arXiv:2610.01699
Efficiently Optimizing the Quantum Value of Bell Inequalities using Batched Gradient Descent
Xinyu Xu, Ping Zhu, Weikang Li, Pierre Pocreau, Dalu Ding, Dawei Ding
Abstract
Exploring the world of Bell inequalities requires numerical methods that scale effectively with the number of parties, inputs, and outputs. Such methods can be used to study large-scale Bell inequalities beyond the reach of analytic methods, to design experiments for violating more complicated Bell inequalities, and to enable applications of Bell inequality violation. Such applications include randomness generation and multi-agent coordination with restricted communication. However, current general purpose optimizers such as the see-saw method struggle to handle Bell inequalities with only a dozen inputs and outputs. We introduce an optimizer for the quantum value of a Bell inequality based on batched gradient descent (BGD). More precisely, our BGD optimizer is a differentiable search over feasible quantum strategies in which states and projective measurements are generated from unconstrained variables. The Bell expression is evaluated by a direct tensor contraction without forming the prohibitively large Bell operator. This formulation makes parallel random restarts inexpensive and is thus naturally implemented by a GPU. We evaluate our optimizer against the see-saw method and find a significant speedup for a wide variety of Bell inequality families. Implemented on a GPU, our BGD optimizer can optimize Bell inequalities in our numerical experiments with more than a thousand inputs or outputs within minutes. For these large-scale Bell inequalities, we also efficiently compute the classical value by enumerating possible strategies using a GPU or expressing the optimization as a mixed-integer linear program. Our optimizers can serve as numerical evaluation tools for Bell experiments and multi-agent coordination problems modeled by Bell inequalities with many inputs and outputs, a common feature of real-world scenarios such as high frequency trading and distributed systems.
AI Summary
The paper addresses numerical lower bounds on the quantum value of Bell inequalities when the numbers of parties, inputs, or outputs make standard see-saw optimization expensive. It formulates a differentiable search over normalized pure states and measurement bases generated by exponentiating skew-Hermitian matrices. Direct tensor contraction evaluates the Bell functional without constructing the dense Bell operator, while batched independent restarts support GPU execution. GPU enumeration with a one-party best-response reduction and a mixed-integer linear program provide complementary classical optimizers. Benchmarks compare CPU/GPU implementations with a modified see-saw implementation, generally using ten restarts and a 300-second budget per restart. On inequalities tailored to maximally entangled states, GPU optimization reaches 1,407 outputs with two inputs, versus 23 outputs for see-saw; both endpoints have 10% restart success. GPU optimization also reaches 13-party MABK and Mermin instances, versus eight parties for see-saw. For a jamming game with 1,820 inputs and four outputs, it obtains winning probability $0.16748$, exceeding the exact classical value $0.15797$ by approximately 6%. Analytic optima and, where tractable, NPA upper bounds assess the numerical results. The method returns feasible lower bounds, not global-optimality guarantees. Its default measurements are rank-one projective; stochastic post-processing does not cover arbitrary fixed-dimensional POVMs. State storage and dense game tensors remain exponential in party number. References for unsolved instances come from separate optimization runs, and the paper does not target very-high-precision optimization. Comparisons concern the specified implementations and resource limits.
Claude-family and GPT-family models were also used to assist with writing and proofreading the manuscript, as well as checking the code used for the numerical experiments.
Error found: The abstract overstates the demonstrated classical-value computation: the headline large-scale benchmarks use analytic classical values, and the 41-input, 41-output load-balancing benchmark has only a classical lower bound because neither enumeration nor MILP certified optimality.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 54 pages, 8 figures