How Stochastic-Local-Volatility Calibration Algorithms Fit Option Surfaces: Leverage Functions, Conditional Variance, Particle Methods, McKean–Vlasov Feedback and Low-Density Failure

Reader question: Local-volatility models can fit a vanilla option surface closely, while stochastic-volatility models often produce more realistic random volatility dynamics. Can one model combine both—and if so, what mathematical problem must be solved?

Stochastic-local-volatility (SLV) models combine a stochastic variance factor with a state-dependent leverage function. The leverage function is calibrated so that the model reproduces a target local-volatility surface while keeping a stochastic volatility state.

This article owns one precise computational job: calibrating the leverage function of a stochastic-local-volatility model through a conditional-variance fixed point. It does not own Dupire local-volatility extraction, Heston parameter calibration, Andreasen–Huge surface fitting, SVI smile parameterisation or generic Monte Carlo simulation.

This is public mathematical and computational education. It is not financial advice, a trading recommendation or a claim that perfect vanilla repricing proves correct exotic-option dynamics.

1. A canonical SLV model

A common Heston-type stochastic-local-volatility model writes:

dS_t = (r_d−r_f)S_tdt + L(t,S_t)√V_tS_tdW_t^S,

dV_t = κ(θ−V_t)dt + ξ√V_tdW_t^V,

with:

dW^SdW^V = ρ dt.

The stochastic-variance parameters are:

• κ — mean reversion;
• θ — long-run variance level;
• ξ — volatility of variance;
• ρ — spot/variance correlation;
• V₀ — initial variance.

The new object is:

L(t,S), the leverage function.

2. Three limiting cases

The SLV architecture is easiest to understand through limits.

If:

L(t,S) ≡ 1,

the model reduces to a pure stochastic-volatility model.

If stochastic variance becomes deterministic, the product:

L(t,S)√V_t

acts like a local-volatility function.

Therefore the SLV model can be viewed as a hybrid whose local component repairs the vanilla surface while the stochastic factor carries additional volatility dynamics.

3. The local-volatility target

Assume an arbitrage-controlled option surface has already been converted into a local-volatility function:

σ_loc(t,s).

That target belongs conceptually to the Dupire layer.

The SLV calibration problem is:

choose L(t,s) so the one-dimensional marginal distribution of S_t matches the local-volatility model.

4. The Gyöngy/mimicking relation

Under the usual conditions, the local variance consistent with the SLV spot marginal satisfies:

σ_loc²(t,s) = L²(t,s) E[V_t | S_t=s].

Therefore:

L²(t,s) = σ_loc²(t,s) / E[V_t | S_t=s].

or, taking the positive square root:

L(t,s) = σ_loc(t,s) / √E[V_t | S_t=s].

This is the central calibration equation.

5. Why calibration is nonlinear

The formula looks explicit, but it is not.

The conditional expectation:

E[V_t | S_t=s]

depends on the joint distribution of (S_t,V_t).

That joint distribution depends on the leverage function L being calibrated.

So L appears on both sides of the problem indirectly:

L → joint distribution → conditional variance → updated L.

This is a fixed-point / McKean–Vlasov-type feedback loop.

6. Why the problem is often described as McKean–Vlasov

The spot diffusion coefficient depends on a conditional expectation of the model’s own distribution.

Schematically:

dS = … + [σ_loc(t,S)/√E(V|S)] √V S dW.

The law of the process enters its coefficients.

This makes the calibrated SLV system a nonlinear distribution-dependent stochastic differential equation rather than an ordinary Markov diffusion with fixed coefficients.

Recent mathematical work studies existence and particle approximations for precisely this calibration problem.

7. Particle calibration: the core idea

Suppose we simulate N particles:

(S_t⁽ⁱ⁾,V_t⁽ⁱ⁾), i=1,…,N.

At a target spot s, estimate:

E[V_t|S_t=s]

from nearby particles.

A kernel estimator is:

m̂(t,s) = [Σ K((S_t⁽ⁱ⁾−s)/h)V_t⁽ⁱ⁾] / [Σ K((S_t⁽ⁱ⁾−s)/h)].

