arXiv subjects

Niraj Kumar

Publications and source records attributed to Niraj Kumar.

At least 19 recordsLinked to original sources

arXiv2026

Functional Tensor Trains for High-Order Greeks and Exposure Tails of Multi-Asset Structured Products

Risk management of worst-of autocallables needs price sensitivities (Greeks) up to third order and the potential future exposure, a tail quantile of the note's future value. In Monte Carlo, each order of differentiation multiplies the sampling noise by a factor that grows with the observation frequency, and nested simulation turns the noise of its inner paths into a bias of the quantile. We instead solve the backward pricing recursion, storing the continuation value as an $hp$ functional tensor train (hp-FTT) whose piecewise-polynomial cores are split at the barriers, so that worst-of barrier events act exactly, with rank one. For Gaussian transitions we show that the polynomial degree sufficient for an error $\varepsilon$ grows only as $\sqrt{\log(1/\varepsilon)}$. One solve then gives every Greek at any spot without sampling noise, by integrating the likelihood-ratio weight against the stored value, and the values it stores at future dates replace inner simulation. Against the strongest Monte Carlo estimators we could build or find published, on worst-of autocallables observed monthly on three to five assets, the order of the Greek and the accuracy decide which method is cheaper. At $1\%$ accuracy Monte Carlo is more than an order of magnitude cheaper for deltas, second-order Greeks go either way, and hp-FTT is cheaper for every third-order Greek, by up to two orders of magnitude. At $0.1\%$ hp-FTT is cheaper for every Greek we could check, under geometric Brownian motion and local volatility alike. For the exposure alone Monte Carlo is cheaper at $1\%$ and hp-FTT at $0.1\%$, and a solve run for the Greeks yields it for little more than scenario sampling. Sampling pays in variance for each derivative and hp-FTT in rank for each asset, so Monte Carlo keeps low-order Greeks at percent accuracy, rarely observed contracts and many assets.

q-fin.CP↗

arXiv2026

From Quantum Tunneling to Classical Optimization through Instantons

We use an instanton-based analysis to show that Path Integral Stochastic Gradient Descent (PISGD), a classical optimization algorithm instantiated from path integral Monte Carlo (PIMC), can inherit the quantum tunneling mechanism. Quantum tunneling is the phenomenon in which a quantum particle passes through a potential barrier that a classical trajectory cannot cross in the same way. This mechanism can contribute to the performance of quantum optimization algorithms, since a common bottleneck in optimization is that the solution becomes trapped in a local minimum, separated from better solutions by a potential barrier. Classical noise-driven optimizers, such as stochastic gradient descent (SGD), have escape times that scale exponentially with the potential barrier height. For quantum tunneling, the escape time is instead exponential in the bounce action, which equals twice the Agmon distance between the two basins. When the barrier is tall but thin, quantum tunneling can allow escaping local minima faster than such classical methods. Remarkably, this scaling is not exclusive to algorithms with genuine quantum dynamics. For PISGD applied to classical nonconvex global optimization, we adapt Kramers' escape-rate theory to path space and obtain the same exponential scaling as quantum tunneling, in terms of the bounce action rather than the barrier height. We confirm this scaling numerically by implementing PISGD for both separable and nonseparable potentials. That said, PISGD retains only incoherent tunneling, since coherence among multiple tunneling paths is absent. Our analysis therefore does not rule out a potential quantum advantage for quantum-tunneling-based algorithms over classical ones in general.

quant-ph↗

arXiv2026

Anytime Training with Schedule-Free Spectral Optimization

Standard neural network training relies on learning-rate schedules tied to a fixed horizon, leading to strong path dependence and costly re-tuning as data availability changes. Schedule-Free (SF) methods address this by removing explicit schedules, yet SF-AdamW, the current state-of-the-art anytime optimizer, consistently underperforms well-tuned AdamW baselines. We propose SF-NorMuon, a schedule-free spectral optimizer that closes this gap: with a single hyperparameter configuration, SF-NorMuon matches or exceeds tuned AdamW on 125M and 772M parameter language models across $1$--$8\times$ Chinchilla horizons, and trails horizon-aware, cosine-scheduled NorMuon by only $\sim 0.03$ nats, substantially narrowing the gap between anytime and scheduled spectral optimizers. On the theoretical side, we prove a stationarity guarantee for schedule-free spectral dynamics and identify weight decay at the fast iterate as essential for long-horizon stability. SF-NorMuon enables practitioners to obtain high-quality checkpoints at any point during training without committing to a horizon in advance. Fully matching horizon-aware spectral optimizers without a schedule remains an open problem, which we highlight as a direction for future work. By closing the performance gap with tuned AdamW baselines, SF-NorMuon makes horizon-free optimization more practical, taking a step towards truly open-ended, continual learning.

