Syamantak Kumar

Research profile

Research summary

Syamantak Kumar is a theory researcher at The University of Texas at Austin whose work spans two roughly equal streams: machine-learning theory on streaming PCA and differential privacy, and theoretical computer science on sampling algorithms, spin glasses, and statistical-computational gaps. The bibliometric record (OpenAlex) is early but steeply accelerating — yearly outputs of 1 (2019), 2 (2021), 1 (2023), 3 (2024), 6 (2025), and 5 in partial 2026, totalling 14 works in the last three years — while citation impact remains nascent (h-index 2, i10-index 1). Evolve has reviewed two 2026 preprints. The first-authored 'High-Magnetization Sampling at Low Temperatures' builds sparsity-aware frameworks that give any-temperature slice samplers for the Sherrington-Kirkpatrick model and k^1.5 complexity for spike-and-slab Bayesian sparse linear regression, with a partly conditional near-Almeida-Thouless result. The middle-authored gap-free streaming PCA paper delivers a general-rank, gap-free analysis of Oja's algorithm with near-matching lower bounds, extends to ePCA and differentially private PCA (Brown's conjecture up to logarithmic factors), and is validated on synthetic data. Topical signatures include statistical methods and inference, stochastic gradient optimization, topic modeling, and sparse sensing. Collaboration is compact — 5 distinct coauthors including Kevin Tian, Anming Gu, Chutong Yang, Purnamrita Sarkar, and Yusong Zhu — a pattern consistent with an early-career theorist publishing in small, tightly coupled teams.

Research keywords

Highlights

• Gap-free streaming PCA reaches near-optimal rates with privacy applications

  A general-rank, gap-free analysis of Oja's algorithm achieves near-matching lower bounds for streaming eigenvector recovery, extends to ePCA and differentially private PCA — resolving Brown's conjecture up to logarithmic factors — and is validated on synthetic data, with medium reproducibility readiness.

• Sparsity-aware samplers work at any temperature in spin glasses

  New sparsity-aware sampling frameworks yield slice samplers for the Sherrington-Kirkpatrick model at any temperature and k^1.5 complexity for spike-and-slab Bayesian sparse linear regression; the near-Almeida-Thouless-line guarantee is partly conditional (first-authored, medium readiness).

Timeline

1. 2019First work recorded in the OpenAlex profile (1 publication).
2. 2021Two works recorded as early-career output begins to accumulate.
3. 2023A single work recorded — the sparsest year before the current ramp-up.
4. 2024Output rises to 3 works as the research agenda expands.
5. 2025Peak year so far, with 6 recorded works.
6. 20265 works so far, including both Evolve-reviewed preprints: the first-authored low-temperature sampling paper and the gap-free streaming PCA paper.

Papers in the Evolve catalog

• 2026-09-22

  Gap-Free Streaming PCA Beyond Rank-One Updates: Near-Optimal Rates and Applications to Differential Privacy

• 2026-09-08first author

  High-Magnetization Sampling at Low Temperatures: Ising Models and Bayesian Sparse Linear Regression

Frequent co-authors