In briefLinking per-workload carbon metrics to provider top-down cloud reports
SummaryThis paper argues that bottom-up cloud carbon metrics, such as Software Carbon Intensity (SCI), and top-down provider reports are currently treated as separate and should be reconciled. It proposes reconciled SCI (rSCI), a per-workload metric that anchors an energy estimate to the provider's report through a residual, which is allocated by its physical drivers. The authors contend that present provider reporting is too coarse and methodologically inconsistent for credible reconciliation, and they outline concrete reporting and standards steps needed.
Philipp Wiesner, Lo\"ic Lannelongue, Alexander Acker et al.
In briefCounterexample showing network coding beats multicommodity flow in undirected multiple-unicast networks
SummaryThe paper disproves a conjecture stating that network coding gives no throughput advantage over multicommodity flow in undirected multiple-unicast networks. The authors build a deterministic linear network code over F_9 on a 182-vertex bipartite subgraph of the point-line incidence graph of PG(2,9). It supports 157 independent unicast sessions at common rate at least 1, whereas every fractional multicommodity flow achieves at most 147/157.
WhyConstructs an explicit linear coding counterexample refuting a long-standing conjecture, with a provable coding gap, though scope is specialized.
In briefProgressive MIMO sensing of multiple targets using variational Bayesian inference with nuisance parameters
SummaryThis paper proposes a multi-stage Bayesian MIMO sensing framework for estimating multiple targets' angles, treating reflection coefficients as unknown nuisance parameters. Each stage uses the previous stage's posterior as its prior and designs transmit beamforming to minimize the sum posterior Cramér-Rao bound. A variational Bayesian inference method yields a closed-form surrogate posterior with polynomial complexity in the number of targets, compared with exponential complexity for existing numerical approaches. Numerical results show the framework progressively improves sensing performance.
In briefSpectral efficiency of massive MIMO-OFDM with double quantization
SummaryThis paper studies uplink spectral efficiency when both ADC quantization and fronthaul quantization are applied in a massive MIMO-OFDM system. The authors derive a linear MMSE channel estimator from double-quantized pilot observations and compute an achievable spectral efficiency using the use-and-then-forget bound. Numerical results show comparable impacts from the two resolutions, with the lower resolution limiting performance when they differ, and six-bit quantization reaching over 90% of the ideal spectral efficiency.
Bengu Bilgic Keskin, A. Oguz Kislal, Ali Gorcin et al.
In briefProbabilistic GC-content constraints for composite DNA storage
SummaryThis paper studies how to encode biochemical GC-content constraints in composite DNA-based data storage. Earlier deterministic models lost much rate by forbidding any strand from leaving the allowed range. The authors introduce an ε-probabilistic constraint framework to account for the stochastic nature of composite DNA, and derive capacity bounds for global constraints using composition types and for local sliding-window constraints using finite-state Markov chains. They also propose a capacity-achieving multi-type enumerative encoder.
Humeyra Bodur, Frederik Walter, Antonia Wachter-Zeh
In briefChannel estimation for AFDM with fractional delay and Doppler, adapting to unknown path count
SummaryThis paper studies channel estimation for affine frequency division multiplexing (AFDM) systems, where channels have fractional delay and Doppler and an unknown number of propagation paths. The authors propose a model-order-adaptive estimator within an improved SAGE framework, using residual tests to add paths and conditional pruning to remove unsupported ones. Simulations report lower NMSE and BER than SAGE, sparse-recovery, and Bayesian benchmarks, with robust performance at terminal speeds up to 600 km/h.
In briefHPC-MQBench: A Slurm benchmark that checks message delivery quality before ranking Kafka speed
SummaryThe paper presents HPC-MQBench, a benchmark that runs messaging experiments on HPC clusters using Slurm, and checks record correctness and delivery before ranking configurations by rate. In a Kafka test with one broker on four nodes, 99 of 120 workload configurations gave qualified observations. The selected workload qualified in all five final-stage observations, while the contribution is an auditable configuration-selection framework; multi-broker scaling is left for future work.
