---
title: Stochastic Volterra Integral System
url: https://www.emergentmind.com/topics/stochastic-volterra-integral-system
type: topic
---

# Stochastic Volterra Integral System

Searching arXiv for recent and foundational papers on stochastic Volterra integral systems to ground the article.
A stochastic Volterra integral system is a stochastic dynamical system in which the state at time \(t\) depends on its past through Volterra-type integral terms, typically with coefficients or kernels depending on both the observation time and the integration time. In the literature summarized here, this class includes forward stochastic Volterra integral equations (SVIEs), backward stochastic Volterra integral equations (BSVIEs), forward-backward systems (FBSVIEs), delay and path-dependent variants, fuzzy-valued systems, and formulations under generalized noises such as Lévy noise, compensated Poisson random measures, fractional Brownian motion, cylindrical Wiener noise, and \(G\)-Brownian motion. The common structural feature is memory: unlike standard SDEs, the present state depends on the entire past trajectory through deterministic or stochastic kernels, so the resulting dynamics are typically non-Markovian [1903.05045].

## 1. Defining structure and principal model classes

A basic forward SVIE studied on a separable Hilbert space \(U\) has the form
\[
X(t)=x_0(t)+\int_0^t \mu(t,s,X(s))\,ds+\int_0^t \sigma(t,s,X(s-))\,dL(s), \qquad t\ge 0,
\]
where \(X\) is an adapted càdlàg \(U\)-valued process, \(x_0\) is an \(\mathcal F_0\)-measurable initial process, \(\mu\) is a drift kernel, \(\sigma\) is a diffusion kernel, and \(L\) is a square-integrable Hilbert-space-valued Lévy process [1903.05045]. In scalar or finite-dimensional Brownian settings, representative models take the form
\[
X(t)=X_0+\int_0^t (t-s)^{-\alpha} b(X(s))\,ds+\int_0^t (t-s)^{-\beta}\sigma(X(s))\,dB_s,
\]
with weakly singular kernels, or
\[
x(t)=\varphi(t)+\int_{0}^{t} f(t,s,x(s))\,ds+\int_{0}^{t} g(t,s,x(s))\,dW(s) +\int_{0}^{t}\int_{\mathbb R_0} h(t,s,x(s),\xi)\,\widetilde N(ds,d\xi),
\]
when jumps are included [2312.03474] [2009.06449].

The term “system” is used in several non-equivalent but related senses. It may denote a vector-valued forward equation; a coupled forward-backward configuration such as
\[
\left\{ \begin{array}{ll} X(t)=\varphi(t)+\int_0^t b(t,s,X(s),u(s))\,ds+\int_0^t \sigma(t,s,X(s),u(s))\,dW(s),\\[0.4em] Y(t)=\psi(t,X(T))+\int_t^T g(t,s,X(s),Y(s),Z(t,s),u(s))\,ds -\int_t^T Z(t,s)\,dW(s), \end{array} \right.
\]
or a path-dependent state-control system with memory in both state and control channels [1602.05661] [2312.13516]. It may also denote a delayed equation, where the coefficients depend on \(X(s-\delta)\), or a symmetric implicit system in fuzzy stochastic analysis in which fuzzy stochastic integrals appear on both sides of the equation [2603.25452] [2410.15169].

Backward formulations introduce additional asymmetry. A generic BSVIE on \([0,T]\) is written as
\[
Y(t)=\psi(t)+\int_t^T g\bigl(t,s,Y(s),Z(t,s),Z(s,t)\bigr)\,ds-\int_t^T Z(t,s)\,dW(s),
\]
and the dependence on \(Z(s,t)\) leads to the adapted M-solution concept rather than a purely adapted solution [1208.2064]. Anticipated BSVIEs extend this further by allowing future-point and future-average terms such as \(Y(s+\delta(s))\) and \(\int_s^{s+\delta(s)}e^{\lambda(s-\theta)}Y(\theta)\,d\theta\) in the generator [2501.14263].

This range of models suggests that the phrase “stochastic Volterra integral system” is best understood as a family of non-Markovian stochastic integral systems unified by Volterra memory kernels rather than as a single canonical equation.

