Strong convergence of path sensitivities

Michael B. Giles

Strong convergence of path sensitivities

T0 review · 0 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Differentiating the Euler-Maruyama approximation with respect to a parameter yields path sensitivity estimates whose strong error is O(h^{1/2}), matching the path approximation itself.

desk verdict Short proof note that closes a real gap: Euler-Maruyama path sensitivities converge strongly with order 1/2 under bounded derivative assumptions. read the letter →

arxiv 2411.15930 v1 pith:ZMM7E4FL submitted 2024-11-24 math.NA cs.NA

Michael B. Giles This is my paper · ORCID

classification math.NAcs.NA MSC 65C3060H3565C05

keywords Euler-MaruyamapathwisesensitivitiesstrongconvergencestochasticdifferentialequationsmultilevelMonteCarloGreeksinfinitesimalperturbationanalysisBurkholder-Davis-Gundyinequality

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 note proves a convergence result that stochastic numerical analysis had left open: if you estimate the derivative of an SDE path with respect to a model parameter by differentiating the Euler-Maruyama scheme, the estimate converges strongly with order $1/2$ on a finite time interval. That is the same strong order as the Euler-Maruyama approximation of the path itself, despite the fact that the differentiated SDE violates the global Lipschitz condition normally required for the standard proof. The theorem covers scalar and vector parameters and paths, and by induction covers second- and higher-order sensitivities whenever the needed derivatives of drift and diffusion exist and are bounded. Because multilevel Monte Carlo error analysis relies on strong orders, the result supplies the missing ingredient for a rigorous treatment of pathwise Greeks in that setting.

What carries the argument

The central object is the pathwise sensitivity SDE, $d\dot S_t = (\dot a_t + a'_t \dot S_t)\,dt + (\dot b_t + b'_t \dot S_t)\,dW_t$, obtained by differentiating the original SDE in the parameter $\theta$. Its Euler-Maruyama discretisation is exactly the derivative of the Euler-Maruyama discretisation of the original SDE, which is what makes the estimator cheap to compute. The proof machinery is the continuous-time interpolation of both schemes followed by a term-by-term decomposition of the error $E_t = \hat{\dot S}_t - \dot S_t$ into twelve integrals, each bounded by Jensen, H\"older, the Burkholder-Davis-Gundy inequality, the known $O(h^{p/2})$ path error from [10], and Gr\"onwall's inequality; boundedness of first and second derivatives supplies the constants $L_a, L_b$ that keep every product controlled.

What would settle it

Run the sensitivity estimator on a scalar SDE with bounded, smooth coefficients, for example $a(\theta,S)=\sin(S+\theta)$, $b(\theta,S)=\cos(S+\theta)$, computing $\dot S$ exactly from the same Brownian increments (or a very fine reference simulation), and estimate $E[\sup_{0<t<T}|\hat{\dot S}-\dot S|^2]$ for $h=2^{-4},...,2^{-8}$. If the error does not decay proportionally to $h$, the theorem's rate would be contradicted. Conversely, replacing the diffusion by an unbounded-derivative coefficient, such as $b(\theta,S)=1+S^2$, and observing a slower rate would confirm that the bounded-derivative assumption is load-bearing.

Watch

Extended reading notes

Core claim

Theorem 2 is the central result: under the assumption that the first and second derivatives of the drift $a$ and diffusion $b$ with respect to $S$ and $\theta$ exist and are uniformly bounded, for any $p \ge 2$ there is a constant $c_p^{(3)}$ such that $E[\sup_{0<t<T}|\hat{\dot S}_t - \dot S_t|^p] \le c_p^{(3)} h^{p/2}$, where $\hat{\dot S}$ is the continuous-time interpolation of the Euler-Maruyama approximation to the sensitivity SDE and $\dot S$ is the exact pathwise sensitivity. In words, the sensitivity estimator has $L^p$ strong order $1/2$, matching the classical Euler-Maruyama path bound. The proof re-traces the standard Euler-Maruyama analysis and controls the extra terms through the boundedness of $a'$, $b'$ and the cross-differences; the same argument extends to vector SDEs, vector parameters, and $k$-th order sensitivities, with Lemma 4 bounding products of differences at each induction step.

Load-bearing premise

The proof assumes the first and second derivatives of the drift and diffusion coefficients with respect to both the state and the parameter exist and are uniformly bounded for all parameter and state values; if any of those derivatives can grow without bound, the constants used to control every term in the error decomposition do not exist and the argument collapses.