In briefPerformance-portable SU(N) lattice gauge simulation using Kokkos
SummaryThis paper presents kwqft, a Kokkos-based implementation of Wilson pure gauge Monte Carlo simulation for SU(N) Yang-Mills theory in arbitrary space-time dimensions. One source code runs on Serial, OpenMP, CUDA, HIP, and SYCL backends, with MPI halo exchange overlapped with interior updates. The authors report results matching known values for gauge groups up to SU(17), performance on an NVIDIA A100 competitive with native CUDA, and a large-scale speedup for an SU(4) lattice on the LineShine supercomputer.
In briefAeroEval: A Staged Validation Framework for AI-Generated Drone Missions
SummaryThis paper presents AeroEval, an agent-assisted middleware that validates drone programs generated by large language models from natural-language mission descriptions. It first checks program syntax, platform API usage, and mission intent, then evaluates realized behavior using execution trajectories, mission requirements, and environmental context. Structured failure information from each stage guides iterative regeneration. In experiments on AirSim and Gazebo, navigation success rose from 55% to 95%.
In briefRendezvous Under Variable Disorientation: Algorithmic Power of Fixed Unit Distance
SummaryThis paper studies deterministic rendezvous of two light-equipped robots under weak geometric and synchronization assumptions. Under variable disorientation, each robot's local coordinate system may change arbitrarily at every observation, and the authors add a fixed unit distance condition to study its effect. They prove impossibility results, establish color-count boundaries, and introduce periodic distance classes that turn distance changes into predictable cyclic phase shifts, yielding algorithms with specified color counts under various schedulers. The results show that a fixed local metric can serve as persistent cyclic memory and clarify how metric stability, visibility, and scheduler atomicity determine solvability and color complexity.
WhyPeriodic distance classes turn metric changes into cyclic phases, halving the full-ASYNCH color bound from twelve to six with proofs across schedulers.
In briefCommunication-aware qubit placement for faster distributed quantum simulation
SummaryThis paper addresses how poor qubit placement across nodes in distributed quantum circuit simulation causes excessive communication and limits performance. It proposes a framework that optimizes qubit layout to minimize inter-node data exchange while dynamically selecting the number of nodes, implemented in the QuCSLO library with OpenQASM parsing and QuEST integration. Tests on synthetic and QASMBench circuits show reduced communication cuts and runtime speedups of over 2x versus identity layouts, especially for highly entangled circuits.
\'I\~nigo Ar\'ejula-A\'isa, Sergio Iserte, Petter Sand{\aa}s et al.
In briefA faster, memory-saving training framework that splits work between GPU and CPU
SummaryTraining large deep learning models can run out of GPU memory, and the existing ZeRO-Offload method moves optimizer states to the CPU but has low GPU utilization and poor overlap of communication and computation. This paper presents MemFerry, which uses Direct Host Access to let the GPU compute on data in CPU memory, with a scheduler and a shadow model that manage parameters split across both memories. It also adds a dynamic programming method to offload gradients, and extends the design to scale-up systems. The authors report up to 1.68x faster training and 1.52x larger trainable models than ZeRO-Offload on one GPU, plus gains on multi-GPU and Huawei CloudMatrix384 nodes.
In briefNear-optimal constant-round approximation for dominating sets in minor-excluded graph classes
SummaryThe paper gives a deterministic LOCAL algorithm for fixed proper minor-closed graph classes that returns a dominating set within a factor of about 2a(C)+1 of the fractional domination number. For planar graphs this yields a guarantee of 7+ε, which together with a known lower bound determines the optimal deterministic constant-round ratio for planar dominating set as 7. It also shows that for r-th powers of paths no deterministic constant-round LOCAL algorithm can beat ratio 2r+1.
WhySettles an open question on deterministic constant-round dominating set approximation for planar graphs, matching upper and lower bounds with complete proofs.
In briefEnvy-free division of indivisible goods using individual payments between agents
SummaryWhen no envy-free allocation of indivisible goods exists, monetary transfers can restore envy-freeness. This paper allows only transfers between agents, with each agent's payments limited by an individual budget, rather than external subsidies. Checking whether a given allocation can be made envy-free under these budgets is polynomial-time solvable, but computing such an allocation from scratch is NP-hard. The authors give parameterized complexity results and efficient algorithms for several special cases.
Robert Bredereck, Eva Deltl, Tanmay Inamdar et al.
