Initialization-Dependent Solution Selection by the Euler-Maruyama Scheme for a Degenerate SDE

arXiv:2610.13800 · math.PR · Submitted 2026-10-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Mathematics. The day in mathematics: new theorems, new proofs, and what they open.

Ada: We're looking at the paper: "Initialization-Dependent Solution Selection by the Euler-Maruyama Scheme for a Degenerate SDE".

Emmy: Here's the gist of the paper. The Euler–Maruyama approximation of the degenerate stochastic differential equation exhibits an initialization-dependent solution-selection phase transition where subcritical initialization selects a non-sticky solution, critical initialization leaves a finite deterministic holding time, and supercritical initialization selects an absorbing solution Distinguished Weak Solutions The paper introduces two basic weak solutions to the degenerate SDE: the identically zero process is an absorbing solution There is also a distinguished non-sticky solution Xns satisfying Z T 0 1Xns t =0 dt = 0 a.s., T < ∞ Moreover, for every deterministic d ≥ 0, one may keep the process at zero up to time d and then start an independent copy of the non-sticky solution These delayed solutions are weak solutions of the same SDE Euler Scheme and Selection Parameter The main point of this paper is that the statement “xh → 0” is far from sufficient to determine the numerical limit The critical spatial scale is rh = h 1/

2(1−α): , and the relevant depth parameter is Θh = h log 1 + log+ rhxh If Θh → 0, the scheme converges to the non-sticky weak solution If Θh → θ ∈ (0, ∞), it converges to the solution that stays at zero for the deterministic time dθ = θ/ log α and then evolves as the non-sticky solution If Θh → ∞, the scheme converges to the absorbing solution Subcritical Regime Analysis If Θh → 0, then Xh ⇒ Xns in Cloc(

0, ∞)) This convergence is established by the occupation estimate showing that for every T < ∞ there is CT < ∞, depending only on T, α and supnxn, such that for 0 < ε ≤ 1, lim sup n→∞ E Z T 0 1Xn t<ε dt ≤ CT ε 1−2α Critical and Supercritical Regimes If Θh → θ ∈ (0, ∞), then Xh ⇒ X(dθ), dθ = θ/cα, in Cloc([0, ∞)) If Θh → ∞, then Xh −→ 0 in probability in Cloc([0, ∞)) The critical convergence holds on every finite horizon Identification of the non-sticky limit The cluster set of the laws is identified as L(X(θ/cα): θ ∈ C(Θ) ∩ [0,∞): ∪ δ0: ∞ ∈ C(Θ) This implies that the Euler laws converge if and only if C(Θ) is a singleton The Euler–Maruyama approximation of dXt = Xt α dWt, 0 < α < 1/2, exhibits an initialization-dependent solution-selection phase transition where subcritical initialization selects the non-sticky solution, critical initialization leaves a finite deterministic holding time, and supercritical initialization selects the absorbing solution The mechanism separates two effects that are invisible if one only records xh → 0 The exact scale invariance used in the analysis is specific to the power-law coefficient A Exit-time estimates for the rescaled chain The exit-time law of large numbers is established by showing that for every fixed R > 0, τR(v) H(v) −→ 1 cα in probability as v → 0, v ≠ 0 The discrete occupation estimates yield the grid occupation estimate E

h Xk<Nh(T) 1Yh k<ε: ≤ CT ε 1−2α B Discrete occupation estimates The discrete Tanaka remainder provides the estimate E N X−1 k=0 Rh(Yh k, a) = EYh N − a − xh − a The annular and mesoscopic occupation estimates yield the grid occupation estimate E

h Xk<Nh(T) 1Yh k<ε: ≤ CT ε 1−2α C Tightness and identification of the non-sticky limit The Kolmogorov–Chentsov criterion gives tightness of the family of processes in C(

