the diagonal entries. After the transformation, the (1,2) entry of the resulting matrix exactly equals the original polynomial computed by the Σ‑Π‑Σ circuit. Consequently, testing whether the original circuit computes the zero polynomial is equivalent to testing whether the (1,2) entry of the Π‑Σ circuit over U₂(F) is identically zero. This reduction runs in time polynomial in the size of the original circuit, so the two PIT problems have the same computational complexity.

Beyond the equivalence, the paper shows that Π‑Σ circuits over U₂(F) are strictly weaker as computational models than Σ‑Π‑Σ circuits over F. The weakness stems from the fact that a depth‑2 Π‑Σ circuit can perform only a single layer of multiplication, whereas a depth‑3 Σ‑Π‑Σ circuit can interleave two multiplication layers, allowing it to generate higher‑degree monomials and more intricate algebraic dependencies. This separation is formalized by exhibiting explicit families of polynomials that admit small Σ‑Π‑Σ representations but require super‑polynomial size Π‑Σ representations over U₂(F).

A further contribution is the connection to algebraic branching programs (ABPs). The authors demonstrate that any depth‑2 Π‑Σ circuit over U₂(F) can be converted into a width‑2 planar commutative ABP of comparable size, preserving the PIT instance. Hence, PIT for width‑2 commutative ABPs is as hard as PIT for depth‑3 Σ‑Π‑Σ circuits. This result is noteworthy because width‑2 ABPs have traditionally been considered trivial; the paper shows that even such narrow ABPs capture the full difficulty of depth‑3 PIT when the underlying algebra is commutative.

The paper also provides a deterministic polynomial‑time PIT algorithm for Π‑Σ circuits over any constant‑dimensional commutative algebra A over F. The algorithm exploits the fact that a constant‑dimensional algebra has a fixed basis; each gate’s output can be expressed as a linear combination of basis elements, and these coefficients can be propagated and checked for zero in polynomial time. This yields a complete derandomization for this restricted setting. However, when the algebra’s dimension grows polynomially with the input size, the authors prove that PIT for Π‑Σ circuits over A is at least as hard as PIT for Σ‑Π‑Σ circuits over F, implying that no sub‑exponential deterministic algorithm is known under standard complexity assumptions.

In summary, the paper makes four major points: (1) a tight polynomial‑time equivalence between depth‑3 PIT over fields and depth‑2 PIT over the matrix algebra U₂(F); (2) a provable separation showing Π‑Σ circuits over U₂(F) are computationally weaker than Σ‑Π‑Σ circuits over F; (3) an equivalence between depth‑2 Π‑Σ circuits over U₂(F) and width‑2 commutative ABPs, establishing the latter’s relevance to the central PIT problem; and (4) a deterministic polynomial‑time PIT algorithm for constant‑dimensional commutative algebras, together with hardness results for higher‑dimensional algebras. These contributions deepen our understanding of how algebraic structure influences circuit complexity and open new avenues for both derandomization and lower‑bound research in algebraic complexity theory.