arxiv: 2605.06349 · v1 · submitted 2026-05-07 · 🧮 math.NA · cs.NA· math.ST· stat.TH

Recognition: unknown

Low-rank kernel methods for American option pricing

Chiara Segala, Michael Multerer, Paul Schneider

Pith reviewed 2026-05-08 06:23 UTC · model grok-4.3

classification 🧮 math.NA cs.NAmath.STstat.TH

keywords low-rank approximationkernel methodsAmerican optionsMonte Carlo simulationoptimal stoppingconditional expectationreproducing kernel Hilbert spacebackward recursion

0 comments

The pith

A low-rank kernel method learns one conditional expectation operator from simulations and reuses it across all exercise dates for American option pricing.

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

The paper shows how to reformulate the estimation of continuation values in American option pricing as learning a linear operator in a reproducing kernel Hilbert space. This operator acts on future payoffs and can be approximated in low rank, allowing it to be trained once on simulated paths and then applied repeatedly during the backward recursion. The approach avoids fitting a fresh regression model at every time step. Convergence of the method is proved and bounds are given on the accumulated approximation error over the exercise dates. If the low-rank structure holds, the pricing procedure becomes faster and more scalable for problems with many time steps or high-dimensional state spaces.

Core claim

The conditional expectation operator that maps future payoffs to present values admits a sufficiently accurate low-rank representation inside a reproducing kernel Hilbert space. This representation can be learned offline from a single set of simulated paths and then reused without modification at every exercise date, turning the usual sequence of separate regressions into a single operator application step.

What carries the argument

The low-rank approximation of the conditional expectation operator in the reproducing kernel Hilbert space, which encodes the mapping from future payoffs to current continuation values and enables the offline-online decomposition.

If this is right

Only one operator needs to be learned regardless of the number of exercise dates.
Regression fitting is performed only once instead of once per time step.
Error bounds can be tracked across the entire backward induction.
The same learned operator can be applied to multiple option contracts that share the same underlying dynamics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The offline-online split could allow the operator to be precomputed on specialized hardware and then deployed on standard pricing engines.
The same low-rank structure might be tested on other optimal stopping problems such as Bermudan swaptions or real-options valuation.
If the kernel is chosen to respect the Markov structure of the state process, the method could reduce memory requirements for storing regression coefficients.

Load-bearing premise

The conditional expectation operator must admit a sufficiently accurate low-rank approximation in the chosen reproducing kernel Hilbert space, and the simulated paths must remain representative of the true distribution without large shifts across exercise dates.

What would settle it

If the observed pricing bias fails to shrink at the predicted rate when the operator rank is increased or when the number of simulated paths is doubled, while keeping all other parameters fixed, the claimed convergence would be contradicted.

Figures

Figures reproduced from arXiv: 2605.06349 by Chiara Segala, Michael Multerer, Paul Schneider.

**Figure 1.** Figure 1: Mean log10 computation time (in µs) as a function of n, for LS (blue solid) and CME-LR (red dashed). the 10 strikes and the 100 replications: εrel(n, T) = 1 100 X 100 r=1 1 10 X 10 k=1 view at source ↗

**Figure 2.** Figure 2: Mean rank of the kernel matrix KY selected by the pivoted Cholesky decomposition as a function of n, across four maturities, for three values of the Cholesky tolerance ε ∈ {10−4 , 10−5 , 10−6}. rank of KY and KX selected by the pivoted Cholesky decomposition as a function of n, for three values of the tolerance ε ∈ {10−4 , 10−5 , 10−6}. While all pricing results reported in this section are obtained with … view at source ↗

**Figure 3.** Figure 3: Mean relative implied volatility error as a function of n, for LS (blue solid) and CME-LR (red dashed). Shaded bands: 95% confidence intervals over 100 replications view at source ↗

**Figure 4.** Figure 4: Mean relative implied volatility error vs. log(K/S0). Left: LS (solid). Right: CMELR offline (dashed). Shaded bands: 95% confidence intervals over 100 replications. 25 view at source ↗

read the original abstract

We propose a scalable and theoretically grounded low-rank conditional expectation model for recursive Monte Carlo optimal stopping problems, in particular American option pricing. Our method reformulates the estimation of continuation values as a learning problem in a reproducing kernel Hilbert space, in which the conditional expectation is represented as a linear operator acting on future payoffs. This perspective yields an offline-online decomposition: the operator is learned once from simulated data and subsequently reused across all exercise dates, eliminating the need to recompute regression models at each step of the backward recursion. We establish convergence guarantees and derive bounds quantifying the approximation errors across exercise dates. Numerical experiments demonstrate the speed and accuracy of the proposed approach relative to extant methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper's offline-online low-rank kernel operator for American options is a practical step forward on scalability but the error bounds look incomplete for the nonlinear recursion.

