Scalar Smoothness Is One Point on a Besov Spectrum
Organizations: The University of Tokyo · McMaster University and Vector Institute
Abstract
Studies of neural network generalization and robustness often summarize smoothness with a single number, such as a Lipschitz constant or a Hessian curvature statistic. However, Lipschitz constants summarize global worst-case sensitivity and Hessian statistics summarize local curvature; neither captures how model responses vary with the size, direction, and location of perturbations encountered in real-world deployment. To measure smoothness under these perturbations, we introduce a finite-difference roughness spectrum that profiles model responses across these conditions. Our roughness spectrum generalizes scalar smoothness measures, bringing Lipschitz sensitivity and directional Hessian curvature into a common framework across perturbation scales. We connect this spectrum to classical Besov theory, which describes how a function's variation changes with scale. Our theoretical analysis also proves that a larger output shift can cause less task damage than a smaller one, and that local curvature alone can miss the blocks most vulnerable to finite perturbations. In image classification, Roughness Source Control (RSC) suppresses margin roughness along noise directions and improves noise robustness in a comparison of 80 checkpoints across four model--dataset combinations. In post-training quantization, we identify blocks that need higher precision by measuring task-loss increases from perturbations matched to the size and direction of each block's quantization error. Our mixed-precision allocation achieves lower perplexity than the Hessian-based HAWQ-V2 baseline under the same GPTQ weights and bit budgets in all 14 settings across seven held-out large language models.
Figures & tables
| at dyadic probe scale | local-curvature baselines | ||||||||
| Model | |||||||||
| Pythia-160M | 12 | 0.94 | 0.90 | 0.87 | 0.76 | 0.58 | 2.1 | 0.43 | 0.41 |
| Pythia-410M | 24 | 0.89 | 0.88 | 0.87 | 0.85 | 0.73 | 2.0 | 0.67 | 0.69 |
| Pythia-1.4B | 24 | 0.98 | 0.94 | 0.83 | 0.56 | 0.40 | 2.0 | 0.60 | 0.76 |
| Pythia-2.8B | 32 | 0.87 | 0.77 | 0.65 | 0.53 | 0.50 | 2.0 | 0.37 | 0.64 |
| Pythia-6.9B | 32 | 0.97 | 0.96 | 0.88 | 0.80 | 0.73 | 1.9 | 0.46 | 0.83 |
| Model | HAWQ-V2 | damage | HAWQ-V2 | damage |
|---|---|---|---|---|
| Phi-3.5-mini (3.8B) | 7.70 | 7.53 | 7.45 | 7.15 |
| OLMo-2-7B | 12.01 | 11.40 | 11.02 | 10.49 |
| Granite-3.1-8B | 6.22 | 6.11 | 6.04 | 5.95 |
| Hermes-3 | 9.78 | 9.50 | 9.26 | 8.99 |
| Mistral-NeMo (12B) | 7.31 | 7.14 | 7.08 | 6.86 |
Appendix figures & tables30 assets
Supplementary material from the paper’s appendix.
Appendix
| Symbol | Meaning |
|---|---|
| Roughness spectrum | |
| Probed function: a margin, an output law, or a loss | |
| , | Base point in and data-induced base-point law |
| , | Order and -th forward difference of at step |
| , | Unit-direction law at scale and a draw from that law |
