In briefRegularizing the Veto Core: adding a small weight to each candidate so voting rules stay Pareto-efficient
SummaryThe paper studies voting rules that pick a candidate close to voters using only their rankings, a setting called metric distortion. It adds a small regularization term to each candidate's weight, which restores Pareto-efficiency to the SimultaneousPluralityVeto rule. The authors also analyze a rank score veto core and bound its distortion. The results address an open question posed by Kızılkaya and Kempe.
In briefA study of the true minimum distances of Goppa codes
SummaryThis paper studies the true minimum distances of binary Goppa codes. It gives necessary and sufficient conditions for these codes to have minimum distance exactly 2r+1 or 2r+2, and constructs several infinite families. It also extends a criterion for q-ary Goppa codes to split auxiliary polynomials, obtaining infinite families that reach their designed distance r+1.
In briefExhaustive Peres Transform: Extraction and Universal Coding
SummaryThe paper studies Peres's iteration of von Neumann's bit-debiasing procedure, run to exhaustion, which bijectively splits each binary string into fair bits and a residual. It proves a deterministic sparsity bound on the residual, yielding a short proof that fair bits per symbol approach the entropy. Two coding schemes are presented, one needing no parameters or arithmetic coder, and the authors show it is universal for stationary ergodic sources.
WhyTurns Peres iteration into a bijection with a deterministic sparsity bound, yielding a parameter-free universal code with rigorous proofs of entropy and conditional-entropy universality.
In briefChoosing encoding basis functions for analog coded computing under differential privacy, preserving exact recovery while balancing privacy and accuracy.
SummaryThis paper studies how to choose encoding basis functions in analog coded computing. The authors note that these functions affect not only reconstruction and numerical stability but also the statistics visible to colluding workers, and thus data privacy. They relax the full orthogonality constraints of the Lagrange basis to obtain a parameterized family of polynomial bases that still allow exact recovery, and derive a basis-dependent differential privacy characterization. Numerical optimization then selects data and noise basis functions to meet a given (ε,δ)-DP requirement. Results show that, at the same privacy level, the optimized bases substantially reduce floating-point reconstruction error compared with conventional ALCC.
In briefOptimizing Waveguide Layout for Pinching-Antenna Communication Systems
SummaryThis paper studies how to optimally deploy waveguides for pinching-antenna systems over a rectangular service region, rather than assuming parallel waveguides. For the two-waveguide case, it derives a closed-form sufficient condition, based on waveguide height and region aspect ratio, under which a deployment parallel to the longer side is globally optimal. For multiple waveguides, it jointly optimizes orientation and offset using an adaptive exchange algorithm. Numerical results show that nonparallel layouts offer clear gains at intermediate heights, while parallel or center-concentrated layouts are preferable in other regimes.
In briefPangolin: TEE-accelerated DAG-based BFT state-machine replication protocol
SummaryThis paper proposes Pangolin, a Byzantine fault-tolerant protocol that uses directed acyclic graphs to propose request blocks in parallel and builds a full state-machine replication framework on top. It uses trusted execution environments so that it tolerates f Byzantine faults with only 2f+1 replicas, and clients verify replies via compact execution proofs. In the authors' evaluation, Pangolin achieves an average 1.27x latency speedup over Mysticeti, with up to 1.52x.
Paul Franke-Bergmann, Maya Kuhlmann, Alexander He{\ss} et al.
In briefStreaming In-Situ HPC Visualization into Unreal Engine 5 via ParaView Catalyst and ANARI
SummaryThis paper presents a disk-free in-situ pipeline that streams HPC simulation data into Unreal Engine 5 through ParaView's ANARI rendering API. A modified ANARI-USD device captures USD scene graphs in memory and sends them via an embedded ZeroMQ broker to clients, where the JuSync middleware and its Unreal plugin parse geometry as it arrives. The design decouples simulation from visualization, so clients can connect or disconnect at any time. The pipeline is demonstrated on the JURECA-DC supercomputer at Forschungszentrum Jülich using Free Electron Laser simulation data with point clouds of tens of millions of particles.