cs.LG↗

arXiv2026

Crushing the Evidence: A Dual-Penalty Evasion Framework for Fooling White-Box Explainable AI Auditors

Post-hoc model explainers such as LIME, SHAP, and Integrated Gradients are widely deployed to audit models in high-stakes sensitive domains, including finance, healthcare, and social welfare. This ensures the model's transparency and acceptability. However, a few studies have examined potential attacks in the explainability pipeline. Adversaries can attempt to conceal algorithmic biases or backdoors using adversarial explanation attacks. These attacks have relied on scaffolding out-of-distribution (OOD) detectors that toggle predictions when queried by an explainer. Consequently, defenses have been developed to successfully neutralize these black-box attacks by identifying their anomalous perturbation footprints. In this paper, we demonstrate a critical vulnerability by introducing a more potent white-box, gradient-regularized evasion attack framework. By employing a continuous-embedding dual-penalty framework, we directly penalize trigger feature gradients during training on in-distribution data. Since our approach embeds the evasion logic natively into the model parameters, without relying on OOD scaffolding wrappers, it generates smooth, in-distribution predictions that leave no anomaly footprint. Empirical evaluations across four benchmark tabular datasets (COMPAS, German Credit, IEEE-CIS, and Communities & Crime) confirm that our method systematically crushes target feature attribution to near-zero (<0.02), maintains >90% Attack Success Rates, and fundamentally bypasses Conditional Anomaly Detection.

cs.LG↗

arXiv2026

Quantum-Informed Portfolio Selection: An End-to-End Pipeline Validated on Trapped-Ion Hardware with Real Market Data

Portfolio diversification - a cornerstone of modern investment management - can be formulated as a Maximum Independent Set (MIS) problem on asset correlation graphs. Solving this problem at scale is computationally challenging, motivating the exploration of quantum algorithms for practical financial optimization. We propose an end-to-end pipeline leveraging qReduMIS, a recursive hybrid quantum-classical algorithm. Rather than using quantum optimization to directly produce a final solution, qReduMIS leverages independent set measurements from the Quantum Approximate Optimization Algorithm (QAOA) to identify frozen nodes - vertices likely to belong to optimal solutions - thereby guiding and unblocking subsequent (provably optimal) classical reductions on the remaining graph. We benchmark qReduMIS on real financial data from four major market indices with up to 225 assets, executing experiments on Quantinuum's 98-qubit trapped-ion Helios system, with QAOA circuits acting on kernels of up to 78 qubits and 1016 two-qubit gates. While standalone QAOA fails to find the optimal solution for two of the largest indices (S&P 100 and Nikkei 225), qReduMIS achieves success probabilities of $0.40$ and $0.95$, respectively, with average approximation ratios $\geq 0.96$ across all four indices. We perform a systematic benchmark on the Quantinuum H2-1 noisy emulator over 73 asset correlation graphs of varying size showing that, for $p=2$ QAOA layers, the optimal time-to-solution scaling exponent of qReduMIS is $3.2$ times smaller than that of standalone QAOA.

quant-ph↗

arXiv2026

Wold-type decomposition for doubly twisted left-invertible covariant representations

In this article, we have introduced the notion of a near-isometric covariant representation of a $C^*$-correspondence. The other objective is to provide a unified approach to several known results for a large class of left-invertible covariant representations of a product system and prove Wold-type decomposition for the case of doubly twisted left-invertible covariant representations and study some applications.

math.OA↗

arXiv2026

Adaptive and Robust Watermark for Generative Tabular Data

In recent years, watermarking generative tabular data has become a prominent framework to protect against the misuse of synthetic data. However, while most prior work in watermarking methods for tabular data demonstrate a wide variety of desirable properties (e.g., high fidelity, detectability, robustness), the findings often emphasize empirical guarantees against common oblivious and adversarial attacks. In this paper, we study a flexible and robust watermarking algorithm for generative tabular data. Specifically, we demonstrate theoretical guarantees on the performance of the algorithm on metrics like fidelity, detectability, robustness, and hardness of decoding. The proof techniques introduced in this work may be of independent interest and may find applicability in other areas of machine learning. Finally, we validate our theoretical findings on synthetic and real-world tabular datasets.

