Peer Revenue Now accepting positions

Papers

Fresh from arXiv, typeset properly. Highlighted words trade on the market.

All of arXiv

Friday 6 November

High Energy Astrophysical Phenomena

A possible connection between the ridge and hypotenuse in the bispectrum of black hole X-ray binaries through quasi-periodic modulation

Quasi-periodic oscillations (QPOs) in black-hole X-ray binaries show nonlinear coupling with broadband variability, producing characteristic ridge, hypotenuse, and web-like structures in the bicoherence. However, the origin of these structures remains unclear. We develop a mathematical model in which a modulator phase-locked to the QPO acts multiplicatively on the underlying variability. The resulting frequency mixing naturally produces both the ridge and hypotenuse associated with the fundamental QPO, as two parts of the same modulation-induced bispectral structure, and the same mechanism extends to the second harmonic. The model further predicts opposite-sign biphases between each ridge and its corresponding hypotenuse. We test these predictions using NICER observations of Swift~J1727.8$-$1613 and MAXI~J1820$+$070. The bispectral maps show QPO-related ridge–hypotenuse structures broadly consistent with the model, with qualitatively opposite-sign biphases between each ridge and its corresponding hypotenuse. In particular, the biphase map provides a complementary diagnostic that reveals weak coupling structures that are difficult to identify from the bicoherence alone. We also compare the measured bispectral modulus with the predictions for isolated ridges and hypotenuses. These predictions broadly reproduce the observed frequency dependence, while localized deviations remain near the fundamental-QPO and second-harmonic frequencies, where different bispectral structures overlap. Our results provide a simple mathematical framework in which the web-like pattern in the bispectrum could arise naturally from multiplicative coupling between a QPO-phase-locked modulator and the underlying variability.
Read the full abstract

All authors Pei Jin, Mariano Méndez, Federico García, Diego Altamirano, Haifan Zhu, Liang Zhang, Guobao Zhang

Computation and Language

Steerable Multi-Objective Group-Relative Preference Optimization with Jointly Trained Low-Rank Adapters

Language model deployment often requires different trade-offs between competing objectives as user preferences change, yet standard post-training methods typically produce a single policy for a fixed objective weighting. We introduce LambdaMix, a new parameter efficient method for learning a preference conditioned family of policies with group-relative online reinforcement learning. LambdaMix jointly trains one low-rank adapter per objective over a frozen base model. One preference vector selects the adapter mixture that generates responses and weights the separately normalized rewards used to update the adapters. A single training run supports changing preferences at inference without further optimization. We experiment with four models on helpfulness and harmlessness and on solving math problems while following instructions. We also evaluate the two largest models on rewriting comments to preserve wording while reducing clues about personal attributes. On the first two tasks, LambdaMix achieves higher hypervolume than all evaluated conditioned and policy merging baselines across all four models. LambdaMix uses fewer training completions and stores fewer adapter parameters than a set of GRPO specialist policies. With architecture and initialization fixed, varying adapter mixtures and objective weights together gives wider score ranges and higher coverage than varying either alone.
Read the full abstract

Combinatorics

Polyhedral Geometry of Degree-Bounded Characteristic Imset Polytopes

Characteristic imset polytopes are convex hulls of a collection of characteristic imset vectors, which are 0/1 vectors that encode the Markov equivalence class of a directed acyclic graph. These polytopes are the feasible region of linear programs which correspond to causal discovery problems. We introduce a new family of these polytopes whose vertices correspond to directed acyclic graphs with in-degree bounded by $k$. This yields a polytope whose dimension is only polynomial in the number of nodes of the underlying graphs instead of exponential. We show that when $k = 1$, this polytope is a matroid independence polytope. For $k = 2$ we provide several large families of facet-defining inequalities, which give a complete $\mathcal{H}$-representation of the polytope on four nodes.
Read the full abstract

Fluid Dynamics

Turbulent wakes behind nine bluff-body shapes from direct numerical simulation and particle image velocimetry