read the letter

The main thing to know is that this work learns a single low-rank approximation to the conditional expectation operator in a reproducing kernel Hilbert space from simulated paths, then reuses that operator at every exercise date instead of refitting a regression model each step backward. That split removes the repeated fitting cost that dominates standard least-squares Monte Carlo for American options. The abstract frames the continuation value as the action of a linear operator on future payoffs, which lets them import low-rank techniques and claim convergence plus explicit error bounds across dates. Numerical tests are reported to beat existing methods on speed and accuracy. Those are the concrete advances. The approach sits on standard RKHS theory and Monte Carlo simulation, so the novelty is mainly in the operator-learning view plus the fixed-rank reuse. The soft spot is the nonlinear max that appears in the dynamic programming step. Once you replace the continuation value with max(exercise payoff, approximated operator applied to value), the result generally leaves the low-rank subspace. The paper states bounds on the linear operator approximation and on errors across dates, but those bounds do not automatically carry over to the composed nonlinear recursion unless extra uniformity or Lipschitz control is shown. If the rank stays fixed while the number of exercise dates grows, accumulated error can increase even when the operator itself is accurate on the training measure. The assumption that one low-rank operator stays representative without meaningful distribution shift across dates is also left as an implicit modeling choice rather than something heavily tested. This paper is for researchers who already work with Monte Carlo optimal stopping or kernel methods in computational finance. A reader who knows least-squares Monte Carlo will see exactly where the speedup comes from and can judge whether the claimed bounds close the gap. It deserves a serious referee. The idea is specific, the theoretical framing is coherent, and the practical claim is testable, so referees can check the missing steps on the nonlinear error propagation and the experimental baselines.

Referee Report

2 major / 2 minor

Summary. The paper proposes a low-rank kernel method for American option pricing in a Monte Carlo setting. It reformulates continuation-value estimation as learning a linear conditional-expectation operator in a reproducing kernel Hilbert space (RKHS), trains this operator once offline on simulated paths, and reuses the same low-rank operator at every exercise date during the backward recursion. Convergence guarantees and explicit approximation-error bounds across dates are derived, and numerical experiments are presented to show gains in speed and accuracy relative to existing regression-based approaches.

Significance. If the central claims hold, the offline-online decomposition offers a genuine scalability improvement for recursive Monte Carlo optimal-stopping problems by eliminating repeated per-date regressions. The RKHS perspective supplies a clean theoretical framework, and the provision of convergence guarantees plus error bounds is a positive feature that distinguishes the work from purely heuristic low-rank approximations.

major comments (2)

[§4, Theorem 4.2] §4 (Error Analysis), Theorem 4.2 and the subsequent recursion bound: The stated error bounds control the linear operator approximation error on the training measure, but the dynamic-programming step replaces the continuation value by the nonlinear map v ↦ max(g, T v). No additional lemma or uniformity argument is supplied showing that the low-rank subspace remains approximately invariant or that the Lipschitz constant of the max operation does not amplify the per-step error with the number of exercise dates. Without this, the claimed bound on total accumulated error across dates is not yet justified.
[§5, Table 1] §5 (Numerical Experiments), Table 1 and Figure 3: The reported pricing errors and timings are given only for the proposed method; no quantitative comparison (bias, standard error, wall-clock time) against the Longstaff-Schwartz least-squares Monte Carlo baseline or against other kernel-regression variants appears. This makes it impossible to verify the claimed accuracy and speed advantages.

minor comments (2)

[Eq. (9)] The notation for the low-rank truncation (Eq. (9)) introduces the rank parameter r without an explicit statement of how r is chosen or adapted to the number of exercise dates.
[Figure 2] Figure 2 caption does not define the symbols used for the different kernel choices, forcing the reader to consult the text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful reading and constructive suggestions. The comments highlight important points for strengthening the theoretical justification and empirical presentation. We address each major comment below and will incorporate revisions accordingly.

read point-by-point responses

Referee: [§4, Theorem 4.2] §4 (Error Analysis), Theorem 4.2 and the subsequent recursion bound: The stated error bounds control the linear operator approximation error on the training measure, but the dynamic-programming step replaces the continuation value by the nonlinear map v ↦ max(g, T v). No additional lemma or uniformity argument is supplied showing that the low-rank subspace remains approximately invariant or that the Lipschitz constant of the max operation does not amplify the per-step error with the number of exercise dates. Without this, the claimed bound on total accumulated error across dates is not yet justified.