Editorial extensions

If this is right

With the strong order $1/2$ established for sensitivity estimators, the standard multilevel Monte Carlo variance and cost analysis applies to pathwise sensitivity estimates, not just to path estimates.
Second- and higher-order sensitivities converge at the same strong order, provided the corresponding derivatives of the drift and diffusion are bounded, so higher-order Greeks inherit the same convergence behaviour.
The argument covers multi-dimensional SDEs and multi-parameter sensitivities, so the result is not limited to scalar test problems.
The author conjectures that the same analysis extends to the Milstein scheme, leaving that extension for future work.

Reading between the lines

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

The uniform boundedness of derivatives is likely stronger than needed; the proof only uses $L^p$ control of products, so polynomial-growth coefficients with finite moments may give the same order. This is an editorial inference, not a claim of the paper.
The same re-tracing argument should transfer to other one-step schemes whose path approximation has strong order $1/2$ or $1$, such as Milstein; a direct test would be to verify the sensitivity error slope numerically on a model with bounded derivatives.
A simple numerical check would settle the practical rate: for a model like $dS_t = \sin(S_t+\theta)\,dt + \cos(S_t+\theta)\,dW_t$, estimate $E[\sup|\hat{\dot S}-\dot S|^2]$ over many paths at $h=2^{-4},...,2^{-8}$ and compare the slope to $1/2$; the paper gives no numerical experiments.

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.

Desk Editor's Note

read the letter

First thing to know: this is a clean, short proof note that closes a real gap. It proves that Euler-Maruyama approximations of path sensitivities (the Greeks) converge strongly with order 1/2, matching the path approximation itself. The obstacle is real—the sensitivity SDE has coefficients that are not globally Lipschitz because b' times Sdot appears—and the proof handles it with the bounded-derivative assumption. That is exactly the sort of result you want before using pathwise sensitivities inside multilevel Monte Carlo.

What's new: prior literature proved strong convergence of the Euler-Maruyama path but not for the pathwise derivative with respect to a parameter. Giles shows that if the first and second derivatives of drift and diffusion are uniformly bounded, then the sensitivity approximation converges in L^p with the usual half-order rate. The proof is a retracing of the Kloeden-Platen argument, using Jensen, Hölder, BDG, and Grönwall. I retraced the twelve-term decomposition in Theorem 2; the strategy works. The stress-test note flagged one notational slip where the text writes |S_u - S_u| but the coefficient depends on the floor value; that is harmless because the floor-value difference is still O(h^{p/2}) by the triangle inequality.

The soft spots are mostly about scope. The bounded-derivative assumption excludes many common financial models (e.g., CIR, Heston) where coefficients have unbounded derivatives near zero. That is stated openly, so it's a limitation, not a flaw. The vector and higher-order extensions are sketched, not fully proved, which is fine for a note whose main target is the scalar first-order theorem. There are no numerical experiments and no machine-checked formalization, but the proof is standard enough that this does not undermine confidence.

I think the result is correct and worth having in the literature. The main theorem is properly proved, and the gap it fills is real. It deserves a serious referee; I'd send it out. For a journal, minor revision would be enough—clarify the floor notation and maybe add a remark on what would be needed to relax the bounded-derivative condition. For someone working on MLMC Greeks, this is a useful citation; for others in numerical SDEs, it's a tidy, correct result that resolves a known open point.

Referee Report

0 major / 5 minor

Summary. This note proves that the Euler-Maruyama discretisation of the pathwise sensitivity SDE (the derivative of the SDE solution with respect to a parameter) has strong L^p error of order 1/2 on finite time intervals, under the assumption that the drift and diffusion coefficients have bounded first and second derivatives. The main result is Theorem 2, which gives E[sup_{0<t<T} |\hat{\dot S}_t - \dot S_t|^p] <= c_p^{(3)} h^{p/2} for any p>=2. The proof follows the classical EM strong convergence proof, decomposing the error into twelve drift and diffusion terms and bounding each with Jensen, Hölder, BDG, and Grönwall inequalities. The paper also sketches extensions to vector SDEs/parameters and higher-order sensitivities.

