In briefResource-aware sub-network allocation for heterogeneous decentralized learning
SummaryThis paper addresses synchronous decentralized learning, where each training round is limited by resource-constrained nodes with different computation, communication, and memory. It proposes RASA-DL, which assigns each node a maximum sub-network width based on hardware and learning-utility profiles, formulating deployment as bottleneck-latency minimization under memory and utility constraints solved by binary search. In experiments with controlled heterogeneity, the minimum-latency deployment improved accuracy over the largest globally feasible uniform deployment by 2.80 points (Moderate) and 2.43 points (Severe) at the same round latency, with larger gains under strong Non-IID data.
In briefWhy alternating updates speed up counterfactual regret minimization
SummaryThis paper examines why alternating the two players' updates makes counterfactual regret minimization with RM+ faster in imperfect-information games. It tests the common intuition that seeing the opponent's fresh strategy acts like optimism, using an analytic witness family of deterministic two-player zero-sum self-play. The main result is a dichotomy theorem linking trajectory separation to recurrent exact-vertex play, with measurements nearly matching the predicted constant.
In briefMinimum-divergence methods for recovering probability distributions from moments
SummaryThis paper studies whether maximum-entropy reconstruction, which recovers a distribution from finitely many generalized moments by minimizing KL divergence to a reference measure, can be generalized. Replacing KL with a smooth f-divergence still yields a reconstruction, but the best-approximation property fails. The authors show that, up to equivalence, KL is the only C² f-divergence with this entropy-adapted property, and illustrate the consequences numerically.
In briefLegitimate Reconstruction Advantage in Wiretapped Remote Monitoring
SummaryThis paper studies joint accuracy and security in remote monitoring where updates meant for a legitimate receiver may also be intercepted by an eavesdropper. It defines legitimate reconstruction advantage (LRA) as the event that the intended receiver reconstructs the source correctly while the eavesdropper does not. Under randomized stationary transmission and zero-order-hold reconstruction, the authors derive closed-form expressions for LRA probability and persistence for finite-state ergodic Markov sources.
In briefExact Completion-Time Bounds for LLM Inference Pipelines on Mixed GPUs
SummaryThis paper gives an exact critical-path formula for the completion time (makespan) of micro-batch pipelines serving large language models on heterogeneous GPU clusters. It shows how the classic homogeneous GPipe formula overestimates or misestimates this time. Results are extended to asynchronous and blocking communication, and the bias is split into a homogeneity term and a communication-ignoring term. The claims are checked numerically on eleven configurations and a large random test.
In briefModel-driven deep learning for low-bit quantized distributed detection
SummaryThis paper studies distributed detection in networks where sensors must send severely quantized messages to a fusion center because of bandwidth limits. The authors propose a model-driven framework with a trainable binary probabilistic quantizer and a tree-structured quantizer for multi-bit messages. When observation statistics are unknown, class-conditional distributions of the messages are estimated from held-out labeled data to form plug-in MAP detectors. Simulations indicate that the method approaches centralized-oracle performance at higher quantization resolutions.
In briefShort-Boundary Signaling for Grant-Free Control Messaging
SummaryThis paper addresses how receivers in asynchronous grant-free access detect the start and end of variable-length transmissions. It notes that existing decoder-aided methods use boundary sequences as long as a codeword, which adds heavy overhead for short commands. The authors treat boundary length as an independent design parameter and propose a convolutional neural network that detects both boundaries from the received waveform and decoder soft output. Simulations show the shortest tested boundary, 16, gives the highest effective payload efficiency at all tested loads.
Massimo Battaglioni, Dania De Crescenzo, Edoardo Carnevali et al.
In briefA study of Fisher information lower bounds for Gaussian mixtures and related estimation limits
SummaryThis paper studies lower bounds on the Fisher information of Gaussian location mixtures and their use in estimation problems. For discrete mixing distributions, it derives a universal positive-semidefinite lower bound based on the order-two Rényi entropy, and dimension-aware trace bounds using Sobolev inequalities and Sibson mutual information. The results are applied to least favorable priors for bounded Gaussian mean estimation, and Brown's identity and the I-MMSE relation translate them into upper bounds on MMSE and mutual information.
