Many neural networks for physics represent a field as a multivector: a single algebraic object that bundles a scalar (pressure), a vector (velocity) and an oriented plane (a bivector, such as a magnetic field) and multiplies them with one rule, the geometric product. Clifford Fourier neural operators are the main example. They learn, for each Fourier mode, a multivector that multiplies the field. These models make fewer errors than channel-stacked baselines. Nobody, however, can currently say how much a trained layer amplifies or loses information, or how to invert it. Those two properties decide whether long autoregressive rollouts blow up and whether the learned operator can be run backwards for inverse problems.
This note proposes to make both properties computable exactly, cheaply, and without converting multivectors into matrices. The argument has two steps. Frame theory, the theory of stable, possibly redundant ways to discretize operators, shows that the conditioning of a discretized operator is controlled by the frame bounds times the extreme singular values of the operator itself. For a layer that multiplies by multivectors mode by mode, those extreme singular values are the singular values of individual multivectors. Recent coordinate-free Clifford algebra supplies formulas for them: characteristic polynomials built from scalar parts, and native inverses. Put together, these give a per-mode, differentiable certificate of Lipschitz constant and invertibility for any metric signature. They also give a native Moore–Penrose pseudo-inverse for singular multivectors, which the algebra literature names as open.
Core Claim
Let $G = T_\Phi^* U T_\Psi$ be the cross Gram matrix of an operator $U$ discretized by frames $\Phi,\Psi$ with bounds $A \le B$. Here $T_\Psi$ synthesizes a field from coefficients and $T_\Phi^*$ analyzes it back. Suppose $U$ acts by multiplying the field by a multivector $u_\xi$ at each Fourier mode $\xi$. Then the nonzero singular values of $G$ are bracketed by the frame bounds and the singular values of the multipliers:
$$ A\,\inf_\xi \sigma_{\min}(u_\xi) \;\le\; \sigma^{+}(G) \;\le\; B\,\sup_\xi \sigma_{\max}(u_\xi). $$
In this display, $\sigma(u_\xi)$ denotes the singular values of left multiplication by $u_\xi$, and $\sigma^{+}$ denotes the nonzero singular values of $G$ restricted to the range of the analysis operator.
The new ingredient is that $\sigma(u_\xi)$ is computed inside the algebra. Use the algebra's own Hermitian conjugation $u^\dagger$ (On SVD and Polar Decomposition in Real and Complexified Clifford Algebras). Form the characteristic polynomial of $u^\dagger u$ through the scalar-part version of the Faddeev–LeVerrier recurrence, using the identity $\mathrm{Tr}(\gamma(v))/2^{\lfloor n/2\rfloor} = \langle v\rangle_0$ (On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension). The roots of that polynomial are the squared singular values. The polynomial has minimal degree, so the forced repeated roots of a regular matrix representation never appear. This matters for gradients, as explained below.
The same coefficients give a native regularized pseudo-inverse $u^{+}_\lambda = (u^\dagger u + \lambda)^{-1}u^\dagger$, built with the basis-free adjugate/determinant inverse. As $\lambda \to 0$, and with rank read off from the highest nonzero coefficient (On Rank of Multivectors in Geometric Algebras), this yields an exact Moore–Penrose pseudo-inverse for multivectors with $\mathrm{Det}(u)=0$.
How This Sits Against Prior Work
Two bodies of work motivate the proposal, and neither touches the other.
The frame side supplies the bounding method. Matrix Representation of Operators Using Frames defines frame matrices $\langle O\psi_k,\phi_m\rangle$. U-cross Gram matrices and their invertibility proves invertibility exactly for Riesz sequences, gives pseudo-inverse formulas through dual frames that lose uniqueness when $U$ is not invertible, and gives perturbation bounds in terms of frame bounds. Representation of Operators Using Fusion Frames extends this to subspace systems and shows that Overlap-Add is a diagonal fusion-frame representation of convolution. All of this assumes scalar Hilbert spaces and computes inverses abstractly. We reuse the bounds, but specialize the operator to multivector multiplication, so that the abstract $\|U\|$ and $\|U^{-1}\|$ become explicit per-mode algebraic quantities.
