arXiv subjects

Kristian Sestak

Publications and source records attributed to Kristian Sestak.

4 recordsLinked to original sources

arXiv2026

Kolmogorov--Arnold Networks as Implicit Regularizers: Noise Robustness and Interpretability for Stellar Classification

This paper tests whether Kolmogorov--Arnold Networks (KAN 2.0) are genuinely more noise-robust than Multi-Layer Perceptrons (MLP) and XGBoost for stellar classification (star/galaxy/quasar, 100,000 SDSS DR17 objects). A naive comparison suggests so: KAN retains +9 percentage points over MLP at SNR=5. But equalizing baseline accuracy via weight decay eliminates the gap -- a properly regularized MLP matches KAN to within 1 p.p. at all SNR levels, both with and without spectroscopic redshift. The same holds on an independent DESI DR1 sample with different photometric bands. KAN's robustness thus traces to implicit regularization by C^2-smooth B-spline activations, not to architecture. Per-class analysis (20 trials) shows that stars degrade fastest (F1: 0.97 to 0.75 at SNR=5), while QSOs remain stable. KAN's native feature importance and SHAP on MLP produce different rankings (Spearman rho = -0.37), capturing complementary aspects of the classification. Colour-index features (u-g, g-r, r-i, i-z) widen KAN's relative advantage, and a hybrid pipeline routing uncertain MLP predictions to KAN improves low-SNR accuracy. KAN is best understood as a convenient auto-regularizer whose genuine advantage is built-in interpretability.

astro-ph.IM↗

arXiv2026

Candidacy and Trigger: A Two-Phase Empirical Model of Hierarchical Collapse

We test a dynamic ODE model of hierarchical asymmetry on a panel of 260 countries over 1960-2023, drawing on World Bank, Penn World Table, V-Dem and World Inequality Database sources. In cross-section the model holds partially: trade openness and bottom-of-distribution health suppress within-country asymmetry. The annual time-evolution equation fails, with out-of-sample R^2 at or below zero across five functional forms. The same state vector, augmented with market, debt and trajectory features, is much more successful as a discriminator: a four-layer leave-one-collapse-out classifier separates 29 historical collapses from 60 stable controls at a nested cross-validated AUC of 0.91. The signal splits into a chronic risk profile visible a decade before the event and an acute inflection three to five years before. Three independent tests reject the endogenous-drift reading of collapse. What remains is a candidacy-and-trigger picture in which structural variables identify the high-risk countries while collapse timing is set by shocks outside the modelled system. A separate strand documents a lagged co-movement between global fertility and global asymmetry on a single n=63 aggregate series; taken alone this would suggest a selection-pool channel. The same pattern is then tested within countries, within demographic strata, inside a two-way fixed-effects panel and through a migration-mediated cross-country interaction model, and the directional reading fails in each. The aggregate co-movement is a compositional effect rather than a causal channel. A global event-study on 7,316 peer-event observations confirms regional spillover in asymmetry and a novel post-collapse degradation of bottom-of-distribution health in regional neighbours. A pre-registered forward-look produces a top-20 / bottom-20 ranking to be evaluated over 2026-2036.

physics.soc-ph↗

arXiv2026

Empirical Confirmation of the Environmental-Dominance Inequality A direct decomposition of Var(ln \r{ho}eff ) across four levels of aggregation

Empirical confirmation of the environmental-dominance inequality Var(ln rho_eff) >> Var(ln k) from arXiv:2605.02985, computed directly from three public datasets (Opportunity Atlas, World Bank GDP per capita PPP, World Inequality Database) at four levels of aggregation: U.S. census tracts, between countries, within-country deciles, and the global pooled-individual distribution. The headline global value Var(ln rho_eff) = 4.33 yields a dominance ratio R in [27, 134] across plausible sigma_ln k in [0.18, 0.40]. The inequality holds with one-to-two orders of magnitude margin at the global and within-country-decile levels, with a single-digit but still dominant margin between countries, and collapses to R in [0.33, 1.61] within already-homogenized U.S. census tracts for income. A 1990-2022 time series shows the global aggregate stable while composition shifts from between-country dispersion (-34%) to within-country dispersion (+26%), consistent with international convergence plus Piketty r > g. Multi-outcome validation shows the inequality is robust for income, infant mortality and incarceration but shrinks toward parity for outcomes targeted by sustained global convergence (life expectancy). Partial-identification and selection-bias bounds (Chetty-style 40-50% selection share) leave R in [14, 80]. All inputs and outputs are SHA-256 hashed in an append-only manifest and fully reproducible from the accompanying notebooks.

physics.soc-ph↗

arXiv2026

The Dominance of Environment over Entity's Capabilities

We present an analytical framework for the probability of individual success based on a single structural asymmetry between the capacity of an entity to explore possibilities, $k$, and the size of the possibility space offered by the environment, $n$, where $k \ll n$. We introduce an effective density $ρ_{\rm eff}$ of favorable possibilities accessible to a given entity, derive the probability of success as $P \approx 1-(1-ρ_{\rm eff})^k$, and decompose its variance across a population. We show that while the elasticities $\varepsilon_ρ$ and $\varepsilon_k$ are comparable, the variance of outcomes is dominated by ${\rm Var}(\ln ρ_{\rm eff})$ whenever it exceeds ${\rm Var}(\ln k)$. A back-of-envelope calibration based on published inequality and productivity data indicates this condition holds by two to three orders of magnitude. The framework provides an analytical complement to the simulation result of Pluchino, Biondo and Rapisarda (2018), and offers a unified structural account of geographic inequality, intergenerational mobility and accessibility-based discrimination as special cases of the narrowing of the accessible set $A(E,P)$.

physics.soc-ph↗