Authors: We agree that the propagation of approximation error through the nonlinear max operator requires explicit control to justify the accumulated bound over multiple exercise dates. The current proof sketch relies on the fact that the Bellman operator is a contraction and that the max(·,·) map is 1-Lipschitz in the uniform norm, which prevents unbounded amplification. However, to make the argument fully rigorous and uniform across dates, we will add a supporting lemma (new Lemma 4.3) that bounds the error after each nonlinear step in terms of the operator approximation error on the training measure. This lemma will also confirm that the low-rank subspace error remains controlled without requiring invariance of the subspace itself. The revised Theorem 4.2 will then cite this lemma to close the recursion. revision: yes
Referee: [§5, Table 1] §5 (Numerical Experiments), Table 1 and Figure 3: The reported pricing errors and timings are given only for the proposed method; no quantitative comparison (bias, standard error, wall-clock time) against the Longstaff-Schwartz least-squares Monte Carlo baseline or against other kernel-regression variants appears. This makes it impossible to verify the claimed accuracy and speed advantages.

Authors: We thank the referee for noting this omission. Although the manuscript text refers to advantages relative to existing regression-based methods, Table 1 and Figure 3 indeed report only the proposed low-rank kernel results. We will revise Table 1 to include side-by-side quantitative comparisons against the Longstaff-Schwartz least-squares Monte Carlo algorithm and at least one additional kernel-regression baseline, reporting bias, standard error, and wall-clock times on the same test instances. Figure 3 will be updated or supplemented with corresponding timing and error plots for the baselines to allow direct visual verification of the claimed gains. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on standard RKHS learning from external simulated data

full rationale

The paper reformulates continuation value estimation as learning a low-rank linear operator approximation to the conditional expectation in an RKHS from simulated transition data. This learned operator is then reused in the backward dynamic programming recursion. Convergence guarantees and error bounds are stated to follow from standard RKHS approximation theory and Monte Carlo sampling, with an explicit offline training step against held-out or representative simulated paths. No step reduces a claimed prediction to a fitted quantity by construction, no self-citation chain is load-bearing for the core claims, and the offline-online split is justified by the data-driven learning procedure rather than by redefinition or renaming of inputs. The method is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract alone supplies insufficient detail to enumerate free parameters, axioms, or invented entities; standard RKHS and Monte Carlo assumptions are implicit but not itemized.

pith-pipeline@v0.9.0 · 5408 in / 1068 out tokens · 25472 ms · 2026-05-08T06:23:08.853092+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

46 extracted references

[1]

Applied Numerical Mathematics , volume =

Harbrecht, Helmut and Peters, Michael and Schneider, Reinhold , title =. Applied Numerical Mathematics , volume =
[2]

J Sullivan , title =

Ilja Klebanov and Ingmar Schuster and T. J Sullivan , title =. SIAM Journal on Mathematics of Data Science , volume =
[3]

Micchelli and Massimiliano A

Charles A. Micchelli and Massimiliano A. Pontil , title =. Neural Computation , pages =. 2005 , volume =

2005
[4]

Adaptive Joint Distribution Learning , journal =

Filipovi\'. Adaptive Joint Distribution Learning , journal =
[5]

and Schwartz, Eduardo S

Longstaff, Francis A. and Schwartz, Eduardo S. , title =. The Review of Financial Studies , volume =
[6]

Brazilian Review of Finance , volume =

Lin, Jingrui and Almeida, Caio , title =. Brazilian Review of Finance , volume =
[7]

Journal of Political Economy , volume =

Fischer Black and Myron Scholes , title =. Journal of Political Economy , volume =
[8]

International Conference on Algorithmic Learning Theory , pages=

Hilbert space embeddings of distributions , author=. International Conference on Algorithmic Learning Theory , pages=
[9]

SIAM/ASA Journal on Uncertainty Quantification , volume =

Higher-Order Quasi-Monte Carlo for Bayesian Shape Inversion , author =. SIAM/ASA Journal on Uncertainty Quantification , volume =
[10]

Foundations of Computational Mathematics , year =

Caponnetto, Andrea and De Vito, Ernesto , title =. Foundations of Computational Mathematics , year =
[11]

2024 , volume =

Li, Zhu and Meunier, Dimitri and Mollenhauer, Mattes and Gretton, Arthur , title =. 2024 , volume =

2024
[12]

, title =

Zanger, Daniel Z. , title =. 2020 , journal =

2020
[13]

, title =

Zanger, Daniel Z. , title =. 2018 , journal =

2018
[14]

, title =

Zanger, Daniel Z. , title =. 2013 , journal =

2013
[15]

, title =

Zanger, Daniel Z. , title =. 2009 , journal =

2009
[16]

Annals of Applied Probability , volume =

Daniel Egloff , title =. Annals of Applied Probability , volume =
[17]

