Challenges in computational lower bounds

We draw two incomplete, biased maps of challenges in computational complexity lower bounds.

💡 Research Summary

The paper by Emanuele Viola, “Challenges in computational lower bounds,” presents a structured overview of open problems in proving lower bounds for explicit Boolean functions. Rather than surveying historical results, the author draws two incomplete, biased “maps” that capture the landscape of challenges across different computational models.

The first map focuses on circuit models with various gate types—majority (Maj), threshold (Thr), symmetric (Sym), AND, and combinations thereof—as well as communication complexity settings such as number‑on‑forehead protocols. Each node in the map corresponds to a concrete challenge: to exhibit an explicit function that cannot be computed within the resources labeled on that node (e.g., size = q, fan‑in = log q, communication rounds = log q). Arrows indicate simulation relationships: if resources A can simulate resources B, then a lower bound for A implies one for B. The paper lists eight challenges (1.1)–(1.8). For quasi‑polynomial size q = n^{c log c}, several equalities hold (e.g., Maj‑Maj‑And log q ↔ Maj‑Sym‑And log q ↔ Sym‑And log q). When q is polynomial, these equalities break down, and only a few “obvious” arrows remain, such as Sym‑And log q → Maj‑And log q and Maj‑Sym → Maj‑Maj (the latter from Goldmann‑Håstad‑Razborov). The paper also notes new arrows derived from techniques in Viola’s own lecture notes, linking certain communication models to circuit classes.

The second map addresses depth‑3 circuits, branching programs, and NC¹‑type circuits. Challenges (2.1)–(2.8) involve parameters such as ε‑correlation, circuit depth, size, fan‑in, and program width/length. For example, (2.1) asks for an ε > 0 such that every depth‑O(1) circuit of size 2^{n^{ε}} fails to compute a given explicit function; the best known bound for depth 3 is 2^{c√n}. Challenge (2.3) seeks improvement over this bound, while (2.4) asks to break the trade‑off between fan‑in k and size for depth‑3 circuits.

A central technical tool connecting the two maps is the “guess‑recurse” method, originally due to Nepomnjašči. Lemma 2.9 shows how a branching program of width w and time t can be simulated by a depth‑3 circuit with at most 2 √t log w + log t wires, output fan‑in w·b, and input fan‑in t/b, where b is a parameter. This yields the arrows (2.1)→(2.5), (2.4)→(2.7), etc., and demonstrates that polynomial‑size, polynomial‑length programs (including those recognizing NC¹) can be compressed into shallow circuits. The proof sketches a “guess middle points” strategy, converting each guessed interval into a CNF and then collapsing layers.

The paper also references a variety of known results that underpin the arrows: boosting and min‑max duality for simulating polynomials by circuits, the discriminator lemma for the reverse direction, Razborov‑Wigderson’s Ω(log n) lower bounds for Maj‑Sym‑And circuits, and the equivalence Maj‑Thr = Maj‑Maj up to polynomial size.

Overall, Viola’s contribution is a meta‑analysis that clarifies which lower‑bound questions are already resolved via known simulations, which are equivalent under current techniques, and which remain open. By explicitly mapping the relationships, the paper provides a roadmap for researchers: to attack a hard challenge, one may either develop new simulation techniques (new arrows) or strengthen existing lower‑bound methods (e.g., boosting, discriminator lemmas, communication complexity arguments). The work underscores that progress on any single node or arrow can cascade through the map, potentially collapsing several open problems at once.