0, T: ) The Brownian time-change construction uniquely determines the law of a non-sticky weak solution This completes the proof of Theorem 6.2 C5 Uniqueness of the non-sticky law The Brownian time-change construction uniquely determines the law of a non-sticky weak solution The Dambis–Dubins–Schwarz theorem gives a Brownian motion B such that Xt = BQt on the stochastic interval

0, Q∞) This construction uniquely determines the law of a non-sticky weak solution Acknowledgments The author used OpenAI ChatGPT (GPT-5.6 Sol, 2026) as an assistive tool for language editing, structural suggestions, literature searches, numerical-code review, exploration of alternative formulations and proof strategies, and critical examination of mathematical arguments The author assumes full responsibility for the originality, correctness, and integrity of the manuscript References [1: Antonov, A., Konikov, M.

Ada: and Spector, M.

Emmy: (2015).

Ada: The free boundary SABR: natural extension to negative rates

2: Bafico, R.

Paper summary: Ada: So we've seen that this paper studies an equation where uniqueness in law fails because of how degenerate the diffusion coefficient is. The main thrust of "Initialization-Dependent Solution Selection by the Euler-Maruyama Scheme for a Degenerate SDE" is showing that starting at zero traps you, but starting near zero lets you choose between different types of solutions.

Emmy: It claims that they can identify this dependence through a critical spatial scale r and a depth parameter theta h. This parameter theta h tells you if the numerical approximation picks the non-sticky solution, gets stuck in a finite delay, or just vanishes.

Kolya: What's interesting is how they link this to an exact scale-invariant Euler chain. They use something called a logarithmic exit-time law of large numbers and a discrete Tanaka occupation estimate to prove these regimes. That means they aren't just guessing; they are using detailed estimates on the occupation of time intervals.

Ada: The paper lays out the trichotomy clearly: if theta h vanishes, you get the non-sticky solution, if it's finite, you get a deterministic delay d related to theta, and if it goes to infinity, you just stay near zero on every compact time interval.

Emmy: The authors emphasize that this separation of behaviors is invisible if you only look at the simple condition that the Euler step size x goes to zero. That initial condition is not enough on its own to determine the numerical limit.

Conclusion: Ada: So to wrap up, this work by Tanaka is about how sensitive numerical solutions can be when the underlying math doesn't have a unique answer. The paper shows that the way you initialize your process matters fundamentally for the resulting model you get from a numerical approximation.

Emmy: Basically, they give us a roadmap based on that depth parameter theta h. If you start close to zero but not too close, you get one type of solution; if you are very far away, or too close in a specific way, you get the absorbing solution. This is about understanding the structure behind why discretization choices matter so much.

Kolya: The implication for us is that when we see numerical results that seem to converge to zero, we need to ask what the initialization looked like relative to that critical scale r. It moves us from just checking convergence to actually characterizing which solution is physically realized by the approximation.

Ada: Exactly. We're moving beyond just saying "the scheme converges" and asking "which convergence path does it take?" This paper gives us a way to categorize those different paths based on simple initial conditions and scale comparisons.

Akihiro Tanaka

math.PR

Submitted: 2026-10-09

Updated: 2026-11-05

Comments: 24 pages. Revised for submission to Electronic Journal of Probability; editorial and disclosure updates

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 80/100

The gist: The Euler–Maruyama approximation of the degenerate stochastic differential equation exhibits an initialization-dependent solution-selection phase transition where subcritical initialization selects…

Key concepts

Degenerate SDE
This is a type of stochastic differential equation where the diffusion term depends on the state variable itself, specifically involving its absolute value multiplied by $\alpha$. This degeneracy causes uniqueness in law to fail, meaning different starting points can lead to completely different long-term behaviors for the process.
Non-sticky solution ($X_{ns}$)
This is a specific weak solution where the process stays at zero for a deterministic time $T$ and then restarts as an independent copy of itself. It represents one possible outcome when the initialization is in the subcritical regime, meaning it converges to this behavior.
Selection Parameter ($\Theta_h$)
This parameter quantifies how far the numerical scheme's behavior is from a simple convergence to zero. It balances the spatial scale of the initialization ($|x_h|$) against the critical spatial scale ($r_h$), determining whether the scheme selects the non-sticky solution, a finite holding time solution, or an absorbing zero solution.

