Pith. sign in

REVIEW 2 major objections 5 minor 51 references

Generalization and Trade-off in Adversarial Training: An RKHS Perspective via Kernel Integral Operators

T0 review · 2 major / 5 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adversarial training in kernel spaces pays a statistical price: noise inside the robustness penalty slows the approximation rate below the usual nonparametric benchmark, and stripping that noise term restores the minimax rate.

desk verdict Solid RKHS theory that pins the AT accuracy loss to noise in the mixed term and fixes it with a debiased estimator that recovers near-minimax rates. read the letter →

arxiv 2607.27995 v1 pith:YCBHLKJI submitted 2026-07-30 stat.ML cs.LG

Yiling Xie , Xiaoming Huo This is my paper
classification stat.MLcs.LG MSC 62G0868T0546E22
keywords adversarialtrainingreproducingkernelHilbertspaceintegraloperatorgeneralizationboundsrobustness-accuracytrade-offsourceconditionnoisedebiasingnonparametricregression
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper asks how adversarial training behaves as a nonparametric estimator inside a reproducing kernel Hilbert space. It shows that the usual robust objective carries a mixed term that multiplies the model’s RKHS norm by the average absolute residual. The irreducible observation noise inside that residual forces a slower approximation rate than ordinary ridge-type methods, so the optimally balanced generalization rate can fall short of the classical minimax prediction rate on a fixed polynomial-spectrum model. The same mixed term also adds curvature that shrinks estimation complexity, but that gain does not cancel the approximation loss. The authors then build a two-stage fix: estimate the mean absolute noise on held-out data and subtract it from the mixed term. The debiased estimator recovers the minimax polynomial rate up to a square-root log factor when the robustness radius shrinks at the stated sample-dependent order, while an exact comparison quantifies the robustness cost of the correction.

What carries the argument

The mixed robustness term 2δ‖f‖_Hk E|Y−f(X)| inside the population adversarial objective, analyzed through the spectral calculus of the kernel integral operator Tk; noise debiasing subtracts an estimate of E|ε| from the empirical absolute residual so the approximation order returns to the ridge scale δ^min{2α+1,2}.

What would settle it

On the paper’s synthetic diagonal-kernel model with known polynomial spectrum and fixed noise law, measure whether full adversarial training’s L2 error tracks the slower balanced exponent while the noise-debiased estimator tracks the faster minimax exponent (up to √log n) at the prescribed radius schedule.

Watch

Extended reading notes

Core claim

On a fixed polynomial-spectrum RKHS model the optimally balanced full adversarial-training estimator obeys a matching lower bound whose leading rate is slower than the minimax L2 prediction benchmark; the slowdown is caused by the irreducible-noise contribution inside the mixed robustness term, and a two-stage noise-debiased estimator removes that contribution and attains the minimax polynomial rate up to a logarithmic factor.

Load-bearing premise

The analysis treats adversarial perturbations as balls in kernel feature space around each k_X, which only upper-bounds ordinary input-space attacks when those attacks induce feature shifts no larger than the chosen radius.

