Volterra Adjustable Decoupling Approximation

• Volterra Adjustable Decoupling Approximation is a method that approximates completely monotone and singular Volterra kernels using a finite exponential mixture for Markovian reformulation.
• It employs a multifactor Euler discretization combined with advanced quadrature rules to balance kernel approximation error and time discretization error with explicit L²-convergence rates.
• The approach reduces computational cost in solving rough stochastic Volterra equations, offering significant speed-ups over direct Euler methods in models like rough volatility.

Volterra Adjustable Decoupling Approximation denotes an adjustable approximation strategy for the $d$-dimensional Stochastic Volterra Equation

$$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$

when the kernel $K$ is completely monotone and may be singular. In the formulation developed by A. Alfonsi and A. Kebaier, the method combines a multifactor approximation of the kernel, an equivalent finite-dimensional SDE representation, and a multifactor Euler discretization whose approximation level can be tuned against the time-discretization level. The resulting procedure provides an $L^2$-quantification of kernel-approximation error, explicit convergence rates for rough kernels, and an asymptotic reduction of computational cost relative to a direct Euler scheme for SVEs (Alfonsi et al., 2021).

1. Governing equation and admissible kernels

The underlying model is a $d$-dimensional Stochastic Volterra Equation with drift $b:\mathbb R^d\to\mathbb R^d$, diffusion coefficient $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$, and driving Brownian motion $W$. The kernel is assumed to satisfy

$$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$

where $\lambda$ is a nonnegative measure on $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$0 such that, for each $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$1,

$$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$2

This hypothesis allows singular kernels while preserving an integral representation by exponentials (Alfonsi et al., 2021).

A typical singular example is the rough kernel

$$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$3

which corresponds to

$$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$4

The coefficients $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$5 and $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$6 are assumed to be globally Lipschitz: $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$7 Under these assumptions there is a unique strong solution $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$8 (Alfonsi et al., 2021).

The completely monotone representation is structurally central. It permits the Volterra memory term to be approximated by finite sums of exponentials, which in turn yields a finite-dimensional Markovian representation. This suggests that the “decoupling” in VADA is not a removal of memory from the original model, but an approximation of memory by a finite collection of factors whose evolution is governed by classical SDE dynamics.

2. Multifactor approximation by Prony series

The standard multifactor approximation is constructed by truncating the measure $$X_t = X_0 + \int_0^t K(t-s)\,b(X_s)\,ds + \int_0^t K(t-s)\,\sigma(X_s)\,dW_s,$$9 at a finite level and discretizing the truncated support. One considers a partition

$K$0

The approximating kernel is then defined by

$K$1

The data also allows more sophisticated quadrature rules, including Simpson, Newton–Cotes, and geometrically graded grids, to choose the pairs $K$2 so as to optimize $K$3 (Alfonsi et al., 2021).

By Proposition 2.1 in Alfonsi–Kebaier, the SVE with kernel $K$4 can be rewritten as a classical SDE in $K$5. If

$K$6

then the factor processes $K$7 solve

$K$8

with $K$9 (Alfonsi et al., 2021).

This representation is the operational core of the method. The original nonlocal kernel is replaced by finitely many exponentially weighted modes, and the approximate Volterra system becomes an $L^2$0-dimensional SDE. A plausible implication is that the accuracy of the approximation is governed primarily by how well the exponential mixture reproduces the $L^2$1 behavior of the target kernel.

3. $L^2$2-error control and rates for rough kernels

The basic non-asymptotic estimate is Theorem 2.7 in Alfonsi–Kebaier. If $L^2$3 and $L^2$4 satisfy

$L^2$5

then there is $L^2$6 such that

$L^2$7

The constant depends on $L^2$8, $L^2$9, $d$0, and related parameters, and the estimate directly quantifies the propagation of kernel-approximation error to the SVE solution (Alfonsi et al., 2021).

For the singular rough kernel $d$1, the paper states a concrete rate under a Riemann-sum construction. With

$d$2

one obtains

$d$3

and consequently

$d$4

The same source states that more elaborate quadratures, such as Simpson, Newton–Cotes, and geometric meshes, drive the exponent from $d$5 arbitrarily close to $d$6, hence $d$7 (Alfonsi et al., 2021).

These bounds clarify the role of multifactor design. A coarse exponential mixture already yields polynomial convergence, while improved quadrature rules sharpen the kernel approximation and therefore the strong approximation of the SVE itself. In the rough-kernel regime, this is particularly significant because the singularity at the origin is precisely the feature that makes direct numerical treatment expensive.

