Trimming a neural operator per input does not buy speed: conformal calibration picks modes or widths, yet measured latency barely moves
Synopsis
Neural operators use the same Fourier mode count and channel width for every input, so cutting computation per input should save time. Conformal calibration selects modes or width per input under a fixed error budget, with an envelope covering all candidates at once. On Burgers and Darcy flow, mode truncation gave no measurable latency drop because full-grid transforms and pointwise channel mixing dominate; narrowing channels lowers Darcy forward cost, but reference-based width selection must first compute the dense prediction, and an input-only binary router that skips that pass is often matched or beaten by static widths.
Interpretation
Cutting a neural operator's computation per input does not by itself shorten inference time, and that gap is measured directly on Fourier neural operators. A Fourier neural operator uses one spectral bandwidth and channel width for every input, while PDE instances differ substantially in complexity. On Burgers, truncating modes gave no measurable latency reduction at batch sizes 1 and 20; on Darcy at batch 20, going from 12 modes to 2 lowered measured latency by 0.7%, and at batch 1 by 0.3%.
The selection rules separate error control from execution time: conformal calibration supplies the error guarantee, and reference passes, scoring, dispatch and candidate inference are accounted for separately. Earlier adaptive-operator work mostly changes spectral computation or parameter cost rather than picking a mode or channel budget per input against a prescribed error tolerance. The triangle-inequality envelope covers 99–100% of Burgers evaluation examples while keeping 31–55% of trained modes; the direct-score rule trades that slack for smaller support, with coverage of 83–99% on Burgers and 85.4–93.1% on Darcy.
An input-only binary width router avoids the dense pass, and its complete pipeline runs at 0.466–0.887 times dense latency on batched Darcy at batch size 64. Reference-based width selection must first compute the dense prediction, so every input pays that pass. On a separately generated 1,000-example Darcy test set, matched static widths had higher coverage in all eight feasible nested cells and in six of nine independent-model cells.
Perspective
The conclusions cover the tested FNO architectures, the Burgers and Darcy benchmarks, batch sizes from 1 to 128, and a single RTX A6000; coverage is computed as absolute error on each benchmark's fixed grid and is not a statement about function-space coverage across resolutions. The authors propose higher-resolution two-dimensional problems, three-dimensional or time-dependent PDEs, and fused input-only routers as the next tests of whether spectral contraction becomes a larger share of execution.
Split-conformal guarantees average over calibration and deployment draws, and with 624 calibration examples a single fitted policy's coverage can plausibly fall a few points below target; under distribution-shift stress tests, changing Burgers viscosity or widening Darcy coefficients to 1–1000 dropped coverage to 0%, because the full model's own error lies far above its calibration quantile and selecting more support or width cannot restore accuracy.
