Searcharxiv⌕ Search

arXiv subjects

Ivan Novikov

Publications and source records attributed to Ivan Novikov.

At least 19 recordsLinked to original sources
arXiv2026

Strategies in POMDPs with Stage Duration

Partially observable Markov decision processes (POMDPs) with stage duration provide a framework for approximating continuous-time behavior by scaling transition probabilities with a stage duration parameter $h \in (0,1]$. While previous literature has primarily focused on the limit of the discounted value as the stage duration $h$ vanishes, this paper investigates the global behavior of the asymptotic value, $V(h)$, across varying stage durations. Our main result demonstrates that any strategy in a POMDP with stage duration $h$ can be mimicked in the base POMDP ($h=1$). Specifically, we provide an explicit construction showing that for any strategy in the POMDP with stage duration $h$, there exists a strategy in the base POMDP that secures the same asymptotic payoff. As a consequence of this theorem, we establish that the value function $V(h)$ is nondecreasing with respect to $h$, and that the continuous-time limit $\lim_{h \to 0} V(h)$ exists.

math.OC↗
arXiv2026

Global Optimization of Atomic Clusters via Physically-Constrained Tensor Train Decomposition

The global optimization of atomic clusters represents a fundamental challenge in computational chemistry and materials science due to the exponential growth of local minima with system size (i.e., the curse of dimensionality). We introduce a novel framework that overcomes this limitation by exploiting the low-rank structure of potential energy surfaces through Tensor Train (TT) decomposition. Our approach combines two complementary TT-based strategies: the algebraic TTOpt method, which utilizes maximum volume sampling, and the probabilistic PROTES method, which employs generative sampling. A key innovation is the development of physically-constrained encoding schemes that incorporate molecular constraints directly into the discretization process. We demonstrate the efficacy of our method by identifying global minima of Lennard-Jones clusters containing up to 45 atoms. Furthermore, we establish its practical applicability to real-world systems by optimizing 20-atom carbon clusters using a machine-learned Moment Tensor Potential, achieving geometries consistent with quantum-accurate simulations. This work establishes TT-decomposition as a powerful tool for molecular structure prediction and provides a general framework adaptable to a wide range of high-dimensional optimization problems in computational material science.

math.OC↗
arXiv2026

On the reconstruction of kinematic distributions computed with Monte Carlo methods using orthogonal basis functions

Reconstruction of one-dimensional kinematic distributions from calculations based on high-dimensional Monte-Carlo integration is a standard problem in high-energy physics. Traditionally, this is done by collecting randomly-generated events in histograms. In this article, we explore an alternative approach, whose main idea is to approximate the target distribution by a weighted sum of orthogonal basis functions whose coefficients are calculated using the Monte-Carlo integration. This method has the advantage of directly yielding smooth approximations to target distributions. Furthermore, in the context of high-order perturbative calculations with local subtractions, it eliminates the so-called bin-to-bin fluctuations, which often severely affect the quality of conventional histograms. We also demonstrate that the availability of a high-quality approximation to the target distribution, for example the leading-order result in the perturbative expansion, can be exploited to construct an optimized orthonormal basis. We compare the performance of this method to conventional histograms in both toy-model and real Monte-Carlo settings, applying it to Higgs boson production in weak boson fusion as an example.

hep-ph↗
arXiv2025

MLPMoE: Zero-Shot Architectural Metamorphosis of Dense LLM MLPs into Static Mixture-of-Experts

Large Language Models (LLMs) are predominantly deployed as dense transformers, where every parameter in every feed-forward block is activated for every token. While architecturally simple, this is computationally inefficient, since inference costs scale linearly with parameter count. Recent upcycling methods such as MoEfication, CMoE, ToMoE, and MoORE reveal that much of the useful computation lives in sparse, semi-modular substructures inside dense feed-forward networks, but these approaches typically rely on clustering, activation profiling, singular value decomposition, or custom routing that requires calibration data. This paper introduces MLPMoE (MLP Mixture-of-Experts), a training-free, deterministic transformation that restructures the dense MLP in transformer blocks into a static, high-cardinality mixture of experts. The transformation uses simple tensor slicing and summation, reinterpreting the algebra of tensor parallelism as a topological conversion rather than a distributed training pattern. We further introduce Fractal Fade (differential branch sparsity) and Compensated Pruning (variance-preserving branch reduction) as lightweight mechanisms for structured sparsity. On Qwen2.5-0.5B-Instruct and DeepSeek-R1-Distill-Llama-8B, the zero-shot MLPMoE transform changes a proxy perplexity metric by less than 0.05 percent while keeping the parameter count effectively constant. On the 8B model, differential sparsity removes about 20 percent of MLP parameters while keeping perplexity within about 2 percent of the dense baseline. The method operates entirely post hoc on existing checkpoints and does not require gradients, calibration sets, or router training. Code is available at https://gist.github.com/iwallarm/fc2ef1eddf226ca7814f9e5e2ae9bad1