## 2. Functional-analytic formulations and well-posedness theory

A central analytic development is the SPDE-lifting method for Hilbert-space-valued SVIEs. The state equation is embedded into a larger function space \(H\) of \(U\)-valued functions on \(\mathbb R_+\), equipped with the right-shift semigroup
\[
(\mathcal S_t h)(x)=h(x+t), \qquad t\ge 0,
\]
whose generator is denoted \(\partial_x\). The lifted coefficients are defined by
\[
a(t,h)=\mu(t+\cdot,t,\delta_0 h), \qquad b(t,h)=\sigma(t+\cdot,t,\delta_0 h),
\]
and the associated first-order SPDE is
\[
dY(t)=\big(\partial_x Y(t)+a(t,Y(t))\big)\,dt+b(t,Y(t))\,dL(t), \qquad Y(0)=x_0.
\]
Under Lipschitz and linear growth conditions on \(a\) and \(b\), the SPDE has a unique adapted càdlàg mild solution, and the original Volterra solution is recovered as the boundary trace \(X(t)=\delta_0 Y(t)\). This yields existence, uniqueness, and finite second moments on finite horizons for the SVIE [1903.05045].

A concrete realization of this lifting is the weighted Filipović-type space \(H_w\), consisting of absolutely continuous \(U\)-valued functions with norm
\[
\|f\|_w^2=\|f(0)\|_U^2+\int_0^\infty w(x)\|f'(x)\|_U^2\,dx<\infty,
\]
where \(w\in C^1(\mathbb R_+)\) is nondecreasing, \(w(0)=1\), and \(w^{-1}\in L^1(\mathbb R_+)\). In this space the shift semigroup is strongly continuous and satisfies
\[
\|\mathcal S_t\|_{\mathrm{op}\le e^{t/2}, \qquad t\ge 0.
\]
The same framework gives directly checkable assumptions on the kernels \(\mu\) and \(\sigma\) [1903.05045].

Well-posedness in other settings is typically obtained by fixed-point or successive approximation arguments. For weakly singular Brownian SVIEs with \(\alpha,\beta\in(0,1/2)\), Lipschitz drift and diffusion imply existence and uniqueness, together with
\[
\sup_{t\in[0,T]}\mathbb E|X(t)|^2\le C_1,\qquad \mathbb E|X(t)-X(s)|^2\le C_2|t-s|^\gamma,
\]
where \(\gamma=\min\{1-2\beta,\;1-\alpha\}\) [2312.03474]. For jump-driven SVIEs under a non-Lipschitz modulus condition governed by a concave \(\kappa\), existence and uniqueness follow from Picard iteration and Bihari’s inequality rather than Grönwall’s inequality [2009.06449]. For equations driven by fractional Brownian motion with Hurst parameter \(H>1/2\), the stochastic integral is interpreted pathwise as a Riemann–Stieltjes integral, and existence and uniqueness are proved in \(W_0^{\alpha,\infty}(0,T;\mathbb R^d)\) via deterministic Volterra estimates and a weighted fixed-point argument [1003.1584].

Path-dependent SVIEs push the dependence from \(X(s)\) to the stopped trajectory \(X_\cdot^s\). In that setting, strong well-posedness is established in an \(L^p\) Bochner-space framework, first for \(p\ge2\) and then locally in time for all \(p>0\), with Hölder regularity estimates and continuity of the flow in the initial condition [2603.20996]. This suggests that the well-posedness theory has evolved from kernel-specific finite-dimensional arguments to a broad operator-theoretic framework capable of handling path dependence, singularity, and nonstandard state spaces.

## 3. Backward, forward-backward, and comparison structures

The backward theory is structurally more delicate than the forward theory. For BSVIEs with generators depending on \(Z(t,s)\) but not \(Z(s,t)\), comparison can be established for adapted solutions under monotonicity in \(y\) and sign conditions on \(\partial_z g\). By contrast, when the generator depends on \(Z(s,t)\), the adapted M-solution framework is required, and the comparison conclusion weakens to an inequality for conditional expectations of time-integrated solutions rather than a pointwise order for \(Y(t)\) [1208.2064].

The infinite-dimensional singular BSVIE theory substantially enlarges the admissible kernel class. The generator satisfies
\[
|g(t,s,y,z_1,z_2)-g(t,s,\bar y,\bar z_1,\bar z_2)|_H \le L_y(t,s)|y-\bar y|_H + L_{z_1}(t,s)\|z_1-\bar z_1\|_{C_2} + L_{z_2}(t,s)\|z_2-\bar z_2\|_{C_2},
\]
with \(L_y,L_{z_1}\in L^2\)-type spaces and \(L_{z_2}\) in a refined local singularity class \(\mathcal L^2(\Delta^*)\). This framework covers fractional kernels, Volterra Heston kernels, and completely monotone kernels, and yields unique adapted M-solutions in natural Hilbert-valued spaces [2312.04094].