Then update:

L(t,s) = σ_loc(t,s)/√m̂(t,s).

The particles are then advanced using this leverage estimate.

8. Bandwidth h controls a bias–variance trade-off

If h is too small:

• few particles contribute;
• conditional-variance estimates are noisy;
• L can become jagged or explode.

If h is too large:

• too many distant spot states are mixed;
• the conditional expectation is over-smoothed;
• smile fit can weaken.

The bandwidth is therefore a calibration parameter and must be logged, stressed and validated.

9. Low-density regions are the principal numerical danger

In a deep tail region, the denominator:

Σ K((S⁽ⁱ⁾−s)/h)

can be tiny.

Then the estimated conditional variance becomes unstable and the leverage function can become very large.

The 2019 inverse-problem literature explicitly focuses on stabilising SLV calibration in regions where the joint probability density is low.

Falsifier: report effective local particle count or kernel denominator across the leverage grid. A beautiful leverage surface in a region with almost no particles is not evidence-driven.

10. Regularisation

To prevent unstable tail behaviour, implementations can use:

• minimum denominator floors;
• conditional-variance floors;
• smoothing penalties;
• tail extrapolation rules;
• Tikhonov-type regularisation;
• adjoint/PDE inverse formulations;
• more particles or adaptive bandwidth.

Every regularisation changes the inverse problem. It should be reported as part of the model, not hidden as numerical housekeeping.

11. A time-slice calibration algorithm

1. Build an arbitrage-controlled implied-volatility surface.
2. Extract a target local-volatility surface.
3. Choose/calibrate stochastic-volatility parameters.
4. Initialize particles at S₀,V₀.
5. At the next time slice, estimate E[V|S=s] on the spot grid.
6. Compute/update L(t,s).
7. Simulate all particles to the next time slice.
8. Repeat through maturity.
9. Reprice vanilla options under the calibrated SLV model.
10. Compare model and target implied-volatility surfaces.

This “march forward while recalibrating the conditional expectation” is the practical heart of many particle methods.

12. Fixed-point iteration

Some implementations iterate within a time slice:

L⁽⁰⁾ → simulate/solve density → m⁽⁰⁾ → L⁽¹⁾ → …

until the leverage function or repricing residual stabilises.

Damping can help:

L_new = (1−ω)L_old + ωL_update, 0<ω≤1.

Falsifier: preserve the residual sequence. A fixed-point algorithm that oscillates or depends strongly on initialization has not converged merely because a maximum iteration count was reached.

13. PDE/Fokker–Planck calibration

Instead of particles, one can evolve the joint density p(t,s,v) using a forward Kolmogorov/Fokker–Planck PDE.

The conditional expectation is then:

E[V|S=s] = [∫ v p(t,s,v) dv] / [∫ p(t,s,v) dv].

This can be smoother than Monte Carlo but creates a two-dimensional density PDE plus leverage feedback.

The PDE grid must resolve:

• low spot tails;
• variance near zero;
• correlation cross-derivatives;
• boundary conditions;
• density conservation.

14. Adjoint/inverse-problem approaches

SLV calibration can also be formulated as an inverse problem in which the leverage function is adjusted to minimize option-price errors subject to a forward PDE.

Adjoint methods compute sensitivities of many calibration instruments with respect to the functional coefficient efficiently.

This is distinct from the direct conditional-expectation formula but targets the same object: a leverage surface consistent with the desired vanilla prices and chosen stochastic-volatility parameters.

15. The stochastic-volatility parameters still matter

Because the leverage function can repair vanilla prices, different stochastic-volatility parameter sets can sometimes all achieve a strong vanilla fit after recalibration.

But they can imply very different:

• forward-smile dynamics;
• spot/volatility correlation;
• variance-of-variance;
• barrier behavior;
• VIX/volatility-product prices;
• hedging behavior.

This is a fundamental identifiability boundary: vanilla spot options identify the marginal spot distribution much more strongly than the joint spot–variance dynamics.

16. Calibrate the stochastic-volatility backbone first

A common practical approach is:

1. choose Heston-like parameters from a stochastic-volatility calibration or dynamics criterion;
2. hold or constrain those parameters;
3. solve for the leverage function that restores vanilla surface fit.

