In briefPaired MDS codes built from elliptic curves, with explicit constructions
SummaryThis paper studies linear complementary pairs (LCPs) of maximum distance separable (MDS) codes arising from elliptic curves. It gives geometric conditions for LCPs of elliptic codes and shows they cannot consist entirely of MDS codes when evaluated over all rational points of the curve. The authors instead restrict evaluation divisors to subgroup cosets to build LCPs of MDS elliptic codes, giving infinite families over Mordell and supersingular curves, including large-length constructions over quadratic extensions.
In briefUlaCache: A KV cache system using unified logical addressing for sparse LLM inference
SummaryThis paper presents UlaCache, a key-value cache management system for large language model inference that spans GPU memory and optional CPU memory under one unified logical address space. Stable logical references let the engine allocate and select KV entries regardless of where they physically reside, and a parallel bucketed LRU policy manages a GPU pool shared across requests. Under matched CPU and GPU memory budgets with speculative decoding enabled, the authors report peak decoding throughput speedups of 1.16x to 2.51x over the state-of-the-art baseline on H200 and B300 GPUs.
In briefMaximum length limits for cyclic MDS codes
SummaryThis paper studies the maximum length of nontrivial simple-root cyclic MDS codes over a finite field F_q. The authors prove that the length n is at most q+1 under certain gcd conditions, including cases where gcd(n, q²−1) exceeds 2, or the dimension is even and the gcd exceeds 1. The bound also holds for lengths whose prime factors lie in {2, 3, 5, 7}, with further results for some larger primes.
In briefWeight Spectrum of Binary Reed-Muller Codes RM(m-7,m): Partial Results
SummaryThis paper studies the weights that occur in high-order binary Reed-Muller codes RM(m-7,m). For every m≥14, it shows that the weight spectrum contains all even weights in [336, 2^m-336], along with a finite set of explicit low weights and their complements. The proof establishes a base inclusion for RM(7,14) and extends it to all m≥14 by a concatenation-based induction. The authors also show that several restricted construction routes cannot produce the four remaining base-code weights 322, 326, 330 and 334.
In briefAlmost MDS codes holding (n-k-1)-designs and an open problem on Witt designs
SummaryThis paper studies almost maximum distance separable (AMDS) codes over finite fields with parameters [n,k,n-k], where the minimum-weight codewords support an (n-k-1)-design. The authors determine the parameters of the dual code and characterize the design and its complementary design formed by its minimum-weight codewords. They give explicit weight enumerator formulas for subset codes of length at least k+1. The results are used to answer an open problem posed by Wang, Tang and Ding in 2023.
In briefExact nucleon computation in matching games: a theoretical game-theory paper on fair payoff allocation
SummaryThis paper studies exact computation of the nucleon, the multiplicative counterpart of the nucleolus, in matching games and two extensions: b-matching games and partitioned matching games. It gives a polynomial-time combinatorial algorithm for ordinary matching games with nonnegative rational weights. It also shows that computing any nucleon allocation is NP-hard for certain restricted b-matching and partitioned matching games. For the unresolved capacity-two case, it proves polynomial-time Turing reducibility to the shortest non-zero cycle problem.
WhyResolves open nucleon problems for matching games with a polynomial algorithm, several NP-hardness results, and a Turing reduction; proofs are rigorous but impact is confined to a niche subfield.
In briefS-PASS: Star-Topology Pinching-Antenna Systems with Pipelined Pre-Positioning
SummaryThis paper proposes S-PASS, a star architecture in which several waveguides share one radio-frequency chain through a switch, along with a time-division protocol called pipelined antenna pre-positioning (PAP). While one waveguide transmits, idle waveguides move their antennas into place for their next users, hiding the slow mechanical repositioning. The authors derive closed-form expressions for scheduling efficiency, the waveguide count at which latency no longer costs transmission time, and the latency bound under which PAP beats fixed placement. Results show this latency bound grows with each added waveguide, with users at the outer waveguide ends benefiting most.