Editorial extensions

If this is right

  • Full adversarial training cannot be expected to match ordinary nonparametric prediction rates when observation noise is non-degenerate.
  • The robustness–accuracy trade-off is an approximation effect driven by noise inside the mixed term, not only an estimation-complexity effect.
  • Estimating and subtracting mean absolute noise restores the classical polynomial rate for source smoothness α ≤ 1/2 when δ shrinks at the stated order.
  • At fixed threat-model radius, debiasing raises the null-solution threshold, so small signals are less likely to be thresholded to zero.
  • The exact objective comparison supplies a quantitative robustness cost of debiasing equal to the correction size times the increase in RKHS norm.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Any robust objective whose penalty multiplies a model-norm factor by an absolute residual will inherit a similar noise-driven approximation slowdown unless that residual is debiased.
  • The feature-space surrogate may systematically mis-rank methods under input-space attacks whenever the kernel map is far from Lipschitz at the attack scale used in practice.
  • Sample-splitting for the noise estimate is a practical bottleneck; a cross-fitted or one-sample version would be the natural next algorithmic test.
  • The same spectral bookkeeping could be tried on other nonparametric robust procedures (e.g., distributionally robust kernel regression) to see whether their rates are likewise noise-limited.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. The paper analyzes adversarial training (AT) in an RKHS via the kernel integral operator. It derives source-uniform approximation, estimation, and generalization bounds for the feature-space AT estimator (equivalent to minimizing the mixed squared loss (|Y−f(X)|+δ‖f‖_Hk)^2). On a fixed polynomial-spectrum model it proves a matching lower bound showing that the optimally balanced AT rate can be slower than the Caponnetto–De Vito minimax prediction benchmark; the slowdown is attributed to the irreducible-noise piece in the mixed term. A two-stage noise-debiased AT (ND-AT) that estimates and subtracts E|ε| is proposed and shown to restore the ridge-order approximation rate and, for α≤1/2 with sample-dependent δ_n, the minimax polynomial rate up to √log n. Robustness cost of debiasing and null-solution thresholds are quantified, and synthetic plus two real-data experiments support the theory.

Significance. If correct, the work gives a clean nonparametric explanation of the robustness–accuracy trade-off for AT: noise in the mixed term slows approximation while adding curvature that helps estimation, and the net effect can be a strictly suboptimal prediction rate. The matching lower bound on a fixed model and the operator-theoretic decomposition are substantial relative to existing parametric AT analyses and to prior RKHS robust-learning formulations that lacked rate characterizations. ND-AT is a principled, implementable fix that recovers minimax rates up to a log factor, with an exact robustness-cost identity. Strengths include full appendix proofs, explicit source-uniform bounds, a disclosed estimator-/model-specific lower bound (Remark 3.5), and experiments with confidence intervals that track the predicted approximation slopes and null-collapse behavior.

major comments (2)
  1. [Section 2.2, Proposition 2.6; Section 5.2] Sec. 2.2 and the inequality after U(X_i): the entire rate theory is for feature-space balls around k_X (Prop. 2.6). The manuscript correctly notes this is a convex surrogate that upper-bounds input-space attack loss only when induced feature shifts stay within δ. For a journal audience this scope limit should be stated more prominently in the abstract/introduction conclusions, and Sec. 5.2 should quantify (or bound) how often the RBF feature-Lipschitz link ‖k_x−k_{x'}‖_Hk ≤ ‖x−x'‖/h is tight under the reported ℓ2 attacks, so readers can judge transfer to the usual threat model.
  2. [Theorem 3.4, Remark 3.5] Theorem 3.4 / Remark 3.5: the matching lower bound is estimator-specific on one diagonal-kernel model on ℕ with a tailored bounded noise law. The polynomial-order match is valuable, but the claim of a “genuine loss of statistical efficiency for adversarial training” (after Thm. 3.4) should be worded to make clear it is not a minimax lower bound over all procedures or over general kernels. A short discussion of which structural features of the construction (atomic design, band J_δ, noise with flat absolute-risk derivative on (−2,2)) are essential would strengthen the interpretation.
minor comments (5)
  1. [Corollary 4.3] Corollary 4.3 uses total sample size on the order of 3n (auxiliary + residual + fitting). Please state explicitly that the displayed rate is in the fitting size n and that the constant factors absorb the fixed split ratios.
  2. [Figure 1] Figure 1(b): report fitted slopes with standard errors (or a small table) next to the theoretical exponents 0.60 and 1.5 so the visual comparison is quantitative.
  3. [Assumption 2.2] Assumption 2.2 is used heavily; a one-line pointer to common noise laws (median-zero bounded density, the constructed noise in Thm. 3.4) already in the text would help non-specialists.
  4. [References; Section 3.3] Typos/notation: “Sch¨ olkopf” spacing in references; ensure N(δ) vs N(δ^2) is always parenthesized consistently in displays after Cor. 3.3.
  5. [Proposition 4.7, Section 5] Proposition 4.7: briefly note that in all reported runs er_δ^2 = a^⊤Ka held to solver tolerance (already in text) and what fallback is used if the S^⊥ component is needed numerically.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AT/ND-AT rates are derived from operator first-order conditions and external minimax benchmarks, not from self-fitted quantities.