Terminology

Summary

The Euler–Maruyama approximation of the degenerate stochastic differential equation exhibits an initialization-dependent solution-selection phase transition where subcritical initialization selects a non-sticky solution, critical initialization leaves a finite deterministic holding time, and supercritical initialization selects an absorbing solution

Distinguished Weak Solutions

The paper introduces two basic weak solutions to the degenerate SDE: the identically zero process is an absorbing solution There is also a distinguished non-sticky solution Xns satisfying Z T 0 1Xns t =0 dt = 0 a.s., T < ∞ Moreover, for every deterministic d ≥ 0, one may keep the process at zero up to time d and then start an independent copy of the non-sticky solution These delayed solutions are weak solutions of the same SDE

Euler Scheme and Selection Parameter

The main point of this paper is that the statement “xh → 0” is far from sufficient to determine the numerical limit The critical spatial scale is rh = h 1/[2(1−α)], and the relevant depth parameter is Θh = h log 1 + log+ rhxh If Θh → 0, the scheme converges to the non-sticky weak solution If Θh → θ ∈ (0, ∞), it converges to the solution that stays at zero for the deterministic time dθ = θ/ log α and then evolves as the non-sticky solution If Θh → ∞, the scheme converges to the absorbing solution

Subcritical Regime Analysis

If Θh → 0, then Xh ⇒ Xns in Cloc([0, ∞)) This convergence is established by the occupation estimate showing that for every T < ∞ there is CT < ∞, depending only on T, α and supnxn, such that for 0 < ε ≤ 1, lim sup n→∞ E Z T 0 1Xn t<ε dt ≤ CT ε 1−2α

Critical and Supercritical Regimes

If Θh → θ ∈ (0, ∞), then Xh ⇒ X(dθ), dθ = θ/cα, in Cloc([0, ∞)) If Θh → ∞, then Xh −→ 0 in probability in Cloc([0, ∞)) The critical convergence holds on every finite horizon

Identification of the non-sticky limit

The cluster set of the laws is identified as L(X(θ/cα): θ ∈ C(Θ) ∩ [0,∞)] ∪ δ0: ∞ ∈ C(Θ) This implies that the Euler laws converge if and only if C(Θ) is a singleton

Conclusion

The Euler–Maruyama approximation of dXt = Xt α dWt, 0 < α < 1/2, exhibits an initialization-dependent solution-selection phase transition where subcritical initialization selects the non-sticky solution, critical initialization leaves a finite deterministic holding time, and supercritical initialization selects the absorbing solution The mechanism separates two effects that are invisible if one only records xh → 0 The exact scale invariance used in the analysis is specific to the power-law coefficient

A Exit-time estimates for the rescaled chain

The exit-time law of large numbers is established by showing that for every fixed R > 0, τR(v) H(v) −→ 1 cα in probability as v → 0, v ≠ 0 The discrete occupation estimates yield the grid occupation estimate E [h Xk<Nh(T) 1Yh k<ε] ≤ CT ε 1−2α

B Discrete occupation estimates

The discrete Tanaka remainder provides the estimate E N X−1 k=0 Rh(Yh k, a) = EYh N − a − xh − a The annular and mesoscopic occupation estimates yield the grid occupation estimate E [h Xk<Nh(T) 1Yh k<ε] ≤ CT ε 1−2α

C Tightness and identification of the non-sticky limit

The Kolmogorov–Chentsov criterion gives tightness of the family of processes in C([0, T]) The Brownian time-change construction uniquely determines the law of a non-sticky weak solution This completes the proof of Theorem 6.2