We present a cross-validated dataset of turbulent wakes behind nine bluff bodies at Reynolds number Re = 10,000: a circular cylinder, a 2:1 elliptical cylinder at 0{\deg} and 90{\deg}, short and long isosceles triangles with the apex upstream or downstream, and a square cylinder at 0{\deg} and 45{\deg}. Each body was studied by wall-resolved direct numerical simulation (DNS) with a spectral-element solver and by time-resolved planar particle image velocimetry (PIV) in the MIT Intelligent Towing Tank. The DNS records contain time- and spanwise-averaged velocity, pressure and Reynolds stresses, 24-bin phase-averaged fields keyed to the lift signal, drag and lift histories, and one three-dimensional instantaneous snapshot per body; the PIV records contain the matching time-mean and phase-averaged fields in the same coordinate frame. The DNS set-up is verified against reference cylinder data at Re = 3900, and DNS and PIV at Re = 10,000 are compared field by field. The dataset supports the development and testing of data-driven turbulence closures, reduced-order models, and wake-stability analyses across geometries with distinct separation behavior.
Read the full abstract

All authors Zhen Zhang, Theo Käufer, Louise Ronglan, Michael S. Triantafyllou, George Em Karniadakis

High Energy Physics (Theory)

Dynamical gluon mass generation, the Schwinger mechanism, spontaneous symmetry breaking, and color confinement in the linear covariant gauges of QCD

We critically re-examine the Kugo-Ojima (KO) scenario for color confinement and identify several inconsistencies and a crucial error in its derivation. Within the KO framework, under customary assumptions, we show that the BRST quartet structure incorporates a unique mechanism for dynamical gluon mass generation – namely, the Schwinger mechanism, in the formulation developed by Aguilar, Binosi, Ferreira, and Papavassiliou (ABFP). We show that the Schwinger mechanism is a Higgs mechanism, in that it occurs together with the spontaneous breaking of the global gauge charge, as it should. The ABFP bound state is the associated (unphysical) composite Goldstone boson. The zero-momentum value of the gluon propagator evaluated in the Pinch Technique–Background-Field Method (PT-BFM) naturally emerges in the derivation, setting a Renormalization-Group invariant scale for mass generation and suggesting a new nonperturbative interpretation of the PT-BFM. BRST symmetry takes over the role of the broken symmetry, while the color symmetry of the spectrum remains unbroken. We obtain a unique, well-defined, unbroken, BRST-exact asymptotic color charge operator that realizes the confinement of color in the sense of the absence of colored asymptotic states from the BRST cohomology. Still within the KO framework, we take the first fundamental step towards a proof of the nonperturbative unitarity of the linear covariant gauges of QCD, and we discuss the experimental evidence in favor of the Schwinger mechanism. We lay down the foundations for a unified theory of gauge-boson mass generation in the linear covariant gauges of Yang-Mills theories.
Read the full abstract

Optics

Layer-Peeling Inversion of Temporal Refractive Index Profiles from Noisy Data

We introduce a direct method for reconstructing the temporal evolution of a medium from electromagnetic field measurements. A sequence of temporal refractive-index changes splits an incident wave into forward- and backward-propagating components, producing a spatial response of finite extent. We show that this finite support, together with causality, enables sequential reconstruction of the temporal interfaces through a time-domain extension of layer peeling. For continuously varying media, we introduce a time-adjusted Goupillaud medium and, through space-time duality, establish the corresponding momentum-conservation and admissibility conditions. The method remains robust under incomplete and noisy measurements, providing a general framework for the metrology, characterization, and design of time-varying electromagnetic systems.
Read the full abstract

Number Theory

Permutation polynomials from quadratic and linear forms

We investigate a family of permutation polynomials of finite fields using the theory of quadratic and linear forms. For an odd prime power $q$, consider polynomials over $\mathbb F_{q^n}$ of the form $\sum_{0\le i\le j<n}a_{ij}x^{q^i+q^j}+L(x)$, where $a_{ij}\in\mathbb F_{q^n}$ and $L$ serves as an $\mathbb F_q$-linear endomorphism of $\mathbb F_{q^n}$. In this paper, we begin with a general treatment over arbitrary fields of characteristic not $2$, and then provide explicit characterizations for several classes of such permutation polynomials.
Read the full abstract