cs.CR↗

arXiv2026

MetaTT: A Global Tensor-Train Adapter for Parameter-Efficient Fine-Tuning

We present MetaTT, a Tensor Train (TT) adapter framework for fine-tuning of pre-trained transformers. MetaTT enables flexible and parameter-efficient model adaptation by using a single shared TT to factorize transformer sub-modules. This factorization indexes key structural dimensions, including layer and matrix type, and can optionally incorporate heads and tasks. This design allows MetaTT's parameter count to scale with the sum, rather than the product, of the modes, resulting in a substantially more compact adapter. Our benchmarks compare MetaTT with LoRA along with recent state-of-the-art matrix and tensor decomposition based fine-tuning methods. We observe that when tested on single-task standard language modeling benchmarks, MetaTT achieves competitive parameter efficiency to accuracy tradeoff. We further demonstrate that MetaTT performs competitively when compared to state-of-the-art methods on multi-task learning. Finally, we leverage the TT decomposition to design a rank adaptive optimizer inspired by the DMRG method from many-body physics. Our results demonstrate that integrating this approach with AdamW enhances optimization performance for a specified target rank.

cs.LG↗

arXiv2026

CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching

Prior-fitted networks (PFNs) are a promising class of tabular foundation models that perform in-context learning, whereby the entire labelled training set is supplied as context, and predictions for test queries are produced in a single forward pass. However, the quadratically scaling self-attention mechanism in many PFN architectures makes inference prohibitive for very large training datasets. We propose CRUMB (Clustered Retrieval Using Minimised-MMD Batching), a three-stage inference wrapper that (i) clusters the test queries, (ii) selects a small, distributionally matched training subset for each cluster by greedily minimising the maximum mean discrepancy (MMD), and (iii) runs exact PFN inference on each reduced-context batch. CRUMB is architecture-agnostic and requires no retraining. On the 51-dataset TabArena benchmark, evaluated across three PFN architectures (TabPFNv2, TabICLv1, TabICLv2), we show that CRUMB outperforms similar state-of-the-art context selection strategies. We also show that CRUMB is resilient to covariate drift, as the MMD-minimisation step naturally helps align the training context distribution to match the current test batch distributions.

cs.LG↗

arXiv2026

Experimental demonstration of quantum advantage in communication complexity for Euclidean distance problem

When considering the complexity of communication protocols, the aim is to perform a certain task with the minimum amount of communication resources, such as time and transmitted information. The use of quantum states may lead to an exponential advantage in the use of such resources. Here, we are interested in the task of calculating the Euclidean distance between two vectors representing real data sets. It has been previously shown that it is possible to obtain an advantage for this task based on quantum fingerprinting. This protocol is defined in the simultaneous message passing model of communication complexity, where the two parties do not communicate with each other but send data to a third party, and exploits practical fingerprints generated using trains of coherent state pulses instead of highly entangled qubit states that are hard to generate for large input sizes needed to demonstrate an exponential advantage. We perform a proof-of-principle experimental demonstration of the Euclidean distance protocol using amplitude modulation techniques for encoding non-binary data sets and high-performance superconducting nanowire single-photon detectors required to increase the accessible input size. We show a quantum advantage in transmitted information surpassing the best classical protocol for an input size of $10^8$, for diverse types of data sets, including those corresponding to real grayscale images, and with reasonable precision and error bounds. Our results highlight the potential of quantum communication complexity for use in a broad set of applications.

quant-ph↗

arXiv2026

Entropy Distribution as a Fingerprint for Hallucinations in Generative Models

Large Language Models (LLMs) often generate factually incorrect outputs, commonly termed hallucinations, that undermine trust and limit deployment in high-stakes settings. Existing hallucination detection methods typically require multiple forward passes, or access to model internals. In this work, we provide theoretical background and empirical evidence that the distribution of token-level entropies, beyond the mean captured by perplexity or length-normalised entropy, serves as a fingerprint of hallucination, with distributional shape and tail behaviour carrying independent signal. We formalize hallucination detection as a statistical hypothesis test and propose the Calibrated Entropy Score (CES), a lightweight algorithm requiring only a single forward pass and black-box access to token logits. CES combines the mean signal with the maximum signal of the generated entropy through a calibrated reference CDF, producing scores that are directly comparable across models and tasks. We establish finite-sample calibration guarantees via a novel random-length Dvoretzky--Kiefer--Wolfowitz inequality, and also prove that CES detects hallucinations with probability converging to one exponentially fast in the generation length. Across eight QA benchmarks and ten generator models spanning open-source and API access models, CES achieves the highest detection performance among all single-pass black-box methods while providing formal error guarantees that existing heuristics lack. Remarkably, CES is statistically indistinguishable from multi-sample methods that require far greater computational cost, closing the gap between lightweight and expensive detection and making it suitable for real-time, large-scale deployment.