The Clifford side supplies the native computations. Shirokov's determinant/inverse paper above, his characteristic-polynomial work (Basis-Free Formulas for Characteristic Polynomial Coefficients in Geometric Algebras), and his rank theorem give exact formulas but no application to operators. His SVD paper proves existence only and states that native numerical algorithms are lacking. We do not attempt a general native SVD iteration. We need only the singular values of a single multivector and a regularized pseudo-inverse, both of which follow from characteristic coefficients of $u^\dagger u$.
The neural-operator side defines the application regime. Clifford Neural Layers for PDE Modeling and Fengbo use multivector Fourier multipliers but impose no stability or invertibility guarantees. Are Neural Operators Really Neural Operators? uses frame theory to characterize aliasing, not conditioning. Transfer Operator Learning with Fusion Frame uses fusion frames as a feature extractor without algebraic guarantees.
One constraint deserves to be stated directly. Faddeev–LeVerrier recurrences are numerically fragile at large degree, and the explicit formulas explode combinatorially (231 summands at $n=8$). We therefore restrict to $n\le 5$, where the polynomial of $u^\dagger u$ has degree at most 4. We test float32 accuracy explicitly rather than assume it.
Research Program
The theoretical part proves the bracket above for Fourier tight frames, Gabor frames and fusion frames. For fusion frames, Overlap-Add on non-periodic domains is the concrete case. The proof also covers the pseudo-inverse: $G^{+}$ is bounded through the native $u^{+}_\lambda$ composed with dual frames. It further derives closed-form singular values for $Cl_{3,0}$, $Cl_{0,3}$, $Cl_{1,3}$ and $Cl_{4,1}$ from the low-degree characteristic polynomials.
The software part is a batched, differentiable PyTorch kernel. Given multipliers $u_\xi$, it returns $\sigma_{\max}$, $\sigma_{\min}$ and $u^{+}_\lambda$ using only geometric products, conjugations and scalar parts.
The modeling part uses the kernel in two ways. The first is a training-time constraint: each multiplier is rescaled so that $\sigma_{\max}\le 1$ and $\sigma_{\min}\ge \epsilon$. The second is a closed-form inverse layer for recovering inputs from outputs.
Baselines
The naive reference is an unconstrained Clifford Fourier neural operator, trained exactly as in Clifford Neural Layers for PDE Modeling on its 2D Navier–Stokes, 2D shallow-water and 3D Maxwell tasks. That paper reports that Clifford models reduce error by "roughly 15-30%" on the 2D fluid tasks relative to standard networks, and that training takes about twice as long. Those are the numbers our constrained models must not degrade. A standard real-channel FNO run on the same data anchors the comparison from below.
The strongest competitor for the computational claim is the matrix route. It maps each multiplier to its minimal matrix representation (complex $2\times2$ for $Cl_{3,0}$, quaternionic for $Cl_{1,3}$ realized as complex $4\times4$) and calls batched torch.linalg.svd and pinv. This is the obvious engineering answer and is already cheap at small $n$. The strongest competitor for the modeling claim is global spectral normalization by power iteration on the whole Fourier layer, a standard way to bound Lipschitz constants.
All methods share the same architecture, data splits, number of optimizer steps and hyperparameter search budget of 20 trials. The only thing that differs is how the multiplier spectrum is computed or constrained.
The ablation that isolates the key ingredient swaps the native kernel for the matrix kernel inside an otherwise identical constrained model. Holding the constraint fixed and changing only the spectrum computation separates the value of certified conditioning from the value of computing it natively.
Beating the baselines means three things on the stated tasks. First, rollout error at the longest horizon is lower than the unconstrained model's, with one-step error no worse than the unconstrained model's. Second, inverse-problem recovery error is lower than gradient-descent inversion of the unconstrained model at equal wall-clock time. Third, for the native-versus-matrix ablation, there are no gradient NaNs or blow-ups in signatures where the matrix route produces them, at no more than 1.5× its runtime.