High Energy Astrophysical Phenomena

Fundamental Noise Processes in Pulsar Timing

The stability of spin rates and emission regions of radio-emitting neutron stars (pulsars) has enabled remarkable advances, including the detection of exoplanets, testing theories of gravity, and stimulating the vigorous exploration of a new window in gravitational wave astronomy an nanohertz frequencies. An ideal astrophysical clock would emit identical pulses at a constant rate, propagating along single geodesics through a perfect vacuum. In reality, the timing precision achievable with pulsars is limited by spin noise, variability in the beam produced in a complex magnetosphere, and multipath propagation through a turbulent interstellar medium, with additional limitations from uncertainties in the changing pulsar – observatory vector and instrumental effects. Here we present the fundamentals of pulsar timing, its current limitations, and prospects for improvements in timing precision. We develop metrics to assess the timing quality of a pulsar and use them to evaluate sources currently included in pulsar timing arrays (PTAs) as a means for characterizing the quality of future pulsar discoveries. With increases in telescope sensitivity, pulsar timing precision will increasingly be limited by propagation effects in the interstellar medium. We identify sources of non-Gaussianity in timing residuals and separate out effects that are chromatic or achromatic, as well as effects that vary on short (seconds to minutes) and longer (days to months) timescales. Modern wide-bandwidth receivers provide simultaneous arrival times across a range of frequencies, enabling the robust mitigation of chromatic interstellar effects and the identification of non-Gaussian events superimposed on random fluctuations. We recommend changes to timing procedures that leverage instrumentation advances to improve timing precision.
Read the full abstract

All authors James M. Cordes, Stella Koch Ocker, Shami Chatterjee, Timothy Dolch, Ross J. Jennings, Michael T. Lam, Jacob E. Turner

Discrete Mathematics

What remains of the planar induced minor conjectures?

Dallard, Milanič, and Štorgel, and independently Gartland and Lokshtanov, conjectured that for every planar graph $H$, Maximum Weight Independent Set and List $3$-Coloring are polynomial-time solvable in $H$-induced-minor-free graphs. As a structural explanation, Gartland and Lokshtanov further conjectured that such graphs admit balanced separators dominated by a constant number of vertices. Bonnet and Chang recently refuted the independent-set part of the algorithmic conjecture, by showing that Maximum Independent Set is $\mathsf{NP}$-hard in graphs excluding the $5\times5$ grid as an induced minor. We revisit both conjectures. We show that the graphs used by Bonnet and Chang exclude much simpler planar graphs, including several $16$-vertex trees, and we introduce a second, independent family of hard instances. We also prove that $3$-Coloring is $\mathsf{NP}$-hard in both settings, thus refuting the algorithmic conjecture in full, even in its quasipolynomial version, unless $\mathsf{NP}\subseteq\mathsf{QP}$. On the structural side, we identify planar graphs $H$ for which the structural conjecture holds, while the algorithmic one fails. Thus, constantly dominated balanced separators cannot by themselves explain the tractability of Maximum Weight Independent Set and List $3$-Coloring.
Read the full abstract

All authors Jadwiga Czyżewska, Tomáš Masařík, Marcin Pilipczuk, Paweł Rzążewski

Systems and Control

A Novel Algorithm to Simulate Conductance-based Neuronal Networks

Conductance-based models are central to neurophysiology and neuromorphic engineering, yet their simulation scales poorly to large networks. We study and analyze an iterative solver that operates in signal space. Each iteration reduces to $N$ decoupled scalar linear time-varying differential equations, one per neuron, which are solved in parallel and in closed form. The solver admits two complementary interpretations: a variable metric forward-backward splitting method, rooted in operator theory, and a waveform relaxation quasi-Newton method, rooted in circuit simulation. We prove uniform and superlinear convergence, with a bound that reflects the conductance structure of the dynamics. The algorithm is illustrated on the classical Hodgkin–Huxley neuron and on networks of up to 5000 neurons.
Read the full abstract

Machine Learning

Encoding Memory in World Models with Test-Time Model-Based Reinforcement Learning

Model-based reinforcement learning (MBRL) approaches have demonstrated strong performance across diverse domains in recent years, learning accurate world models from agent trajectories. However, despite this progress, current approaches still struggle to surpass human performance on tasks that require long-term memory and planning. In this work, we address the problem of long-term memory in MBRL by moving beyond the commonly studied paradigm of architecture design. We introduce Test-Time Model-Based Reinforcement Learning (TT-MBRL), using gradient descent at test time to encode the memory of episode observations into world model parameters. This enables the agent to learn a distinct world model for each episode, improving memory capabilities compared to traditional approaches that solely rely on contextual memory. We first establish a proof of concept on a simplified two-dimensional variant of the Memory Maze environment, which we refer to as Memory Maze 2D, showing consistent improvements across different world model architectures. We then demonstrate the effectiveness of TT-MBRL in the 3D domain, achieving new state-of-the-art results on the Memory Maze benchmark. Finally, we show that TT-MBRL maintains strong performance on non-memory benchmarks such as DeepMind Control and Atari 100k.
Read the full abstract

