{
 "artifacts": null,
 "category": "architecture",
 "description": "Replace a finite-dimensional latent prior for fields or trajectories with a continuous flow acting on function coefficients in a separable Hilbert space. A Gaussian reference with an explicit covariance spectrum is transformed by a learned vector field, so the model can represent smooth, non-isotropic function distributions while sharing parameters across discretization resolutions.",
 "formulas_latex": [
  "$$\\frac{dv(t)}{dt}=h(v(t),t;\\theta),\\qquad v(0)=u_0,\\qquad u_0\\sim\\mu_0=\\mathcal{N}(0,\\mathcal{C}_0)\\text{ on }\\mathcal{H}.$$",
  "$$u_K=\\sum_{k=1}^{K}z_ke_k,\\qquad z_k\\sim\\mathcal{N}(0,\\lambda_k),\\qquad \\dot z(t)=H_\\theta(z(t),t),\\quad H_{\\theta,k}(z,t)=\\left\\langle h\\left(\\sum_{j=1}^{K}z_je_j,t;\\theta\\right),e_k\\right\\rangle.$$",
  "$$\\log p_1(z(1))=\\log p_0(z(0))-\\int_{0}^{1}\\operatorname{tr}\\!\\left(\\frac{\\partial H_\\theta}{\\partial z}(z(t),t)\\right)dt,$$",
  "$$\\mathcal{L}(\\theta)= -\\frac{1}{B}\\sum_{b=1}^{B}\\left[\\log p_0(z_0^{(b)})-\\int_0^1\\operatorname{tr}\\!\\left(\\frac{\\partial H_\\theta}{\\partial z}(z^{(b)}(t),t)\\right)dt\\right].$$"
 ],
 "id": 3096,
 "implementation": "(1) Integration point: use this as the prior or latent distribution in a neural operator, field VAE, trajectory model, or PDE surrogate. Store each field as coefficients $z\\in\\mathbb{R}^{K}$ in a fixed basis such as Fourier, cosine, finite elements, or PCA modes, rather than as raw grid pixels. The vector field $H_\\theta$ can be a small time-conditioned MLP for a first test, then a Fourier neural operator for resolution transfer. Initialize $\\lambda_k$ with a smooth spectrum such as $\\lambda_k=c(1+k^2)^{-\\alpha}$.\n\n(2) Pseudocode: sample $z_0[k]\\leftarrow\\sqrt{\\lambda_k}\\,\\epsilon_k$, with $\\epsilon_k\\sim\\mathcal{N}(0,1)$; set $s_0=\\log\\mathcal{N}(z_0;0,\\operatorname{diag}(\\lambda))$; integrate $(z,s)$ from $t=0$ to $1$ using an adaptive ODE solver with $\\dot z=H_\\theta(z,t)$ and $\\dot s=-\\operatorname{tr}(J_H(z,t))$, where $J_H=\\partial H_\\theta/\\partial z$; train by minimizing $-s_1$ on observed function coefficients. For generation, sample $z_0$ and integrate only $\\dot z$ forward.\n\n(3) Compute the Gaussian base density analytically. Estimate the trace with Hutchinson probes if $K$ is large: $\\operatorname{tr}(J_H)\\approx R^{-1}\\sum_{r=1}^{R}\\xi_r^TJ_H\\xi_r$, where $\\xi_r\\sim\\{\\pm1\\}^K$, using automatic differentiation. The paper supplies the Hilbert-space ODE construction; the truncation level $K$, solver tolerance, covariance spectrum, and whether the learned flow is stable under refinement are empirical design choices.\n\n(4) First experiment: train on Burgers or Darcy-resolution field snapshots at $K=32$, comparing a diagonal Gaussian prior, a standard finite-dimensional MLP-CNF, and this spectral CNF at equal parameter count. Evaluate held-out negative log likelihood, downstream conditional-generation error, and samples decoded at $K=32,64,128$. Success means improved held-out likelihood and visibly smoother, statistically consistent samples, with prediction error remaining nearly unchanged when the evaluation grid is refined. Also record ODE function evaluations and likelihood cost to determine whether resolution invariance is obtained without an unacceptable sampling slowdown.",
 "math_summary": "The paper defines the infinite-dimensional ODE $\\frac{dv(t)}{dt}=h(v(t),t;\\theta)$ with $v(0)=u_0$, where $u_0$ is sampled from the Gaussian measure $\\mu_0=\\mathcal{N}(0,\\mathcal{C}_0)$ on a separable Hilbert space $\\mathcal{H}$, $h:\\mathcal{H}\\times[0,1]\\to\\mathcal{H}$ is a neural vector field, and $\\mathcal{C}_0$ is the reference covariance operator. To obtain an implementable model, choose orthonormal basis functions $\\{e_k\\}_{k=1}^{K}$, covariance eigenvalues $\\lambda_k\u003e0$, and represent $u_K=\\sum_{k=1}^{K}z_ke_k$ with $z_k\\sim\\mathcal{N}(0,\\lambda_k)$. The truncated flow is $\\dot z(t)=H_\\theta(z(t),t)$, where $H_{\\theta,k}=\\langle h(\\sum_jz_je_j,t;\\theta),e_k\\rangle$. Its change-of-variables identity is $\\log p_1(z(1))=\\log p_0(z(0))-\\int_0^1\\operatorname{tr}(\\partial H_\\theta/\\partial z)(z(t),t)\\,dt$, with $p_0=\\mathcal{N}(0,\\operatorname{diag}(\\lambda_1,\\ldots,\\lambda_K))$. The functional transfer is useful only if the vector field is stable under basis refinement, for example by applying a decaying spectral parameterization or a bounded neural-operator kernel.",
 "math_tags": [
  "functional-analysis",
  "probability",
  "dynamical-systems",
  "numerical-analysis"
 ],
 "ml_areas": [
  "mlp",
  "training",
  "sampling",
  "world-model"
 ],
 "paper": {
  "arxiv_id": "2609.03343",
  "arxiv_url": "https://arxiv.org/abs/2609.03343",
  "summary_what_math_gives_to_ml": "The paper's transferable asset is a continuous normalizing flow defined directly on a separable Hilbert space, allowing a simple Gaussian measure over functions to be transported into a learned, informative function-space prior. This is more than an ordinary finite-dimensional CNF when the target object is a field, trajectory, or operator input: the representation can remain resolution-independent and can encode correlations across discretization levels. The most practical neural-network transfer is a spectrally truncated Hilbert-space CNF whose vector field is implemented by a neural operator or basis-space MLP, with likelihood correction computed from the Jacobian trace and tested for mesh-resolution invariance.",
  "title": "Learning Informative Prior with Infinite-Dimensional Continuous Normalizing Flow for Bayesian Inverse Problem",
  "year": "2026"
 },
 "ratings": {
  "difficulty": 6,
  "novelty": 6,
  "usefulness": 6
 },
 "solves": [
  "accuracy",
  "sample-efficiency",
  "scalability"
 ],
 "title": "Resolution-Invariant Hilbert-Space CNF Prior",
 "url": "https://synthcore.org/idea/3096/resolution-invariant-hilbert-space-cnf-prior",
 "verification": {
  "peer_reviewed": false,
  "status": "unverified",
  "status_label": "Unverified",
  "verdict_source": "deterministic test code (paired-seed permutation statistics)",
  "verification_axes": {
   "benchmark_mechanism": {
    "confirmed": null,
    "tested": false
   },
   "practical_benchmark": {
    "beats_baseline": null,
    "tested": false
   },
   "toy_mechanism_gate": {
    "confirmed": null,
    "tested": false
   }
  }
 }
}