Resolution-Independent Contractive Fourier Blocks for Stable Invertible Operators

Invertible networks let a single model run forward (parameters to observations) and backward (observations to parameters). This matters for physical inverse problems and for normalizing flows, which are generative models that turn data into Gaussian noise through a reversible map. One clean way to build them is the invertible residual network. Each layer computes $y = x + F(x)$, and $F$ is forced to shrink distances, meaning its Lipschitz constant (the largest factor by which it can stretch the gap between two inputs) is below one. Under that condition the layer can be undone by repeatedly applying $x \leftarrow y - F(x)$. So far, every such network builds $F$ from small local convolutions or from quadratic-cost attention. The result sees only a small neighborhood, and its stability guarantee depends on the input size.

We propose building $F$ from a global Fourier filter: multiply the signal's frequency components by a learned gain, then transform back. The gain is a continuous function of frequency with a known ceiling. This choice has a property that neither ingredient has on its own. A filter bounded by $c$ at every continuous frequency is also bounded by $c$ at whatever frequencies a particular grid happens to sample. So the contraction guarantee, and with it a guarantee that the inverse is well conditioned, holds at every resolution, including ones never seen in training. This gives one block with a global receptive field at $O(N\log N)$ cost, a stability bound that does not grow with pixel count, and an inverse whose sensitivity to noise is provably limited at any discretization.

Core Claim

A residual block whose nonlinear part is a frequency-parameterized, magnitude-capped spectral gate is bi-Lipschitz with constants $1-L$ and $1+L$, where $L<1$ is independent of grid size. This yields invertible operators that stay stable when inverted under noise and at unseen resolutions. Coupling-based invertible Fourier operators are also exactly invertible at any resolution, but their inverse stretch is unbounded, and that stretch is what amplifies noise in ill-posed inverse problems.

The claim is framed as stability rather than invertibility for a specific reason. Invertible Fourier Neural Operators (iFNO) splits channels into halves and lets one half update the other through a Fourier layer. That construction is exactly invertible on any grid, so "we can invert at new resolutions" would not be new. What iFNO lacks is a bound on how much its inverse can amplify perturbations. Its coupling scales pass through Softplus/exponential transforms with no ceiling, and iFNO never tested the inverse pass across discretizations. Our block supplies that missing bound.

The Block

Let $x$ be a $C$-channel field on a grid and $\hat{x}(\xi)$ its orthonormal discrete Fourier transform at frequency $\xi$. We define

$$F(x) = \sigma\!\left(\mathcal{F}^{-1}\Big[\textstyle\sum_{k=1}^{K} g_k(\xi;\theta_k)\, W_k\, \hat{x}(\xi)\Big]\right),$$

where the terms are as follows:

  • $g_k(\xi;\theta_k)\in[0,1]$ is an anisotropic Butterworth-style gate, a smooth low/band-pass profile with a learned cutoff, order and orientation, taken from Butterworth as Attention.
  • $W_k$ are $C\times C$ channel-mixing matrices.
  • $\sigma$ is LipSwish, the 1-Lipschitz activation from Residual Flows.

The orthonormal transform is an isometry. The multiplier's operator norm is therefore the supremum over sampled frequencies of $\|\sum_k g_k W_k\|_2$. This is at most $\sum_k \sup_\xi |g_k|\,\|W_k\|_2 \le \sum_k \|W_k\|_2$, a quantity that does not depend on $N$. We rescale so that

$$\mathrm{Lip}(F) \le \sum_{k} \|W_k\|_2 \le L < 1,$$

using exact $C\times C$ spectral norms. These matrices are tiny, so no power-iteration estimate over a full convolution is needed.

Several facts follow from this construction. Zero padding and cropping, which FNO-style models use on non-periodic domains, do not increase norms, so the bound survives them. The layer $x+F(x)$ is strongly monotone with constant $1-L$. That is exactly the condition Can neural operators always be continuously discretized? identifies as escaping its no-go theorem on discretizing diffeomorphisms. Their result therefore constrains the design: we must keep the strongly monotone structure and not relax it to general diffeomorphisms. It also tells us what to test.

The block also fixes two limitations of The Lipschitz Constant of Self-Attention. That paper built contractive $L^2$ attention, but its bound grows as $O(\log N)$, its cost is $O(N^2)$, and it requires tied query and key weights. Our bound is flat in $N$, and our gates are anisotropic, so we do not inherit the symmetry restriction. The theory is modest and honest about its origin. The key inequality is the elementary fact that sampling a bounded function cannot exceed its bound. The contribution is to exploit it architecturally and test it.

Research Program

We will implement the block inside two host systems, reusing everything else unchanged.

The first host is an invertible PDE operator, the primary setting, built on the iFNO codebase. iFNO's latent lifting, its $\beta$-VAE posterior head and its data pipeline stay the same. Only its coupling Fourier blocks are replaced by our contractive residual blocks. We keep iFNO's lifting because raw fields cannot be inverted without information loss, and we deliberately retain its multi-stage training so that the comparison isolates the block.