Computational Engineering

TPIC: Energy-Conserving Space Plasma Simulations on GPU Supercomputers

Artificial heating and numerical damping affect particle energization in kinetic space plasma simulations, even when the numerical scheme remains stable. Implicit and semi implicit energy conserving Particle-in-Cell (PIC) methods eliminate these errors in the discrete energy balance while relaxing explicit stability constraints. We present TPIC, a massively parallel implementation of the Energy Conserving Semi-Implicit Method (ECSIM) for GPU supercomputers. We extend the exact energy conservation in the ECSIM formulation to closed domains with conducting boundaries by deriving and canceling a spurious term in the discrete Poynting flux theorem. TPIC executes the complete particle and field cycle on NVIDIA and AMD GPUs, combining matrix-engine acceleration with asynchronous communication and computation. We validate TPIC in a variety of electrostatic and electrodynamic test cases relevant for space plasma physics applications, including quasineutral plasma expansion, two-stream instability, Whistler waves, Weibel instability, Whistler turbulence, and three-dimensional magnetic reconnection with conducting walls and 12.3 billion particles. We show that in all scenarios, TPIC agrees with the corresponding analytical theory or reference numerical results, and preserves the total system energy to machine precision. On 4,096 AMD GPUs, TPIC advances 1.9 trillion particles with 96.8% weak scaling efficiency.
Read the full abstract

Human-Computer Interaction

Entering a New Domain in the Era of Generative AI: An Autoethnography of Solo Game Development as a Bioinformatics Educator

The rapid improvement of generative AI (GenAI) not only increased the efficiency of practice but also empowered individuals with capabilities to explore new domains. However, identity formation in the wake of such empowerment remains understudied. We use an autoethnography to understand identity formation in the context of solo game development as a bioinformatics educator. With the goal of developing a create-your-own-adventure (CYOA) game, we collected data for 73 days through journal entries, interacting with communities of practices. The autoethnographer had continuously reassessed the situation, and closely interacted with communities of practice. Their prior knowledge was used to provide in-game context and to uniquely interpret game development processes. These findings suggest that the autoethnographer made efforts to belong to communities of practice and engaged extensively in sensemaking through the lens of a bioinformatics educator. Although GenAI puts power at hand, attaining a new identity took place at a human pace.
Read the full abstract

Computational Complexity

On Positive Computation

We introduce new notions of reducibility for monotone classes and show that a monotone version of Circuit Satisfiability and the Hamiltonian Cycle Problem are complete for $mNP$, even under positive reductions (which are not the same as monotone reductions). If this holds for all (monotone) $mNP$-complete problems - that they are also so complete for $mNP$ -, then $P\neq NP$. We introduce positive models of relativized computation and show that the separation $posP\subsetneq mNP$ does not relativize, even when oracles are restricted to monotone sets. Yet, there is a (weak) diagonalization proof that $posP\subsetneq mNP$.
Read the full abstract

Machine Learning

BiLoGo: Reducing the smoothing of localised responses in Transformer-based neural operators for sheet forming

Trial-and-error FE simulation of sheet forming could be computationally expensive, neural-network-based learned simulators are used to accelerate design. These simulators should accurately capture localised responses, such as thinning around sharp radii or near ribs, which determine formability. Transformer-based neural operators propagate information efficiently across the mesh but tend to smooth such localised responses. This study develops BiLoGo, a Bipartite Local-Global learned simulator with an FE-guided design that connects local node-element message passing with a Transformer-based neural operator for long-range information exchange. BiLoGo has three modules: a bidirectional node-element GraphSAGE model that learns nodal-elemental coupling on a bipartite graph and predicts each quantity on its native FE entity without interpolation; a dual-stream Transformer-based neural operator that processes blank deformation and tool contact with separate token sets, preventing localised contact information from being averaged with the whole-blank state; and a fusion and element update module that combines global and local features via gated residual fusion and updates element features conditioned on element-node kinematics. Evaluated on a dome-shaped cold forming case and two corner-shaped hot stamping cases (one with ribs), BiLoGo achieves the lowest positional and threshold-based thinning errors on all three cases compared with single-level and hierarchical graph-based models and Transformer-based neural operators. Near the ribs, the smoothing index of predicted thinning increases from 0.41-0.56 for Transformer-based neural operators to 0.94 for BiLoGo-U. Compared with the hierarchical graph network CAtt-BiUGNN, BiLoGo-T reduces mean positional error by 15.5% on the Corner case and BiLoGo-U by 8.3% on the Corner+Rib case, with 66-69% less training memory.
Read the full abstract

