Exact Regularized Determinants for Low-Rank Clifford Integral Operators

Many physical fields mix quantities of different geometric type. Fluid flow couples a scalar pressure to a vector velocity, and electromagnetism couples a vector electric field to a bivector magnetic field. Clifford (geometric) algebra packs these into one object, a multivector, with a single product that respects how each part transforms. Neural operators that work this way, meaning networks that map whole functions to functions instead of grid values to grid values, have been more accurate than channel-stacked networks on fluid and Maxwell benchmarks. They have one practical gap: nobody can cheaply measure their spectrum, meaning the eigenvalue-like quantities that decide whether a learned solution operator amplifies or damps errors over repeated application. Scalar function-space operators have a mature tool for this, the Hilbert–Carleman determinant. It is a version of the determinant that stays finite even when the ordinary trace of the operator diverges. That tool has never been written for operators whose kernels are multivector-valued.

The proposal is to derive and use one identity. If a Clifford integral operator is built from a low-rank kernel, a sum of $r$ products of multivector-valued functions, then its regularized determinant and all its power traces equal those of a small $r\times r$ matrix of multivectors. That small matrix can be evaluated using only products and the scalar-part projection of the algebra. The resulting quantities are exact, independent of the mesh, the same across metric signatures, and differentiable. We then test whether a grade-preserving low-rank operator trained with a penalty on this spectral quantity beats both Clifford Fourier operators and channel-stacked low-rank operators on long rollouts.

Core Claim

Three bodies of work meet here. Shirokov and coauthors showed that characteristic polynomials, determinants, inverses and rank of a single multivector can be computed without matrices. The key step is that the matrix trace of a representation equals, up to a constant, the scalar (grade-0) part $\langle u\rangle_0$. This lets the Faddeev–LeVerrier recurrence, which builds characteristic polynomial coefficients from traces of powers, run entirely inside the algebra (Shirokov et al., Basis-Free Formulas for Characteristic Polynomial Coefficients, On Rank of Multivectors). That work is strictly finite-dimensional. Hà Quang Minh extended log-determinant divergences to Hilbert–Schmidt operators through the Hilbert–Carleman determinant $\det_2(I+A)=\det(I+A)e^{-\mathrm{Tr}A}$. That work is strictly scalar. SVD-NO learns PDE solution operators with a continuous low-rank kernel $\kappa(z,z')\approx\Phi(z)\Sigma\Psi(z')^\top$, but it treats fields as unstructured channels.

We should be honest about how deep the first step goes. An operator on multivector-valued functions is, abstractly, an operator on an ordinary vector-valued Hilbert space. The existence of a Clifford Carleman determinant is therefore classical once a representation is fixed. The new content is threefold: a representation-free formula, an exact finite reduction for low-rank kernels, and a use for it.

Take the kernel to be $k(x,y)=\sum_{i,j=1}^{r}\phi_i(x)\,S_{ij}\,\psi_j(y)^\dagger$. Here $\phi_i,\psi_j$ are multivector-valued functions on the domain $\Omega$, $S$ is an $r\times r$ matrix of multivectors (a diagonal $S$ recovers an SVD-like form), and $\dagger$ is Shirokov's internal Hermitian conjugation built from reversion and grade involution (On SVD and Polar Decomposition). Define the multivector Gram matrix $G_{ji}=\int_\Omega \psi_j(x)^\dagger\phi_i(x)\,dx$. We will prove:

$$\mathrm{Tr}(\mathcal{K}^m)=c_{p,q}\,\Big\langle \mathrm{tr}\big((SG)^m\big)\Big\rangle_0,\qquad \log\det{}_2(I+\mathcal{K})=\sum_{m\ge 2}\frac{(-1)^{m+1}}{m}\,\mathrm{Tr}(\mathcal{K}^m),$$

where $\mathrm{tr}$ sums the diagonal entries of the $r\times r$ multivector matrix and $c_{p,q}$ is the representation dimension, which we remove by working with the normalized trace. The right-hand series is the Plemelj–Smithies expansion.

