In briefSpectra of multivariate Vandermonde matrices and error bounds for high-dimensional super-resolution
SummaryThe paper studies rectangular multivariate Vandermonde matrices whose columns correspond to irregularly spaced points on a high-dimensional torus. For points forming well-separated clusters, the column subspaces of different clusters are shown to be nearly orthogonal, so the spectral analysis reduces to each cluster. This yields sharp scaling for all singular values and min-max error bounds for recovering clustered point sources, matching known optimal one-dimensional rates.
In briefOptimizing discrete phase shifts of smart reflecting surfaces with evenly spaced reference directions
SummaryThis paper proposes a low-complexity algorithm for choosing discrete phase shifts on reconfigurable intelligent surfaces (RISs) that assist wireless communication. It evaluates candidate phase settings generated from uniformly distributed reference directions within the feasible optimality sector. The number of reference directions acts as a tunable parameter that trades computational cost against optimality. Simulations show large speedups with a small optimality gap, and the gap shrinks as more reference directions are used.
In briefDeterministic, variance-free zeroth-order optimisation using spherical designs
SummaryThis paper proposes a zeroth-order optimisation method that estimates gradients using central differences along directions given by a fixed spherical t-design, rather than random directions. The estimator is deterministic and therefore introduces no variance, and the authors also analyse a general finite difference method under higher-order smoothness and its bias. Convergence is established using the concept of input-to-state stability. Numerical experiments are reported to show the method outperforms existing random zeroth-order methods.
Amir Ali Farzin, Yuen-Man Pun, Philipp Braun et al.
In briefSigmaTrace: A Library-Agnostic Profiler for CPU- and GPU-Initiated RDMA Communication
SummaryRDMA networking offers low-latency, high-bandwidth communication in large systems, but existing observability tools either target specific libraries or report only aggregated data. SigmaTrace is a library-agnostic framework that traces individual RDMA operations at the libibverbs/mlx5dv layers and uses memory interposition to observe GPU-initiated RDMA without source patching. The paper reports host-side tracing overhead of 1.3% on NCCL AllReduce at 128 GPUs and under 2% on RCCL at 1024 GPUs. In experiments on a persistent-kernel BFS and Mixture-of-Experts inference, it reveals in-kernel barrier spread and 2.13x inter-node traffic skew that existing profilers miss.
In briefScalable Multi-Slot First-Price Pacing Equilibrium via Entropic Smoothing
SummaryThis paper studies pacing equilibria, which describe how budget-constrained advertisers clear an ad market when each request fills several slots. The authors regularize allocations with binary entropy, obtaining a smooth, strictly convex dual that splits into independent one-dimensional threshold solves per request. The smoothed equilibrium converges to the hard one at an explicit O(√T) rate, and a duality gap measures how far bids are from it. The authors implement a multi-GPU solver, test it on a production ad workload, and use it to compute each campaign's hindsight equilibrium cost per key result as a price anchor.
In briefPilot-Projected Nonlinear Self-Interference Estimation with Simultaneous Uplink Training
SummaryThis paper studies nonlinear digital self-interference cancellation and uplink channel estimation for single-antenna in-band full-duplex radios when a desired uplink pilot is present during self-interference training. The authors treat the unknown uplink channel as a nuisance parameter and eliminate it, reducing the problem to estimation on the orthogonal complement of the pilot convolution subspace. They propose two estimators, a pilot-projected ridge estimator with BIC tuning and a pilot-projected group-sparse Bayesian estimator. Simulations show gains over non-projected and hard-selection alternatives, with the latter suppressing residual self-interference below the noise floor at moderate training.
In briefOrientation diversity provably boosts angular-domain rank in 3D MIMO arrays.
SummaryThe paper proves that antenna element orientation is an effective spatial resource complementary to position. For collinear or coplanar arrays, angular-domain rank is capped by the number of projection-ambiguity classes of scattering directions. Orientation-diverse elements with sufficiently rich analytic patterns restore the maximal rank min(N,L) for almost every orientation draw. A closed-form correlation expression, monotone in orientation spread under isotropic scattering, is given, and Monte Carlo simulations corroborate the exact-rank and capacity behavior.
SummaryLegislative seats are allocated by rounding population or vote shares to whole numbers, and randomized rounding can leave a larger party or state with fewer seats than a smaller one, which the authors call envy. The paper designs apportionment rules that minimize the expected number of such envious pairs and shows that rules ignoring lower quotas can have worst-case envy of Θ(n²). It also introduces a polynomial-time linear programming framework for objectives such as expected and variance of envy. Empirically, envy-free allocations exist in every historical US decennial census instance.
WhyIntroduces envy-minimizing randomized apportionment with tight worst-case bounds, a polynomial-time LP framework, and validation on historical US census data.