In briefWave Processing Unit: A Programmable Light-Based Circuit for Fast Matrix Math in Wireless Communication
SummaryThis paper proposes the Wave Processing Unit (WPU), a programmable network of Mach-Zehnder interferometers that performs matrix-vector multiplications for MIMO communications. It presents a closed-form method to decompose unitary matrices into a rectangular arrangement of MZI elements, extended to arbitrary complex operators via SVD. For 64-antenna, 8-user downlink precoding, the authors report 87.5% fewer RF chains than a fully digital architecture with the same sum rate under ideal hardware, and about 380 times faster configuration than iterative baselines.
In briefMatrix Φ-entropic dependence measures: extending scalar Φ-ribbons to matrices and characterizing SDPI constants
SummaryThe paper introduces matrix Φ-ribbons, extending scalar Φ-ribbons to the matrix setting. For Φ in the subadditivity class, it establishes data processing and tensorization properties and proves scalar and matrix Φ-entropy inequalities. These are used to show both kinds of Φ-ribbons characterize the corresponding Φ-strong data processing inequality (SDPI) constants, with exact values obtained for several sources. Examples show the matrix SDPI constant can exceed the scalar one, and inner bounds derived from prescribed independence structures are proven optimal when Φ(t)=t log t.
In briefRISK-KV: Quality-Aware Selective Rebuilding of KV Cache for Distributed LLM Inference
SummaryThis paper addresses the cost of keeping or replaying historical key-value caches when migrating or recovering requests in distributed large language model inference. It proposes RISK-KV, which uses cross-layer ridge regression to predict missing layers' attention inputs and measures reconstruction risk by the change in negative log-likelihood. On three language models, representative configurations omit 18.2% to 33.3% of historical KV with perplexity changes below 0.5%, and a local recovery experiment with a 512-token history showed a 1.51 to 3.31 times speedup over full Token Replay.
In briefBroadcast beam shaping for centimeter-wave gigantic MIMO systems
SummaryThis paper proposes a beam shaping technique for designing environment-aware broadcast beams in centimeter-wave gigantic MIMO systems, using a flat "prostrate" beam in the vertical plane to spread power more evenly across the ground. The authors develop a geometric model linking base station height, beam spread, maximum service distance, and downtilt angle, and formulate ground coverage maximization as a non-convex optimization problem. They solve it with convexification and successive convex approximation, and provide complexity and convergence analyses. Simulations show the intended beam pattern is achieved and the prostrate beam gives substantially extended coverage distance over conventional wide-beam benchmarks.
In briefReinforcement learning scheduling for hard and soft real-time DNN inference on heterogeneous accelerators
SummaryThe paper presents a scheduling framework for heterogeneous multi-accelerator systems running both hard and soft real-time DNN inference workloads. It pairs an online reinforcement learning scheduler, which selects sub-job priorities and accelerator assignments, with a deterministic runtime safety layer that guarantees hard real-time deadlines are met. Experiments on heterogeneous DNN accelerator platforms report that the framework preserves hard real-time schedulability while improving QoS-aware deadline satisfaction, client-level service quality, and accelerator utilization over state-of-the-art approaches.
Francesco Giulio Blanco, Enrico Russo, Giuseppe Ascia et al.
In briefReorienting the Vector Axis: Restructuring FlashAttention for Long-Vector Processors
SummaryThe paper ports PyTorch's CPU FlashAttention kernel to a RISC-V CPU with 2KB vector registers and finds it reaches only 16.2% of peak floating-point throughput. The authors transpose the score tile and running output so queries lie along the contiguous axis, and move the softmax polynomial coefficients into scalar registers to remove spilling and cut vector memory instructions by 78.7%. On silicon, the two changes together cut whole-kernel cycles by 1.84 to 2.08x at head dimension 64, raising utilization to 29.7% of peak.
In briefWireless point cloud transmission with per-patch rate adaptation
SummaryThis paper proposes a joint source-channel coding framework for efficient wireless point cloud transmission. The point cloud is split into local patches, encoded into compact latent representations, and mapped to symbols sent over a noisy channel. An entropy model uses spatial dependencies between patches to estimate coding costs and allocate a different number of symbols to each patch. Experiments on ModelNet40 report overall rate-distortion gains over existing baselines across channel conditions.