cs.LG↗
arXiv2025

A2AS: Agentic AI Runtime Security and Self-Defense

The A2AS framework is introduced as a security layer for AI agents and LLM-powered applications, similar to how HTTPS secures HTTP. A2AS enforces certified behavior, activates model self-defense, and ensures context window integrity. It defines security boundaries, authenticates prompts, applies security rules and custom policies, and controls agentic behavior, enabling a defense-in-depth strategy. The A2AS framework avoids latency overhead, external dependencies, architectural changes, model retraining, and operational complexity. The BASIC security model is introduced as the A2AS foundation: (B) Behavior certificates enable behavior enforcement, (A) Authenticated prompts enable context window integrity, (S) Security boundaries enable untrusted input isolation, (I) In-context defenses enable secure model reasoning, (C) Codified policies enable application-specific rules. This first paper in the series introduces the BASIC security model and the A2AS framework, exploring their potential toward establishing the A2AS industry standard.

cs.CR↗
arXiv2025

Parton-shower and fixed-order QCD effects in Higgs-boson production in weak-boson fusion and its decays to bottom quarks

Recently, it was observed [arXiv:2407.09363] that an aggressive cut on the $b$-jets' transverse momenta applied to Higgs-boson production in weak-boson fusion followed by the decay $H \to b \bar b$, leads to very large QCD corrections to the fiducial cross section. In this paper we show that these corrections are caused by soft and collinear QCD radiation and, therefore, can be efficiently treated by a parton shower. We combine the parton-shower description of the decay $H \to b \bar b$ with NNLO QCD corrections to Higgs production in weak-boson fusion and its subsequent decay, and demonstrate that the quality of the theoretical prediction is markedly improved even if $b$-jets with rather high transverse momenta are selected. The remaining uncertainty of the theoretical prediction, mainly driven by imprecise modelling of $H \to b \bar b$ decay, is estimated to be of the order of $\mathcal{O}(5{-}7\%)$.

hep-ph↗
arXiv2024

Enhanced Parking Perception by Multi-Task Fisheye Cross-view Transformers

Current parking area perception algorithms primarily focus on detecting vacant slots within a limited range, relying on error-prone homographic projection for both labeling and inference. However, recent advancements in Advanced Driver Assistance System (ADAS) require interaction with end-users through comprehensive and intelligent Human-Machine Interfaces (HMIs). These interfaces should present a complete perception of the parking area going from distinguishing vacant slots' entry lines to the orientation of other parked vehicles. This paper introduces Multi-Task Fisheye Cross View Transformers (MT F-CVT), which leverages features from a four-camera fisheye Surround-view Camera System (SVCS) with multihead attentions to create a detailed Bird-Eye View (BEV) grid feature map. Features are processed by both a segmentation decoder and a Polygon-Yolo based object detection decoder for parking slots and vehicles. Trained on data labeled using LiDAR, MT F-CVT positions objects within a 25m x 25m real open-road scenes with an average error of only 20 cm. Our larger model achieves an F-1 score of 0.89. Moreover the smaller model operates at 16 fps on an Nvidia Jetson Orin embedded board, with similar detection results to the larger one. MT F-CVT demonstrates robust generalization capability across different vehicles and camera rig configurations. A demo video from an unseen vehicle and camera rig is available at: https://streamable.com/jjw54x.

cs.CV↗
arXiv2024

Limit Value in Zero-Sum Stochastic Games with Vanishing Stage Duration and Public Signals