The second host is a Residual Flow on CIFAR-10. Here some local convolutional residual blocks are swapped for spectral-gate blocks. We use the unbiased log-determinant estimator and memory-saving gradient from Residual Flows. This checks that the block is competitive in density estimation and is not tuned to PDEs alone.

Three analyses accompany the builds. The first measures the empirical inverse Lipschitz constant, estimated by adversarial perturbation of outputs, as grid resolution varies. The second counts the fixed-point iterations needed to reach a fixed reconstruction error; contraction at rate $L$ predicts roughly $\log(\epsilon)/\log(L)$ iterations, independent of $N$. The third examines the learned gates, to see whether global low-frequency routing is actually used or whether the model collapses to local behavior.

For the inversion latency problem we will also test one inexpensive idea. The linear part of the block is diagonal in frequency, so $I + \sum_k g_k W_k$ can be inverted exactly per frequency and used as a warm start for the fixed-point iteration.

Baselines

The naive reference point is a plain FNO trained separately in each direction. It has no invertibility and no stability guarantee. Invertible Fourier Neural Operators reports a relative error of about 0.17 for it on the Darcy D-LINE inverse task at 20% noise. Any invertible method must beat this to justify its constraints.

The strongest existing competitor is iFNO itself. On the same Darcy D-LINE task at 20% noise it reports a relative error of about 0.11, and 30–50% lower inverse error than FNO, iDON and NIO across its benchmarks. It is the closest method: Fourier routing plus exact invertibility, differing from ours only in using coupling instead of contraction. We will run it from the official code, with our block matched in latent width, depth and parameter count, trained on identical data splits and noise levels, with the same hyperparameter search budget per method. Beating it means lower inverse relative error at matched noise on the training resolution. Just as important, we must show a clearly smaller degradation when training at one resolution and testing inverse recovery at 2× coarser and 2× finer grids, a regime iFNO never reported.

For density estimation the reference is Residual Flows at 3.28 bits/dim on CIFAR-10, with i-ResNet at 3.45 (Invertible Residual Networks). We rerun Residual Flows at matched parameters and wall-clock. Matching 3.28 with fewer blocks or fewer inversion iterations would count as success. We are not counting on a large likelihood gain.

Three ablations isolate the key ingredient. The first replaces the spectral gate with a spectrally normalized local 3×3 convolution at equal parameters, which tests whether the global receptive field matters. The second replaces the continuous parametric gates with free per-mode weights capped only on the training grid. This tests whether continuous-frequency parameterization is what buys resolution transfer, because free per-mode weights have no defined value at new frequencies. The third removes the magnitude cap, which tests whether contraction, and not merely Fourier routing, drives the stability gains.

Why not just…?

Why not just use iFNO's coupling layers, which are already invertible at any resolution? This is the most serious threat, and we concede that exact invertibility across grids is not our advantage. The question is conditioning. Coupling inverses divide by learned scales with no lower bound, so noise in the observation can be amplified arbitrarily. Our block's inverse stretch is at most $1/(1-L)$ per layer at every resolution. The cross-resolution noisy-inversion experiment settles this. If iFNO's empirical inverse Lipschitz constant and error stay as flat as ours across resolutions, our central claim fails, and we would report that.

Why not just stack deeper spectrally normalized local convolutions? Each contractive layer both limits signal growth and costs roughly 15–20 fixed-point iterations to invert, the range The Lipschitz Constant of Self-Attention reports for its blocks. Reaching global context through depth therefore multiplies inversion cost. The local-convolution ablation at matched parameters and matched inversion budget measures this directly. Darcy flow, where the solution at a point depends on the whole coefficient field, is a setting where we expect locality to hurt.

Why not just use contractive $L^2$ self-attention? It is the only existing contractive global operator, but its $O(\log N)$ bound forces weaker blocks as grids grow, and its $O(N^2)$ cost is prohibitive on PDE grids. Its evaluation was limited to synthetic data and Penn Treebank. We will include it on the coarsest Darcy grid where memory allows. Its bound-versus-resolution curve alongside ours is itself a result.

Why not just train an unconstrained Fourier model and clip gradients? Without a norm ceiling the fixed-point inverse has no convergence guarantee. The Lipschitz-attention paper found that unconstrained dot-product attention "cannot be reliably inverted." Our no-cap ablation tests the Fourier analogue.

Decisive Checks

Everything runs on the iFNO benchmark suite (Darcy flow, Navier–Stokes, wave propagation) at clean, 10% and 20% noise. We add a resolution-transfer protocol: train at one grid, then invert at coarser and finer grids.

The experiment that most directly decides the proposal is a plot of inverse relative error and empirical inverse Lipschitz constant against test resolution for our block, iFNO and the per-mode ablation. The proposal succeeds if our curve is flat while the others rise. It fails if iFNO's is equally flat, or if our block's contraction cap costs so much capacity that its in-distribution error exceeds iFNO's roughly 0.11 by a wide margin.

Secondary checks are three. First, the fixed-point iteration count should stay constant across resolutions, as the $N$-independent bound predicts. Second, CIFAR-10 bits/dim should be within noise of the 3.28 reference at matched compute. Third, gate inspection should show nontrivial low-frequency, anisotropic routing, not collapse to a near-identity filter.