full rationale

The central claims (Cor. 3.3, Thm. 3.4, Cor. 4.3) compare generalization rates of the full-AT and noise-debiased estimators to the external Caponnetto–De Vito / Mendelson–Neeman / Fischer–Steinwart minimax polynomial benchmark. Approximation orders follow from population FOCs of the mixed objective (Thm. 3.1, 4.1); estimation uses standard kernel operator concentration; the matching lower bound is built on an independently specified diagonal-kernel model and fixed noise law (Thm. 3.4), not on quantities fitted from the upper-bound analysis. Hyperparameters δ_n and λ_a are theory-prescribed or CV-chosen for experiments and do not define the claimed exponents. Self-citations (Xie–Huo parametric AT) are background only and are not load-bearing for the nonparametric rate chain. No step reduces a claimed prediction to its own definition or fit by construction.

Assumptions & free parameters 2 free parameters · 8 assumptions · 2 invented entities

The central rate claims rest on a standard nonparametric regression setup plus kernel spectral assumptions, a quadratic absolute-risk condition that makes noise nondegenerate, and the paper’s feature-space adversarial formulation. No physical constants or data-fitted universal scales enter the theorems; δ and ridge λ are hyperparameters. The main modeling invention is feature-space AT and the noise-debiased objective, analyzed rather than postulated as new physics.

free parameters (2)
  • adversarial radius δ (and sample-dependent δ_n) = theory: δ_n ≍ n^{-β(α+1)/[β(2α+1)+α+1]} (AT), δ_n = n^{-β/[2(2β(α+1/2)+1)]} (ND-AT); experiments use grids / δ=0.1
    Robustness level treated as free in bounds; rate-optimal δ_n is prescribed by balancing theory, not fitted to prove the exponent, but experiment grids and real-data δ=0.1 are chosen by hand.
  • preliminary KRR regularization λ_a = theory λ_a = N_a^{-β/[2β(α+1/2)+1]}; synthetic λ_a=N_a^{-1/2}
    First-stage ridge parameter; theory gives an explicit rate-optimal order, experiments use CV or λ_a=N_a^{-1/2}.
assumptions (8)
  • domain assumption Additive model Y=f*(X)+ε with f* in H_k, ε independent, mean-zero, square-integrable (Assumption 2.1).
    Standard nonparametric regression; required for population objectives and noise separation.
  • domain assumption Quadratic growth of excess absolute risk: 0 ≤ E|ε−t|−E|ε| ≤ κ t^2 (Assumption 2.2), implying nondegenerate E|ε|.
    Controls mixed-term noise contribution and approximation proofs; holds for median-zero noise with bounded density but is stronger than mere mean-zero noise.
  • domain assumption Sub-exponential noise moments E|ε|^p ≤ (p!/2) σ^p (Assumption 2.3).
    Enables Bernstein-type operator and scalar concentration in estimation bounds.
  • domain assumption Source condition f*=T_k^α g with ‖g‖_Hk≤R, 0<α≤1 (Assumption 2.4).
    Classical kernel-learning smoothness; rates are source-uniform over this ball.
  • domain assumption Polynomial eigenvalue decay σ_j ≍ j^{-β}, β>1 (Assumption 2.5), used for polynomial rates and effective dimension N(δ)≲δ^{-1/β}.
    Standard spectral complexity assumption for explicit minimax comparisons.
  • ad hoc to paper Adversarial training is formulated as feature-space perturbations of k_X within RKHS balls of radius δ, equivalent to the mixed squared loss (|Y−f(X)|+δ‖f‖_Hk)^2 (Prop. 2.6).
    Central modeling choice following robust kernel learning literature; makes the problem convex and analyzable but is only a surrogate for input-space attacks.
  • standard math Bounded kernel: sup_x k(x,x)≤1.
    Normalizes operator norms and concentration constants throughout.
  • standard math External minimax benchmark: optimal L2 rate n^{-β(α+1/2)/[2β(α+1/2)+1]} for 0<α≤1/2 under source and polynomial spectra (Caponnetto & De Vito; related works).
    Used as the comparison target for claiming statistical loss and recovery; not re-derived from scratch.
invented entities (2)
  • Noise-debiased adversarial training objective (two-stage estimator subtracting estimated E|ε| from the mixed term) independent evidence
    purpose: Remove irreducible-noise contribution that slows AT approximation while retaining a robustness penalty.
    Defined in Sec. 4 eqs. (7)–(8); analyzed as a new procedure rather than an empirical black box.
  • Fixed diagonal-kernel lower-bound model on ℕ with tailored bounded noise ε=S(2+U)
    purpose: Witness matching polynomial lower bounds for full AT approximation and estimation terms.
    Constructed in the proof of Thm. 3.4; valid as a hardness example but not a universal data-generating law.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalization and Trade-off in Adversarial Training: An RKHS Perspective via Kernel Integral Operators." pith.science (2026). https://pith.science/paper/YCBHLKJI

