# Cohomological Jacobian Flattening

- ID: 35
- Canonical URL: https://synthcore.org/idea/35/cohomological-jacobian-flattening
- API JSON: https://synthcore.org/api/idea/35.json
- API Markdown: https://synthcore.org/api/idea/35.md
- Verification status: mech_ok_no_baseline
- Source: [arXiv:2512.02307](https://arxiv.org/abs/2512.02307)
- Category: dynamics
- Solves: stability, generalization
- ML areas: training-dynamics, rnn, ssm, regularization
- Math tags: dynamical-systems, differential-geometry, optimization, ergodic-theory
- Ratings: usefulness 6/10; difficulty 5/10; novelty 7/10

## Idea description

Regularize a neural dynamical map so that its log-volume expansion is cohomologous to a constant rather than forcing the Jacobian determinant to be constant at every state. Learn a scalar potential that explains transient expansion and penalize only the non-telescoping component, which should reduce long-horizon gradient explosion or collapse while retaining useful average expansion.

## Mathematical statement

The paper defines the unstable Jacobian as $\operatorname{J}^{u}_{x}=|\det(Df(x)|_{E^{u}})|$ and its $k$-step product by $\operatorname{J}^{u}_{x}(k)=\prod_{\ell=0}^{k-1}\operatorname{J}^{u}_{f^{\ell}(x)}$. Constant unstable Jacobian periodic data are characterized by the cohomological equation $u\circ f(x)-u(x)+c=\log\operatorname{J}^{u}_{x}$, where $f$ is the dynamical map, $u:M\to\mathbb{R}$ is a Hölder-continuous potential, and $c\in\mathbb{R}$ is a constant expansion rate. Along an orbit this implies $\log\operatorname{J}^{u}_{x}(k)=u(f^{k}(x))-u(x)+kc$, so the potential terms telescope and long-run average expansion is $c$. For a neural map $F_\theta$, use $\ell_\theta(x)=\log|\det DF_\theta(x)|$ or, for a selected expanding subspace $E^u$, $\ell_\theta(x)=\log|\det(DF_\theta(x)|_{E^u})|$, and minimize the residual $\ell_\theta(x)-u_\phi(F_\theta(x))+u_\phi(x)-c$. This differs from pointwise volume regularization: state-dependent Jacobian variation is allowed when it is a coboundary.

## Key formulas

- $$\operatorname{J}^{u}_{x}(k)\stackrel{\mathrm{def}}{=}\prod_{\ell=0}^{k-1}\operatorname{J}^{u}_{f^{\ell}(x)}$$
- $$u\circ f(x)-u(x)+c=\log\operatorname{J}^{u}_{x}$$
- $$\log\operatorname{J}^{u}_{x}(k)=u(f^{k}(x))-u(x)+kc$$
- $$\mathcal{L}_{\mathrm{coh}}(\theta,\phi,c)=\frac{1}{B}\sum_{i=1}^{B}\left[\ell_\theta(x_i)-u_\phi(F_\theta(x_i))+u_\phi(x_i)-c\right]^2$$