If every backbone parameter and every leverage node is fit simultaneously without structure, the inverse problem can become extremely underidentified.

17. Feller behavior and variance near zero

For a CIR/Heston variance process:

dV = κ(θ−V)dt + ξ√V dW.

The classical Feller condition is:

2κθ ≥ ξ².

If it is violated, the process can approach or hit zero depending on parameter regime/interpretation.

This matters because the leverage equation divides by:

√E[V|S=s].

Very small conditional variance can create extreme leverage values and density singularities.

The SIAM hybrid calibration literature explicitly addresses difficulties when zero variance is attainable.

18. Inputs and outputs

Inputs can include:

• discount/dividend or domestic/foreign curves;
• arbitrage-controlled implied-volatility surface;
• target local-volatility surface;
• stochastic-volatility parameters;
• spot/variance correlation;
• particle count or PDE grid;
• time steps;
• kernel and bandwidth;
• regularisation parameters;
• leverage grid and interpolation/extrapolation;
• simulation scheme for V.

Outputs can include:

• leverage surface L(t,s);
• estimated conditional variance E[V|S=s];
• local effective particle counts;
• vanilla repricing errors;
• model implied-volatility surface;
• tail regularisation flags;
• fixed-point convergence history;
• density normalization errors;
• exotic-price sensitivity to backbone parameters;
• simulation diagnostics.

19. Evidence polarity

Evidence for confidence includes:

• vanilla prices reprice within intended tolerance;
• leverage surface is stable under increased particle count/grid refinement;
• conditional-variance estimates are supported by adequate local density;
• fixed-point residuals decrease;
• density integrates to one in PDE methods;
• particle and PDE implementations agree on benchmarks;
• tail regularisation changes only low-information regions;
• exotic values are stable across reasonable calibration controls.

Evidence against confidence includes:

• leverage spikes in low-density tails;
• strong bandwidth sensitivity;
• particle count materially changes central-region leverage;
• vanilla residuals remain structured;
• fixed-point iteration oscillates;
• variance-zero behavior destabilises the denominator;
• exotic prices vary widely across equally good vanilla calibrations.

20. Counterexample: perfect local-vol fit, unstable leverage tail

Suppose the local-vol target is smooth but only three particles lie near a deep downside spot node.

The kernel conditional variance there is noisy, producing a huge L.

Vanilla options may barely notice because that state has low probability, yet a barrier or tail exotic can be very sensitive.

Falsifier: map leverage stability against local probability mass and stress tail regularisation.

21. Counterexample: too-narrow kernel

A very small bandwidth makes m̂(t,s) follow particle noise.

The leverage surface becomes jagged, and the next simulation step feeds that noise back into the distribution.

Falsifier: increase h and N independently. If the leverage surface changes materially while vanilla error barely changes, the conditional estimate is under-smoothed.

22. Counterexample: too-wide kernel

A wide kernel stabilises the denominator but averages variance across spot states that have genuinely different conditional volatility.

Falsifier: compare implied-volatility residuals by strike and conditional-variance cross-sections. Persistent smile flattening indicates over-smoothing.

23. Counterexample: low particle count creates self-reinforcing error

An inaccurate conditional variance changes L; the changed L alters future particles; those particles then produce the next conditional estimate.

This feedback can turn ordinary Monte Carlo noise into a path-dependent calibration error.

Falsifier: repeat calibration with independent seeds and larger N. The leverage surface, not only final price, should be stable within an uncertainty band.

24. Counterexample: a bad local-vol target poisons the SLV model

If the input implied surface violates static arbitrage or the Dupire derivatives are unstable, σ_loc can become negative/erratic.

SLV cannot repair a corrupted target.

Falsifier: audit the local-vol surface before SLV calibration. The upstream surface is a dependency, not a detail.

25. Counterexample: two SLV models fit all vanillas but disagree on barriers

Backbone A has mild variance-of-variance; Backbone B has high ξ and different ρ. Each gets its own leverage surface and both reprice the same vanilla surface.

A path-dependent barrier price can differ materially.