C5 Uniqueness of the non-sticky law

The Brownian time-change construction uniquely determines the law of a non-sticky weak solution The Dambis–Dubins–Schwarz theorem gives a Brownian motion B such that Xt = BQt on the stochastic interval [0, Q∞) This construction uniquely determines the law of a non-sticky weak solution

Acknowledgments

The author used OpenAI ChatGPT (GPT-5.6 Sol, 2026) as an assistive tool for language editing, structural suggestions, literature searches, numerical-code review, exploration of alternative formulations and proof strategies, and critical examination of mathematical arguments The author assumes full responsibility for the originality, correctness, and integrity of the manuscript

References

[1] Antonov, A., Konikov, M. and Spector, M. (2015). The free boundary SABR: natural extension to negative rates [2] Bafico, R. p. 1 and Baldi, P. p. 3 (1982). Small random perturbations of Peano phenomena [3] Bass, R.F. (2014). A stochastic differential equation with a sticky point [4] Bass, R. p. 0F., Burdzy, K. and Chen, Z.-Q. (2007). Pathwise uniqueness for a degenerate stochastic differential equation [5] Bou-Rabee, N. p. 0 and Holmes-Cerfon, M. p. 1C. (2020). Sticky Brownian motion and its numerical solution [6] Dambis, K. p. 1E. (1965). On the decomposition of continuous submartingales [7] Delarue, F. p. 3 and Flandoli, F. p. 3 (2014). The transition point in the zero noise limit for a 1D Peano example [8] Dubins, L. p. 3E. and Schwarz, G. (1965). On continuous martingales [9] Engelbert, H. p. 1-J. and Peskir, G. p. 1 (2014). Stochastic differential equations for sticky Brownian motion [10] Engelbert, H. p. 1-J. and Schmidt, W. p. 1 (1985). On solutions of one-dimensional stochastic differential equations without drift [11] Gyöngy, I. p. 1 and Rásonyi, M. p. 3 (2011). A note on Euler approximations for SDEs with Hölder continuous diffusion coefficients [12] Peskir, G. p. 3 (2022). Sticky Bessel diffusions [13] Ramírez, M. p. 1 and Uribe Bravo, G. p. 1 (2024). The sticky Lévy process as a solution to a time change equation [14] Sharma, A. p. 1 (2025). Weak approximation of stochastic differential equations with sticky boundary conditions [15] Suganuma, K. p. 1 and Yamada, T. p. 1 (2017). Forward-rate models allowing for negative interest rates and market expectations of future interest rates [16] Taguchi, D. p. 1 and Tanaka, A. p. 0 (2019). On the Euler–Maruyama scheme for degenerate stochastic differential equations with non-sticky condition

--- Page 3 ---

Initialization-Dependent Solution Selection by the Euler–Maruyama Scheme for a Degenerate SDE Akihiro Tanaka October 2026 Abstract We study the Euler–Maruyama approximation of the degenerate stochastic differential equation dXt = Xt α dWt, 0 < α < 1/2, X0 = 0, for which uniqueness in law fails. The Euler scheme started exactly from zero is trapped by the absorbing solution, whereas a nonzero initialization xh → 0 can select a different weak solution We identify the initialization-dependent limiting regimes The critical spatial scale is rh = h 1/[2(1−α)], and the relevant depth parameter is Θh = h log 1 + log+ rhxh If Θh → 0, the scheme converges to the non-sticky weak solution If Θh → θ ∈ (0, ∞), it converges to the solution that stays at zero for the deterministic time dθ = θ/ log α and then evolves as the non-sticky solution If Θh → ∞, the scheme converges to the absorbing solution

--- Page 4 ---

Let h > 0 and let xh ≠ 0 with xh → 0. We consider Xh t = xh + Z t 0 Xh ηh(s) α dW, ηh(s) = ⌊s/h⌋ h. p. 1 (2) If xh = 0, the scheme remains identically zero It is therefore natural to perturb the initialization The main point of this paper is that the statement “xh → 0” is far from sufficient to determine the numerical limit The first relevant scale follows from balancing one Euler increment with the current position If x is of order rh, then x α√h is of the same order as x precisely when rh = h 1/[2(1−α)], rα h√h = rh