We consider the behaviour of $\lambda$-discounted zero-sum games as the discount factor $\lambda$ approaches 0 (that is, the players are more and more patient), in the context of games with stage duration. In stochastic games with stage duration h, players act at times 0, h, 2h, and so on. The payoff and leaving probabilities are proportional to h. When h tends to 0, such discrete-time games approximate games played in continuous time. The asymptotic behavior of the values (when both $\lambda$ and h tend to 0) was already studied in the case of stochastic games with perfect observation of the state and in the state-blind case.We consider the same question for the case of stochastic games with imperfect observation of the state. More precisely, we consider a particular case of such games, stochastic games with public signals, in which players are given at each stage a public signal that depends only on the current state. Our main result states that there exists a stochastic game with public signals, with no limit value (as the discount factor $\lambda$ goes to 0) if stage duration is 1, but with a limit value when stage duration h and discount factor $\lambda$ both tend to 0. Informally speaking, it means that the limit value in discrete time does not exist, but the limit value in continuous time (i.e. when h approaches 0) exists. Such a situation is impossible in the case of stochastic games with perfect observation of the state.

math.OC↗
arXiv2024

QCD corrections to Higgs boson production and $H \to b \bar{b}$ decay in weak boson fusion

We study QCD corrections to the process where a Higgs boson is produced in weak boson fusion and then decays into a pair of massive $b$ quarks. We find that typical experimental criteria used to identify $b$ jets in this process affect QCD corrections to the decay, making it necessary to account for them in the proper description of this process. Indeed, if corrections to the production and decay are combined, the fiducial cross section of the weak boson fusion process $p p \to H(\to b \bar{b}) j j$ is reduced by about $40\%$ relative to leading-order predictions, compared to just about $8\%$ if only corrections to the production process are considered. We investigate the origin of these large corrections through next-to-next-to-leading-order and conclude that they appear because a number of independent moderately-large effects conspire to significantly reduce the fiducial cross section for this process.

hep-ph↗
arXiv2024

Asymptotic Value in Zero-Sum Stochastic Games with Vanishing Stage Duration and Public Signals

We study $\lambda$-discounted zero-sum games as the discount factor $\lambda$ approaches $0$ (that is, the players are more and more patient), in the context of games with stage duration. In stochastic games with stage duration $h$, players act at times $0, h, 2h$, and so on. The payoff and leaving probabilities are proportional to $h$. When $h$ tends to $0$, such discrete-time games approximate games played in continuous time. The asymptotic behavior of the values (when both $\lambda$ and $h$ tend to $0$) has already been studied for stochastic games with full state observation and for state-blind games. We consider the same question for the case of stochastic games with deterministic public signals on the state. We construct a stochastic game with public signals, with no asymptotic value (as the discount factor $\lambda$ goes to $0$) if the stage duration is $1$, but with an asymptotic value when the stage duration $h$ and the discount factor $\lambda$ both tend to $0$. Informally, this means that the asymptotic value in discrete time does not exist, whereas it does exist in continuous time. This situation cannot occur in stochastic games with full state observation.

math.OC↗
arXiv2024

Zero-Sum State-Blind Stochastic Games with Vanishing Stage Duration

In stochastic games with stage duration h, players act at times 0, h, 2h, and so on. The payoff and leaving probabilities are proportional to h. As h approaches 0, such discrete-time games approximate games played in continuous time. The behavior of the values when h tends to 0 was already studied in the case of stochastic games with perfect observation of the state. We examine the same question for the case of state-blind stochastic games. Our main finding is that, as h approaches 0, the value of any state-blind stochastic game with stage duration h converges to the unique viscosity solution of a partial differential equation.

math.OC↗
arXiv2023

MLIP-3: Active learning on atomic environments with Moment Tensor Potentials

Nowadays, academic research relies not only on sharing with the academic community the scientific results obtained by research groups while studying certain phenomena, but also on sharing computer codes developed within the community. In the field of atomistic modeling these were software packages for classical atomistic modeling, later -- quantum-mechanical modeling, and now with the fast growth of the field of machine-learning potentials, the packages implementing such potentials. In this paper we present the MLIP-3 package for constructing moment tensor potentials and performing their active training. This package builds on the MLIP-2 package (Novikov et al. (2020), The MLIP package: moment tensor potentials with MPI and active learning. Machine Learning: Science and Technology, 2(2), 025002.), however with a number of improvements, including active learning on atomic neighborhoods of a possibly large atomistic simulation.