@misc{pith2026260727995,
  author       = {Pith},
  title        = {Pith review of: Generalization and Trade-off in Adversarial Training: An RKHS Perspective via Kernel Integral Operators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YCBHLKJI}},
  note         = {Machine review of arXiv:2607.27995}
}
read the original abstract

Adversarial training has emerged as a powerful approach for protecting models against adversarial attacks in a broad range of real-world applications. In this paper, we study adversarial training in the reproducing kernel Hilbert space (RKHS) framework through the associated kernel integral operator. We first derive source-uniform generalization error bounds for the RKHS adversarial training estimator in terms of the robustness level, sample size, source smoothness, and kernel spectrum. On a fixed polynomial-spectrum model, we further establish a matching lower bound showing that the optimally balanced generalization rate can be slower than the minimax prediction benchmark. This result reveals a loss of statistical accuracy in adversarial training. Our analysis shows that this loss arises from the interaction between adversarial robustness and observation noise: the noise contribution in the mixed robustness term slows the approximation rate, although the same term reduces the estimation complexity. To address this limitation, we propose a two-stage noise-debiased procedure that estimates and removes the noise contribution from the mixed term. The resulting estimator improves the generalization rate and attains the minimax polynomial rate, up to a logarithmic factor, when the robustness level is selected at the stated sample-dependent order. Our results characterize the generalization behavior of adversarial training in a nonparametric framework and provide a new interpretation and a principled solution for the trade-off between adversarial robustness and generalization. Numerical experiments support the theoretical findings and demonstrate the effectiveness of the proposed method.

Figures

Figures reproduced from arXiv: 2607.27995 by the authors.

Figure 1
Figure 1. Synthetic experiment over the radius grid [PITH_FULL_IMAGE:figures/full_fig_p016_1.png] view at source ↗
Figure 2
Figure 2. Test-time input attacks for estimators trained at [PITH_FULL_IMAGE:figures/full_fig_p017_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 2 linked inside Pith

  1. [1]

    Annals of Probability , pages=

    Optimum bounds for the distributions of martingales in Banach spaces , author=. Annals of Probability , pages=. 1994 , publisher=

    1994

  2. [2]

    Foundations and trends

    An introduction to matrix concentration inequalities , author=. Foundations and trends. 2015 , publisher=

    2015

  3. [3]

    Statistics & Probability Letters , volume=

    On some extensions of Bernstein’s inequality for self-adjoint operators , author=. Statistics & Probability Letters , volume=. 2017 , publisher=

    2017

  4. [4]

    Foundations of Computational Mathematics , volume =

    Caponnetto, Andrea and De Vito, Ernesto , title =. Foundations of Computational Mathematics , volume =. 2007 , doi =

    2007

  5. [5]

    , title =

    Sutherland, Danica J. , title =. arXiv preprint arXiv:1702.02982 , year =

    arXiv

  6. [6]

    Constructive approximation , volume=

    Learning theory estimates via integral operators and their approximations , author=. Constructive approximation , volume=. 2007 , publisher=

