Nullspace-coordinate constrained operator blocks
Source paper: Geometry-aware LegONet for PDE Learning on Arbitrary Domains arXiv:2607.23069 ⓘ · analyzed Sep 2, 2026
AI-generated research hypothesis, automatically tested. Not peer-reviewed.
Idea description
Build a neural operator from frozen ambient mechanism blocks and a geometry-specific algebraic constraint adapter. The adapter parameterizes all outputs in the affine set satisfying sampled linear constraints exactly, so the network never produces boundary-violating states and does not require a penalty coefficient or post-step projection.
Formulas
Mathematical statement
The paper describes a target-domain admissible affine manifold defined by linear trace constraints and evolves dynamics in its mass-orthonormal tangent coordinates. Let $u\in\mathbb{R}^n$ be the discretized field, $C\in\mathbb{R}^{m\times n}$ the sampled boundary or trace matrix, $b\in\mathbb{R}^m$ the prescribed boundary values, and $M\succ0$ the mass matrix. Choose a particular state $u_p$ satisfying $Cu_p=b$ and a basis $V\in\mathbb{R}^{n\times(n-r)}$ for the nullspace of $C$, so $CV=0$. The required mass-orthonormal tangent coordinates satisfy $V^TMV=I$. Every admissible field is $u=u_p+Vz$, where $z\in\mathbb{R}^{n-r}$. If $F_i^\theta$ are reusable ambient mechanism blocks and $F=\sum_i\alpha_iF_i^\theta$ is their assembled vector field, the reduced dynamics are $\dot z=V^TMF(u_p+Vz)$, equivalently $\dot u=V\dot z$ on the tangent space. The construction exploits the exact invariant $Cu(t)=b$; unlike alternating unconstrained evolution and projection, it evaluates the mechanism directly in admissible coordinates.
Implementation notes
Integrate this at the output and recurrent-update interface of a neural operator or latent PDE world model. Discretize the field on a fixed ambient grid with $n$ values. For each geometry, construct the sampled trace matrix $C$ and boundary vector $b$; compute a least-squares particular solution $u_p=C^TM^{-1}(CM^{-1}C^T)^{-1}b$ and a nullspace basis $V$ using an SVD or QR factorization of $C$, followed by weighted orthonormalization so that $V^TMV=I$. Pretrain ambient mechanism networks $F_i^\theta:\mathbb{R}^n\to\mathbb{R}^n$ on source geometries, then freeze them. At inference or fine-tuning time, encode the initial field as $z_0=V^TM(u_0-u_p)$, assemble $F(u)=\sum_i\alpha_iF_i^\theta(u)$, and update only $z$: u = up + V @ z; du = sum(alpha[i] * Fi(u) for i); z = z + dt * V.T @ M @ du. Decode with u=up+V@z at every step. Learn scalar or vector mechanism weights $\alpha_i$ in reduced coordinates if desired, but do not learn a soft boundary penalty. Compute $u_p$ and $V$ from the algebraic construction; estimate numerical conditioning, rank, and truncation tolerances empirically. First test on a small Fourier/MLP operator or graph neural operator for a 2-D diffusion or advection equation, pretraining on rectangles and evaluating on circles, annuli, and irregular masks. Compare against the same frozen blocks with unconstrained outputs plus projection, and against soft Dirichlet penalties, using equal parameters, rollout steps, and FLOPs. Pre-register these mechanism tests: (1) the maximum boundary residual $\max_k\|Cu_k-b\|_2$ must remain below $10$ times the SVD tolerance, while the penalty baseline should vary with its penalty coefficient; (2) at equal rollout length, reduced-coordinate updates should have lower long-horizon energy drift than projection-after-step updates, measured by $|\|u_k\|_M-\|u_0\|_M|/\|u_0\|_M$ when the test PDE conserves mass; (3) zeroing the nullspace adapter while retaining the same mechanism blocks must increase boundary residuals by at least two orders of magnitude. Falsify the transfer if exact residuals are not achieved, or if reduced coordinates provide no stability or unseen-geometry benefit after accounting for the QR or SVD overhead.