Thomas George, Severin A. T. Klapproth, Jens Henrik G\"obbert et al.
In briefCorrections to the Method of Gaps paper: a counterexample and a generalized formulation
SummaryThis paper gives a counterexample to Lemma 5 of the original paper on the method of gaps for supervised learning generalization error. It shows the data-driven gap method needs an extra assumption about identical data-point marginals. The authors present a generalized formulation that avoids Lemma 5 and does not require independent, identically distributed training samples, extending the method to broader distributions and randomized algorithms.
In briefA Hyperplane Bound and Ward's Divisibility Conjecture
SummaryThe paper proves that every q-ary linear code (q=p^f) that meets the Griesmer bound, has positive dimension and minimum distance d≥1, is p^{max{0,ν_p(d)−f+1}}-divisible, confirming Ward's conjecture. Its main geometric tool is a dimension-independent hyperplane bound on how many linear hyperplanes a nonzero scalar-invariant function can vanish on. The author also distinguishes this argument from an earlier unpublished proof.
WhyProves Ward's divisibility conjecture for Griesmer-bound linear codes via a dimension-free hyperplane bound and inductive arguments, with a complete proof.
In briefGeoTrain: Energy-aware automatic parallelism for training AI models on mixed, geographically spread GPUs
SummaryThis paper presents GeoTrain, a system for training foundation models across geographically distributed GPUs with differing compute, memory, power, and network capabilities. It jointly selects GPUs, partitions model layers, and sets pipeline configurations while optimizing both training time and energy use. Experiments with decoder-only Transformer models report that GeoTrain reduces iteration time and energy consumption compared with existing state-of-the-art approaches.
In briefCEMA-Router: Infrastructure-Aware Batch Routing for LLM Inference Across Cloud and Edge
SummaryThis paper presents CEMA-Router, a system that routes batches of large language model inference requests across cloud and edge resources. It jointly balances response quality, monetary cost, energy use, and batch completion time using a memetic search algorithm informed by node profiles, backlog estimates, and model-quality signals. The authors implement it as an end-to-end system and evaluate it on RouterBench traces with GSM8K and MMLU workloads on a heterogeneous testbed. Reported results show reductions in the weighted objective of up to 86% and 80%, energy of up to 33% and 19%, and completion time of up to 27% and 30%.
In briefBuoyancy-Driven Holistic Resource Management for Cloud-Native Workloads
SummaryIn multi-tenant clouds, workloads compete for shared resources such as last-level cache and memory bandwidth, and congestion shifts as workload mixes change. The paper uses workload buoyancy, a runtime framework that captures each workload's sensitivity to individual resources and application performance, to guide scheduling, node-local allocation, and rescheduling across the workload lifecycle. Implemented as a Kubernetes-native prototype and tested on a multi-node cluster, it raised the share of workloads meeting SLOs from 7.74% to 87.8% on average in a high-congestion scenario, while reducing congestion and fragmentation with limited rescheduling.
Oliver Larsson (1), Thijs Metsch (2), Cristian Klein (1) et al.
In briefFault-tolerant task offloading for real-time apps at the mobile edge
SummaryThis paper studies deadline-constrained task offloading and resource allocation in mobile edge computing systems with backhaul networks, aiming to maximize total inference accuracy improvement. The problem is formulated as an integer linear program and shown to be NP-hard. The authors propose FastSA, a linear-time approximation algorithm, and a fault detection and recovery framework for backhaul link and edge server failures. Experiments use image classification workloads and real-world taxi traces.
