Communication lower bounds via Advisor-Verifier-Prover games

One model that three unrelated primitives reduce to, and one hypothesis about it

Motivation

Information-theoretic cryptography has a long-standing embarrassment. For its most basic primitives — private information retrieval, general secret sharing, conditional disclosure of secrets, garbling — nobody can prove a super-linear communication lower bound, let alone the exponential ones everyone expects. The best bounds date from the mid-to-late 1990s. And the questions are isolated from one another: a breakthrough on secret sharing would not transfer to PIR, or the reverse.

That isolation is the real obstacle, because it means every technique has to be invented once per primitive. A model that all of them reduce to would let a single argument pay off several times over — and would let partial progress accumulate instead of scattering.

Provenance and history

Benny Applebaum and Oded Nir, Advisor-Verifier-Prover Games and the Hardness of Information Theoretic Cryptography (IACR ePrint 2023/1378), supply such a model, by relating the questions to query/space trade-offs lifted to interactive proof systems.

In an AVP game an advisor sees a function \(f\), computes a randomized advice string, hands it privately to the verifier, and leaves. The verifier then receives an input \(x\), sends one query to an untrusted prover, and decides whether \(f(x) = 1\) from the prover’s single reply. The advice is hidden from the prover and is not reusable; all parties are unbounded; the cost is the sum of the three message lengths.

The paper gives communication-efficient transforms from sd-PIR, general secret sharing and fully-decomposable randomized encodings into AVPs with a laconic prover — one whose answer is a constant number of bits. Its Hypothesis 1.2 asserts that prover-laconic AVPs need super-polynomial communication, and its Theorem 1.3 converts that into super-polynomial lower bounds for all three primitives at once.

Two facts locate the hypothesis. It is not vacuous: the transforms also yield a prover-laconic AVP at \(2^{\tilde{O}(\sqrt{n})}\) with a single-bit prover, so such protocols exist sub-exponentially. And it is false without laconicity: the paper’s Example 1.1 gives a non-laconic AVP at polynomial cost, derived from an \(n\)-server PIR protocol.

What the paper proves toward the hypothesis is deliberately modest. Counting-based lower bounds adapt to the model, re-deriving several existing bounds in a unified way, and the paper states plainly that it does not reach the best-known bound for any of the primitives. It offers Hypothesis 1.2 “both as a working hypothesis or as an (ambitious) target for future works.”

Parameter lattice

Where each primitive stands, and what the hypothesis would give.

Primitive Best upper bound Best lower bound Under Hypothesis 1.2
sd-PIR, \(2^n\)-bit database \(2^{\tilde{O}(\sqrt{n})}\) \(Cn\), \(C > 1\) (late 1990s) super-polynomial
Secret sharing, \(n\)-bit predicates \((3/2)^{n+o(n)}\) \(\Omega(n/\log n)\) (mid 1990s) super-polynomial
Fully-decomposable RE, \(n\)-bit predicates no super-linear bound super-polynomial

And the AVP model itself, by prover length.

Prover Communication achieved Status
Laconic (\(C\) bits, \(C\) constant) \(2^{\tilde{O}(\sqrt{n})}\), single-bit prover proved
Laconic \(\mathrm{poly}(n)\) c/0089 — conjectured impossible
Non-laconic \(\mathrm{poly}(n)\) proved, Example 1.1 — so laconicity is essential

Statements in this hub