Kernel Dynamics under Path Entropy Maximization
Jnaneshwar Das · 2026
Read Kernel Dynamics under Path Entropy Maximization by Jnaneshwar Das (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-27880.txt.
Introduction
Das treats the kernel itself as a moving object under MaxCal rules. A kernel encodes which distinctions an agent can draw. Path entropy maximization steers its change over time. Each kernel induces its own geometry on probability space. Fixed-point rules mark self-consistent kernels. Renormalization flow appears as a structured special case. Neural tangent kernel drift may fit the frame. Work bounds tie effort to new mutual information.
Catalog record
| CREATOR | Jnaneshwar Das |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 36713 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.27880
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Kernel Dynamics under Path Entropy Maximization Jnaneshwar Das Abstract We propose a variational framework in which the kernel function k : 𝒳 × 𝒳 → ℝ k:\mathcal{X}\times\mathcal{X}\to\mathbb{R} , interpreted as the foundational object encoding what distinctions an agent can represent, is treated as a dynamical variable subject to path entropy maximization (Maximum Caliber, MaxCal). Each kernel defines a representational structure over which an information geometry on probability space may be analyzed; a trajectory through kernel space therefore corresponds to a trajectory through a family of effective geometries, making the optimization landscape endogenous to its own traversal. We formulate fixed-point conditions for self-consistent kernels, propose renormalization group (RG) flow as a structured special case, and suggest neural tangent kernel (NTK) evolution during deep network training as a candidate empirical instantiation. Under explicit information-thermodynamic assumptions, the work required for kernel change is bounded below by δ W ≥ k B T δ I k \delta W\geq k_{B}T\,\delta I_{k} , where δ I k \delta I_{k} is the mutual information newly unlocked by the updated kernel. In this view, stable fixed points of MaxCal over kernels correspond to self-reinforcing distinction structures, with biological niches, scientific paradigms, and craft mastery offered as conjectural interpretations. We situate the framework relative to assembly theory and the MaxCal literature, separate formal results from structured correspondences and conjectural bridges, and pose six open questions that make the program empirically and mathematically testable. I Introduction Every inference engine—biological, computational, or learned—operates on a substrate that determines which differences in the world it can represent. We call this substrate the kernel : a positive-definite function k:\mathcal{X}\times\mathcal{X}\to\mathbb{R} whose choice induces an inner product on a reproducing kernel Hilbert space (RKHS) \mathcal{H}_{k} , and—via the representational structure it defines—an effective geometry on the space of probability distributions \mathcal{P} and a bound on extractable work via the Sagawa-Ueda generalized second law [ 1 ] . Throughout the paper, “kernel” is used in a broadened but still technical sense: not merely as a computational device for nonlinear learning, but as the mathematical object that determines which differences in the world are representable, comparable, or actionable for an agent. This representational reading motivates treating kernel change as a change in the agent’s effective distinction-making capacity, which is the quantity the MaxCal construction is designed to model. In this paper we adopt the kernel as the primitive object from which geometry, dynamics, and information-thermodynamic bounds are derived. At fixed kernel, this static viewpoint recovers a common exchange rate k_{B}\ln 2 per bit under standard assumptions of reversible bookkeeping. The present manuscript focuses on what that static viewpoint leaves open: a given agent operates within a kernel but the question of how kernels themselves change over time was deferred. The present paper addresses that deferral directly. We ask: if kernels are dynamical variables, what variational principle governs their trajectories? The answer we develop is Maximum Caliber (MaxCal) [ 2 ] —path entropy maximization over trajectories through kernel space. The resulting framework has three properties that distinguish it from prior treatments of kernel learning: 1. The optimization landscape is endogenous : each kernel k deforms the geometry of \mathcal{P} , so the landscape through which kernels evolve is itself a function of the trajectory. 2. Fixed points of the dynamics are self-consistent kernels: distinction structures that are self-reinforcing under their own effective geometry. We interpret RG fixed points as the most structurally direct special case, and biological niches, scientific paradigms, and craft mastery as candidate higher-level analogues of self-consistent kernels. 