# Frozen-Diffusion Parametrix Preconditioner

- ID: 2976
- Canonical URL: https://synthcore.org/idea/2976/frozen-diffusion-parametrix-preconditioner
- API JSON: https://synthcore.org/api/idea/2976.json
- API Markdown: https://synthcore.org/api/idea/2976.md
- Verification status: unverified
- Source: [arXiv:2609.00529](https://arxiv.org/abs/2609.00529)
- Category: training
- Solves: stability, speedup, accuracy
- ML areas: training, optimizer, world-model
- Math tags: pde, stochastic-processes, harmonic-analysis, numerical-analysis, dynamical-systems
- Ratings: usefulness 7/10; difficulty 6/10; novelty 7/10

## Idea description

Replace the raw HJB residual loss of a neural PDE solver with a parametrix-preconditioned fixed-point target. At each local space-time patch, analytically propagate terminal values and source terms through a Gaussian kernel whose covariance uses a frozen diffusion matrix, while asking the network to learn only the variable-coefficient correction. This should reduce the burden on the network to represent stiff high-frequency diffusion dynamics and improve short-horizon convergence.

## Mathematical statement

The paper studies the HJB equation $\partial_t V+\inf_u\{L^uV+\ell+u^\top Ru\}=0$ with terminal condition $V(T,x)=\varphi(x)$, where $L^uV=(f+gu)\cdot\nabla V+\frac12 A:D^2V$ and $A=\sigma\sigma^\top$ is the state-dependent diffusion covariance. The transferable parametrix operation freezes $A$ at a patch center $(\tau,x_0)$, writing $A_0=A(\tau,x_0)$, and uses the constant-coefficient Gaussian semigroup with Fourier multiplier $\widehat{P_h^{A_0}q}(\xi)=\exp[-\frac12h\,\xi^\top A_0\xi]\hat q(\xi)$ for step size $h>0$. The variable-coefficient remainder is $\frac12(A(t,x)-A_0):D^2V$. With $H(t,x,p)=\inf_u\{(f(t,x)+g(t,x)u)\cdot p+u^\top Ru\}$, the adapted backward fixed-point equation is $V(t)=P_{T-t}^{A_0}\varphi+\int_t^T P_{s-t}^{A_0}[\ell(s,\cdot)+H(s,\cdot,\nabla V(s,\cdot))+\frac12(A(s,\cdot)-A_0):D^2V(s,\cdot)]\,ds$. Here $P_h^{A_0}$ performs exact diffusion smoothing, $D^2V$ is the Hessian, and the correction vanishes when the diffusion is locally constant.

## Key formulas

- $$\partial_{t}V(t;x)+\inf_{u}\left\{L^{u}V(t;x)+\ell(t;x)+u^{\top}Ru\right\}=0,\qquad V(T;x)=\varphi(x),$$
- $$L^{u}v(t;x)=\bigl(f(t;x)+g(t;x)u\bigr)\cdot\nabla_{x}v(t;x)+\frac{1}{2}\,\sigma(t;x)\sigma(t;x)^{\top}:D_{x}^{2}v(t;x),$$
- $$\widehat{P_h^{A_0}q}(\xi)=\exp\!\left(-\frac12h\,\xi^{\top}A_0\xi\right)\hat q(\xi),\qquad A_0=\sigma(\tau,x_0)\sigma(\tau,x_0)^{\top},$$
- $$V(t)=P_{T-t}^{A_0}\varphi+\int_t^T P_{s-t}^{A_0}\!\left[\ell(s,\cdot)+H(s,\cdot,\nabla V(s,\cdot))+\frac12\bigl(A(s,\cdot)-A_0\bigr):D^2V(s,\cdot)\right]ds.$$

## Implementation notes

Integrate the method into a neural HJB solver at the training-target and residual-evaluation stage, not into the network architecture itself. Use an MLP $V_\theta(t,x)$ and a backward time grid $t_0<\cdots<t_N=T$. Partition sampled states into local patches with centers $(\tau_j,x_j)$; estimate $A_j=\sigma(\tau_j,x_j)\sigma(\tau_j,x_j)^\top$, symmetrize it, and add $\epsilon I$ if needed. For each patch and timestep, compute the network quantities $p=\nabla_xV_\theta(t_{n+1},x)$ and $B=D_x^2V_\theta(t_{n+1},x)$ using autodiff. Evaluate the Hamiltonian $H=\min_u[(f+gu)^\top p+u^\top Ru]$; for quadratic control with invertible $R$, use $u^*=-\frac12R^{-1}g^\top p$ and substitute it into $H$. Form the source correction $q=\ell+H+\frac12(A(t_{n+1},x)-A_j):B$. Apply the frozen Gaussian propagator $P_{\Delta t}^{A_j}$ to $q$ by either Monte Carlo samples $x'=x+L_j\sqrt{\Delta t}\,z$, where $L_jL_j^\top=A_j$, averaging $q(t_{n+1},x')$, or a truncated cosine/Fourier feature approximation using the multiplier $\exp[-\frac12\Delta t\xi^\top A_j\xi]$. Construct the backward target $y=P_{\Delta t}^{A_j}V_\theta(t_{n+1},\cdot)+\Delta tP_{\Delta t}^{A_j}q$ and minimize $\|V_\theta(t_n,x)-y\|^2$ over sampled points, optionally adding a small raw PDE residual term. Run one Picard sweep per epoch, detach the target for stability, and refresh patch centers every few epochs. The paper supplies the frozen-diffusion/variable-coefficient decomposition; Monte Carlo error, patch assignment, and timestep size are empirical choices. First test on a 10-dimensional linear-quadratic diffusion-control problem with known Riccati solution, comparing a vanilla PINN residual baseline against the same MLP with raw backward targets. Measure value-function relative error, gradient error, loss decrease per optimizer step, and sensitivity to $\Delta t$. Success is faster decrease and lower error at equal network evaluations, especially when $A(x)$ varies strongly across the state space.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.