Significance. The result is significant because the combined process (S, \dot S) does not satisfy the standard global Lipschitz condition (the product b'(S) \dot S is not Lipschitz in the joint variable), so the usual EM convergence theorem cannot be applied directly. The note fills this gap with a self-contained proof that uses only standard tools and explicitly stated bounded-derivative assumptions. The main theorem gives a rigorous foundation for the use of pathwise sensitivities in multilevel Monte Carlo and Greeks computations. The paper is honest about what is proved completely (the scalar first-order case) and what is only sketched (vector and higher-order extensions); the sketched extensions do not affect the central claim. The proof contains no fitted parameters or circular arguments.

minor comments (5)

[Section 3, Lemma 3] The proof of Lemma 3 is given as a single sentence ('The proof follows the same approach used with Theorem 1'); since this lemma is used in Theorem 2 to control moments of \hat{\dot S}, please expand the proof to show the Gronwall argument for the piecewise-constant coefficient process explicitly.
[Section 3, proof of Theorem 2] The displayed inequalities for the terms involving (\hat a'_u - a'_u) \hat{\dot S}_u and similar write E|\hat S_u - S_u|^p where the mean value theorem actually gives |\hat S_{\underline u} - S_u|; the two are interchangeable up to constants by the triangle inequality, but the notation should be made precise.
[Section 2, opening paragraph] The assumptions list bounded derivatives but do not state the finite-moment assumptions on S0 and \dot S0; please add explicit conditions such as E|S0|^p < ∞ and E|\dot S0|^p < ∞ (for the relevant p) to make the Gronwall arguments fully rigorous.
[Abstract and Introduction] There are a few typos ('a n autonomous', 'pro ves') that should be corrected.
[Section 4, Extensions] The extension to vector SDEs and higher-order sensitivities is only sketched; if the journal allows, a remark stating that the details are omitted for brevity would be helpful, or alternatively provide a brief indication of the inductive step.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the strong convergence proof for Euler-Maruyama path sensitivities is self-contained and relies on standard external results.

full rationale

The paper proves that the Euler-Maruyama approximation of the pathwise sensitivity SDE converges with strong order 1/2. The derivation chain is fully self-contained: it starts from the standard Euler-Maruyama strong convergence bound for the underlying path, E[sup |S_hat - S|^p] = O(h^{p/2}) from Kloeden and Platen, and combines it with explicit bounds on the sensitivity moments (Theorem 1, Lemma 2) and on the discretized sensitivity moments (Lemma 3). The core of Theorem 2 decomposes the error into twelve terms, each bounded by either the known path-error rate, the known increment bound E[|S_s - S_s|^p] = O(h^{p/2}), the new Lemma 2 increment bound for the sensitivity, or Holder/Cauchy-Schwarz estimates using finite moments. The final Gronwall step yields Z_t <= c1 h^{p/2} + c2 ∫ Z_s ds, so no fitted parameter or self-referential input is used. The only assumptions are the stated uniform boundedness of first and second derivatives of a and b, which are explicit and not derived from the target result. Citations to Burdzy-Davis-Gundy, Kloeden-Platen, and Gronwall are standard external mathematical facts, not self-citations that carry the argument. The paper does note that the higher-order and vector extensions are sketched rather than fully proved, but that is a stated limitation, not circularity. No circular step is present.

Assumptions & free parameters 0 free parameters · 7 assumptions · 0 invented entities

No free parameters are fitted to data; the paper introduces only existential constants arising from inequalities. The proof relies on standard stochastic analysis theorems and on explicit smoothness and boundedness assumptions on the coefficients. No new entities, forces, or dimensions are postulated.

assumptions (7)

standard math Itô's lemma
Used in Lemma 1 to derive the differential inequality for E[|dot S_t|^p].
standard math Burkholder-Davis-Gundy inequality
Used in Theorem 1, Lemma 2, and Theorem 2 to bound stochastic integrals.
standard math Grönwall's inequality
Used repeatedly to convert integral inequalities into finite-time bounds.
standard math Hölder's and Jensen's inequalities
Used to bound product terms and to pass from even integer moments to general p.
standard math Known Euler-Maruyama strong convergence for the underlying SDE (Kloeden-Platen Theorem 10.6.3)
Assumes E[|hat S_u - S_u|^p] = O(h^{p/2}) and E[|S_u - underline S_u|^p] = O(h^{p/2}), which are inputs to the main theorem.
domain assumption Uniform boundedness of derivatives of drift and diffusion up to second order
Stated at the start of Section 2; provides the constants L_a and L_b used throughout the proof.
domain assumption Globally Lipschitz drift and diffusion for the original SDE
Required for the standard Euler-Maruyama strong convergence theorem and for existence of the relevant moments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Strong convergence of path sensitivities." pith.science (2026). https://pith.science/paper/ZMM7E4FL