    2007

  7. [7]

    Foundations of computational mathematics , volume=

    Learning rates of least-square regularized regression , author=. Foundations of computational mathematics , volume=. 2006 , publisher=

    2006

  8. [8]

    Annals of Statistics , volume=

    Regularization in kernel learning , author=. Annals of Statistics , volume=. 2010 , publisher=

    2010

Show all 51 references
  1. [9]

    Journal of Machine Learning Research , volume=

    Sobolev norm learning rates for regularized least-squares algorithms , author=. Journal of Machine Learning Research , volume=

  2. [10]

    Less is more: Nystr

    Rudi, Alessandro and Camoriano, Raffaello and Rosasco, Lorenzo , journal=. Less is more: Nystr. 2015 , note=

    2015

  3. [11]

    and Shlens, Jonathon and Szegedy, Christian , title =

    Goodfellow, Ian J. and Shlens, Jonathon and Szegedy, Christian , title =. International Conference on Learning Representations , year =

  4. [12]

    International Conference on Learning Representations , year=

    Towards Deep Learning Models Resistant to Adversarial Attacks , author=. International Conference on Learning Representations , year=

  5. [13]

    International Conference on Learning Representations , year=

    Certifying Some Distributional Robustness with Principled Adversarial Training , author=. International Conference on Learning Representations , year=

  6. [14]

    Advances in neural information processing systems , volume=

    Neural tangent kernel: Convergence and generalization in neural networks , author=. Advances in neural information processing systems , volume=

  7. [15]

    Foundations and Trends

    Kernel Mean Embedding of Distributions:: A Review and Beyond , author=. Foundations and Trends. 2017 , publisher=

    2017

  8. [16]

    Advances in Neural Information Processing Systems , volume=

    Encoding robustness to image style via adversarial feature perturbations , author=. Advances in Neural Information Processing Systems , volume=

  9. [17]

    International conference on computational learning theory , pages=

    A generalized representer theorem , author=. International conference on computational learning theory , pages=. 2001 , organization=

    2001

  10. [18]

    Steinwart, Ingo and Christmann, Andreas , title =

  11. [19]

    Asymptotic Behavior of Adversarial Training Estimator under

    Xie, Yiling and Huo, Xiaoming , journal=. Asymptotic Behavior of Adversarial Training Estimator under. 2025 , publisher=

    2025

  12. [20]

    Advances in Neural Information Processing Systems , volume=

    High-dimensional (group) adversarial training in linear regression , author=. Advances in Neural Information Processing Systems , volume=

  13. [21]

    arXiv preprint arXiv:2601.06807 , year=

    Adversarially Perturbed Precision Matrix Estimation , author=. arXiv preprint arXiv:2601.06807 , year=

  14. [22]

    Sensors and Actuators B: Chemical , volume=

    Chemical gas sensor drift compensation using classifier ensembles , author=. Sensors and Actuators B: Chemical , volume=. 2012 , publisher=

    2012

  15. [23]

    Advances in Neural Information Processing Systems , volume=

    Kernel learning with adversarial features: Numerical efficiency and adaptive regularization , author=. Advances in Neural Information Processing Systems , volume=

  16. [24]

    Advances in Neural Information Processing Systems , volume=

    Regularization properties of adversarially-trained linear regression , author=. Advances in Neural Information Processing Systems , volume=

  17. [25]

    Annals of Statistics , volume=

    Precise statistical analysis of classification accuracies for adversarial training , author=. Annals of Statistics , volume=. 2022 , publisher=

    2022

  18. [26]

    International Conference on Learning Representations , year =

    Robustness May Be at Odds with Accuracy , author=. International Conference on Learning Representations , year =

  19. [27]

    International conference on machine learning , pages=

    Theoretically principled trade-off between robustness and accuracy , author=. International conference on machine learning , pages=. 2019 , organization=

    2019

  20. [28]

    International Conference on Machine Learning , pages=

    Understanding and Mitigating the Tradeoff between Robustness and Accuracy , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  21. [29]