In briefHellinger conjecture for antipodal Boolean functions
SummaryThis paper studies how much two output distributions overlap when a Boolean decision is made under binary symmetric noise. It proves that, across the whole noise interval, a single input coordinate minimizes this overlap among antipodal decisions, whose values flip when all input signs are reversed. The proof uses a normalized nuclear norm of a cross-class Gram matrix block of square-root channel densities, and the bound is extended to arbitrary independent binary-input channels.
WhyProves a single coordinate minimizes overlap for antipodal Boolean decisions, extending the sharp bound to arbitrary independent binary-input channels.
In briefA distributed accelerated ADMM method for quadratically regularized optimal transport on bipartite graphs
SummaryThis paper addresses optimal transport problems with graph-structured connections and unbalanced mass, which classical models handle poorly. It introduces DAP-ADMM, a fully distributed algorithm in which each node communicates only with its graph neighbors. The authors prove global convergence and non-ergodic O(1/k) bounds for fixed local penalty parameters, and add a solver-free projection routine and a self-adaptive penalty update that needs no global aggregation. Numerical experiments show it runs faster than standard distributed ADMM.
In briefError-correcting codes protecting function values for symbol-pair Lee-metric channels
SummaryThis paper studies function-correcting codes that protect specified function values of messages against errors while reducing redundancy, aimed at symbol-pair read channels used in high-density storage. The authors generalize the symbol-pair Lee distance from Z_4 to Z_q and target 2^m-ary PSK constellations. They derive Plotkin-type and Gilbert–Varshamov-type bounds on optimal redundancy, along with explicit constructions for several function classes and a Plotkin-type lower bound for linear functions.
In briefSecond-order image bounds with distortion and secrecy constraints
SummaryThe paper studies how coding constraints relate to the probability of source cells and their output images, extending single-output posterior bounds to simultaneous posterior and conditional-composition image bounds. It adds actual-distortion marks and posterior-mass constraints to handle secrecy. Applications include second-order regions for multi-source helper intersections, lossy reconstruction with informed lossless reconstructions, Gaussian common/private descriptions, and privacy amplification, with explicit rate bounds for general cases.
In briefPhysics-aware power optimization for Rydberg quantum arrays using Barankin bound minimization
SummaryThis paper proposes a power optimization framework for Rydberg-atom quantum linear arrays used in terahertz direction-of-arrival estimation. It derives a closed-form stochastic Barankin bound under low-SNR conditions and incorporates an electromagnetically induced transparency readout model with physical effects such as photon shot noise. Simulations show the framework predicts threshold breakdown more accurately than the Cramér-Rao bound and enlarges the reliable operating region under severe THz attenuation.
In briefNetwork coding can outperform routing in undirected multiple-unicast networks
SummaryThe paper disproves a 2004 conjecture by Li and Li that network coding gives no advantage over fractional routing in undirected networks. It constructs a finite undirected network where a binary linear block code achieves a common rate strictly above the maximum fractional routing rate. It also gives an unbounded family of networks where coding approaches rate one while routing is bounded, and the finite counterexample and family theorem are formalized in Lean.
WhyDisproves a 2004 conjecture on undirected network coding by constructing a finite counterexample and an unbounded family, formally verified in Lean.
In briefvLLM-Omni Technical Report: A Unified Serving Runtime for Omni-Modality Generation
SummaryIntelligent systems increasingly generate text, audio, images, video, and robot actions, which existing single-architecture inference stacks handle poorly. This report presents vLLM-Omni, a unified serving runtime that organizes each workload as a multi-stage pipeline under one orchestrator, using specialized engines, stage replicas, a data-plane connector, and session-oriented control. It covers the architecture and OpenAI-compatible and OpenPI APIs for omni, TTS, image/video, world-model, robot, and duplex workloads. The system is evaluated in multimodal nightly CI on H100 and H200 hardware, focusing on Qwen3-Omni.
In briefSDP tightness and error exponents when community count grows logarithmically
SummaryThis paper studies a semidefinite programming relaxation for community recovery when the number of communities grows logarithmically with network size. It gives the sharp asymptotic boundary for tightness away from critical cases. Even when rare vertices cause tightness to fail, the normalized matrix error of every near-optimal solution still vanishes, with matching high-probability exponents. In the spectrally stable region, one SDP solve plus rounding and refinement achieves exact recovery.