Falsifier: treat exotic disagreement as model-risk evidence. Vanilla fit alone cannot select the joint dynamics.

26. Counterexample: particle simulation mishandles variance positivity

A naive Euler step can produce negative V.

Clipping V after the fact changes the stochastic process and therefore the conditional expectation used to calibrate L.

Falsifier: compare full truncation/QE or a validated positivity-aware scheme and rerun leverage calibration. Simulation choice is part of the calibrated model.

27. SLV versus Dupire local volatility

Dupire local-volatility algorithms own the one-factor state-dependent volatility surface extracted from vanilla option prices.

SLV uses that local-vol surface as a marginal target while adding stochastic variance.

28. SLV versus Heston calibration

Heston calibration algorithms own the pure stochastic-volatility parameter fit.

SLV adds a functional degree of freedom L(t,S), specifically to recover a desired spot marginal surface.

29. SLV versus Andreasen–Huge

Andreasen–Huge local-volatility algorithms own a one-maturity-at-a-time local-volatility calibration approach.

SLV is a different inverse problem: given a local-vol target and stochastic variance law, recover the leverage function consistent with their joint dynamics.

30. SLV versus SVI

SVI volatility-smile algorithms own static smile/surface parameterisation.

SVI can be an upstream smoothing/arbitrage layer; it is not the SLV dynamics or leverage calibration.

31. Alternatives

Pure local volatility: exact marginal fit with deterministic state-dependent volatility, simpler calibration.

Pure stochastic volatility: fewer functional degrees of freedom, potentially more coherent volatility dynamics but imperfect smile fit.

Rough-volatility models: different temporal structure for volatility dynamics.

Local-volatility with stochastic rates: adds rate dynamics without stochastic variance.

Adjoint/PDE SLV calibration: alternative numerical owner for the same leverage function.

32. Weak links

• arbitrage-contaminated input surface;
• unstable Dupire local volatility;
• poorly identified stochastic-volatility backbone;
• low local particle density;
• kernel bandwidth choice;
• variance process near zero;
• simulation discretisation bias;
• fixed-point non-convergence;
• tail extrapolation/regularisation;
• assuming vanilla fit identifies exotic dynamics.

33. What would falsify confidence?

Confidence should be withdrawn if vanilla repricing fails; if leverage changes materially with particle count/bandwidth/grid refinement; if low-density tails dominate exotic results; if fixed-point residuals do not converge; if particle and PDE implementations disagree beyond numerical error; or if equally good vanilla calibrations produce materially different target-use outputs without a governance choice.

34. Verification and update triggers

Preserve the raw option surface, arbitrage-cleaning method, local-vol extraction, stochastic-volatility parameters, particle/PDE configuration, kernel, bandwidth, regularisation, leverage interpolation and random seeds.

Revalidate when:

• the implied-volatility surface changes materially;
• the stochastic-volatility backbone is recalibrated;
• tail liquidity changes;
• particle count/time step changes;
• variance-zero behavior becomes material;
• new exotic products are added;
• the calibration algorithm/regularisation changes.

35. Primary and high-quality references

• The Calibration of Stochastic Local-Volatility Models: An Inverse Problem Perspective, Computers & Mathematics with Applications, 2019.
• Andrei Cozma, Matthieu Mariapragassam and Christoph Reisinger, Calibration of a Hybrid Local-Stochastic Volatility Stochastic Rates Model with a Control Variate Particle Method, SIAM Journal on Financial Mathematics.
• Mao Fabrice Djete, Nonregular McKean–Vlasov Equations and Calibration Problem in Local Stochastic Volatility Models, Mathematics of Operations Research, published online 17 March 2026.
• A Reproducing Kernel Hilbert Space Approach to Singular Local Stochastic Volatility McKean–Vlasov Models, Finance and Stochastics, 2024.
• An Adjoint Method for the Exact Calibration of Stochastic Local Volatility Models.

Educational boundary: SLV calibration is powerful because it separates a vanilla marginal target from a stochastic volatility backbone. That flexibility creates a nonlinear inverse problem and does not remove the need to test joint dynamics, tails and numerical stability.