Verification
Stage 1 · Toy mechanism gate: Passed ✓
Stage 2 · Mechanism transferred to benchmark: Confirmed ✓
Stage 2 · Practical benchmark result: Significantly worse ✗
Stage 1 — Mechanism check agent confidence 9/10
Built a weighted nullspace-coordinate adapter and deterministic constrained linear rollout in experiment.py. The algebraic claims were verified numerically: ||Cu_p-b||=5.6e-17, ||CV||=5.0e-17, and ||VᵀMV-I||=4.6e-15; reduced and M-metric projected one-step updates differed by only 2.7e-16. The exact-invariance effect is real, but no stability advantage over projection was observed: both had identical large Euler energy drift, while soft penalties retained substantial residuals and unconstrained evolution violated the constraints severely.
- Agent confidence
- 9/10
- Baseline
- M-metric projection: max constraint residual 1.49e-15; max relative mass-norm drift 43.27. Soft penalty λ=30: residual 2.47; unconstrained ambient: residual 2.12e3.
- Idea
- Reduced coordinates: max constraint residual 4.46e-15; max relative mass-norm drift 43.27. One-step difference from M-metric projection 2.74e-16.
Limitations: Only a small deterministic linear toy system was tested; no neural operator, learned frozen mechanism blocks, unseen geometries, FLOP/runtime comparison, or PDE accuracy experiment was performed. Explicit Euler was unstable for the chosen skew dynamics, so the test does not establish long-horizon stability benefits beyond exact constraint preservation.
How to run: python3 experiment.py
Stage 2 — Benchmark vs. tuned baseline
Significantly WORSE than the tuned baseline (+245.1%, p=0.0081, wins 0 / 8 paired seeds).
- Benchmark
- Poisson–Dirichlet process
- Model
- mlp_tiny
- Paired seeds
- 8
- Baseline mean
- 0.0002
- Idea mean
- 0.0008
- Effect (Δ)
- 0.0006 (+245.1%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 0.0081 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- ±72.4%
- Mechanism
- Confirmed ✓
- Practical effect
- Hurts
- Baseline tuning
- swept over 3 configs
Only the registered poisson_dirichlet PDE track and mlp_tiny model were tested. Unseen geometries, recurrent rollouts, nonhomogeneous affine constraints, frozen multi-block mechanisms, runtime/FLOP overhead, and larger neural operators were not evaluated.
How to run: python3 bench_stage2.py
Verdict computed by deterministic test code from paired-seed statistics — not by the language model.
Stage 2 — Benchmark latest Failed ✗
Agent confidence: 10/10
The nullspace-coordinate MLP exactly enforced the Dirichlet constraints on trained benchmark models, confirming the mechanism signature, but it significantly worsened the independent test MSE relative to the same MLP predicting the full field. The registered PDE benchmark therefore rejects the claimed overall win.
- Baseline
- test MSE mean 0.00022453273413702846, std 0.00002826627803497815
- Idea
- test MSE mean 0.0007747690397081897, std 0.00018247493264257159
- Paired delta
- +0.0005502363055711612; permutation p=0.0081; idea wins 0/8
Benchmark result
Significantly WORSE than the tuned baseline (+245.1%, p=0.0081, wins 0 / 8 paired seeds).
- Benchmark
- Poisson–Dirichlet process
- Model
- mlp_tiny
- Paired seeds
- 8
- Baseline mean
- 0.0002
- Idea mean
- 0.0008
- Effect (Δ)
- 0.0006 (+245.1%; negative = idea better)
- Wins
- 0 / 8 paired seeds
- p-value
- 0.0081 (permutation test, 20 000 shuffles)
- Smallest detectable effect
- ±72.4%
- Mechanism
- Confirmed ✓
- Practical effect
- Hurts
- Baseline tuning
- swept over 3 configs
Only the registered poisson_dirichlet PDE track and mlp_tiny model were tested. Unseen geometries, recurrent rollouts, nonhomogeneous affine constraints, frozen multi-block mechanisms, runtime/FLOP overhead, and larger neural operators were not evaluated.
How to run: python3 bench_stage2.py
Artifacts
- bench_report.json 3.6 KB View Raw JSON
- bench_stage2.py 5.4 KB View Raw JSON
- experiment.py 4.3 KB View Raw JSON
- report.md 1.5 KB View
- report_bench_2026-09-02T121428.md 5.2 KB View
- results.json 2.2 KB View Raw JSON