It is known that there exist multi-prover interactive protocols (MIP protocols) for the complexity class NEXP, succinct MIP protocols for NP and multi-prover interactive protocols with shared entanglement (MIP^∗ protocols) for RE. This extraordinary power of multi-prover interactive proof systems comes from the assumption that provers do not communicate with each other during the protocols. If they are allowed to communicate freely, the setting is the same as in the single-prover case, and the computational power of the system becomes significantly weaker. In this paper, we investigate for the first time the setting where communication (i.e., leakage of information) between provers is allowed but bounded. We introduce two techniques to approach this question and show that multi-prover interactive proof systems are robust against some amount of leakage. Our first technique is based on parallel repetition theorems. We apply it to show that for any polynomial p, we can construct two-prover one-round MIP and MIP^∗ protocols for NEXP and RE, respectively, that are robust against p(n) bits of leakage. We further derive our second technique to convert any low-soundness PCP construction to a two-prover one-round MIP protocol for NP robust against leakage. We also discuss the relation between robustness against leakage in multi-prover interactive proof systems and the Sliding Scale Conjecture in the PCP literature.

Stochastic Gradient Descent (SGD) is the foundation of modern machine learning (ML). In privacy-sensitive settings, gradients can reveal details about individual data points. Differential Privacy (DP) protects sensitive data during ML training by clipping gradients and adding calibrated Gaussian noise. However, existing frameworks assume semi-honest participants, which fails in adversarial or federated environments where malicious actors can bypass or alter the noise addition process, breaking privacy guarantees. We present VeriDP, a framework for verifiable differentially private training that cryptographically enforces and proves the correct execution of differentially private stochastic gradient descent (DP-SGD) in zero knowledge. VeriDP integrates Zero-Knowledge Proofs (ZKPs) with polynomial commitments, sumcheck and GKR-based proofs, and incrementally verifiable computation (IVC) to generate compact proofs of correct gradient computation, clipping, averaging, and Gaussian noise generation—without revealing private data or randomness. Unlike previous systems that only verify the final privacy budget, VeriDP enables per-iteration verifiability of each model update, providing strong privacy assurances even in adversarial settings. This establishes a novel and complete Zero-Knowledge Proof of Differentially Private Stochastic Gradient Descent (ZK-DPSGD), uniting differential privacy and verifiable computation for secure and auditable ML. Our evaluation shows that prover time increases linearly with the number of input samples, while both verifier time (2–5 ms) and proof size (3–4 KB) remain compact and effectively constant.

A non-local quantum computation (NLQC) replaces an interaction between two quantum systems with a single simultaneous round of communication and shared entanglement. We study two classes of NLQC, f-routing and f-BB84, which are of relevance to classical information theoretic cryptography and quantum position-verification. We give the first non-trivial lower bounds on entanglement in both settings, but are restricted to lower bounding protocols with perfect correctness. Within this setting, we give a lower bound on the Schmidt rank of any entangled state that completes these tasks for a given function f(x,y) in terms of the rank of a matrix g(x,y) whose entries are zero when f(x,y) = 0, and strictly positive otherwise. This also leads to a lower bound on the Schmidt rank in terms of the non-deterministic quantum communication complexity of f(x,y). Because of a relationship between f-routing and the conditional disclosure of secrets (CDS) primitive studied in information theoretic cryptography, we obtain a new technique for lower bounding the randomness complexity of CDS.

We study the problem of constructing worst-case algorithms from average-case algorithms. Prior to this work, such reductions were only known for a small number of specific problems or restricted computational models. In contrast, we show that for quantum computation, all linear problems admit worst-case to average-case reductions. Specifically, we provide an explicit and efficient transformation of quantum algorithms that are only correct on a small (even sub-constant) fraction of their inputs into ones that are correct on all inputs. En route, we obtain a tight Ω(n2) lower bound on the average-case quantum query complexity of the Matrix-Vector Multiplication problem.

We present a new framework for designing worst-case to average-case reductions. For a large class of problems, it provides an explicit transformation of algorithms running in time T that are only correct on a small (subconstant) fraction of their inputs into algorithms running in time O(T) that are correct on all inputs.

Locally decodable codes (LDCs) are error-correcting codes C: Σ^k → Σⁿ that admit a local decoding algorithm that recovers each individual bit of the message by querying only a few bits from a noisy codeword. An important question in this line of research is to understand the optimal trade-off between the query complexity of LDCs and their block length. Despite importance of these objects, the best known constructions of constant query LDCs have super-polynomial length, and there is a significant gap between the best constructions and the known lower bounds in terms of the block length. For many applications it suffices to consider the weaker notion of relaxed LDCs (RLDCs), which allows the local decoding algorithm to abort if by querying a few bits it detects that the input is not a codeword. This relaxation turned out to allow decoding algorithms with constant query complexity for codes with almost linear length. Specifically, [{Ben-Sasson} et al., 2006] constructed a q-query RLDC that encodes a message of length k using a codeword of block length n = O_q(k^{1+O(1/√q)}) for any sufficiently large q, where O_q(⋅) hides some constant that depends only on q. In this work we improve the parameters of [{Ben-Sasson} et al., 2006] by constructing a q-query RLDC that encodes a message of length k using a codeword of block length O_q(k^{1+O(1/{q})}) for any sufficiently large q. This construction matches (up to a multiplicative constant factor) the lower bounds of [Jonathan Katz and Trevisan, 2000; Woodruff, 2007] for constant query LDCs, thus making progress toward understanding the gap between LDCs and RLDCs in the constant query regime. In fact, our construction extends to the stronger notion of relaxed locally correctable codes (RLCCs), introduced in [Tom Gur et al., 2018], where given a noisy codeword the correcting algorithm either recovers each individual bit of the codeword by only reading a small part of the input, or aborts if the input is detected to be corrupt.