Forward Stochastic Integral

• Forward stochastic integral is defined via pathwise regularization using forward finite differences, extending classical Itô integration.
• It accommodates non-adapted, anticipating processes and varied Gaussian drivers, including cylindrical and fractional Brownian motions.
• This integral offers integration by parts formulas and exact L²-isometries, facilitating applications in SPDEs, finance, and infinite-dimensional analysis.

A forward stochastic integral is a generalization of classical stochastic integration that is defined via pathwise regularization by forward finite differences, rather than through semimartingale or martingale techniques. The construction is robust, extending Itô integration to include non-adapted, anticipating processes and a wide range of Gaussian drivers including cylindrical Brownian processes and fractional Brownian motion (fBM) with Hurst parameter $H > 1/2$. Forward stochastic integrals preserve key properties of Itô calculus on adapted processes, offer powerful new integration by parts formulas for non-adapted multipliers, and provide exact $L^2$-isometries (sometimes with additional trace corrections) under minimal integrability or smoothness assumptions. Applications encompass SPDEs with non-adapted coefficients, insider trading in mathematical finance, and pathwise analysis of stochastic systems.

1. Definition and Construction

The archetypal forward integral with respect to a real-valued continuous process $X$ is defined by regularizing via forward increments: $$\int_0^t Y_s\,d^-X_s := \lim_{\epsilon \to 0^+}\int_0^t Y_s\,\frac{X_{s+\epsilon} - X_s}{\epsilon}\,ds,$$ provided the limit exists in a suitable sense (u.c.p., in probability, or $L^2$). In the classical Itô setting, for predictable $Y$ and $X$ a semimartingale, this limit recovers the Itô integral. More generally, the forward integral admits extensions to non-adapted (anticipating) integrands and processes $X$ that are not semimartingales, such as fBM with $H > 1/2$ and infinite-dimensional drivers.

A canonical infinite-dimensional setting involves a Hilbert space $H$, a UMD Banach space $X$, and a cylindrical Brownian motion $W$ on $L^2(0,T; H)$. For $H$-strongly measurable, weakly $L^2$-integrable $G \colon [0,T] \times \Omega \rightarrow L(H, X)$, the regularized forward integral approximates

$$I^-(G, n) = \sum_{k=1}^n \int_0^T G(s) h_k\, [W(s + \tfrac{1}{n})h_k - W(s)h_k]\,ds,$$

where $(h_k)$ is an orthonormal basis of $H$. $G$ is forward integrable if $I^-(G,n)$ converges in probability as $n \rightarrow \infty$, with the limit denoted $\int_0^T G(s)\,d^-W(s)$. This construction extends directly to V-valued (e.g., cylindrical) Wiener noises (Pronk et al., 2013, Olivera, 2012).

2. Regularization, Convergence, and Sobolev Norms

The forward integral is based on Riemann-type approximations. The convergence of the forward Riemann sums to a limiting integral can be analyzed in fractional Sobolev–Slobodeckij spaces $W^{\alpha,p}(0,T; X)$, where fractional regularity is essential—especially in infinite dimensions or for rough drivers. Given $G$ in $L^p(\Omega; V^{\beta,p}(0,T; H, X))$ for $0 < \alpha < \beta < 1/2$, the pathwise sequence of approximants converges in $L^p(\Omega; W^{\alpha,p}(0,T; X))$ to the Itô integral when $G$ is adapted, and thus the forward integral recovers Itô for adapted, square-integrable processes (Pronk et al., 2013).

The analysis of convergence uses γ–radonifying norms, embedding results for $W^{\alpha,p}$, and multiplier theorems for convolution operators over operator-valued functions—demonstrating the fine regularity of the integral process and precisely quantifying pathwise Hölder continuity of order $< 1/2$.

3. Forward Integrals for Fractional Brownian Motion

For fractional Brownian motion (fBM) of Hurst index $H > 1/2$, the forward integral admits a clean construction and exact $L^2$-isometry formulas. For a $d$-dimensional fBM $B=(B^{(1)},...,B^{(d)})$ with covariance $R(s, t)$, the forward integral for (possibly non-adapted) integrands $Y_t = g(t, B_t)$ is defined as

$$\int_0^T g(t, B_t)\,d^-B_t = \lim_{\epsilon \to 0} \frac{1}{\epsilon}\int_0^T \langle g(s, B_s), B_{s+\epsilon} - B_s \rangle\,ds,$$

in $L^2(\mathbb{P})$. The $L^2$-norm of the forward integral is given by

$$\| \int_0^T Y_t\,d^-B_t \|_{L^2}^2 = \mathbb{E}\int_{[0,T]^2}\langle Y_s \otimes Y_t, \Lambda(s, t) \rangle\,ds\,dt,$$

where $$\Lambda(s, t) = \frac{\partial^2}{\partial s \partial t} R(s, t) I_d + \mathcal{W}(s, t)$$, and $\mathcal{W}(s, t)$ is a quadratic field in $(B_s, B_t)$ encoding the "anticipating" correction. The RKHS inner-product term is recovered for the adapted case, and for adapted or Hölder regular integrands, the forward integral coincides with the Young integral (Ohashi et al., 2023).