Francesco La Commare, Jamie Tucker-Foltz, Victor Verdugo
In briefKVCompass: A discovery layer for cross-region prefix cache sharing in LLM serving
SummaryThis paper observes that KV cache reuse for large language model inference is mostly limited to a single cluster, so geographically distributed regions often recompute the same popular prefixes. The authors present KVCompass, a discovery plane for Internet-scale KV cache sharing that supports variable-length prefix lookup at the edge and forwards each request to the worker that best balances reuse, load, and network distance. Evaluation on a 64-worker geo-distributed simulation and a 14-worker three-continent testbed reports 1.3–14× lower P99 TTFT, up to 3.7× lower P99 end-to-end latency, and up to 4.7× higher throughput.
In briefEncoding information on DNA with patterns from enzymatic labeling
SummaryThis paper presents a formal framework for DNA-based data storage that models DNA labeling in strings, using a known DNA template and patterns induced by designed labels to represent information. One hypothetical implementation uses CRISPR-Cas9 and gRNA reagents. The authors analyze the general labeling channel, including fixed-length labels, give upper bounds on code size, and construct an encoder-decoder pair proven optimal in maximum code size under specific conditions. Fixed-length label codes are tied to de Bruijn sequences and constrained coding.
Daniella Bar-Lev, Eitan Yaakobi, Zohar Yakhini et al.
In briefAgentic AI for Designing Heterogeneous Systems Across Chip, Scale-Up, and Scale-Out
SummaryThis paper presents ArchAgent, a hierarchical multi-agent framework in which an LLM-based planner splits system design into chip, scale-up, and scale-out subproblems handled by specialist agents. A deterministic analytical evaluator computes all performance estimates, rather than the LLM generating them. On a mixed AI and HPC workload under a 1 MW power budget, it finds heterogeneous designs with a 1.93x weighted geometric mean speedup over a homogeneous GPU baseline using 18 evaluations, versus 100 for hill climbing, which reached 1.65x.
In briefPower of Alternating Local Certification with Constant-Size Certificates
SummaryThis paper studies alternation hierarchies for distributed decision in anonymous networks, where a prover and disprover alternately assign constant-size certificates to nodes over a constant number of rounds. Nodes decide acceptance based on their input and neighboring certificates, under three models: SET, MULTISET, and PORTS. The authors show the models form strict hierarchies and that SET coincides with MSO1 over finite alphabets. They also show PORTS cannot capture MSO2, for example Hamiltonicity or perfect matching, and cannot verify network size over unbounded alphabets.
Pedro Montealegre, Mart\'in R\'ios-Wilson, Jukka Suomela et al.
In briefRe-examining fairness of pure Nash equilibria in Round-Robin allocation
SummaryThe paper shows that, for additive valuations, not every pure Nash equilibrium of the Round-Robin mechanism yields an allocation that is envy-free up to one good (EF1) under agents' true valuations, correcting an earlier claim by Amanatidis et al. On the positive side, every pure Nash equilibrium of any fixed recursively balanced picking sequence mechanism is 1/2-EF1 and proportional up to one good (PROP1). The authors also give Round-Robin examples showing that no guarantee strictly better than 2/3-EF1 holds for all equilibria.
Georgios Amanatidis, Haris Aziz, Georgios Birmpas et al.
In briefOn the Length of Maximum Sum-Rank Distance Codes
SummaryThis paper extends the classical geometric theory of arcs in finite projective spaces to maximum sum-rank distance (MSRD) codes, studying how their block structure limits code length. In dimension three, it determines the possible intersection patterns of lines with associated linear sets and gives formulas for the codes' characters and weight distributions. It shows that a block of rank at least two reduces the maximum total length by at least q-1, that sufficiently long MSRD codes are classical, and that three-dimensional one-weight MSRD codes do not exist. An exceptional three-dimensional code over F_9 is also constructed.
In briefFour New Binary Codes Correcting Single Errors
SummaryThe paper constructs four binary codes: a (16, 3072, 3) code, a (17, 6144, 3) code, an (18, 12288, 3) code, and a (19, 24576, 3) code. These yield four new lower bounds on A(n,3) for n = 16 to 19. Each code is a union of cosets of a linear code, and the constructions were found by a computer search using a tabu search algorithm.
In briefReview of integrated sensing, communications, and identification for low-altitude economy
SummaryThis paper reviews integrated sensing, communications, and identification (ISCAI) for the low-altitude economy. It explains that low-altitude information networks must provide reliable communication, environmental sensing, and target identification together. The review covers communication, sensing, and identification approaches, along with system architectures, channel modeling, signal models, and performance metrics. It highlights a backscatter communication-based identification mechanism that uses backscatter devices as electronic license plates for aircraft.
In briefSpatial Distribution of Mean Peak Age of Information in LEO Satellite IoT Networks
SummaryThis paper studies a large-scale, fully loaded IoT network covered by LEO satellites, focusing on the uplink mean Peak Age of Information (PAoI). Using stochastic geometry, it derives the spatial distribution of this metric by first obtaining the meta distribution of the uplink conditional coverage probability. Numerical results show how the number and altitude of LEO satellites affect the coverage probability meta distribution and the spatial distribution of mean PAoI, offering insights for choosing satellite beamwidth to minimize conditional mean PAoI.