In briefHow Probabilistic Constellation Shaping Controls Ranging Sidelobe Levels in ISAC
SummaryThis paper studies probabilistic constellation shaping (PCS), a technique originally meant to bridge finite uniform constellations and capacity-achieving continuous distributions. The authors show that a class of type-based PCS methods can precisely control the codeword-wise integrated sidelobe level ratio (ISLR) in OFDM systems, offering stronger sensing guarantees than controlling only the expected ISLR. They further argue that OFDM is nearly the only unitary modulation basis where such precise codeword-wise ISLR control is possible. Based on these results, the paper proposes practical type-based PCS methods compatible with probabilistic amplitude shaping (PAS), which achieve higher communication rates than existing ISAC-oriented PCS methods under the same ISLR constraints.]
In briefTwo Moments Suffice: Limits of Continuous Encoders for Over-the-Air Computation
SummaryThis paper studies over-the-air computation, where transmitters send locally transformed observations over one use of a complex multiple-access channel so a receiver can compute functions such as empirical variance from two separately postprocessed sums. With four or more transmitters and exact noiseless recovery, the authors show that every continuous encoder must share a common affine structure preserving both sums, reducing encoder design to choosing finitely many coefficients. They define a robustness score, cast robustness maximization as an exact convex semi-infinite program, and derive closed-form optimal encoders for several function classes, including empirical variance.
In briefStructure of polynomial ideal codes and additive Reed–Solomon codes
SummaryThis paper studies polynomial ideal (PI) codes and additive Reed–Solomon (aRS) codes in the folded Hamming metric. It determines the maximum possible length of PI codes, shows that the dual of every PI code is a generalized PI code, and derives their dually QMDS property and an explicit weight distribution. It also links PI codes from irreducible polynomials to aRS codes with pairwise nonconjugate evaluation points and computes the permutation group of these aRS codes.
In briefApproximating the Value of Time-Bounded Weighted Timed Games
SummaryThis paper studies the approximation of the value in weighted timed games, which extend timed automata with weights and adversarial players. Computing the exact value is known to be undecidable, and approximation is also undecidable in general. The authors show that for weighted timed games whose executions are time-bounded, the approximation problem is decidable in 4-ExpTime.
In briefHow Decoding Failure Skews IBLT Count Estimates: A Gaussian Selection Law at the Peeling Threshold
SummaryThe paper studies why count measurements used to size a second round in a self-sizing set-reconciliation protocol become biased after an IBLT peeling decoding failure. In the ideal peeling model, the measurement conditional on failure converges to a bivariate normal truncated to a half-plane, which the authors call the Gaussian-selection law. Based on this law, they derive a corrected threshold and upper confidence bound for the difference count.
WhyProposes a Gaussian-selection law with closed-form constant, proven via a quenched CLT and matching empirical quantiles; a substantive but subfield-focused contribution.
In briefIterative bounded distance decoding with a 2-bit reliability marker for product-like codes
SummaryThis paper proposes a soft-aided hard-decision decoding scheme for product codes that uses two-bit quantized reliability markers. One bit supports a two-level tolerance check that validates bounded distance decoding (BDD) results, and the other identifies bits to be randomly flipped when BDD fails or miscorrects. List decoding is added to improve error performance, and density evolution is used to design the quantization thresholds. Simulations show the scheme outperforms several unquantized soft-aided decoders while keeping memory costs and BDD attempts low.
In briefAdaBool: Sparse Boolean Matrix Processing with Adaptive Data Representation
SummaryThis paper presents AdaBool, a framework for Sparse Boolean Matrix Multiplication (SpBMM), which determines whether logical relationships exist between two sets of entities. It splits matrices into fine-grained blocks and encodes each block with the representation best suited to its local sparsity. A heterogeneous GPU execution model then groups blocks by encoding type and maps them to matching processing paths. The authors report that AdaBool outperforms representative baselines by over an order of magnitude on production and public workloads.
In briefAsymptotic rate of Delsarte's bound for binary and spherical codes, solved via potential theory and variational methods
SummaryThe paper gives the asymptotic rate of Delsarte's linear programming bound for binary codes and spherical codes. The lower bound uses potential theory to reduce the problem to an obstacle problem, which is solved in closed form using elliptic integrals. The upper bound uses interlacing spectra and the Markov-Krein correspondence to build a variational problem that has an exact solution. The two bounds match.