| , | Dyadic probe radius and the finite grid of scales, |
| construction | target | slope | ||
| lacunary (interior), | 0.304 | 0.004 | 5.57 | |
| lacunary (interior), | 0.500 | 0.000 | 6.22 | |
| lacunary (interior), | 0.699 | 0.001 | 7.46 | |
| lacunary (grid-edge), | 0.276 | 0.024 | 6.38 | |
| lacunary (grid-edge), | 0.442 | 0.058 | 7.14 | |
| lacunary (grid-edge), | 0.614 | 0.086 | 8.51 |
| Accuracy: clean and under perturbations | |||||
|---|---|---|---|---|---|
| Method | Clean | Corr. | Corr.(noise) | int4 | Gauss |
| Baseline | 95.31 0.03 | 74.92 0.12 | 63.36 0.27 | 94.88 0.15 | 69.39 0.31 |
| No weight decay | 92.47 0.37 | 72.27 0.26 | 66.99 1.89 | 91.49 0.53 | 76.26 2.47 |
| Label smoothing | 95.28 0.02 | 74.34 0.37 | 62.76 0.38 | 94.98 0.17 | 68.69 1.92 |
| Mixup | 95.88 0.12 | 79.46 0.49 | 67.22 1.85 | 95.39 0.15 | 73.06 3.03 |
| GroupNorm | 94.30 0.17 | 82.48 0.19 | 74.14 0.25 | 94.09 0.11 | 78.63 0.27 |
| Diagnostics on clean data | |||
|---|---|---|---|
| Method | |||
| Baseline | 0.48 0.05 | 12.39 2.85 | 12.4 1.2 |
| Label smoothing | 0.51 0.05 | 29.28 1.81 | 14.7 2.2 |
| Mixup | 0.51 0.01 | 46.70 7.37 | 14.0 0.9 |
| SAM | 0.47 0.04 | 4.85 0.52 | 10.5 0.4 |
| SAM (wall-clock-matched) | 0.48 0.01 | 4.28 0.01 | 11.3 0.0 |
| Accuracy: clean and under perturbations | |||||
|---|---|---|---|---|---|
| Method | Clean | Corr. | Corr.(noise) | int4 | Gauss |
| Baseline | 96.25 0.04 | 76.20 0.25 | 63.42 0.70 | 96.08 0.11 | 66.68 0.94 |
| No weight decay | 94.03 0.05 | 72.72 0.20 | 63.30 0.91 | 93.79 0.01 | 69.56 1.30 |
| Label smoothing | 96.01 0.03 | 75.58 0.23 | 63.48 0.94 | 95.68 0.02 | 67.28 2.33 |
| Mixup | 96.71 0.07 | 80.98 0.07 | 68.50 0.88 | 96.45 0.03 | 74.26 1.45 |
| SAM | 96.86 0.01 | 77.56 0.30 | 65.30 0.47 | 96.64 0.02 | 69.25 0.95 |
| Accuracy: clean and under perturbations | |||||
|---|---|---|---|---|---|
| Method | Clean | Corr. | Corr.(noise) | int4 | Gauss |
| Baseline | 90.31 0.22 | 79.00 0.34 | 73.01 0.20 | 89.85 0.26 | 77.28 0.33 |
| Label smoothing | 90.02 0.04 | 78.36 0.03 | 72.40 0.06 | 89.74 0.14 | 77.07 0.40 |
| Mixup | 89.68 1.00 | 77.75 1.27 | 71.84 1.26 | 89.21 0.98 | 77.14 1.00 |
| No stochastic depth | 89.21 0.23 | 77.66 0.23 | 71.83 0.24 | 89.03 0.02 | 75.86 0.73 |
| SAM | 88.53 0.82 | 77.08 0.85 | 71.72 0.42 | 88.35 0.74 | 76.38 0.41 |
| weight and activation | Hessian curvature | ours | |||||
|---|---|---|---|---|---|---|---|
| Model | weight MSE | salience | salience MSE | H trace | HAWQ-V2 | drift | damage |
| Pythia-160M | 0.71 | 0.68 | 0.68 | 0.43 | 0.41 | 0.59 | 0.98 |
| Pythia-410M | 0.15 | 0.83 | 0.83 | 0.67 | 0.69 | 0.72 | 0.95 |
| Pythia-1.4B | 0.44 | 0.63 | 0.63 | 0.60 | 0.76 | 0.84 | 0.99 |
| Pythia-2.8B | 0.50 | 0.61 | 0.61 | 0.37 | 0.64 | 0.92 | |
| Pythia-6.9B | 0.96 | 0.96 | 0.96 | 0.46 | 0.83 | 0.99 | |
| reference | baselines | |||||||
|---|---|---|---|---|---|---|---|---|
| Model | fp16 | int4 | rand | first | last | MSE | salience | |