Forward-backward stochastic Volterra systems arise naturally in control. For optimal control problems of FBSVIEs with closed, possibly non-convex control regions, the state system couples a forward SVIE with a backward SVIE, and first-order necessary conditions are obtained through a duality principle between linear backward stochastic Volterra equations and linear stochastic Fredholm-Volterra equations with conditional expectation. The resulting maximum principle uses one adjoint system and only first-order differentiability of the coefficients, with the adjacent cone replacing spike variation [1602.05661].

A related but distinct forward-backward structure appears in linear-quadratic control of forward SVIEs. There the optimality system is a coupled FSVIE and Type-II BSVIE, and its decoupling requires a causal auxiliary state, a Type-III BSVIE, and a path-dependent Riccati equation on path space. The resulting open-loop optimal control admits a causal state feedback representation, and when the control appears only in the diffusion term, the feedback reduces to a Markovian state feedback despite the non-Markovian state equation [2204.08694].

These developments show that stochastic Volterra integral systems do not merely extend SDEs by memory terms; they require distinct notions of solution, comparison, duality, and decoupling.

## 4. Optimal control, maximum principles, and games

The control theory of stochastic Volterra systems is dominated by Pontryagin-type principles rather than dynamic programming, because the Volterra memory destroys the usual semigroup flow property. In the linear-quadratic setting for forward SVIEs, the state equation
\[
X(s)=x_t(s)+\int_t^s \big[A(s,r)X(r)+B(s,r)u(r)\big]\,dr +\int_t^s \big[C(s,r)X(r)+D(s,r)u(r)\big]\,dW(r)
\]
is paired with a quadratic cost, and the open-loop optimal control is represented as
\[
\bar u(s)=\mathcal{O}(s)\mathfrak{X}(\cdot,s),
\]
where \(\mathfrak{X}\) is a causal auxiliary state and \(\mathcal O(s)\) is derived from a path-dependent Riccati equation. Under \(Q(s)\ge 0\), \(R(s)\ge \alpha I\), and \(G\ge 0\), the Riccati equation has a unique strongly regular solution [2204.08694].

For generalized Volterra control systems formulated as stochastic integral-differential equations with memory in both state and control,
\[
dX_t = b\Big(t,X_t,\int_0^t k(t,s)X_sds,u_t,\int_0^t l(t,s)u_sds\Big)dt + \sigma\Big(t,X_t,\int_0^t k(t,s)X_sds,u_t,\int_0^t l(t,s)u_sds\Big)dW_t,
\]
the adjoint equation is a generalized anticipated BSDE whose drift contains future conditional expectations of integrals against the kernels \(k\) and \(l\). The stochastic maximum principle then yields the first-order condition
\[
\Big[H_u^*(t)+E^{\mathcal F_t}\Big[\int_t^T l(s,t)H_v^*(s)ds\Big]\Big]\cdot(\alpha_t-u_t^*)\ge 0,
\]
with a corresponding explicit optimal control in the linear-quadratic example [2312.13516].

Delay changes the structure again. For delayed SVIEs,
\[
X^{u}(t) = x_{0}(t) + \int_{0}^{t} b\big(t,s,X^{u}(s-\delta),u(s)\big)\,ds + \int_{0}^{t} \sigma\big(t,s,X^{u}(s-\delta),u(s)\big)\,dB(s),
\]
Hida–Malliavin calculus is used to derive an adjoint anticipated backward stochastic Volterra integral equation (ABSVIE). The representation
\[
q(t,s)=\mathbb E\!\left[D_s p(t)\mid\mathcal F_s\right]
\]
is central in both necessary and sufficient maximum principles. The sufficient principle is expressed through conditional maximization of a Volterra Hamiltonian, and the necessary principle is the conditional stationarity of \(\partial H/\partial u\) [2603.25452].

Game-theoretic extensions use ABSVIEs as adjoint equations for nonzero-sum stochastic differential games of delayed Volterra systems. In that setting, the generator may contain both pointwise time-advanced and average time-advanced terms, and well-posedness, comparison, and Malliavin regularity are established before deriving the Nash equilibrium conditions. In the linear-quadratic SDVIE game, the equilibrium controls take the explicit form
\[
u_1^*(t)= -\big[r_1(t)+\tilde r_1(t+\delta_1)\big]^{-1}Y_{0_1}(t),\qquad u_2^*(t)= -\big[r_2(t)+\tilde r_2(t+\delta_2)\big]^{-1}Y_{0_2}(t)
\]
[2501.14263].