All authors Yingxue Zhao, Haoran Li, Haosu Zhou, Tobias Pfaff, Nan Li

Probability

Fourier Positivity and Sharp Poincaré Inequalities on Graphs

We prove sharp dimension-free $L^1$-Poincaré inequalities on Hamming cubes, a sharp quadratic $L^1$-Poincaré inequality on product Poisson spaces, and a sharp pseudo-Poincaré inequality on finite abelian Cayley graphs. These inequalities use a generalization of the Beckmann boundary introduced by Ivanisvili, Xie, and Zhang. In particular, we prove that the optimal dimension-free constant in the classical $L^1$-Poincaré inequality on the Hamming cube is $\sqrt{\pi/2}$, answering a question of Ivanisvili et al. Semigroup duality and explicit Malliavin weights reduce these estimates to integrals of first absolute moments of centered Bernoulli and Poisson sums. The key ingredient is a Fourier positivity principle, which shows that replacing the coefficients by their absolute values can only increase these integrals. For the underlying Bernoulli functional at subquadratic exponents, we also characterize the maximizers and determine their exact phase transitions.
Read the full abstract

Signal Processing

Beyond Feasibility: Efficient and Robust On-Device Localization with UWB Concurrent Transmissions

Concurrent transmissions (CTX) enable an ultra-wideband (UWB) tag to range to, or self-localize from, multiple nodes with a single frame reception, significantly improving update rate and energy efficiency compared to traditional approaches. Prior work demonstrated the concept feasibility via prototypes whose real-world applicability is however hindered by many limitations, notably including their inability to directly execute on UWB tags and to cope with interfering UWB signals. We address these limitations in two ways. First, we redesign the core CTX techniques to enable a fast and efficient execution on resource-constrained MCUs without sacrificing accuracy. Second, we leverage the new features in the IEEE 802.15.4z standard to achieve robustness to interference from other UWB devices and enable tags to dynamically select which anchors to use, two crucial and unexplored capabilities in CTX-based systems. We quantitatively evaluate our approach experimentally in two real-world testbeds with different characteristics. We first analyze our new techniques individually, offering an in-depth discussion of design choices and performance trade-offs. Then, we embed all our techniques into a CTX-based system enabling a tag to self-localize entirely on-device, unlocking gains up to 44% in runtime and 85% in energy efficiency compared to traditional approaches.
Read the full abstract

All authors Enrico Soprana, Michael Stocker, Davide Vecchia, Matteo Trobinger, Carlo Alberto Boano, Gian Pietro Picco

Quantum Gases

Quantum-walk measurements with sequential imaging of the same atomic sample

We demonstrate sequential single-atom-resolved imaging of the same ultracold atomic sample and apply this technique to quantum-walk measurements in an optical lattice. In quantum gas microscope experiments, fluorescence imaging provides microscopic information on atomic distributions but induces motional heating through photon scattering, making it difficult to reuse the same sample for subsequent quantum dynamics. To overcome this limitation, we apply Raman sideband cooling with parameters optimized for post-imaging cooling, enabling sequential single-atom-resolved measurements of the same sample. By optimizing the cooling parameters with machine learning, we achieve high fidelity over multiple imaging cycles. To demonstrate the capability of this technique, we measure one-dimensional quantum walks of single atoms in an optical lattice. By determining the initial atomic positions from fluorescence images, we extract quantum-walk transition probabilities without requiring deterministic preparation of the initial atomic positions. The observed dynamics agree with numerical simulations of quantum walks of single atoms. Our method provides a route toward reducing experimental overhead in quantum gas microscope studies requiring repeated acquisition of spatially resolved correlation data.
Read the full abstract

Computer Vision

CRTrack-Cap: A New Benchmark for Cross-View Dense Video Object Captioning

Dense video object captioning (DVOC) localizes, tracks, and describes object trajectories within a video. Existing methods process each camera view independently, producing separate camera-local tracklets and captions for the same person across cameras. These outputs lack a unified description and leave complementary appearance evidence across views unused. We therefore formulate cross-view DVOC (CVDVOC), which associates these tracklets into scene-level global trajectories and generates one caption for each trajectory from synchronized views. To support this task, we construct CRTrack-Cap from CRTrack by auditing text-identity links and converting usable referring expressions into canonical captions. It retains 425 global trajectories and 838 local tracklets across 13 scenes, with canonical captions for 335 identities. We adapt CHOTA to the cross-view setting as CV-CHOTA to jointly evaluate tracking and caption quality. Our reference framework combines View-Exclusive Cross-Camera Association (VECA) for grouping tracklets with Conflict-Evidence Gating (CEG) for integrating caption evidence using its reliability and disagreement across views. Against three single-view methods adapted with MvMHAT for cross-view association, our framework achieves CV-CHOTA scores of 83.43 and 33.48 on the scene-disjoint in-domain and cross-domain tests, respectively. Together, CRTrack-Cap and our framework provide a benchmark and a strong baseline for describing the same person consistently across camera views.
Read the full abstract