## Implementation notes

Integrate this into the transition function of an invertible RNN, state-space model, neural ODE discretization, or normalizing-flow-like recurrent map. Let the state be $x_t\in\mathbb{R}^d$ and define $x_{t+1}=F_\theta(x_t,a_t)$; initially omit the action/input from the Jacobian calculation or treat it as fixed. Use a small MLP $u_\phi(x)$ to represent the scalar potential and a learned scalar $c$. For each minibatch of states $x_i$, compute $y_i=F_\theta(x_i,a_i)$, the Jacobian log-volume $\ell_i=\log|\det DF_\theta(x_i,a_i)|$, and the residual $r_i=\ell_i-u_\phi(y_i)+u_\phi(x_i)-c$. Add $\lambda\mathcal{L}_{\mathrm{coh}}=\lambda B^{-1}\sum_i r_i^2$ to the task loss. For small $d$, compute the determinant by autograd Jacobian construction and slogdet; for larger $d$, use an architecture with triangular Jacobian, or estimate the trace of $\log|DF|$ only when a reliable matrix-log estimator is available. Alternate ordinary gradient updates of $(\theta,\phi,c)$, and constrain the gauge of $u_\phi$ by subtracting its minibatch mean. The mathematical quantity is the one-step coboundary residual; empirically estimate whether it telescopes by measuring $R_k(x)=\log|\det DF_\theta^k(x)|-[u_\phi(F_\theta^k(x))-u_\phi(x)+kc]$ over unrolled trajectories. Also compare against the stronger baseline $\mathbb{E}[\ell_\theta(x)-c]^2$, which forces pointwise log-Jacobian constancy, and against no Jacobian regularizer. First test on a 4- or 8-dimensional invertible residual map trained to predict the next state of a synthetic chaotic system, using 32- or 64-step training and 512-step rollout evaluation. Success means lower variance of $R_k/k$, fewer exploding or collapsed trajectories, bounded gradient norms over long unrolls, and equal or better one-step prediction at the same task loss. A useful target is at least a 2x increase in stable rollout horizon or a 30% reduction in gradient variance relative to the pointwise baseline.

## Verification

- Status: mech_ok_no_baseline
- Mechanism evidence: yes
- Mechanism confirmed: yes
- Practical verdict: inconclusive
- Verdict: Built and numerically verified cohomological Jacobian flattening on a nonlinear conjugate expanding map. The exact potential produced max finite-horizon residuals from 4.2e-17 at k=1 to 1.3e-15 at k=32; pointwise Jacobian variance scaled approximately as eps^2 for small eps, with fitted log-log slope 2.60; and scaling the potential by alpha gave residual ratios exactly matching 1-alpha. A Fourier potential reduced one-step residual MSE from 1.61e-3 for the pointwise-constant baseline to 3.42e-11, but no neural-network optimization or long-horizon stability experiment was run, so this validates the mechanism rather than an end-to-end ML win.

### Mechanism check

- Verdict: Built and numerically verified cohomological Jacobian flattening on a nonlinear conjugate expanding map. The exact potential produced max finite-horizon residuals from 4.2e-17 at k=1 to 1.3e-15 at k=32; pointwise Jacobian variance scaled approximately as eps^2 for small eps, with fitted log-log slope 2.60; and scaling the potential by alpha gave residual ratios exactly matching 1-alpha. A Fourier potential reduced one-step residual MSE from 1.61e-3 for the pointwise-constant baseline to 3.42e-11, but no neural-network optimization or long-horizon stability experiment was run, so this validates the mechanism rather than an end-to-end ML win.
- Confidence: 8/10
- Limitations: Only a one-dimensional analytically conjugate expanding toy map was tested. The experiment did not train an invertible RNN, estimate Jacobians with autograd in higher dimensions, compare task loss or FLOPs, or measure long-horizon gradient explosion/collapse and rollout horizon.

### Practical benchmark

- Paired seeds: 8
- Baseline mean: 0
- Idea mean: 0
- p-value: 0.07245
- Paired wins: 7/8
- Benchmark verdict: no significant win

## Artifacts

- [bench_report.json](https://synthcore.org/code/690/bench_report.json)
- [coh_jacobian_experiment.py](https://synthcore.org/code/690/coh_jacobian_experiment.py)
- [report.md](https://synthcore.org/code/690/report.md)
- [report_bench_2026-08-31T221417.md](https://synthcore.org/code/690/report_bench_2026-08-31T221417.md)
- [results.json](https://synthcore.org/code/690/results.json)
- [stage2_coh_bench.py](https://synthcore.org/code/690/stage2_coh_bench.py)
- [Download all files as ZIP](https://synthcore.org/download/690)

## Disclaimer

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