A consistent pattern across these results is that memory terms migrate into the adjoint equation rather than disappearing under augmentation. This suggests that, for stochastic Volterra integral systems, the adjoint object is often itself a Volterra-type backward equation.

## 5. Numerical approximation and computational schemes

Numerical analysis of stochastic Volterra systems is shaped by kernel singularity and non-Markovian dependence. For weakly singular SVIEs with non-differentiable drift, the randomized Milstein scheme avoids Taylor expansion of the drift by introducing i.i.d. auxiliary variables \(\tau_j\sim U(0,1)\) into the drift quadrature. The principal strong error estimate is
\[
\max_{n\in\mathbb S_N}\|X_h^n - X(t_n)\|_{L^2} \le C\, h^{\min\{1-2\beta,\;1-\alpha\},
\]
and the numerical experiments recover the predicted slopes \(0.7\) for \((\alpha,\beta)=(0.3,0.1)\) and \(0.4\) for \((\alpha,\beta)=(0.2,0.3)\) [2312.03474]. The same paper addresses the practical simulation of multiple singular stochastic integrals via a Riemann–Stieltjes discretization of nested integrals.

For path-dependent SVIEs, an interpolated \(K\)-integrated Euler–Maruyama scheme is designed to respect the Volterra structure by integrating the kernels over each mesh cell rather than using point evaluations. Under Hölder-Lipschitz assumptions on the coefficients and regularity assumptions on the kernels, the fixed-time strong error satisfies
\[
\sup_{t\in[0,T]} \|X_t-\bar X_t^h\|_{L^p} \le C\left(1+\|\,\|x_0\|_T\,\|_p\right) \left(h^\gamma + h^{\delta\wedge\theta\wedge\widehat\theta}\right),
\]
and the uniform-in-time error is controlled by
\[
\left\|\sup_{t\in[0,T]}\|X_t-\bar X_t^h\|_{\mathbb H}\right\|_p \le C_\varepsilon \left(1+\|\,\|x_0\|_T\,\|_p\right) h^{(\delta\wedge\theta\wedge\widehat\theta\wedge\gamma)(1-\varepsilon)}
\]
[2603.20996].

Other numerical frameworks exploit the specific Volterra structure differently. The Walsh-function method approximates the linear SVIE
\[
x(t)=f(t)+\int_{0}^{t}k_1(s,t)x(s)\,ds+\int_{0}^{t}k_2(s,t)x(s)\,dB(s)
\]
by expanding the solution and kernels in a Walsh basis and converting the equation into a finite algebraic system
\[
X-F-m\hat H_1-m\hat H_2=0.
\]
Under Lipschitz assumptions on \(f\), \(k_1\), and \(k_2\), the basis approximation error is \(O(h)\) and the full solution error satisfies
\[
\|x(t)-x_m(t)\|_2^2=O(h^2)
\]
[2305.00823].

A more global high-order method is the cubature method for Stratonovich SVIEs. Because the solution is generally not a semimartingale, the method first constructs a stochastic Taylor expansion via a functional Itô formula on an auxiliary two-time process \(\Theta_t^{i,s}\), then matches the relevant Volterra signatures by a discrete cubature measure. In the multi-block case the global approximation error is of order
\[
|E^Q[G(X_T)]-E[G(X_T)]| \le C_N\,\frac{T^{\frac{N+1}{2}{M^{\frac{N-1}{2},
\]
up to explicit coefficient-dependent factors [2110.12853].

These schemes illustrate two recurring numerical themes: memory must be discretized rather than ignored, and kernel singularity usually determines the convergence rate.

## 6. Generalized noises, values, and integration frameworks

The concept of a stochastic Volterra integral system extends beyond standard Itô equations. In white-noise analysis, Brownian-driven Volterra processes with volatility modulation are treated on the Potthoff–Timpel distribution space \(G^*\). For
\[
X_\sigma(t)=\int_0^t g(t,s)\sigma(s)\,dB(s),
\]
integration with respect to \(X_\sigma\) is defined by
\[
\int_0^t \Phi(s)\,dX_\sigma(s) = \int_0^t K_g(\Phi)(t,s)\sigma(s)\,\delta B(s) + \int_0^t D_s\left\{K_g(\Phi)(t,s)\right\}\sigma(s)\,ds,
\]
where
\[
K_g(\Phi)(t,s) = \Phi(s)g(t,s) + \int_s^t \left(\Phi(u)-\Phi(s)\right)\,g(du,s).
\]
A generalized volatility modulation via Wick product is also introduced, and the Wick formulation coincides with the pointwise product under strong independence [1303.4625].