In briefNew maximum Hermitian codes and invariants from skew polynomial rings
SummaryThis paper studies maximum rank-distance codes in the space of Hermitian matrices. The authors construct additive maximum Hermitian rank-metric codes for every admissible length and minimum distance, settling a final open existence case. The construction uses central simple algebras with unitary involution obtained as quotients of skew polynomial rings. New invariants are introduced to test equivalence of additive Hermitian codes, and computing them shows that the constructed codes include many pairwise inequivalent ones.
WhySettles the last open existence case for Hermitian d-codes using skew polynomial constructions and new invariants, with proofs; impact stays largely within coding theory.
In briefFair, Sustainable Allocation of Renewable Resources Over Multiple Periods
SummaryThis paper studies how to fairly allocate renewable resources to agents over multiple periods, where resource quantities evolve over time according to renewal functions. Allocations must satisfy viability and capacity constraints and remain envy-free over the whole time horizon. The authors analyze the computational and parameterized complexity of finding such allocations across the numbers of agents, resources, and periods. They show the problem is NP-hard under unary encoding even when any one of these dimensions is fixed to a constant, with other cases depending on the dimension pair, renewal function class, and numerical encoding.
In briefAdaptive Covering Complexity: Linking Dictionary-Based and Self-Referential String Compression
SummaryThe paper introduces adaptive covering complexity, the minimum number of contiguous blocks needed to partition a string, where each block may come from a fixed external dictionary or the already decoded prefix. It proves that dynamic self-reference can reduce the phrase count by at most an exponential factor compared with static dictionary covering. The authors establish equivalences with relative Lempel-Ziv and non-overlapping LZ77, show the longest-match greedy parse is optimal, and give an O(M + n) implementation using dual suffix automata. Dictionary selection is shown to be NP-hard.
In briefCompressing Markov chain data with a stationary-distribution shortcut
SummaryThis paper studies the information bottleneck problem for data from discrete-time ergodic finite-state Markov chains. It proposes treating observations from the time-varying chain as draws from its unique stationary distribution, and compressing them using the stationary solution. The resulting error is upper bounded by several information gap terms, and the authors give sufficient conditions for these gaps to vanish and analyze their convergence rates.
In briefSingle-level reformulations of Wardrop-equilibrium network tolling with separable affine costs
SummaryThis paper studies how to set tolls on some arcs of a congested network to maximize a network authority's revenue, assuming users route according to a Wardrop equilibrium. For separable affine arc costs, it uses Beckmann's convex characterization to derive three single-level reformulations and an upper bound on optimal revenue for certifying optimality gaps. Computational tests on Sioux Falls and Eastern Massachusetts instances show the arc-based formulation solves more cases to proven optimality within 1800 s than the existing benchmark, while path-based formulations give near-optimal heuristic solutions.
In briefEvent-driven architecture for a 15-minute urban flood digital twin: Hanoi case study
SummaryThis paper presents a prototype automated urban flood digital twin that updates flood depth information on a 15-minute cycle under limited data. It chains rainfall-triggered 3Di hydrodynamic simulations in decoupled, event-driven workflows and publishes road-edge depths through spatial services for routing. In a 57-run storm replay in Hanoi driven by satellite rainfall estimates, every cycle completed within the 15-minute deadline. The authors state the results establish system feasibility, not forecast accuracy or route safety, since simulated depths were not validated against field gauges.
In briefA matroid-constrained version of the k-submodular cover problem
SummaryThis paper studies the k-submodular cover problem under a matroid constraint on an n-element ground set. It seeks k disjoint subsets with minimum total cost whose union is independent in a given matroid and whose k-submodular utility exceeds a threshold. The authors propose a matroid extended greedy algorithm, analyze its bi-criteria approximation ratio, and derive an improved ratio for unit costs.
In briefBest priors for betting-based tests with bounded payoffs: an exact expected-time result
SummaryThis paper studies hypothesis testing by betting, where initial wealth is spread over fixed bets according to a prior distribution. Under technical conditions, the authors prove a sharp formula for the expected rejection time that holds uniformly over all bounded alternatives as the error probability goes to zero. They characterize the priors achieving the best leading constant and show that the uniform prior can require arbitrarily more observations than suitable priors.