Why not just…?
Why not just map to matrices and use LAPACK? For $Cl_{3,0}$, this probably works as well. The minimal representation is complex $2\times2$, so singular values are generically distinct and closed form, and we expect parity there. The real threat is therefore confined to that signature. Native computation is expected to pay off elsewhere. For signatures whose minimal representation is quaternionic, or a direct sum, the real or complex embedding forces singular values to repeat. SVD gradients scale like $1/(\sigma_i-\sigma_j)$, so the backward pass becomes unstable precisely at these forced repeats. The matrix route also has to be rewritten for each signature, as the SVD literature notes (On Singular Value Decomposition and Polar Decomposition in Geometric Algebras). The ablation runs both routes in $Cl_{1,3}$ and $Cl_{4,1}$. If the matrix route shows no gradient pathology there, we drop the "native" claim and keep only the conditioning bound and the pseudo-inverse.
Why not just spectral-normalize with power iteration? Power iteration bounds only the largest singular value of the whole layer. It says nothing about the smallest, which is what governs invertibility and information loss. It also needs several iterations per step, and those iterations are only approximate. The inverse-problem experiment tests this directly: a power-iteration-normalized model has no closed-form inverse and must be inverted by optimization.
Why not just train unconstrained and rely on data? It is possible that conditioning never limits accuracy on these benchmarks. The rollout experiment will show this if constrained and unconstrained models diverge at the same horizon. In that case the contribution narrows to inverse problems and to certification as a diagnostic.
Proposed Experiments
The first experiment is a numerical micro-benchmark. We draw random and adversarially near-singular multivectors in $Cl_{3,0}$, $Cl_{0,3}$, $Cl_{1,3}$ and $Cl_{4,1}$. For each, we compare native and matrix singular values, inverses and pseudo-inverses against float64 references. We report float32 relative error, the norm of gradients with respect to $u$, and throughput per million multipliers. Near $\mathrm{Det}=0$ we specifically check the native pseudo-inverse against a float64 SVD pseudo-inverse.
The second experiment checks the bound empirically. We discretize trained multipliers with a Fourier ONB, a redundant Gabor frame, and an Overlap-Add fusion frame on a non-periodic crop. We then confirm that measured singular values of the assembled $G$ fall inside the predicted bracket, and report how tight it is.
The third experiment covers the modeling tasks. On the Navier–Stokes, shallow-water and Maxwell tasks from Clifford Neural Layers for PDE Modeling, we train constrained and unconstrained models and measure one-step and long-rollout relative $L_2$ error. On Maxwell, we additionally use the native pseudo-inverse layer to recover the initial fields from later states, and compare against gradient-based inversion. Ahmed Body and ShapeNet Car (Fengbo) serve as a scale check. For these, no reported numeric error is found on Lacuna for Fengbo itself, so we rerun its released configuration as the reference.
The decisive checks are two. If the matrix route matches the native route on accuracy, gradient stability and speed across every signature, the native component is unnecessary and the work reduces to the bound and the pseudo-inverse. If constrained models neither extend stable rollout horizons nor enable better inversion, certified conditioning is not useful at these scales, and the result should be reported as a negative finding.
References
- On SVD and Polar Decomposition in Real and Complexified Clifford Algebras
- On computing the determinant, other characteristic polynomial coefficients, and inverse in Clifford algebras of arbitrary dimension
- On Rank of Multivectors in Geometric Algebras
- Matrix Representation of Operators Using Frames
- U-cross Gram matrices and their invertibility
- Representation of operators using fusion frames
- Basis-Free Formulas for Characteristic Polynomial Coefficients in Geometric Algebras
- Clifford Neural Layers for PDE Modeling
- Fengbo: a Clifford Neural Operator pipeline for 3D PDEs in Computational Fluid Dynamics
- Representation Equivalent Neural Operators: a Framework for Alias-free Operator Learning
- Transfer Operator Learning with Fusion Frame
- On Singular Value Decomposition and Polar Decomposition in Geometric Algebras