The series then collapses to a finite Faddeev–LeVerrier recurrence on $SG$. The cost is $O(r^3)$ multivector products after one pass over the mesh to form $G$. The Hilbert–Schmidt continuity of $\det_2$ (Hà Quang Minh) gives the second theorem: if the truncated kernels converge in Hilbert–Schmidt norm, the computed determinants converge to the full-rank operator's value.

Because every step uses only the algebra's product and $\langle\cdot\rangle_0$, the same code covers Euclidean $C\ell_{3,0}$, spacetime signatures, and degenerate projective algebras without the signature-dependent case analysis that matrix representations need. The grade-0 part of a product also splits into contributions from matching grade pairs. This gives something no channel-based operator offers: a readout of how much of the operator's spectral mass flows through its scalar, vector and bivector parts.

Research Program

The theoretical part has two pieces. The first is the core reduction above together with its truncation-error bound. The second extends Shirokov's intrinsic rank theorem, which says rank equals the index of the highest nonzero characteristic coefficient of $M^\dagger M$ (On Rank of Multivectors), to the operator $\mathcal{K}^\dagger\mathcal{K}$. This yields an intrinsic operator rank computed from $(G_\phi S G_\psi S^\dagger)$, where $G_\phi$ and $G_\psi$ are the two Gram matrices.

We will not try to solve Shirokov's open problem of a constructive native SVD in general. The architecture does not need one: orthonormality of $\phi,\psi$ is imposed softly through an SVD-NO-style Gram penalty, now measured with the internal inner product $\langle a^\dagger b\rangle_0$.

The computational part is a grade-preserving low-rank Clifford neural operator. Its basis functions $\phi_i,\psi_j$ are produced by Clifford convolutions or Clifford-FNO blocks (Clifford Neural Layers), so grades never mix through arbitrary linear maps, which is the problem Geometric Clifford Algebra Networks identified. The model applies the kernel at cost linear in mesh size, as SVD-NO does.

The training objective adds a spectral penalty on the one-step map. The penalty is the normalized Schatten-$2m$ norm $\big(\mathrm{Tr}((\mathcal{K}^\dagger\mathcal{K})^m)\big)^{1/2m}$, which approaches the operator norm as $m$ grows. It is computed exactly from the core and is optionally complemented by $\log\det_2$ to discourage near-singular layers. The working hypothesis is that controlling this quantity mesh-independently improves long autoregressive rollouts. The Clifford structure should make the control cheaper and better targeted, because it can be applied grade by grade.

Baselines

The naive reference point is the persistence forecast, which copies the last state forward, together with the plain FNO of Li et al.. The file reports only relative gains for FNO over earlier learned solvers (30% lower error on Burgers, 60% on Darcy); no absolute reported error on our target tasks was found on Lacuna. FNO is the reference because PDEBench found it consistently ahead of U-Nets and PINNs (PDEBench).

The strongest structure-aware competitor is the Clifford FNO of Brandstetter et al.. It reported roughly 15–30% lower MSE than matched-parameter standard baselines on 2D Navier–Stokes and shallow water, with about 2× training time. The strongest low-rank competitor is SVD-NO. It reported state-of-the-art results on four of five PDENNEval tasks, including 17.8% lower error than the best baseline on 2D shallow water and 25% on 1D Allen–Cahn, and an epoch time drop from 176 s to 2.3 s on diffusion–sorption. Absolute errors for these baselines are not in the file, so we will reproduce all of them ourselves.

Every model gets the same training trajectories, a parameter count matched within 5%, the same GPU-hour budget, and the same number of hyperparameter trials drawn from a shared search protocol.