In briefPinching-antenna-based NOMA for covert communication networks
SummaryThis paper applies pinching antenna systems (PASS) to NOMA covert communication, where activated pinching antennas send superimposed signals to two randomly located users while two passive wardens monitor the downlink. The authors derive closed-form detection error probabilities for the wardens under bounded noise uncertainty and determine optimal detection thresholds. Covert rates for both users are characterized under ideal and non-ideal successive interference cancellation. Numerical results indicate PASS-NOMA achieves higher covert rates than conventional OMA schemes.
In briefDeepest-error points for full-length Reed–Solomon codes
SummaryThis paper resolves the Cheng–Murray deep-hole problem for full-length Reed–Solomon codes over finite fields in every dimension from 2 to q−2, with explicit statements for characteristic two. For odd q, deep holes are exactly the words of degree k. For even q with k≥3 the same conclusion holds, except for an extra family when k=q−3. The paper also treats normal rational curve rigidity and projective Reed–Solomon covering radii.
WhySettles the Cheng–Murray deep-hole problem for full-length Reed–Solomon codes with complete proofs, impacting coding theory.
In briefPrecise Coexistence: Waveform Design for MIMO Sensing and Communication That Protects Legacy Systems
SummaryThis paper proposes a dual-function waveform design framework for MIMO integrated sensing and communication systems in congested electromagnetic environments. It suppresses beamformed energy only within the angle-frequency regions used by legacy systems such as positioning sensors, maritime radar, and satellite links, while preserving spatial degrees of freedom elsewhere. The design is formulated as a nonconvex optimization problem and solved with a low-complexity MM-ADMM algorithm. Simulations show precise legacy protection alongside robust radar sensing and multi-user communication performance.
Syed Ali Haider Shirazi, Taejoon Kim, David J. Love et al.
In briefExact GSP autobidding equilibria when advertisers or slots are few
SummaryThe paper studies autobidding equilibria under budget and ROI constraints in sponsored-shopping auctions with generalized second-price pricing. It proves that exact equilibria can be computed in polynomial time when either the number of advertisers or the number of display slots is fixed. The algorithms reduce the problem to polynomially many polynomial systems in bounded dimension.
WhyGives polynomial-time algorithms for GSP autobidding equilibria with fixed advertisers or slots via novel structural reductions and complete proofs, significant for auction theory.
In briefSupport periods and support-primitive decomposition of constacyclic codes over Galois rings (a paper on polynomial support periods and code decomposition)
SummaryThe paper studies support periods of polynomials and constacyclic codes over Galois rings GR(p^e,r). It relates coefficient compression of monic divisors of X^N-λ to support-primitive decomposition of constacyclic codes. For simple-root lengths it gives formulas for support period, intrinsic period, distance and duality, and an arithmetic Singleton bound. It also treats repeated-root cases and Hensel-type lifting of compressed factors.
In briefProof of Ward's conjecture on Griesmer codes
SummaryThis paper proves a conjecture about Griesmer codes, a class of linear error-correcting codes that meet the Griesmer bound. Writing q=p^f with p prime and e for the p-adic valuation of d, Ward conjectured that such codes are p^{max{0,e-f+1}}-divisible. The paper establishes this divisibility result.
In briefPareto inefficiency of EFX fair allocation
SummaryThis paper studies allocating indivisible goods to agents with additive valuations while seeking envy-freeness up to any good (EFX) and Pareto optimality. Earlier work showed these may conflict when agents can value some goods at zero, but whether this holds when all values are positive was open. The authors present a simple instance where every good has positive value to every agent, yet no Pareto optimal EFX allocation exists. They also generalize to a family of instances showing Pareto optimal α-EFX allocations may fail to exist for any α greater than 2/3.
WhyResolves an open question on EFX and Pareto optimality with positive values, giving a construction generalizing to any alpha above 2/3.
Sepideh Mahabadi, Prasanna Ramakrishnan, Jakub Tarnawski
In briefRepeated best replies in random two-player games form cycles whose structure depends on the game type.
SummaryThis paper studies best-response dynamics in two-player bimatrix games with random payoffs, where players alternately respond optimally to each other. It examines symmetric, antisymmetric, common-interest, zero-sum, and independent-payoff games. The authors analyze the number of cycles, cycles of given length, and longest cycle length in large random games, and show that game structure strongly affects typical cycle structure.