WhyProves a sharp uniform expected-rejection-time formula and characterizes optimal priors, substantially refining analysis of adaptation cost in betting tests.
In briefFinite BSC decomposition of binary-input symmetric-output channels, with applications to channel simulation
SummaryThe paper proposes a finite BSC-decomposition that represents any binary-input symmetric-output channel as a mixture of binary symmetric channels with at most ℓ+1 distinct crossover probabilities. It bounds how much the weighted average capacity of the component BSCs exceeds the original channel's capacity, and shows the leading term is asymptotically optimal in the worst case. Building on this, the authors design a channel simulation scheme that, combined with efficient BSC simulation algorithms, achieves redundancy O((log n / n)^{2/3}) over capacity when simulating n copies.
In briefFrobenius profiles of MRD codes built from polynomial multiples
SummaryThe paper studies Frobenius intersection profiles of maximum rank distance (MRD) codes built from polynomial multiples over finite fields. It shows that every numerically admissible profile can be realized in this family over all sufficiently large extensions, using a root-exponent pattern independent of the extension degree. A lower bound valid for all polynomial generators is also established, and along an infinite parameter family the upper and lower bounds differ by only O_q(log n).
In briefPMR-aided shifted-pruning SCL-flip decoder for CRC-aided polar codes
SummaryThe paper proposes a posterior-mass rescue (PMR)-aided shifted-pruning SCL-flip decoder for CRC-aided polar codes. When CA-SCL and differential-metric shifted-pruning recovery fail to find a CRC-valid candidate, PMR continues discarded prefixes from the initial pass, ranked by estimated posterior mass, without extra shifted-pruning retries. Simulations on (128, 64) and (256, 128) codes show lower observed FER than SCLF-Diff (T = 15) and literature baselines, close to SCLF-Diff (T = 20), though the worst-case path-step bound is larger.
Niloufar Hosseinzadeh, Mohammad Morsali, Mohsen Moradi et al.
In briefHardness amplification for approximate Nash equilibrium with a constant number of players
SummaryThe paper shows that, for fixed tolerances ε and η, computing an (ε,η) weak approximate Nash equilibrium requires quasipolynomial time under deterministic PPAD-ETH, even in symmetric binary-payoff games with a constant number of players. Its main tool is an error-amplification reduction from bimatrix games to symmetric binary-payoff games. Combined with Rubinstein's bimatrix hardness result, it yields a deterministic worst-case lower bound of N^{(log N)^{1-o(1)}}.
WhyAn error-amplification reduction extends bimatrix hardness to constant-player symmetric games, yielding a quasipolynomial lower bound with rigorous proof.
In briefHeuristic optimization of OpenCL channel depths for FPGA decoupled memory channels
SummaryThis paper extends an LLVM-based static analysis framework with a hardware-aware refinement layer that sets OpenCL channel depths for decoupled access-execute architectures on FPGAs. The authors report that grid-based kernels with high baseline memory stalls reach peak performance at minimal channel depths and lose up to 61% throughput when depths are increased. On the HotSpot application, the approach achieves a 2.52x improvement in resource efficiency while maintaining peak speedup.
Arnab A. Purkayastha, Shobhit Aggarwal, Jay Tharwani
In briefFrom Problem Setup to Explanation: AI-Assisted Research in Information Theory
SummaryThis article examines how AI contributes to information-theory research in different ways, such as suggesting proofs, taking part in mathematical dialogue, coordinating computation, or acting as part of a search framework. It analyzes recent cases through a four-stage pipeline: problem formulation, AI's role in finding solutions, verification of resulting claims, and translation of machine outputs into human-readable mathematical exposition. The authors recommend that publications disclose AI's role and extent at each stage, report verification of analytical reasoning separately from numerical computation certification, and emphasize articulating the underlying ideas, techniques, and interpretations.