cs.AI↗

arXiv2025

A Unified Framework for Provably Efficient Algorithms to Estimate Shapley Values

Shapley values have emerged as a critical tool for explaining which features impact the decisions made by machine learning models. However, computing exact Shapley values is difficult, generally requiring an exponential (in the feature dimension) number of model evaluations. To address this, many model-agnostic randomized estimators have been developed, the most influential and widely used being the KernelSHAP method (Lundberg & Lee, 2017). While related estimators such as unbiased KernelSHAP (Covert & Lee, 2021) and LeverageSHAP (Musco & Witter, 2025) are known to satisfy theoretical guarantees, bounds for KernelSHAP have remained elusive. We describe a broad and unified framework that encompasses KernelSHAP and related estimators constructed using both with and without replacement sampling strategies. We then prove strong non-asymptotic theoretical guarantees that apply to all estimators from our framework. This provides, to the best of our knowledge, the first theoretical guarantees for KernelSHAP and sheds further light on tradeoffs between existing estimators. Through comprehensive benchmarking on small and medium dimensional datasets for Decision-Tree models, we validate our approach against exact Shapley values, consistently achieving low mean squared error with modest sample sizes. Furthermore, we make specific implementation improvements to enable scalability of our methods to high-dimensional datasets. Our methods, tested on datasets such MNIST and CIFAR10, provide consistently better results compared to the KernelSHAP library.

cs.LG↗

arXiv2025

Realization of a Quantum Streaming Algorithm on Long-lived Trapped-ion Qubits

Large classical datasets are often processed in the streaming model, with data arriving one item at a time. In this model, quantum algorithms have been shown to offer an unconditional exponential advantage in space. However, experimentally implementing such streaming algorithms requires qubits that remain coherent while interacting with an external data stream. In this work, we realize such a data-streaming model using Quantinuum Helios trapped-ion quantum computer with long-lived qubits that communicate with an external server. We implement a quantum pair sketch, which is the primitive underlying many quantum streaming algorithms, and use it to solve Hidden Matching, a problem known to exhibit a theoretical exponential quantum advantage in space. Furthermore, we compile the quantum streaming algorithm to fault-tolerant quantum architectures based on surface and bivariate bicycle codes and show that the quantum space advantage persists even with the overheads of fault-tolerance.

quant-ph↗

arXiv2025

Provably faster randomized and quantum algorithms for $k$-means clustering via uniform sampling

The $k$-means algorithm (Lloyd's algorithm) is a widely used method for clustering unlabeled data. A key bottleneck of the $k$-means algorithm is that each iteration requires time linear in the number of data points, which can be expensive in big data applications. This was improved in recent works proposing quantum and quantum-inspired classical algorithms to approximate the $k$-means algorithm locally, in time depending only logarithmically on the number of data points (along with data dependent parameters) [q-means: A quantum algorithm for unsupervised machine learning, Kerenidis, Landman, Luongo, and Prakash, NeurIPS 2019; Do you know what $q$-means?, Cornelissen, Doriguello, Luongo, Tang, QTML 2025]. In this work, we describe a simple randomized mini-batch $k$-means algorithm and a quantum algorithm inspired by the classical algorithm. We demonstrate that the worst case guarantees of these algorithms can significantly improve upon the bounds for algorithms in prior work. Our improvements are due to a careful use of uniform sampling, which preserves certain symmetries of the $k$-means problem that are not preserved in previous algorithms that use data norm-based sampling.

quant-ph↗

arXiv2025

A simple analysis of a quantum-inspired algorithm for solving low-rank linear systems