@misc{pith2026241115930,
  author       = {Pith},
  title        = {Pith review of: Strong convergence of path sensitivities},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZMM7E4FL}},
  note         = {Machine review of arXiv:2411.15930}
}

abstract

It is well known that the Euler-Maruyama discretisation of an autonomous SDE using a uniform timestep $h$ has a strong convergence error which is $O(h^{1/2})$ when the drift and diffusion are both globally Lipschitz. This note proves that the same is true for the approximation of the path sensitivity to changes in a parameter affecting the drift and diffusion, assuming the appropriate number of derivatives exist and are bounded. This seems to fill a gap in the existing stochastic numerical analysis literature.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

11 extracted references · 11 canonical work pages

[1]

Broadie and P

M. Broadie and P. Glasserman. Estimating security price derivativ es using simulation. Management Science , 42(2):269–285, 1996

work page 1996
[2]

S. Burgos. The computation of Greeks with multilevel Monte Carlo . DPhil thesis, University of Oxford, 2014. 9

work page 2014
[3]

Burgos and M.B

S. Burgos and M.B. Giles. Computing Greeks using multilevel path sim ulation. In L. Plaskota and H. Wo´ zniakowski, editors,Monte Carlo and Quasi-Monte Carlo Meth- ods 2010 , pages 281–296. Springer, 2012

work page 2010
[4]

Burkholder, B

D.L. Burkholder, B. Davis, and R.F. Gundy. Integral inequalities f or convex functions of operators on martingales. In Proc. Sixth Berkeley Symposium Math. Statist. Prob., Vol II, pages 223–240. University of California Press, Berkeley, 1972

work page 1972
[5]

Capriotti and M.B

L. Capriotti and M.B. Giles. 15 years of adjoint algorithmic diﬀeren tiation in ﬁnance. Quantitative Finance , 24(9):1353–1379, 2024

work page 2024
[6]

M.B. Giles. Multilevel Monte Carlo methods. Acta Numerica, 24:259–328, 2015

work page 2015
[7]

Giles and P

M.B. Giles and P. Glasserman. Smoking adjoints: fast Monte Carlo G reeks. RISK, January 2006

work page 2006
[8]

Glasserman

P. Glasserman. Monte Carlo Methods in Financial Engineering . Springer, New York, 2004

work page 2004

Show all 11 references

[9]

Heinrich

S. Heinrich. Multilevel Monte Carlo methods. In Multigrid Methods , volume 2179 of Lecture Notes in Computer Science , pages 58–67. Springer, 2001

2001
[10]

Kloeden and E

P.E. Kloeden and E. Platen. Numerical Solution of Stochastic Diﬀerential Equations . Springer, Berlin, 1992

1992
[11]

L’Ecuyer

P. L’Ecuyer. A uniﬁed view of the IPA, SF and LR gradient estimat ion techniques. Management Science , 36(11):1364–1383, 1990. 10

1990

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

Broadie and P

M. Broadie and P. Glasserman. Estimating security price derivativ es using simulation. Management Science , 42(2):269–285, 1996

work page 1996

[2] [2]

S. Burgos. The computation of Greeks with multilevel Monte Carlo . DPhil thesis, University of Oxford, 2014. 9

work page 2014

[3] [3]

Burgos and M.B

S. Burgos and M.B. Giles. Computing Greeks using multilevel path sim ulation. In L. Plaskota and H. Wo´ zniakowski, editors,Monte Carlo and Quasi-Monte Carlo Meth- ods 2010 , pages 281–296. Springer, 2012

work page 2010

[4] [4]

Burkholder, B

D.L. Burkholder, B. Davis, and R.F. Gundy. Integral inequalities f or convex functions of operators on martingales. In Proc. Sixth Berkeley Symposium Math. Statist. Prob., Vol II, pages 223–240. University of California Press, Berkeley, 1972

work page 1972

[5] [5]

Capriotti and M.B

L. Capriotti and M.B. Giles. 15 years of adjoint algorithmic diﬀeren tiation in ﬁnance. Quantitative Finance , 24(9):1353–1379, 2024

work page 2024

[6] [6]

M.B. Giles. Multilevel Monte Carlo methods. Acta Numerica, 24:259–328, 2015

work page 2015

[7] [7]

Giles and P

M.B. Giles and P. Glasserman. Smoking adjoints: fast Monte Carlo G reeks. RISK, January 2006

work page 2006

[8] [8]

Glasserman

P. Glasserman. Monte Carlo Methods in Financial Engineering . Springer, New York, 2004

work page 2004