, title =

Barone-Adesi, Giovanni and Whaley, Robert E. , title =. Journal of Finance , volume =
[18]

Heston , journal =

Steven L. Heston , journal =. A Closed-Form Solution for Options with Stochastic Volatility with Applications to Bond and Currency Options , urldate =
[19]

Journal of Risk and Financial Management , volume =

Becker, Sebastian and Cheridito, Patrick and Jentzen, Arnulf , title =. Journal of Risk and Financial Management , volume =
[20]

Journal of Machine Learning Research , volume =

Becker, Sebastian and Cheridito, Patrick and Jentzen, Arnulf , title =. Journal of Machine Learning Research , volume =
[21]

European Journal of Applied Mathematics , volume =

Becker, Sebastian and Cheridito, Patrick and Jentzen, Arnulf and Welti, Timo , title =. European Journal of Applied Mathematics , volume =
[22]

and Van Roy, Benjamin , title =

Tsitsiklis, John N. and Van Roy, Benjamin , title =. IEEE Transactions on Neural Networks , volume =
[23]

An Analysis of a Least Squares Regression Method for

Cl. An Analysis of a Least Squares Regression Method for. Finance and Stochastics , volume =
[24]

Valuation of the Early-Exercise Price for Options Using Simulations and Nonparametric Regression , journal =

Carri. Valuation of the Early-Exercise Price for Options Using Simulations and Nonparametric Regression , journal =
[25]

, title =

Tilley, James A. , title =. Transactions of the Society of Actuaries , volume =
[26]

Glasserman, Paul , title =
[27]

Pricing of High-Dimensional

Kohler, Michael and Krzy. Pricing of High-Dimensional. Mathematical Finance , volume =
[28]

Journal of Computational Physics , volume =

Sirignano, Justin and Spiliopoulos, Konstantinos , title =. Journal of Computational Physics , volume =
[29]

Proceedings of the National Academy of Sciences , volume =

Han, Jiequn and Jentzen, Arnulf and E, Weinan , title =. Proceedings of the National Academy of Sciences , volume =
[30]

Deep Backward Schemes for High-Dimensional Nonlinear

Hur. Deep Backward Schemes for High-Dimensional Nonlinear. Mathematics of Computation , volume =
[31]

Neural Network Regression for

Lapeyre, Bernard and Lelong, J. Neural Network Regression for. Monte Carlo Methods and Applications , volume =
[32]

Mathematical Finance , volume =

Rogers, Leonard Christopher Gordon , title =. Mathematical Finance , volume =
[33]

and Kogan, Leonid , title =

Haugh, Martin B. and Kogan, Leonid , title =. Operations Research , volume =
[34]

Management Science , volume =

Stentoft, Lars , title =. Management Science , volume =
[35]

and Schwartz, Eduardo S

Brennan, Michael J. and Schwartz, Eduardo S. , title =. Journal of Finance , volume =
[36]

and Ross, Stephen A

Cox, John C. and Ross, Stephen A. and Rubinstein, Mark , title =. Journal of Financial Economics , volume =
[37]

Mathematical Finance , volume =

Keller-Ressel, Martin , title =. Mathematical Finance , volume =
[38]

, title =

Lee, Roger W. , title =. Mathematical Finance , volume =
[39]

, title =

Dudley, Richard M. , title =
[40]

, title =

Karatzas, Ioannis and Shreve, Steven E. , title =
[41]

Conditional mean embeddings as regressors , year =

Gr\". Conditional mean embeddings as regressors , year =. Proceedings of the 29th International Coference on International Conference on Machine Learning , pages =
[42]

Proceedings of the 26th Annual International Conference on Machine Learning , pages =

Song, Le and Huang, Jonathan and Smola, Alex and Fukumizu, Kenji , title =. Proceedings of the 26th Annual International Conference on Machine Learning , pages =. 2009 , publisher =

2009
[43]

Using the Nystr\"

Williams, Christopher and Seeger, Matthias , booktitle =. Using the Nystr\"
[44]

Proceedings of the 36th International Conference on Neural Information Processing Systems , series =

Li, Zhu and Meunier, Dimitri and Mollenhauer, Mattes and Gretton, Arthur , title =. Proceedings of the 36th International Conference on Neural Information Processing Systems , series =. 2022 , publisher =

2022
[45]

Hou, Boya and Sanjari, Sina and Dahlin, Nathan and Bose, Subhonmesh , title =. Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium on Educational Advances in Artificial Intelligence , series =. 2023 , publisher =

2023
[46]

A Measure-Theoretic Approach to Kernel Conditional Mean Embeddings , address =

Park, Junhyung and Muandet, Krikamol , booktitle =. A Measure-Theoretic Approach to Kernel Conditional Mean Embeddings , address =