All authors Xiangqun Zhang, Likai Wang, Zekun Qian, Rui Zhou, Ruize Han, Wei Feng

Numerical Analysis

SparseModesNet for coupled thermal-hydraulic-mechanical system in heterogeneous porous media

Coupled thermo-hydro-mechanical (THM) models describe interactions among displacement, fluid pressure, and temperature in geothermal systems. Fine-mesh simulations over many time steps are costly, particularly with heterogeneous permeability and thermal conductivity or fracture networks that alter fluid and heat transport. Proper orthogonal decomposition (POD) reduces the number of unknowns but can require many modes because its reconstruction is linear. We apply SparseModesNet to reconstruct coupled THM solutions using a small set of selected modes and a neural-network nonlinear decoder. Starting from a larger collection of candidate modes, SparseModesNet selects modes to reduce reconstruction error rather than simply retaining those with the highest energy. We compare SparseModesNet with POD on three problems: a two-dimensional heterogeneous medium, a two-dimensional connected fracture network, and a three-dimensional geometry based on Utah FORGE fractures. We evaluate reconstructed displacement, pressure, and temperature fields, as well as pressure- and temperature-based quantities of interest. We also test temporal interpolation by training on a subset of snapshots and reconstructing omitted snapshots within the same time interval. For the same reduced dimension, SparseModesNet achieves smaller reconstruction errors than POD and accurately recovers the solution fields and selected geothermal quantities of interest.
Read the full abstract

All authors Kapil Chawla, Tomoki Koike, Sanghyun Lee, Julie Bessac

Game Theory

Contextual Search and Bilateral Trade with a Trace-Square-Root Potential

We present a unified potential method for contextual search and bilateral trade with one-bit feedback and obtain order-tight algorithms for both problems. Our framework starts from a Bayesian perspective in which the unknown target (for contextual search) and the unknown pair of valuations (for bilateral trade) are sampled from a known prior distribution. In that model, the decision maker bases its prediction on the posterior distribution, obtained by conditioning on feedback received in previous rounds. For Bayesian contextual search, we present an algorithm with a total loss bounded by $O(\mathrm{Tr} \sqrt \Sigma)$, where $\Sigma$ is the covariance matrix of the target's prior; and for contextual bilateral trade, our algorithm's regret is bounded by $O(\mathrm{Tr} \sqrt{\Sigma_S}+\mathrm{Tr} \sqrt{\Sigma_B})$, where $\Sigma_S$ and $\Sigma_B$ are the covariance matrices of the seller's and buyer's valuations. We then use ideas from this Bayesian setting and adapt a “dual” of this trace-square-root potential to the classical worst-case setting, where the target and valuations are chosen adversarially from the unit Euclidean ball. For both contextual search and bilateral trade, we obtain tight $O(\sqrt d)$-regret algorithms. These results match the existing lower bounds up to a constant factor and improve on the previous guarantees of $O(d\log d)$ for contextual search and of $O(d6^d)$ for bilateral trade.
Read the full abstract

All authors Romain Cosson, Federico Fusco, Anupam Gupta, Stefano Leonardi, Renato Paes Leme, Matteo Russo

Machine Learning

Know What You Need: Efficient Activation Checkpointing

Large Language Model training is increasingly constrained by memory capacity. Activation Checkpointing reduces activation memory by avoiding storing selected activations during the forward pass and rematerializing them during backpropagation, thereby trading memory savings for additional computation. Although this technique is essential for scaling model training, rematerialization can be prohibitively expensive, particularly for Transformer linear layers. Approximate computation could reduce this overhead, but conventional approaches either introduce errors that accumulate over the backward pass or fail to exploit information available from the initial forward pass. We introduce a novel approximate rematerialization method that can yield unbiased gradient estimates: using information collected during the initial forward pass, the method selectively recomputes only the most important layer components, either through greedy selection or randomized sketching. Ablation studies show that the resulting gradients closely match the exact gradients. When applied to feed-forward layers, attention layers, or both, our method can recompute only two-thirds of the layer components in a Llama-488M model, while increasing perplexity by just 1.7%, 1.2%, and 2.5%, respectively.
Read the full abstract