WhyUses potential theory and an obstacle problem to pin down the exact asymptotic rate of Delsarte bounds, with matching upper and lower bounds.
In briefCertifying outage risk for fluid antenna receivers using limited historical data
SummaryThis paper addresses fluid antenna receivers that choose a service position from partial port observations, where limited historical data leave uncertainty in the covariance used for that choice. The authors propose a before-deployment reliability certification for a fixed partial-observation rule, using a history-based parameter region and covariance, conditional-probability and selector bounds for simultaneous risk limits. Under the stated stationary model and numerical assumptions, erroneous positive certification is limited to 0.05 for one qualification procedure. In a structured Gaussian example with 32 candidate positions, all 282 risk groups were qualified.
In briefOn the Subadditivity of Mutual f-Information
SummaryThis paper addresses an open problem posed by Polyanskiy and Wu on whether mutual f-information is subadditive. It studies two generalized information measures: mutual f-information, based on f-divergence, and variational (G,f)-information, defined by minimizing transformed f-divergences. For mutual f-information, the authors give a sufficient condition on g(x)=x²f''(x) that ensures subadditivity under conditionally independent observations, yielding new explicit families. For the variational measure, they extend known sufficient conditions to additional transforms G.
In briefMoving-antenna near-field localization: optimizing antenna placement to reduce positioning error
SummaryThis paper studies near-field target localization using movable antennas (MAs), where a transmitting target's location is estimated by a receiver. The authors use the squared position error bound (SPEB) as the performance metric and optimize antenna placement at the receiver to minimize it. They relax the antenna spacing constraints, reformulate the problem using degree-4 truncated moments as a semidefinite program, and then construct a clustered placement policy from the recovered distribution. Numerical results show a three-cluster optimal structure that outperforms conventional array geometries and closely approaches exhaustive search.
In briefLow-complexity decoder using four-level quantized soft information for product-like codes
SummaryThis paper proposes iBDD-GSRF, a low-complexity soft-aided hard-decision decoding scheme for product-like codes. It uses four-level quantized soft information, a graded scaled reliability rule for bounded-distance decoding results, and bit-flipping based on bit-level and component-code reliability. Simulations on product codes and staircase codes show a gain of more than 0.4 dB over conventional iBDD.
In briefHelix: A Unified Programmable FPGA Accelerator for Many Krylov Solvers
SummaryThis paper presents Helix, a programmable FPGA accelerator that supports eight different Krylov subspace solvers for sparse linear systems. It extracts common computational primitives across solvers and uses programmable datapaths and a DSL compiler to map each solver onto one shared architecture. On a Xilinx Alveo U55C, Helix sustains 47.5% to 96.9% of an NVIDIA RTX PRO 6000 GPU's throughput, with programmability adding only 0.25% to 0.57% execution-time overhead.
In briefProbabilistic Optical-Layer Surrogate for Risk-Aware Control-Plane Decisions: a generative model for fast estimation of path feasibility and violation risk
SummaryThe paper proposes a generative surrogate model for the optical layer, trained on traces from validated optical simulators. It outputs full conditional distributions over physical-layer metrics such as OSNR, Q-margin, nonlinear penalty, and feasibility. This lets a controller estimate violation probability and tail risk at any threshold at runtime without retraining per decision. The surrogate is compared with simulators, deterministic learned predictors, and quantile-regression estimators across nine research questions.
In briefFast, quadratically convergent method for recovering spectrally sparse signals
SummaryThis paper studies recovering spectrally sparse signals from linear compressive measurements. It exploits the low-rank Hankel structure of the signal and formulates a nonconvex fixed-rank optimization problem with Hankel regularization. The authors propose the Superfast Iterative Hard Thresholding (SIHT) algorithm, the first nonconvex method with provable quadratic convergence for this problem. For random Gaussian measurements, it achieves exact recovery without noise and stable recovery with noise, with near-optimal sample complexity.