| Pythia-160M | 3/12 | 50.57 | 156.81 | 93.11 | 138.30 | 114.83 | 96.56 | 114.83 |
| 6/12 | 73.82 | 102.32 | 70.62 | 67.48 | 70.62 | |||
| Pythia-410M | 6/24 | 18.43 | 42.47 | 37.74 | 38.66 | 24.99 | 30.94 | 24.99 |
| 12/24 | 27.15 | 34.16 | 21.92 | 25.93 | 21.63 | |||
| Pythia-1.4B | 6/24 | 13.11 | 19.69 | 17.79 | 17.91 | 16.84 | 17.46 | 16.84 |
| perplexity | ||||||
| Model | HAWQ-V2 | ours | CI ( ) | result | ||
| and : committed damage rule, Table 10 | ||||||
| 160M | 3/12 | 97.70 | 89.43 | ours | ||
| 6/12 | 70.86 | 66.63 | ours | |||
| 410M | 6/24 | 24.99 | 24.69 | ours | ||
| 12/24 | 21.59 | 21.33 | ours | |||
| reference | ||||||
| Model | fp16 | int4 | HAWQ-V2 | drift- greedy | HAWQ-V2 | drift- greedy |
| Pythia-1.4B | 13.11 | 19.69 | 15.09 | 14.55 | 16.71 | 16.06 |
| Pythia-2.8B | 11.34 | 15.45 | 12.60 | 12.51 | 13.93 | 13.66 |
| Pythia-6.9B | 10.56 | 19.40 | 12.01 | 12.09 | 14.62 | 14.47 |
| Held-out LLMs | ||||||
| Phi-3.5-mini (3.8B) | 6.79 | 11.20 | 8.78 | 7.82 | 10.06 | 8.75 |
| allocation stability | WikiText-103 perplexity | ||||
| Rule and HAWQ calibration budget | (truth) | top-8 | |||
| HAWQ-V2, | 0.61 | 0.93 | 0.88 | 13.93 | 12.62 |
| HAWQ-V2, | 0.66 | 1.00 | 1.00 | 13.92 | 12.60 |
| HAWQ-V2, | 0.66 | 1.00 | 1.00 | 13.92 | 12.60 |
| damage-greedy, 96 seq. | NA | 13.54 | 12.31 | ||
| uniform int4 | NA | 15.45 | |||
| Block | drift@inj | drift@final | damage | damage/ | |
|---|---|---|---|---|---|
| 1 | 0.084 | 0.252 | 0.054 | 0.0015 | 0.03 |
| 3 | 0.087 | 0.255 | 0.056 | 0.0055 | 0.10 |
| 10 | 0.077 | 0.244 | 0.051 | 0.0055 | 0.11 |
| 16 | 0.077 | 0.257 | 0.058 | 0.0135 | 0.23 |
| 21 | 0.081 | 0.248 | 0.054 | 0.0241 | 0.45 |
| 26 | 0.084 | 0.192 | 0.031 | 0.0227 | 0.73 |
| Preregistered prediction | Observed | Threshold |
|---|---|---|
| P1 low drift–damage correlation | ||
| P2 high damage–damage correlation | ||
| P3 dissociation persists | gap | |
| P4 flip tracks drift | ||
| P5 damage allocation best | , perplexity | of the best rival |
| reference | baselines | |||||||
|---|---|---|---|---|---|---|---|---|
| Model | fp16 | int4 | rand | first | last | MSE | salience | |
| Qwen2.5-0.5B | 6/24 | 13.88 | 31.77 | 25.94 | 23.76 | 23.56 | 23.37 | 23.56 |
| 12/24 | 17.70 | 21.10 | 20.85 | 20.10 | 20.75 | |||
| Qwen2.5-1.5B | 7/28 | 9.97 | 16.88 | 14.54 | 14.39 | 13.11 | 13.21 | 13.11 |
| 14/28 | 12.94 | 13.50 | 12.02 | 12.03 | 12.02 | |||
| Qwen2.5-7B | 7/28 | 7.47 | 12.89 | 12.14 | 12.27 | 8.47 | 8.53 | 8.47 |
| weight and activation | Hessian curvature | ours | |||||
|---|---|---|---|---|---|---|---|
| Model | weight MSE | salience | salience MSE | H trace | HAWQ-V2 | drift | damage |
| Qwen2.5-0.5B | 0.33 | 0.14 | 0.17 | 0.88 | 0.89 | 0.94 | 0.97 |
| Qwen2.5-1.5B | 0.21 | 0.35 | 0.34 | 0.83 | 0.84 | 0.84 | 0.96 |
| Qwen2.5-3B | 0.50 | 0.27 | 0.26 | 0.72 | 0.52 | 0.73 | 0.98 |
| Qwen2.5-7B | 0.50 | 0.56 | 0.56 | 0.76 | 0.72 | 0.79 | 0.86 |
| weight and activation | Hessian curvature | ours | |||||
|---|---|---|---|---|---|---|---|