Algebraic Geometry

Some normal abelian surfaces in $\mathbf{P}^3$

Every complex projective surface is birational to a singular surface $S \subseteq \mathbf{P}^3$, but it is unclear whether one can ask that $S$ have only isolated singularities. In particular, it was a folk question as to whether an abelian surface could admit such a model. We present examples to show that there do exist normal abelian surfaces in $\mathbf{P}^3$. One of these examples deforms in a one-parameter family.
Read the full abstract

Machine Learning (Statistics)

Kernel-Based Learning of Analytic Operators and Parametric PDEs

We study kernel-based learning of analytic operators from finite samples, with a focus on how the regularity of the target should determine the choice of the kernel. We establish embeddings of analytic operators into vector-valued reproducing kernel Hilbert spaces and derive finite-sample, high-probability guarantees that quantify the uncertainty due to random sampling: in the noiseless setting, suitable analytic kernels yield rates exponential in a fractional power of the sample size, while Sobolev-type kernels yield polynomial rates. We apply the theory to the solution operator of a parametric elliptic PDE, where the training data are themselves finite-element approximations, and obtain bounds that separate the sampling and discretization contributions to the error and prescribe how the mesh size should be coupled to the sample size. The results are illustrated numerically.
Read the full abstract

All authors Giovanni S. Alberti, Markus Holzleitner, Lorenzo Rosasco, Matteo Santacesaria

Data Structures and Algorithms

Online Timely Bin Packing Revisited: A Simpler Algorithm with Optimality in Special Cases

Online Timely Bin Packing is a variant of classical online bin packing in which items arrive over time with sizes and deadlines. Packing an item may be deferred until its deadline, but once a collection of items is packed into a bin, the bin is closed and cannot be used again. The objective is to minimize the total number of bins used. The problem was introduced by Khattar in 2023. In 2025, Aminian proposed Discrete-Greedy, a deterministic online algorithm with a competitive ratio of exactly $2$, which remains the best-known upper bound. Its analysis is particularly simple; however, the algorithm requires maintaining a fractional simulation internally and may leave usable space in some bins unfilled. We propose Thrifty-Greedy, a simple deterministic online algorithm that does not require a fractional simulation. We prove that its competitive ratio is exactly $2$, matching the best-known upper bound. At the same time, the algorithm automatically achieves the optimal competitive ratio on two natural classes of inputs: it is $1$-competitive when all items have the same size, and $1.5$-competitive when every item has size in $\{1/3,2/3\}$. In contrast, Discrete-Greedy has a competitive ratio of at least $2$ on both classes.
Read the full abstract

All authors Pei-Lin Chiang, Wing-Kai Hon, Ying-Chia Huang, Li-Tzu Yang

Algebraic Topology

New persistent chain spaces: identifying and computing the relevant persistent Laplacians of a filtration

Persistent homology (PH) summarizes the topological features of data that persist across scales, but it is insensitive to many geometric properties. Persistent Laplacians (PLs) address this limitation: for a pair of simplicial complexes in a filtration, the kernel of the persistent Laplacian recovers the persistent Betti number, while its non-zero spectrum carries additional geometric information. However, PLs are defined for individual pairs of filtration values, so analysing a whole filtration requires evaluating many operators that largely contain the same information, and the result is hard to summarize and interpret. In this paper, we show that this redundancy has a precise structure. For each pair of filtration values, we decompose the space of chains underlying the persistent up-Laplacian into the contributions inherited from the neighboring pairs and a remaining part, which we call the new persistent chain space. We show that the new persistent chain spaces of all earlier pairs form a direct sum decomposition of this space, so that every persistent up-Laplacian in the filtration is determined by them. Moreover, the new persistent chain spaces are non-trivial only at pairs given by the persistence pairing in the same and in the next homological dimension, so that the persistence pairing determines exactly where the persistent up-Laplacian can change. Experiments on several types of point clouds show that evaluating persistent Laplacians only at these pairs reduces the computation time, both for our implementation and for existing software.
Read the full abstract