An infinite-dimensional analogue studies Hilbert-valued volatility-modulated Volterra processes driven by a cylindrical Wiener process. The stochastic integral
\[
\int_0^t  Y(s)\,d X(s)
\]
is defined as a Skorohod integral plus a correction term involving the Malliavin derivative of \(K_g(Y)(t,s)\), and the resulting calculus includes an Itô formula and a random-field formulation relevant to SPDE kernels [1303.7143].

Other generalizations change the state space rather than the noise. Symmetric fuzzy stochastic Volterra integral equations with constant retardation consider fuzzy-valued processes \(U:[-\tau,T]\times\Omega\to F\), with fuzzy stochastic Lebesgue–Aumann integrals and Hukuhara differences on both sides of the equation. Under hypotheses \((H0)\)–\((H3)\), the initial value problem has a unique solution and depends continuously on the initial fuzzy path, kernels, and nonlinearities [2410.15169].

A different extension replaces classical probability by sublinear expectation. For \(G\)-SVIEs of the form
\[
X(t)=\varphi(t)+\int_0^t b(t,s,X(s))\,ds+\int_0^t h(t,s,X(s))\,d\langle B\rangle_s+\int_0^t \sigma(t,s,X(s))\,dB_s,
\]
existence, uniqueness, mean-square continuity, and pathwise continuity of a modification are established in \(M_G^2\)-type spaces. A comparison theorem is proved for the separable diffusion form \(H(t)\sigma(s,x)\), and the method avoids classical assumptions on partial derivatives of the coefficients by combining quasilinearization with a two-step approximation [2504.21293].

These examples show that the Volterra structure is compatible with generalized state spaces, generalized stochastic integration, and uncertainty in the driving law.

## 7. Long-time behavior, stationarity, and asymptotics

Long-time analysis for stochastic Volterra systems differs sharply from the Markovian case. In the SPDE-lifting framework for homogeneous Hilbert-space-valued SVIEs, limiting distributions are obtained by importing invariant-measure criteria from the lifted SPDE. In the weighted space \(H_w\), if the coefficients vanish at infinity and the dissipativity condition
\[
L_b^2+2L_a<\alpha_w
\]
holds on the vanishing-at-infinity subspace \(H_w^0\), then \(P^{X(t)}\) converges weakly to a limit \(\Gamma\) on \(U\) [1903.05045]. A second theorem handles the persistent constant component through a one-sided monotonicity condition on the drift at infinity [1903.05045].

A more recent stationarity theory emphasizes a negative conclusion. For forward SVIEs with convolution kernel \(K(t-s)\), strong stationarity is essentially impossible unless the kernel is constant or the system is degenerate. In the affine mean-reverting case
\[
X_t = X_0\phi(t) +\int_0^t K(t-s)\big(\mu(s)-\lambda X_s\big)\,ds +\int_0^t K(t-s)\sigma(s,X_s)\,dW_s,
\]
the paper shows that one can nevertheless induce a “fake stationary regime” by choosing a deterministic stabilizer \(\varsigma\) in \(\sigma(t,x)=\varsigma(t)\sigma(x)\). The stabilizer solves
\[
c\lambda^2\Big(1-(\phi-f_\lambda * \phi)^2(t)\Big) = (f_\lambda^2 * \varsigma^2)(t),\qquad t\ge0,
\]
and forces constant first two moments and constant expected squared diffusion in type-I fake stationarity [2511.03474].

The same work studies \(L^p\)-confluence across initial values and long-run functional weak limits of the shifted process \(X_{t+\cdot}\). Under suitable assumptions, the family is \(C\)-tight in \(C(\mathbb R_+,\mathbb R)\), and any limit process is weakly \(L^2\)-stationary when the original process is in fake stationary regime of type I. The framework applies to fractional kernels
\[
K_\alpha(t)=\frac{t^{\alpha-1}{\Gamma(\alpha)}
\]
and exponential-fractional kernels
\[
K_{\alpha,\rho}(t)=e^{-\rho t}\frac{t^{\alpha-1}{\Gamma(\alpha)},
\]
covering both rough behavior for \(\alpha<1\) and long-memory or persistence for \(\alpha>1\) [2511.03474].

This suggests a useful correction to a common intuition: memory does not automatically produce stationary long-run dynamics. In the Volterra setting, stationarity usually has to be reformulated, weakened, or induced by additional deterministic structure rather than inherited from the kernel alone.

Source: https://www.emergentmind.com/topics/stochastic-volterra-integral-system