In briefBandwidth–subpacketization tradeoff for MDS array codes in linear exact repair
SummaryThis paper studies the repair bandwidth of MDS array codes in distributed storage when the sub-packetization level is fixed. It considers centralized linear exact repair of h failures from d helper nodes over a finite field. The authors show that the rank profile of any linear repair scheme is constrained by two mixed-alphabet codes, yielding explicit repair-bandwidth lower bounds via classical coding-theoretic arguments.
WhyReduces linear repair rank profiles to mixed-alphabet codes, deriving unified repair-bandwidth lower bounds with rigorous coding-theoretic arguments.
In briefEncoding Trees: A Variable-Length Prefix Code for Arbitrary-Size Symbol Sets
SummaryThis paper defines a new class of binary trees called encoding trees, extended through labelling and pruning into pruned labelled encoding trees. Using these trees, the authors build one code that converts symbol strings drawn from a sequence of non-empty sets into binary strings, and a reverse code that goes from binary strings back to symbols. Both codes are proven to be variable-length prefix codes. The paper also gives an equation for the minimum sequence length needed to encode a binary string of a given length.
In briefCorrelated Quantal Response Equilibrium: A New Equilibrium Notion for Predicting Human-AI Strategic Play
SummaryThe paper proposes Correlated Quantal Response Equilibrium (CQRE), a solution concept that combines and generalizes Quantal Response Equilibrium and Correlated Equilibrium to model games mixing AI agents and human players. It establishes existence results and presents a method to compute the CQRE that best fits a given joint action distribution. The authors report empirical effectiveness across diverse complex general-sum games and a large-scale extension of the Selten–Chmura setting covering AI–AI and human–AI play.
In briefOptimal Rényi and log-additive entropy estimation sample complexity
SummaryThe paper determines the minimax sample complexity of estimating Rényi entropy from independent samples on an alphabet of size at most k, where the known order α>1 may vary with k. For fixed additive accuracy and success probability, it proves matching upper and lower bounds with constants independent of k and α. The analysis is extended to logarithms of positive additive functionals, with corresponding sample complexity rates.
WhyGives matching minimax sample-complexity bounds for Rényi entropy and log-functional estimation with k- and α-independent constants, a strong theoretical advance mainly affecting information theory an
In briefUniversality of approximate message passing under a fourth-moment condition
SummaryThe paper proves state evolution for approximate message passing with rectangular random matrices whose independent, centered, variance-normalized entries need only have uniformly integrable fourth powers. Empirical averages of pseudo-Lipschitz functions converge to Gaussian state evolution limits at every fixed iteration. The authors also show that the required independence property characterizes Gaussian entries, and that average variance normalization alone does not guarantee homogeneous limits.
WhyProves AMP state evolution for non-Gaussian matrices under a fourth-moment condition, plus a Gaussianity characterization and a variance counterexample, a clear theoretical advance.
In briefKoala: A Non-Disruptive Reconfiguration Protocol for Long-Running Stateful Dataflows
SummaryStateful dataflow systems must decide where state is stored relative to processing tasks, trading adaptability against performance. Koala is a reconfiguration protocol for distributed dataflow systems that migrates state lazily and at fine granularity, using a lightweight ownership transfer mechanism to separate event routing from state location. Compared with the state-of-the-art protocol DRRS, Koala drains backlog up to 23 times faster and transfers 8.7 times less state on average during reconfiguration.
In briefDeterministic radio gossip: new round bounds via pruning and selectors
SummaryThe paper studies deterministic gossiping in unknown strongly connected directed radio networks without collision detection. It proposes a neighborhood-clearing procedure and an adaptive pruning procedure, with path and permutation selectors reducing completion cost. The authors give round bounds such as O(m log²n log log n) under a degree condition and Õ(min{n^{4/3}, N+(mn)^{3/5}}) for every graph. The unrestricted polynomial-identifier density extension is stated as open.
WhyIntroduces new pruning and path-selector techniques for deterministic radio gossiping with several structure-dependent round bounds, a clear theoretical improvement, though limited to a specific netwo
In briefA Matroidal Framework for Random Access in Linear Codes: a paper on how evenly random sampling recovers each coordinate of linear codes.