physics.comp-ph↗
arXiv2022

The photon energy spectrum in $B\to X_s\gamma$ at N$^3$LL$'$

We present predictions for the photon energy spectrum in inclusive $B\to X_s\gamma$ decays mediated by the electromagnetic penguin operator $O_7$ to N$^3$LL$'$. We use soft-collinear effective theory (SCET) to resum the singular contributions in the peak region at large photon energy. In the tail region the resummed predictions are matched to fixed order at N$^3$LO, where we include the known fixed-order contributions for $O_7$ up to $\mathcal{O}(\alpha_s^2)$. We develop a method to suitably parametrize the still unknown $\mathcal{O}(\alpha_s^3)$ nonsingular corrections in terms of theory nuisance parameters, whose variations provide an estimate of the associated theory uncertainty. In this context, we also study different ways to treat higher-order cross terms in the matching. Another important aspect of our analysis is the short-distance scheme used for the $b$-quark mass $m_b$. We find that in the present context, the 1$S$ mass scheme, which was previously used up to 2-loop order, fails to work at 3-loop order, because the mass scheme enters at a soft scale much smaller than $m_b$ here, for which the 1$S$ scheme was not devised. Using instead the MSR mass scheme with $R\sim 1\,\mathrm{GeV}$, we obtain stable results with good perturbative convergence up to N$^3$LL$'$.

hep-ph↗
arXiv2022

AI-accelerated Materials Informatics Method for the Discovery of Ductile Alloys

In computational materials science, a common means for predicting macroscopic (e.g., mechanical) properties of an alloy is to define a model using combinations of descriptors that depend on some material properties (elastic constants, misfit volumes, etc.), representative for the macroscopic behavior. The material properties are usually computed using special quasi-random structures (SQSs), in tandem with density functional theory (DFT). However, DFT scales cubically with the number of atoms and is thus impractical for a screening over many alloy compositions. Here, we present a novel methodology which combines modeling approaches and machine-learning interatomic potentials. Machine-learning interatomic potentials are orders of magnitude faster than DFT, while achieving similar accuracy, allowing for a predictive and tractable high-throughput screening over the whole alloy space. The proposed methodology is illustrated by predicting the room temperature ductility of the medium-entropy alloy Mo-Nb-Ta.

cond-mat.mtrl-sci↗
arXiv2021

Magnetic Moment Tensor Potentials for collinear spin-polarized materials reproduce different magnetic states of bcc Fe

We present the magnetic Moment Tensor Potentials (mMTPs), a class of machine-learning interatomic potentials, accurately reproducing both vibrational and magnetic degrees of freedom as provided, e.g., from first-principles calculations. The accuracy is achieved by a two-step minimization scheme that coarse-grains the atomic and the spin space. The performance of the mMTPs is demonstrated for the prototype magnetic system bcc iron, with applications to phonon calculations for different magnetic states, and molecular dynamics simulations with fluctuating magnetic moments.

physics.atom-ph↗
arXiv2020

Parton Distribution Functions of the Charged Pion Within The xFitter Framework

We present the first open-source analysis of parton distribution functions (PDFs) of charged pions using xFitter, an open-source QCD fit framework to facilitate PDF extraction and analyses. Our calculations are implemented at next-to-leading order (NLO) using APPLgrids generated by MCFM generator. Using currently available Drell-Yan and photon production data, we find the valence distribution is well constrained; however, the considered data are not sensitive enough to unambiguously determine sea and gluon distributions. Fractions of momentum carried by the valence, sea and gluon components are discussed, and we compare with the results of JAM collaboration and the GRV group.

hep-ph↗
arXiv2019

Percolation of three fluids on a honeycomb lattice

In this paper, we consider a generalization of percolation: percolation of three related fluids on a honeycomb lattice. K. Izyurov and A. Magazinov proved that percolations of distinct fluids between opposite sides on a fixed hexagon become mutually independent as the lattice step tends to 0. This paper exposes this proof in details (with minor simplifications) for nonspecialists. In addition, we state a few related conjectures based on numerical experiments.

math.PR↗