Learning theorist at Universitat de València: Brownian-kernel RKHS representations and optimization geometry
2
papers here
2
reviewed by Evolve
2
first-author
0
corresponding
—
works · OpenAlex
—
h-index · OpenAlex
Bibliometric figures come from OpenAlex (2026-10-15) and cover the author's entire public record, not just the papers listed here. Reproducibility signal comes from Evolve's own reviews.
Research profile
Output—
Output volume is not scored yet: the available record covers only two 2026 arXiv preprints, too few to compute publication counts or trends. This axis would measure total research production across venues and years.
Impact—
Impact is not scored yet: the input data contain no citation counts or h-index for these preprints. This axis would measure citation-based influence and the uptake of the work by other researchers.
Focus0/100
The focus axis scores 0. The raw field distribution splits evenly (0.5 and 0.5) between machine learning theory (optimization and kernel methods) and statistical learning theory across the two reviewed papers. In practice the spread is narrow: both topics sit inside kernel-based learning theory and share the Brownian-kernel RKHS theme.
Collaboration32/100
Collaboration scores 32, computed from 4 distinct coauthors across the reviewed papers. Both preprints are first-authored rather than corresponding, and the recurring partners are Gustau Camps-Valls and Nicole Mücke, each appearing twice in the frequency list. The small pool points to a compact, theory-driven working circle rather than large multi-institution teams.
Activity—
Activity is not scored yet: the record holds no publication history beyond the two 2026 preprints. This axis would measure publishing cadence and the continuity of research output over time.
Reproducibility78/100
Reproducibility scores 78, the mean readiness across 2 reviewed papers (reproMean 78, reproPapers 2). The optimization-geometry preprint rates high with well-verified results, while the Brownian-heads paper is rated medium because its code remains unreleased despite rigorous derivations. Releasing that code is the clearest path to lifting this axis.
Research summary
Mahdi Mohammadigohari is a learning-theory researcher at Universitat de València whose Evolve profile rests on two first-authored arXiv preprints, both dated 2026. The work forms a coherent program around Brownian-motion-kernel RKHS representations. The first paper, 'Optimization Geometry of Equivalent Brownian RKHS Representations', sits in machine learning theory (optimization and kernel methods) and proves that equivalent Brownian-RKHS coordinates leave gradient descent and SGD trajectories identical while making Adam's trajectories separable; Evolve's review calls it rigorous and well-verified, with modest novelty and high reproducibility readiness. The second, 'Brownian Heads for Deep ReLU Representations', belongs to statistical learning theory and derives sharp activation-mass Rademacher complexity bounds for Brownian heads, together with the cost of same-sample selection; it is judged convincing, though its code remains unreleased, giving medium reproducibility readiness. Bibliometrically the record is narrow: the field distribution splits evenly (0.5/0.5) between the two theory fields, four distinct coauthors appear, with Gustau Camps-Valls and Nicole Mücke the recurring partners, and the radar scores collaboration at 32 and reproducibility at 78 while output, impact and activity lack sufficient data. This is an early-stage profile that will sharpen as citation and publication history accumulate.
Collaboration pattern
Collaboration is compact: 4 distinct coauthors across the reviewed papers, and both preprints are first-authored (non-corresponding). Gustau Camps-Valls and Nicole Mücke are the recurring partners, each appearing twice in the coauthor frequency list. The radar collaboration score is 32, consistent with a small, theory-driven working circle rather than large multi-institution teams.
Research keywords
Highlights
GD and SGD stay invariant under equivalent Brownian-RKHS coordinates — Adam does not
Sharp activation-mass Rademacher bounds for Brownian heads in deep ReLU representations
A two-paper Brownian-kernel program spanning optimization and generalization
The two first-authored 2026 preprints split the field distribution evenly (0.5 optimization and kernel methods, 0.5 statistical learning theory) around one theme: Brownian-kernel RKHS representations, analyzed first through optimization geometry, then through Rademacher-complexity generalization bounds.
Timeline
- 2026Publishes the first-authored arXiv preprint 'Optimization Geometry of Equivalent Brownian RKHS Representations' in machine learning theory (optimization and kernel methods).
- 2026Publishes the first-authored arXiv preprint 'Brownian Heads for Deep ReLU Representations: Activation Mass and the Cost of Same-Sample Selection' in statistical learning theory.
- 2026Both preprints enter the Evolve catalog: the two papers average a reproducibility-readiness score of 78/100, with collaboration scored 32 across 4 distinct coauthors.
Papers in the Evolve catalog
Frequent co-authors
About this profile
This profile is AI-structured by Evolve.Pub from public bibliographic data (OpenAlex) and Evolve's own paper reviews. It is not written or endorsed by the researcher unless the profile is claimed. Radar dimensions are computed from real bibliometrics with fixed formulas — raw figures are shown alongside; missing dimensions mean insufficient data, never zero. Errors? Claim the profile to correct or hide it.
- Bibliometric dimensions are computed from a narrow sample: 2 arXiv preprints reviewed by Evolve (both 2026); output, impact and activity axes have insufficient underlying data.
- No citation counts, h-index, or publication history beyond the two reviewed preprints is available in the input data.
- The country field is empty in the source record; the affiliation is given as Universitat de València.
- Identity is resolved from public bibliographic data linked to ORCID 0000-0002-4811-6585.