WhyProposes the first nonconvex spectral compressed sensing algorithm with provable quadratic convergence and near-optimal sample complexity guarantees.
In briefLWM-TF: A Wireless Foundation Model for 5G NR Time-Frequency Grids
SummaryThis paper introduces LWM-TF, a 9.58M-parameter factorized masked autoencoder pre-trained on 5G NR uplink OFDM resource grids taken after the FFT and before channel estimation. It attends and masks along both OFDM symbols and resource blocks, reconstructing the normalized log-power of masked tokens and predicting their scale separately. With the encoder frozen, linear probes perform modulation classification, CRC success prediction, and mobility detection, outperforming a supervised CNN and engineered receiver statistics, especially with few labels.
In briefAn autocomplete model for elections: a Markov method to summarize and generate synthetic ranked-choice ballots (paper)
SummaryThis paper studies ranked choice election preference profiles and proposes a Markov model that records which candidate tends to follow which on observed ballots, stored in a succession matrix. The model is used to generate synthetic profiles with tunable ballot length and is compared with methods based on positional frequencies and head-to-head comparisons. Experiments on Scottish ranked-choice data report that the succession-based approach captures held-out ballot structure, performs well at completing partial ballots, and often preserves election outcomes.
In briefCyclotron: A Recurrence Language for Interprocessor Communication
SummaryThis paper presents Cyclotron, a compiler that describes streaming dataflow algorithms with recurrence equations and compiles them into distributed programs running on interconnected processors. Users specify communication between points in the iteration space directly in the recurrence language, and the compiler generates per-processor sends, receives, and computations. The authors evaluate distributed matrix multiplication, TRSM, Cholesky, and fused Cholesky–TRSM on up to 16 CPU nodes and 16 GPUs, comparing against ScaLAPACK, SLATE, and cuSolverMp. They report that Cyclotron is competitive with these hand-optimized libraries and often outperforms them.
Shiv Sundram, Akhilesh Balasingam, Nathan Zhang et al.
In briefBreaking Down URLLC Deadline Misses in Beam-Managed 6G Networks
SummaryThis paper examines why ultra-reliable low-latency communication packets miss their deadlines in 6G millimeter-wave networks. It classifies misses into queueing-limited, transmission-limited, and beam-tracking-limited causes, using scenarios with 64 to 4096 antennas. The authors report that beam tracking becomes the dominant cause beyond about 1825 antennas, and that shortening the beam-sweep period substantially reduces these misses.
In briefExact joint MAP decoder estimating active user count and message counts in orthogonal unsourced random access
SummaryThe paper studies jointly recovering the number of active users (load) and the message-count vector in type-based unsourced random access with very short messages. It proposes an exact joint maximum-a-posteriori decoder for arbitrary priors, showing that optimal solutions for all candidate loads are nested prefixes of one greedy chain, enabling joint estimation in O(M+K_max log M) time. In the reported setting (K uniform over 0 to 60 at 2 dB), the decoder markedly reduces the no-activity false-alarm probability and normalized count error compared with the Bayesian-count two-stage receiver. The authors also prove that the sum-only statistic is insufficient.
In briefPlacement-controlled microbenchmarks of NVIDIA Grace-Blackwell unified memory
SummaryThis paper studies CUDA Unified Memory performance on NVIDIA GB200 using microbenchmarks that control where data is placed, with GH200 as a baseline and GB10/DGX Spark for validation. It separates cases such as local HBM access, host/system memory access, explicit data movement, prefetching, reuse, and eviction under capacity pressure. The authors argue that unified addressability does not mean uniform memory behavior, so a single bandwidth figure is misleading, and they summarize results with a measurement-calibrated cost model.
In briefCache-aided linear function computation: how caching cuts the data sent to compute linear combinations
SummaryThis paper studies distributed systems where each source stores its own data and a receiver requests a linear combination of these data. Sources use pre-request caches to reduce the communication needed for exact recovery. The authors derive a lower bound on worst-case communication rate over any finite field and propose a parity-aided placement (PAP) scheme for binary files. For every N≥3, PAP's rate meets the lower bound when the cache size is at most 1/(2N-1) or at least (N-1)/N, establishing the optimal memory-rate tradeoff in those ranges.