3. Kernel change has a thermodynamic cost bounded below by k_{B}T per bit of newly unlocked mutual information—a Landauer principle for conceptual change. Relative to prior work, the contribution here is not a new kernel-learning algorithm, nor a reformulation of classical information geometry, nor a standard MaxCal model on a fixed state space. Rather, the paper treats the kernel itself as the evolving object of inference and asks what variational principle governs trajectories through the space of distinction-making structures. The result is a framework that is partly formal and partly programmatic: it introduces the kernel-space MaxCal construction, formulates self-consistency and stability conditions, and uses these to organize a set of candidate correspondences across physics, learning, biology, and embodied craft. The paper is organized as follows. Section II reviews the necessary background in MaxCal, RKHS geometry, and the information-thermodynamic assumptions used here. Section III defines the space of kernels and its topology. Section IV lifts MaxCal from state space to kernel space. Section V derives fixed-point conditions and their stability. Section VI instantiates the framework across RG flow, NTK evolution, adaptive sample-return planning, biological evolution, and craft mastery. Section VII situates the framework relative to assembly theory. Section VIII separates formal results from conjectural bridges. Section IX poses six open questions. Claim status in this paper. Formal content includes the definition of kernel space \mathcal{K} , the lifting of MaxCal to path measures on kernel trajectories, the self-consistency condition, and the frozen-kernel stability criterion. Structured correspondences include the mappings to RG flow and finite-width NTK evolution. Conjectural bridges include the interpretations in biological evolution, craft mastery, adaptive field sampling, and scientific paradigm shifts. II Background II-A Maximum Caliber Let \Gamma denote the space of trajectories \gamma:[0,T]\to\mathcal{X} over a state space \mathcal{X} . MaxCal selects the path distribution p[\gamma] that maximizes the path entropy \mathcal{S}[p]=-\sum_{\gamma}p[\gamma]\ln\frac{p[\gamma]}{q[\gamma]} (1) subject to dynamical constraints \langle f_{i}[\gamma]\rangle=F_{i} , where q[\gamma] is a reference path measure [ 2 ] . The resulting distribution takes the form p[\gamma]\propto q[\gamma]\exp\!\left(-\sum_{i}\lambda_{i}f_{i}[\gamma]\right) (2) where \lambda_{i} are Lagrange multipliers. MaxEnt (Jaynes) is recovered when trajectories reduce to single configurations. MaxCal derives Green-Kubo relations, Onsager reciprocity, Prigogine’s minimum entropy production, and master equations as special cases [ 2 ] . II-B RKHS Geometry and the Kernel Primitive A Mercer kernel k induces an RKHS \mathcal{H}_{k} via the feature map \phi:\mathcal{X}\to\mathcal{H}_{k} , \phi(x)=k(\cdot,x) . The kernel is treated as the primitive representational object, while Fisher-Rao supplies the canonical information geometry on the associated statistical manifold; kernel change therefore alters the effective geometry of inference by altering the representational substrate on which that manifold is defined. Concretely, the square-root embedding p\mapsto\sqrt{p}\in L^{2}(\mathcal{X},\nu) pulls back the L^{2} inner product to the Fisher-Rao metric [ 4 ] g^{F}_{ij}(p)=\int\frac{\partial\ln p}{\partial\theta^{i}}\frac{\partial\ln p}{\partial\theta^{j}}\,p\,d\nu (3) The Hellinger kernel k_{H}(p,q)=\int\sqrt{p\,q}\,d\nu is the unique kernel whose induced geometry respects sufficient statistics [ 5 ] . The kernelized Stein discrepancy identifies the score function \nabla_{x}\log p(x) as the Riesz representative of the Stein operator in \mathcal{H}_{k} [ 6 ] . The central modeling assumption of this paper is that k is prior to all of this structure: the metric, the score function, the mutual information I , and the Sagawa-Ueda work bound W_{\text{extracted}}\leq\Delta F+k_{B}T\cdot I (4) are treated as derived objects once a kernel is fixed. II-C Standing Assumptions The following assumptions are used throughout the paper. (A1) \mathcal{X} is a Polish (complete, separable, metrizable) space. (A2) \nu is a \sigma -finite reference measure on \mathcal{X} . (A3) Every kernel k\in\mathcal{K} is Mercer (continuous, symmetric, positive semi-definite) and Hilbert-Schmidt: \iint k(x,x^{\prime})^{2}\,d\nu(x)\,d\nu(x^{\prime})<\infty . (A4) The agent–environment system admits a joint distribution P_{\mathrm{agent,env}} from which the mutual information I_{k}=I(A;E\mid k) is computed by restricting the agent’s representation to the RKHS