--- Page 5 ---

At the grid points Yh k = Xh kh, the rescaled process Vh k = Yh k /rh satisfies the exact recursion Vh k+1 = Vh k + Vh k α Zk+1, (4) which no longer contains h The numerical selection problem is therefore reduced to the exit behavior of a single scale-invariant Markov chain started from vh = xh/rh

--- Page 6 ---

The chain has a logarithmic depth. We use H(v) = log 1 + log+ 1v, v ≠ 0, (5) For very small v, one step transforms L = log(1/v) approximately into αL; consequently H decreases by approximately log α per step This suggests the physical exit time hτR(vh) ≈ hH(vh) log α This leads to the selection parameter Θh = hH(xh/rh) = h log 1 + log+ rhxh

--- Page 7 ---

Our main result is the following trichotomy If Θh → 0, the microscopic trapping period disappears and the Euler approximation selects Xns If Θh → θ ∈ (0, ∞), it converges to X(dθ), dθ = θ/ log α If Θh → ∞, the scheme remains close to zero on every compact time interval

Pruning redundant reasoning tokens

The first relevant scale follows from balancing one Euler increment with the current position The exact reduction is special to the power-law coefficient and is deliberately exploited throughout the paper

How the training signal is built

The chain has a logarithmic depth. We use H(v) = log 1 + log+ 1v, v ≠ 0, (5) For very small v, one step transforms L = log(1/v) approximately into αL; consequently H decreases by approximately log α per step This suggests the physical exit time hτR(vh) ≈ hH(vh) log α This leads to the selection parameter Θh = hH(xh/rh) = h log 1 + log+ rhxh

Where the gains come from

The main result is the following trichotomy If Θh → 0, the microscopic trapping period disappears and the Euler approximation selects Xns If Θh → θ ∈ (0, ∞), it converges to X(dθ), dθ = θ/ log α If Θh → ∞, the scheme remains close to zero on every compact time interval The critical regimes can still be investigated without underflow by evolving the logarithmic depth variable in the exact representation of Appendix A

If xh = 0, the scheme remains identically zero It is therefore natural to perturb the initialization The main point of this paper is that the statement “xh → 0” is far from sufficient to determine the numerical limit

The chain has a logarithmic depth.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Development of an initialization-dependent solution selector for numerical approximations of degenerate stochastic differential equations based on a critical depth parameter Theta h. The improved system can distinguish between convergence to non-sticky solutions, solutions with deterministic time delays, and convergence to the absorbing zero solution by analyzing the behavior of this parameter as it approaches zero, a positive constant, or infinity.

  2. Implementation of a scale-invariant Euler chain analysis to determine physical time from numerical steps using the formula S R h = Theta h tau R H(v). This allows the system to predict when an initialization will lead to a specific macroscopic delay d theta = theta / log alpha, enabling the selection of delayed weak solutions.

  3. Automated classification of initialization sequences based on their rate of vanishing, specifically distinguishing between subcritical initializations (like polynomial or algebraic ones) and supercritical ones (requiring double-exponential depth). This provides a rigorous method for selecting the correct numerical solution law when direct convergence to zero is insufficient.

  4. Numerical verification of the cluster set of Euler laws by analyzing the cluster set of the selection parameter Theta h. The system can determine if different initialization rates lead to distinct limiting processes, confirming that the Euler laws converge if and only if C(Theta) is a singleton.

  5. Computation of occupation estimates in degenerate SDEs using discrete Tanaka arguments applied directly to the Euler chain, replacing standard continuous interpolation formulas. This allows for the rigorous proof of convergence to the non-sticky law under subcritical conditions where previous methods failed due to lack of control over the lagged state.

Sources

Related papers