For the Riemann-Liouville fBM, the forward integral for square-integrable adapted $F$ admits a canonical martingale representation involving the Nelson stochastic derivative and the Volterra kernel: $$\int_0^T F_t\,d^-B^H_t = \int_0^T\!\!\int_0^t \mathbb{E}[\varphi^{(1)}_F(t,s)]\,\frac{\partial K}{\partial t}(t,s)\,ds\,dt + \int_0^T \mathcal{K}F(T,r)\,dW_r,$$ together with an exact $L^2$-isometry (Costa et al., 14 Dec 2025).

4. Anticipating Integrands and Financial Applications

A defining virtue of the forward integral is its natural treatment of anticipating integrands, for which classical Itô or Skorokhod–Malliavin approaches can yield economically or mathematically inadequate models. For instance, in the context of financial markets with insider trading, the forward (Russo–Vallois) integral provides SDE solutions exhibiting desirable sample-path continuity and higher expected wealth for the insider compared to adapted or Skorokhod/Ayed–Kuo formulations. The explicit SDE for an insider’s wealth $S_1$ driven by $d^-B_t$ has a closed-form solution, and the resulting expected wealth strictly exceeds that of Itô and other anticipating calculi (Bastons et al., 2018).

Integral	Pathwise continuity	Expected insider wealth vs. Itô	Model consistency
Forward (RV)	Yes	Strictly greater	Yes
Skorokhod	No (discrete jumps)	Possibly smaller or negative	No
Ayed–Kuo	No (discrete jumps)	Smaller or negative	No

The economic interpretation is that only the forward integral models the anticipated profit of a bona fide insider in a way compatible with financial theory. This is a key differentiator in stochastic control and mathematical finance.

5. Infinite-Dimensional Extensions and SPDEs

The pathwise forward regularization approach seamlessly extends to infinite-dimensional drivers such as cylindrical Wiener processes, yielding a rigorous integral for operator-valued or Banach-space valued integrands. The construction proceeds by expanding the integrand in an orthonormal basis and summing the forward integrals of its coordinates, subject to convergence in probability.

This method supports the solution theory of SPDEs with non-adapted or anticipating coefficients. Notably, forward integration is crucial in the construction of mild solutions to stochastic PDEs where coefficients, initial data, or propagators are non-adapted or only $\mathcal{F}_t$-measurable. The integration by parts formula for non-adapted operator multipliers allows the solution of non-autonomous evolution equations for which Itô integration is ill-posed (Olivera, 2012, Pronk et al., 2013).

6. Integration by Parts and Multiplier Results

A central technical advance is the forward integration by parts formula for operator-valued non-adapted multipliers $M(t)$. For $M$, not necessarily adapted but $C^1$ with controlled singularities, and adapted forward integrable $G$,

$$\int_0^T M(s)G(s)\,d^-W(s) = M(0)\int_0^T G(s)\,dW(s) + \int_0^T M'(s)\left(\int_s^T G(r)\,dW(r)\right)ds,$$

with almost sure equality and in $X$-valued settings. This non-adapted multiplier result greatly extends the reach of stochastic analysis in infinite dimensions, facilitating analysis in non-Markovian or time-inhomogeneous frameworks (Pronk et al., 2013).

7. Relation to Existing Integration Theories and Limitations

The forward integral recovers the Itô integral in the adapted $L^2$-setting, the pathwise Stieltjes integral for bounded variation $X$, and the Young integral for sufficiently regular sample paths. It contrasts with the Skorokhod (divergence) approach in fBM integration, which imposes more stringent regularity and has different correction terms.

Applicability to fBM when $H < 1/2$ remains subtle; the existing framework requires $H > 1/2$ for local integrability of the second derivative of the covariance kernel, though extensions via renormalization have been investigated.

The forward stochastic integral unifies, extends, and pathwise regularizes stochastic integration for a broad range of driving processes and allows for the analysis and solution of stochastic equations with anticipating structure, both in finite and infinite dimensions (Pronk et al., 2013, Olivera, 2012, Costa et al., 14 Dec 2025, Ohashi et al., 2023, Bastons et al., 2018).