WhyGives sharp asymptotic tightness boundaries and matching error exponents for SDP in logarithmic-community stochastic block models, a clear theoretical advance with subfield impact.
In briefDecentralized multi-agent reinforcement learning for sharing spectrum in mission-critical uses
SummaryThis paper proposes a decentralized multi-agent reinforcement learning model that lets secondary users learn how to access shared radio frequency bands with fewer collisions. Each agent uses lightweight linear actor-critic learners suited to resource-constrained edge devices, framed as a Markov potential game. Compared with random and forecast-aware heuristic baselines, the learned policies preserve throughput while greatly reducing collisions, by up to 96.8%.
Dimitrios Pylorof, Imtiaz Nasim, Humberto E. Garcia et al.
In briefFull-duplex massive MIMO network balancing radar sensing, communication, and edge computing
SummaryThis paper studies a full-duplex massive MIMO network where users offload tasks to an edge server while an access point simultaneously probes a target with radar. The authors derive a rate bound for zero-forcing uplink reception and Cramer-Rao lower bounds for target angle estimation. They jointly optimize offloading fractions, transmit powers, and edge computing resources to minimize average latency using a successive convex approximation algorithm. Simulations show that joint allocation reduces radar transmit power by about 86% compared with equal communication-user power at a -40 dB CRLB threshold, and achieves lower latency than fixed-allocation and half-duplex schemes.
Vo Phi Son, Van-Dinh Nguyen, Nhan Thanh Nguyen et al.
In briefA paper on safely lowering numerical precision in scientific computing to save compute while keeping results accurate.
SummaryThe paper argues that hardware is shifting toward low-precision formats built for AI, while discretization errors in PDE-based simulations mask the lower bits of FP64, making them unnecessary. The authors propose a Signal-to-Noise Ratio (SNR) criterion, based on spatial and temporal refinement, to judge the impact of safe precision reduction. They also present an automated, dataflow-centric workflow for lowering precision in mixed-precision subgraphs. Large-scale tests on weather and climate applications report up to 1.8x speedup on GPUs without compromising physical fidelity.
Pratyai Mazumder, Alexandru Calotoiu, Marcin Copik et al.
In briefEfficient Algorithms for Online Subadditive Combinatorial Allocation
SummaryThis paper studies online combinatorial allocation, where indivisible items are assigned to buyers with subadditive valuations. Earlier work by Correa and Cristi showed a 6-competitive online algorithm but only proved its existence. This paper gives an expected polynomial-time (6+ε)-competitive algorithm and recovers Feige's (2+ε)-approximation offline. For identically distributed valuations, it achieves a (60/11+ε)-competitive algorithm. The key technique is a recursive Bundle Score Generator that is efficiently computable.
WhyResolves an open question by giving an efficient polynomial-time 6-competitive algorithm via a new recursive bundle score generator with full proofs.
Calum MacRury, Rian Neogi, Kanstantsin Pashkovich et al.
In briefOptimal encodings and systematicity in function-correcting codes
SummaryThis paper studies function-correcting codes, which protect the value of a function of a message against errors. It shows that non-systematic encodings can need less redundancy than systematic ones. For linear functions and a fixed linear code, the authors give an algorithm that builds an optimal generator matrix. For two-valued functions, they characterize when non-systematic codes exist and show redundancy 1 suffices in several cases.
WhyShows non-systematic encodings reduce redundancy from 2t to 1, with optimal linear generator matrix algorithm and full characterization for two-valued functions; strong theory for a coding subfield.
Charul Rajput, Kanchana Lokshmii Jagatti, B. Sundar Rajan
In briefLarge player classes and approximately weighted voting: a paper showing that simple games become close to weighted majority voting when each class of equivalent players is large
SummaryThe paper studies whether simple games admit weighted majority representations. It is known that completeness of the coalitional desirability relation does not guarantee such a representation, so the authors show that weightedness holds in an approximate sense. When all classes of equally desirable players grow large while their number stays bounded, the approximation error is O(1/n_min), a bound the authors show is tight.