4. Multifactor Euler discretization

After the finite-dimensional SDE reformulation, the approximation is discretized on a uniform time grid

$d$8

Let $d$9 denote the factor processes in the $b:\mathbb R^d\to\mathbb R^d$0-dimensional SDE representation. The multifactor Euler update is

$b:\mathbb R^d\to\mathbb R^d$1

where

$b:\mathbb R^d\to\mathbb R^d$2

This is the discrete scheme used to approximate the original Volterra solution via the factorized system (Alfonsi et al., 2021).

Theorem 4.1 states that, under $b:\mathbb R^d\to\mathbb R^d$3 and $b:\mathbb R^d\to\mathbb R^d$4, there is a constant $b:\mathbb R^d\to\mathbb R^d$5 independent of $b:\mathbb R^d\to\mathbb R^d$6 such that

$b:\mathbb R^d\to\mathbb R^d$7

where $b:\mathbb R^d\to\mathbb R^d$8 is the Hölder exponent of the resolvent of $b:\mathbb R^d\to\mathbb R^d$9. For the rough kernel, $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$0 (Alfonsi et al., 2021).

Taken together with the $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$1-error estimate, this yields

$\sigma:\mathbb R^d\to\mathbb R^{d\times d}$2

The notable feature is uniformity with respect to the number of factors. This means that refinement of the multifactor approximation does not degrade the discretization estimate through an $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$3-dependent constant, which is essential for balancing the two sources of error.

5. Adjustable balancing of approximation and discretization

The adjustable aspect of VADA appears in the joint selection of the number of factors and the number of time steps. The data summarizes the total mean-square error in the form

$\sigma:\mathbb R^d\to\mathbb R^{d\times d}$4

for some $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$5, with $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$6 in the simplest Riemann case. Balancing both contributions leads to

$\sigma:\mathbb R^d\to\mathbb R^{d\times d}$7

and therefore to a computational cost

$\sigma:\mathbb R^d\to\mathbb R^{d\times d}$8

This is the adjustment rule that tunes approximation versus discretization in the multifactor Euler construction (Alfonsi et al., 2021).

The same source states that, since $\sigma:\mathbb R^d\to\mathbb R^{d\times d}$9 for any $W$0, the procedure yields an asymptotic speed-up over $W$1: $W$2 It also states that finer quadrature can make $W$3, in which case $W$4 and the cost is approximately $W$5, so there is no gain; for moderate $W$6, one may instead choose mid-range rules such that $W$7, leading to cost $W$8 instead of $W$9 (Alfonsi et al., 2021).

In this sense, VADA is “adjustable” because it is not tied to a single kernel quadrature or a single scaling law. The method permits a deliberate compromise between a more accurate kernel approximation and a lower-dimensional factor system. The paper’s formulation explicitly describes the practical rule as selecting $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$0 by equating the two contributions in the error bound.

6. Relation to rough volatility and numerical practice

The abstract emphasizes two rough-volatility applications. For the rough Bergomi model, the paper proposes various approximating multifactor kernels, states their rates of convergence, and illustrates their efficiency. For the rough Heston model, it states that the new multifactor Euler scheme outperforms the Euler scheme for SVEs for option pricing (Alfonsi et al., 2021).

The direct benchmark is the Volterra-Euler scheme on the original SVE, referred to in the data as the Richard–Talay–Yang scheme. That direct method requires, at each time step, summing $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$1 past kernels and therefore has cost $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$2. By contrast, the multifactor Euler scheme has cost $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$3. The performance difference is therefore not merely an implementation detail; it follows from replacing a history-dependent recursion by a factor recursion whose state dimension is controlled by the chosen multifactor approximation (Alfonsi et al., 2021).

A common misconception is that improved quadrature always implies improved overall efficiency. The data does not support that conclusion in general. It states that quadratures that push $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$4 close to $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$5 can remove the asymptotic gain by forcing $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$6, while intermediate choices may preserve a computational advantage. This suggests that, within VADA, kernel-approximation quality and computational complexity must be evaluated jointly rather than separately.

The broader significance of the construction lies in its synthesis of three ingredients: completely monotone kernel representation, multifactor approximation by exponentials, and a strong-error analysis uniform in the number of factors. Within that framework, VADA is best understood as a structured approximation principle for singular-kernel SVEs, especially in the rough-kernel setting $$K:(0,T]\to\mathbb R \text{ is completely monotone, i.e. } K(t)=\int_0^\infty e^{-\rho t}\,\lambda(d\rho),$$7, where memory effects are strong and direct discretizations are costly (Alfonsi et al., 2021).