SummaryThis paper studies random access in linear codes when codeword symbols are sampled independently and uniformly with replacement. A code is called recovery balanced if every coordinate needs the same expected number of samples to be recovered. The authors express these expectations through the associated matroid and derive a coordinatewise identity linking a code and its dual. This shows a linear code is recovery balanced if and only if its dual is, confirming a conjecture of Gruica et al. The paper also defines recovery balance for arbitrary matroids and shows that asymptotically almost every sparse paving matroid is not recovery balanced.
Daniella Bar-Lev, Hiram H. L\'opez, Gretchen L. Matthews et al.
In briefLarge-scale SAT sampling using GPU-based Boolean constraint propagation
SummaryThis paper presents Gpap-S, the first GPU sampler built on Boolean constraint propagation (BCP), for sampling solutions of Boolean formulas used in hardware constrained-random verification. Its core engine, Gpap, propagates many independent partial assignments concurrently on a shared formula. The authors report it is faster than prior GPU and CPU samplers across benchmarks, including rare-target cases, and passes the Barbarik statistical test on every instance tested.
WhyFirst GPU sampler based on Boolean constraint propagation, with about two-orders-of-magnitude speedups at rare targets against strong baselines; impact mainly limited to constrained-random verificatio
In briefPlacing weights alone is not enough: synchronization limits multi-node CPU decoding of ternary LLMs
SummaryThe paper finds that ternary LLMs read only a few gigabytes per decoded token, yet on a dual-socket, four-NUMA-node server the fastest ternary runtime it tested uses less than half the memory bandwidth. Even with each weight placed on the node of the thread that reads it, threads spend much of each token waiting at synchronization points, mostly for the slowest threads on each node. On identical kernels, a Megatron-style column split beat a row-owned split in every configuration measured, with most of the gain from reduced waiting between threads. Where decoding is DRAM-bound, the resulting engine is 1.5-2.1x faster than ik_llama.
In briefBirthday estimator proven uniformly optimal for estimating Rényi entropy
SummaryThis paper studies the natural birthday estimator, which counts collisions to estimate Rényi entropy, a measure of uncertainty. The authors prove a Bernstein-type concentration bound for it and show it is minimax optimal across alphabet size, order, accuracy, and confidence, without sample splitting, clipping, or median aggregation. This settles the sample complexity of integer-order Rényi entropy estimation up to absolute constants, closing the gap between known upper and lower bounds.
WhyProves the birthday estimator is minimax optimal for integer-order Rényi entropy with Bernstein concentration, closing the known upper-lower bound gap.
In briefBeamfocusing vs. Airy Beams: Is Bending Beams Needed to Avoid Antenna Blockage?
SummaryThe paper examines whether Airy beams, often proposed to bend array beams around obstacles, are necessary for antenna blockage. The authors prove that conventional beamfocusing maximizes received power for a single-antenna receiver regardless of how many antennas are blocked, and extend this to multi-antenna receivers under a phase-separability condition. Numerical results show unmodified beamfocusing is optimal for physically small receivers, while larger receivers deviate from it to illuminate unobstructed parts, without becoming Airy beams.
In briefExact envy-free cake cutting with polynomial expected queries
SummaryThe paper studies how to divide a cake-like divisible resource among several agents so that no agent prefers another's share. The authors propose a randomised protocol that uses only independent, uniformly sampled points from the cake as randomness. For any n agents with nonnegative, additive, nonatomic valuations, the protocol terminates with probability one, returns complete and exactly envy-free allocations, and has expected Robertson-Webb query complexity Õ(n^30), polynomial in n.
WhyA randomized protocol achieves polynomial expected queries for exact envy-free cake cutting, breaking the longstanding exponential barrier; proofs are theoretical.
In briefBudgeted admission control in shared clusters using delayed capacity records
SummaryThis paper studies a cluster admission gate that may receive capacity records no longer matching current schedulability, under a bound on the joint probability of admission and unavailability. It derives a stationary admission policy for a fixed observation delay, applies finite likelihood allocation to delay-tagged records, and adds a deterministic guard limiting unavailable-state admissions. The authors report exact binomial calibration and an independent audit of the controlled system.