We describe and analyze a simple algorithm for sampling from the solution $\mathbf{x}^* := \mathbf{A}^+\mathbf{b}$ to a linear system $\mathbf{A}\mathbf{x} = \mathbf{b}$. We assume access to a sampler which allows us to draw indices proportional to the squared row/column-norms of $\mathbf{A}$. Our algorithm produces a compressed representation of some vector $\mathbf{x}$ for which $\|\mathbf{x}^* - \mathbf{x}\| < \varepsilon \|\mathbf{x}^* \|$ in $\widetilde{O}(κ_{\mathsf{F}}^4 κ^2 / \varepsilon^2)$ time, where $κ_{\mathsf{F}} := \|\mathbf{A}\|_{\mathsf{F}}\|\mathbf{A}^{+}\|$ and $κ:= \|\mathbf{A}\|\|\mathbf{A}^{+}\|$. The representation of $\mathbf{x}$ allows us to query entries of $\mathbf{x}$ in $\widetilde{O}(κ_{\mathsf{F}}^2)$ time and sample proportional to the square entries of $\mathbf{x}$ in $\widetilde{O}(κ_{\mathsf{F}}^4 κ^6)$ time, assuming access to a sampler which allows us to draw indices proportional to the squared entries of any given row of $\mathbf{A}$. Our analysis, which is elementary, non-asymptotic, and fully self-contained, simplifies and clarifies several past analyses from literature including [Gilyén, Song, and Tang; 2022, 2023] and [Shao and Montanaro; 2022].

cs.DS↗

arXiv2025

New Improvements in Solving Large LABS Instances Using Massively Parallelizable Memetic Tabu Search

Low Autocorrelation Binary Sequences (LABS) is a particularly challenging binary optimization problem which quickly becomes intractable in finding the global optimum for problem sizes beyond 66. This aspect makes LABS appealing to use as a test-bed for meta-heuristic optimization solvers to target large problem sizes. In this work, we introduce a massively parallelized implementation of the memetic tabu search algorithm to tackle LABS problem for sizes up to 120. By effectively combining the block level and thread level parallelism framework within a single Nvidia-A100 GPU, and creating hyper optimized binary-valued data structures for shared memory among the blocks, we showcase up to 26 fold speedup compared to the analogous 16-core CPU implementation. Our implementation has also enabled us to find new LABS merit factor values for sixteen different problem sizes between 92 and 120. Crucially, we also showcase improved values for five odd-sized problems {99, 107, 109, 113, 119} whose previous best known results coincided with the provably optimal skew-symmetric search sequences. Consequently, our result highlights the importance of a focus on general-purpose solver to tackle LABS, since leveraging its skew-symmetry could lead to sub-optimal solutions.

cs.DC↗

arXiv2025

Energy-Efficient Real-Time Job Mapping and Resource Management in Mobile-Edge Computing

Mobile-edge computing (MEC) has emerged as a promising paradigm for enabling Internet of Things (IoT) devices to handle computation-intensive jobs. Due to the imperfect parallelization of algorithms for job processing on servers and the impact of IoT device mobility on data communication quality in wireless networks, it is crucial to jointly consider server resource allocation and IoT device mobility during job scheduling to fully benefit from MEC, which is often overlooked in existing studies. By jointly considering job scheduling, server resource allocation, and IoT device mobility, we investigate the deadline-constrained job offloading and resource management problem in MEC with both communication and computation contentions, aiming to maximize the total energy saved for IoT devices. For the offline version of the problem, where job information is known in advance, we formulate it as an Integer Linear Programming problem and propose an approximation algorithm, $\mathtt{LHJS}$, with a constant performance guarantee. For the online version, where job information is only known upon release, we propose a heuristic algorithm, $\mathtt{LBS}$, that is invoked whenever a job is released. Finally, we conduct experiments with parameters from real-world applications to evaluate their performance.

cs.DC↗

arXiv2025

GPU-Parallelizable Randomized Sketch-and-Precondition for Linear Regression using Sparse Sign Sketches

A litany of theoretical and numerical results have established the sketch-and-precondition paradigm as a powerful approach to solving large linear regression problems in standard computing environments. Perhaps surprisingly, much less work has been done on understanding how sketch-and-precondition performs on graphics processing unit (GPU) systems. We address this gap by benchmarking an implementation of sketch-and-precondition based on sparse sign-sketches on single and multi-GPU systems. In doing so, we describe a novel, easily parallelized, rejection-sampling based method for generating sparse sign sketches. Our approach, which is particularly well-suited for GPUs, is easily adapted to a variety of computing environments. Taken as a whole, our numerical experiments indicate that sketch-and-precondition with sparse sign sketches is particularly well-suited for GPUs, and may be suitable for use in black-box least-squares solvers.

cs.DS↗