In briefGoDDaR: Automatically Detecting and Fixing Go Deadlocks
SummaryGo's goroutines can deadlock, and its runtime only catches cases where all threads are blocked, missing partial deadlocks where some goroutines stay blocked. This paper presents GoDDaR, which converts Go source into an intermediate representation of message-passing events and uses symbolic analysis to find global and partial deadlocks. It also transforms this representation to remove problematic synchronization and outputs fix suggestions as a diff for the original code. Evaluation on a benchmark of deadlock examples reports competitive partial-deadlock detection and improved automated fixing compared with prior tools.
In briefA unified analytical framework for general MIMO channels across diverse architectures
SummaryThis paper studies a generally correlated Rayleigh MIMO channel whose vectorized form follows a Gaussian distribution with an arbitrary covariance matrix. The authors prove that the channel matrix's spectral norm is bounded as transmit and receive antenna numbers grow proportionally, and they derive a deterministic approximation of ergodic mutual information with an explicit convergence rate. The approximation is given by matrix-valued self-consistent equations whose solution existence, uniqueness, and an iterative algorithm are established. Applications to downlink distributed MIMO and uplink heterogeneous MIMO, with numerical results, confirm the approximation's accuracy.
WhyProposes a unified matrix self-consistent framework for arbitrary-correlation MIMO mutual information, generalizing prior results with numerical validation.
In briefA partial Strassen method that speeds up matrix multiplication with less memory than BLAS
SummaryThe paper introduces the Partial Strassen algorithm, which applies Strassen steps to only part of a matrix multiplication. The authors state that its memory footprint remains small in both single-threaded and multi-threaded settings. In tests, the three-level version ran in about 80% of BLAS's time using 2.25 times the memory, and it outperformed a traditional Strassen implementation as matrix size grew.
In briefFictitious Play Can Converge Arbitrarily Slowly in Polynomial Rates
SummaryThe paper shows that fictitious play, a learning dynamic in two-player zero-sum games, can converge at arbitrarily slow polynomial rates. For each integer k≥2, it constructs a payoff matrix whose empirical-strategy duality gap decays as Θ(t^{-1/k}) after t steps. For k≥3 these games are counterexamples to Karlin's conjectured O(t^{-1/2}) rate, extending Wang's (2025) Θ(t^{-1/3}) construction.
WhyConstructs recursive game families where fictitious play converges arbitrarily slowly polynomially, refuting Karlin's O(t^{-1/2}) conjecture and extending prior Θ(t^{-1/3}) results.
In briefSampling algorithms for self-orthogonal and totally self-orthogonal codes over odd-characteristic fields
SummaryThe paper presents an algorithm that uniformly samples random linear codes of any hull dimension and type over finite fields of odd characteristic, giving the first method to sample uniformly random self-orthogonal codes of any rate, including self-dual codes. It revisits an algorithm by Albrecht, Benčina and Lai using mass formulae from Li, Shi and Ling. The authors also study total orthogonality, orthogonality with respect to all induced Galois geometries at once, characterise it for vectors and codes, and give an algorithm sampling linear codes with a total hull of prescribed dimension.
Martin R. Albrecht, Benjamin Ben\v{c}ina, Russell W. F. Lai
In briefRegret rates for profit maximization in multilateral trade, depending on support size
SummaryThis paper studies online profit maximization in multilateral trade, a setting that generalizes dynamic pricing and bilateral trade. An intermediary posts prices to buyers and offers payments to sellers, and trade happens only if every agent accepts. The authors design an algorithm with regret Õ(√(K^d T)) for constant dimension d, where K bounds the support size of each agent's marginal valuation distribution. A normalized zooming procedure avoids over-refining nearly flat regions.
Francesco Bacchiocchi, Matteo Castiglioni, Anna Lunghi et al.
In briefReed-Solomon codes reach capacity: list-decoding and proximity gaps explained
SummaryThis paper examines the limits of list-decodability for Reed-Solomon codes, an important open problem in algebraic coding theory. Brakensiek, Chen, Putterman, Zhang, and Zheng showed that RS codes over fields of large characteristic are algorithmically list-decodable up to capacity. Jeronimo then extended these techniques to the proximity-gaps question for RS codes. The article offers a unified and clear exposition of these results.