Two ablations isolate the key ingredients. The first, a channel-stacked core, is identical to our model except that the multivector core and basis functions are flattened into real channels with ordinary matrices. It isolates grade structure. The second removes the spectral penalty, or replaces it with standard spectral normalization on discretized weights. It isolates the value of exact, mesh-independent control. A third check, which is not a competitor, computes the same spectral quantities through an explicit matrix representation and LAPACK. It must agree to numerical precision, and we report its memory and speed.

Beating the field means lower one-step error than Clifford FNO and SVD-NO on at least the multivector-native tasks (Navier–Stokes with smoke, shallow water, Maxwell). It also means clearly lower error growth over long rollouts than the channel-stacked ablation at equal cost. Training cost must be closer to SVD-NO than to Clifford FNO's 2× overhead.

Why not just…?

Why not just discretize and take the matrix determinant of the grid operator? That quantity depends on the mesh. It costs $O((N\,2^{n/2})^3)$ for $N$ grid points and representation size $2^{n/2}$, and it has to be recomputed at every resolution. The core formula is exact in function space and costs $O(r^3)$. Our first check measures both paths on refined grids. The grid determinant should converge to the core value, which validates the theorem while showing the grid route is only a slow approximation of it.

Why not just build the matrix representation of the small core and call LAPACK? This is a real alternative, and numerically it gives the same answer. We treat it as the correctness oracle, not a rival. The case for the basis-free route is not speed. It is that one implementation covers every signature, that it avoids materializing real, complex or quaternionic representations case by case, and that it exposes the grade-resolved decomposition of the trace. If the representation route turns out faster in practice, we will use it inside training and keep the basis-free form for theory and diagnostics. The proposal's value does not depend on that choice.

Why not just add spectral normalization to channel-stacked SVD-NO? This is the most serious threat to the empirical claim. A channel-stacked low-rank kernel already has an $r\times r$ core, so a Sylvester-identity reduction applies to it too. What it lacks is grade structure. The channel-stacked ablation with the same exact penalty settles this directly. If it matches our model in rollout stability and accuracy, then the benefit comes from low-rank spectral control alone, and the Clifford claim fails. The theorem would still stand as mathematics.

Why not just use classical Fredholm determinants? They require trace-class operators. Many learned and physical kernels, including Green's-function-like ones, are only Hilbert–Schmidt, and for those the plain product diverges (Hà Quang Minh). The Carleman form is the natural target, and we verify empirically that the learned kernels' power traces behave accordingly.

Proposed Experiments

Verification comes first. On analytically chosen multivector kernels in $C\ell_{3,0}$, $C\ell_{1,3}$ and the projective $\mathbb{G}_{3,0,1}$, we compare the core-based $\log\det_2$ and Schatten norms against grid-matrix computations at resolutions from $32$ to $512$ per axis. We also measure truncation error as a function of $r$ against the Hilbert–Schmidt bound.

The learning experiments use Brandstetter et al.'s 2D Navier–Stokes with smoke, 2D shallow water and 3D Maxwell tasks. They add PDEBench 2D shallow water and compressible Navier–Stokes, scored with RMSE, conservation RMSE and frequency-spectrum RMSE (PDEBench), and the PDENNEval tasks used by SVD-NO, so the comparison is like for like. Rollout stability is the headline metric, measured as error growth over trajectories several times longer than the training horizon. We will also test zero-shot resolution transfer, training at $64^2$ and evaluating at $256^2$, where mesh-independence of the penalty should matter most.

As a scale-up, we will run on ShapeNet Car against Fengbo. The file reports only that Fengbo beat five of six literature baselines with 42M parameters, 60% fewer than GINO, and no absolute error is available.

The decisive check is the pair of comparisons between the full model and the channel-stacked ablation, and between the exact spectral penalty and standard spectral normalization. The project also reports the grade-resolved spectral mass of trained operators on Maxwell, where the physics predicts that vector-to-bivector coupling dominates. This tests whether the diagnostic sees what the equations say it should.