    2007 , publisher=

    UCI machine learning repository , author=. 2007 , publisher=

    2007

  22. [30]

    Theory of Probability & Its Applications , volume=

    Remarks on inequalities for large deviation probabilities , author=. Theory of Probability & Its Applications , volume=. 1986 , publisher=

    1986

  23. [31]

    Training Robust

    Lucas, Keane and Lin, Weiran and Bauer, Lujo and Reiter, Michael K and Sharif, Mahmood , booktitle=. Training Robust

  24. [32]

    Efficient adversarial training in

    Xhonneux, Sophie and Sordoni, Alessandro and G. Efficient adversarial training in. Advances in Neural Information Processing Systems , volume=

  25. [33]

    2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) , pages=

    Domain adaptation for biomedical image segmentation using adversarial training , author=. 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) , pages=. 2018 , organization=

    2018

  26. [34]

    International Conference on Machine Learning , pages=

    Rademacher complexity for adversarially robust generalization , author=. International Conference on Machine Learning , pages=. 2019 , organization=

    2019

  27. [35]

    Adversarially Robust Generalization Requires More Data , volume =

    Schmidt, Ludwig and Santurkar, Shibani and Tsipras, Dimitris and Talwar, Kunal and Madry, Aleksander , booktitle =. Adversarially Robust Generalization Requires More Data , volume =

  28. [36]

    International Conference on Machine Learning , pages=

    More data can expand the generalization gap between adversarially robust and standard models , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    2020

  29. [37]

    Advances in Neural Information Processing Systems , volume=

    Adversarial Robustness of Nonparametric Regression , author=. Advances in Neural Information Processing Systems , volume=

  30. [38]

    arXiv preprint arXiv:2506.01267 , year=

    Adversarial learning for nonparametric regression: Minimax rate and adaptive estimation , author=. arXiv preprint arXiv:2506.01267 , year=

    arXiv

  31. [39]

    Transactions of the American mathematical society , volume=

    Theory of reproducing kernels , author=. Transactions of the American mathematical society , volume=

  32. [40]

    2002 , publisher=

    Learning with kernels: support vector machines, regularization, optimization, and beyond , author=. 2002 , publisher=

    2002

  33. [41]

    Journal of Machine Learning Research , volume=

    Regularization via mass transportation , author=. Journal of Machine Learning Research , volume=

  34. [42]

    Journal of Machine Learning Research , volume=

    Optimal rates of kernel ridge regression under source condition in large dimensions , author=. Journal of Machine Learning Research , volume=

  35. [43]

    European Journal of Operational Research , volume=

    A data-driven software tool for enabling cooperative information sharing among police departments , author=. European Journal of Operational Research , volume=. 2002 , publisher=

    2002

  36. [44]

    The population biology of abalone (haliotis species) in tasmania. i. blacklip abalone (h. rubra) from the north coast and islands of bass strait , author=. Sea Fisheries Division, Technical Report , volume=

  37. [45]

    Confidence regions in

    Blanchet, Jose and Murthy, Karthyek and Si, Nian , journal=. Confidence regions in. 2022 , publisher=

    2022

  38. [46]

    Biometrika , volume=

    Square-root lasso: pivotal recovery of sparse signals via conic programming , author=. Biometrika , volume=

  39. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Reliably fast adversarial training via latent adversarial perturbation , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  40. [48]

    Comptes Rendus Mathematique , volume=

    A Bennett concentration inequality and its application to suprema of empirical processes , author=. Comptes Rendus Mathematique , volume=

  41. [49]

    Algorithmic Learning Theory , pages=

    A vector-contraction inequality for Rademacher complexities , author=. Algorithmic Learning Theory , pages=. 2016 , publisher=

    2016

  42. [50]

    Trudy Moskovskogo Matematicheskogo Obshchestva , volume=

    Subdifferentials of convex functions , author=. Trudy Moskovskogo Matematicheskogo Obshchestva , volume=

  43. [51]

    1998 , publisher=

    Variational Analysis , author=. 1998 , publisher=

    1998

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.