| Model | weight MSE | salience | salience MSE | H trace | HAWQ-V2 | drift | damage |
| Mamba-130M | 0.59 | 0.65 | 0.67 | 0.67 | 0.77 | 0.85 | 0.99 |
| Mamba-370M | 0.49 | 0.70 | 0.74 | 0.69 | 0.72 | 0.88 | 0.96 |
| Mamba-790M | 0.64 | 0.59 | 0.62 | 0.74 | 0.75 | 0.87 | 0.94 |
| Mamba-1.4B | 0.35 | 0.59 | 0.61 | 0.53 | 0.60 | 0.79 | 0.96 |
| Mamba-2.8B | 0.31 | 0.62 | 0.66 | 0.40 | 0.43 | 0.87 | 0.90 |
| reference | baselines | ours | ||||||
|---|---|---|---|---|---|---|---|---|
| Model | fp16 | AWQ4 | last | salience | HAWQ-V2 | drift | damage | |
| Pythia-2.8B | 8/32 | 11.34 | 15.11 | 13.47 | 13.47 | 13.41 | 13.19 | 13.17 |
| 16/32 | 12.26 | 12.26 | 12.24 | 12.59 | 12.08 | |||
| Pythia-6.9B | 8/32 | 10.56 | 19.27 | 15.20 | 15.20 | 14.51 | 15.67 | 14.12 |
| 16/32 | 11.57 | 11.57 | 11.57 | 14.82 | 11.51 | |||
| Qwen2.5-7B | 7/28 | 7.47 | 13.84 | 8.29 | 8.29 | 8.34 | 8.29 | 8.29 |
| baselines | ours | ||||
|---|---|---|---|---|---|
| Model | salience MSE | HAWQ-V2 | drift | damage | |
| Pythia-2.8B | 32 | 0.76 | 0.72 | 0.29 | 0.94 |
| Pythia-6.9B | 32 | 0.90 | 0.86 | 0.96 | |
| Qwen2.5-7B | 28 | 0.66 | 0.83 | 0.86 | 0.89 |
| Phi-3.5-mini (3.8B) | 32 | 0.05 | 0.21 | 0.63 | 0.87 |
| Hermes-3 | 32 | 0.20 | 0.18 | 0.89 | |
| reference | baselines | ours | ||||||
|---|---|---|---|---|---|---|---|---|
| Model | fp16 | SpQR4 | last | salience | HAWQ-V2 | drift | damage | |
| Pythia-2.8B | 8/32 | 11.34 | 12.38 | 12.00 | 12.00 | 12.03 | 12.16 | 11.96 |
| 16/32 | 11.77 | 11.77 | 11.76 | 11.88 | 11.70 | |||
| Pythia-6.9B | 8/32 | 10.56 | 11.94 | 11.32 | 11.32 | 11.54 | 11.71 | 11.33 |
| 16/32 | 10.87 | 10.87 | 10.87 | 11.51 | 10.87 | |||
| Qwen2.5-7B | 7/28 | 7.47 | 8.17 | 7.83 | 7.83 | 7.83 | 7.83 | 7.83 |
| baselines | ours | ||||
|---|---|---|---|---|---|
| Model | salience MSE | HAWQ-V2 | drift | damage | |
| Pythia-2.8B | 32 | 0.54 | 0.31 | 0.73 | |
| Pythia-6.9B | 32 | 0.85 | 0.61 | 0.85 | |
| Qwen2.5-7B | 28 | 0.58 | 0.73 | 0.71 | 0.80 |
| Phi-3.5-mini (3.8B) | 32 | 0.34 | 0.81 | 0.89 | |
| Hermes-3 | 32 | 0.50 | 0.82 | ||
| zero-shot task accuracy | |||||||
|---|---|---|---|---|---|---|---|
| Model | cfg | PIQA | ARC-E | ARC-C | HellaS. | WinoG. | Avg |
| Qwen2.5-7B | fp16 | 0.796 | 0.747 | 0.522 | 0.659 | 0.682 | 0.681 |
| int4 | 0.765 | 0.607 | 0.415 | 0.630 | 0.595 | 0.602 | |
| HAWQ-V2 | 0.774 | 0.679 | 0.495 | 0.652 | 0.626 | 0.645 | |
| damage | 0.781 | 0.711 | 0.482 | 0.654 | 0.635 | 0.652 | |
| greedy | 0.782 | 0.704 | 0.495 | 0.649 | 0.614 | 0.649 | |
| int8 | int8 | |||
| Allocation | V2 | IN-C | V2 | IN-C |
| a1_in1k: fp16 69.0, uniform int8 68.5, int4 0.4 | ||||
| roughness (ours) | 57.9 | 17.1 | 65.6 | 29.6 |
| HAWQ-V2 sensitivity | 56.7 | 15.6 | 65.6 | 29.6 |
| salience MSE (AWQ-like) | 42.4 | 12.1 | 50.2 | 17.1 |
| weight magnitude | 0.9 | 0.3 | 2.3 | 1.0 |
Noise Robustness
Noise robustness studies how to preserve reliable predictions when data, labels, signals, or computing hardware are corrupted.
See all 99 papers