Artificial Intelligence

Not Your Diary, Just Your Topic: Grading the Wrong-User Control

A common test of personalization replaces a user's history with someone else's and measures the drop in accuracy. But this swap also changes the topics in the history, so the drop may reflect topic mismatch. We examine this problem with a frozen language model on LaMP-2 news categorization, using histories from similar users and synthetic histories assembled from other users' articles. The user's own history improves accuracy by 2.3 percentage points over a model trained without history. Removing history only at evaluation makes this gain appear larger, at 4.8 points. Category labels alone reproduce 86% of the accuracy gap between the user's own history and a random user's history. With synthetic histories matched on labels and text similarity, the gap falls to 0.9 points. Across two user splits and three training seeds, the advantage over a similar user's history is 3.2 points for users seen during training and 0.65 points for unseen users, where the evidence for an advantage remains inconclusive. A small movie-tagging experiment with one training seed also finds no advantage over other users' histories. These findings show why personalization should be evaluated on unseen users, with histories of similar content and a model trained without history as comparisons. We also find that adding gated cross-attention improves on a prefix-only model under our training setup; whether this holds more broadly remains open.
Read the full abstract

All authors Karina Romanova, Igor Kiselev, Vladimir Platonov, Viliana Devbunova, Gregory Leleytner, Adeliya Leleytner, Lina Veltman, Aleksandr Nikolich

Rings and Algebras

Topological Mundici Adjunction and Its Applications

We establish a canonical topological lifting of Mundici equivalence, representing every topological MV-algebra as the unit interval of a topological $\ell$-group and extending the correspondence to a categorical adjunction. As applications, we give a direct group-theoretic proof of the equivalence of $T_0$ and $T_{3\frac12}$ and show that every topological MV-algebra is a retral space. We further extend the lifting to topological Riesz MV-algebras and identify the canonical topology with Wiweger's mixed topology under suitable assumptions, recovering the strict topology as a special case. Finally, we obtain a unit-interval formulation of Koopman representations for continuous-time dynamical systems.
Read the full abstract

Cryptography and Security

ReplayAudit: Auditing Outsourced LLM Fine-Tuning via Sealed Records and Selective Replay

Auditing outsourced LLM fine-tuning requires checking an untrusted provider's training while controlling verification cost. Uniform sampling ignores training-time signals, while checking a suspicious record can require replaying several preceding steps. We present ReplayAudit, a telemetry-guided auditing system that seals training records before challenges and independently replays sampled intervals from sealed checkpoints. Loss changes and parameter-update norms guide stratified random sampling; replay determines whether the selected computation agrees with its records. We relate sampling probabilities to replay work and evaluate the system on real LoRA training. A 7B BF16 prototype rejects a false update-norm report during a random audit; targeted checks also reject false loss reports and skipped updates. We calibrate sampling budgets offline to at least a 90% probability of directly selecting an injected step. For two learning-rate deviation settings, executing these budgets on the same honest 1.5B trajectory reduces median complete-audit wall time by 49% and 67% relative to uniform sampling. When skipped updates evade the high-anomaly stratum, the corresponding audit instead costs 38% more. These results show how deviation capture and actual replay work jointly determine the cost of telemetry-guided auditing.
Read the full abstract

General Relativity and Quantum Cosmology

Quasinormal modes of Kerr–Newman–de Sitter black hole: Linear stability and strong cosmic censorship conjecture

We present an alternative formulation of coupled gravitational and electromagnetic perturbations of Kerr–Newman–de Sitter black holes, yielding decoupled radial and angular equations with a well-defined $\Lambda\rightarrow 0$ limit. We compute the quasinormal mode spectra using spectral and direct integration methods, finding excellent agreement. All of these modes are consistent with the linear stability of the Kerr–Newman–de Sitter black hole. We investigate the photon sphere and near-extremal modes, revealing a non-trivial interplay between rotation and electric charge. We further explore the implications for the strong cosmic censorship conjecture and find that the photon sphere modes predominantly determine its validity within the parameter space considered. While increasing the rotation of the black hole generally favours the strong cosmic censorship conjecture, increasing the electric charge can promote its violation. We also find that the region of parameter space exhibiting a violation of the conjecture is comparatively smaller for gravitational perturbations than for electromagnetic perturbations. These results highlight the important role of rotation, electric charge, and the nature of perturbations in determining the validity of the strong cosmic censorship conjecture in Kerr–Newman–de Sitter spacetimes.
Read the full abstract

All authors Subhodeep Sarkar, Suvikranth Gera, Avijit Chowdhury, Chiranjeeb Singha, Sreelakshmi Mundakkal Dasan, Sumanta Chakraborty