Stochastic Calculus for Pathwise Observables

• Stochastic calculus for pathwise observables is a framework that defines integration, differentiation, and change-of-variable rules directly on entire trajectories.
• Rough path calculus lifts stochastic signals to enhanced spaces, enabling deterministic construction of integrals and pathwise solutions of differential equations.
• Functional Itô calculus and Föllmer integration offer alternative approaches to handle non-anticipative functionals, local times, and jump processes with practical computational applications.

Stochastic calculus for path-wise observables studies functionals of entire trajectories rather than only state variables at fixed times. Its central problem is to define integration, differentiation, and change-of-variable rules directly on path space, or on suitable enhancements of paths, so that quantities such as stochastic area, signatures, first-passage times, running extrema, thermodynamic currents, and solution maps of differential equations become analytically tractable. In rough path theory, a stochastic signal is lifted to a rough path and the Itô–Lyons map turns integrals, signatures, Jacobians, and differential-equation solutions into continuous functionals of the lift; in functional Itô calculus, non-anticipative functionals on stopped paths are differentiated horizontally and vertically; in recent treatments of nonequilibrium systems, path-wise observables such as currents, densities, and transition-path times are taken as primary objects (Inahama, 2016, Cont et al., 2010, Stutzer et al., 6 Aug 2025, Dieball et al., 2024).

1. Path-space observables and non-anticipative structure

A recurrent formalization is the non-anticipative functional. On $D([0,T],\mathbb{R}^d)$, a map

$F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$

is non-anticipative when $F(t,x)=F(t,x_{\wedge t})$, so its value at time $t$ depends only on the past path up to $t$ (Cont et al., 2010). This is the basic object in functional Itô calculus. In rough path theory the same idea appears after enhancement: once a process $w$ is lifted to a random rough path $W$, every pathwise construction that factors through $W$ becomes a functional on rough-path space, and continuity of the Itô–Lyons map implies continuity of the resulting observable (Inahama, 2016).

The class of observables is broad. For rough paths it includes Lévy area, signatures, Jacobians, and solutions of rough differential equations. In the Cont–Fournié framework it includes functionals of stopped càdlàg paths such as running maxima and other path-dependent transforms (Cont et al., 2010, Cont et al., 2010). For Markov-jump dynamics, time-integrated observables split into currents

$$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$

and densities

$\rho_t=\int_0^t\sum_i V_i(s)\,d\tau_i(s),$

providing a direct discrete-state analogue of diffusion functionals (Stutzer et al., 6 Aug 2025). In driven-colloid experiments, local velocities, one-sided and two-sided first-passage times, and transition-path times are treated explicitly as observables of individual paths rather than of ensemble averages such as the mean-squared displacement (Dieball et al., 2024).

This path-space viewpoint changes what is meant by “state information.” A path-wise observable may depend on the whole trajectory up to time $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$0, yet remain continuous, differentiable, or representable once the appropriate pathwise structure—quadratic variation, rough enhancement, local time, or jump counting noise—is specified. A plausible implication is that stochastic calculus for path-wise observables is less a single formalism than a family of compatible calculi indexed by the regularity and algebraic structure of the underlying trajectories.

2. Rough-path calculus and deterministic reconstruction of stochastic integrals

For $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$1, a level-2 geometric $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$2-rough path over $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$3 is a pair $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$4 on $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$5 satisfying Chen’s relations

$F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$6

together with finite $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$7-variation for $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$8 and finite $F:[0,T]\times D([0,T],\mathbb{R}^d)\to\mathbb{R}$9-variation for $F(t,x)=F(t,x_{\wedge t})$0, and approximability by smooth lifts (Inahama, 2016). The Sewing Lemma then turns an almost-additive two-parameter increment into a genuine integral. This yields a deterministic construction of rough integration and, for sufficiently smooth vector fields, a pathwise solution theory for rough differential equations

$F(t,x)=F(t,x_{\wedge t})$1

Lyons’ universal limit theorem gives existence, uniqueness, and local Lipschitz continuity of the solution map in rough-path metric (Inahama, 2016).

This framework recovers classical stochastic integrals by choosing appropriate lifts. For Brownian motion, dyadic piecewise-linear lifts converge almost surely in $F(t,x)=F(t,x_{\wedge t})$2-variation to the enhanced Brownian rough path. Feeding this lift into the rough differential-equation machinery yields the Stratonovich solution. The Itô solution is recovered by modifying the second level according to

$F(t,x)=F(t,x_{\wedge t})$3

that is, by subtracting the bracket (Inahama, 2016). The probabilistic distinction between Itô and Stratonovich is thus encoded by the second-order enhancement rather than by an external integration convention.

Keller and Zhang formulate an Itô-style rough-path calculus for controlled paths driven by possibly non-geometric rough paths. For an $F(t,x)=F(t,x_{\wedge t})$4-Hölder rough path $F(t,x)=F(t,x_{\wedge t})$5, the bracket

$F(t,x)=F(t,x_{\wedge t})$6

plays the rôle of quadratic variation, and controlled paths admit first- and second-order path derivatives. Their pathwise Itô–Ventzell formula and rough-PDE framework extend this calculus to coefficients depending on the rough path itself, which corresponds to stochastic PDEs with random coefficients (Keller et al., 2014).

A more singular regime appears for fractional Brownian motion of low Hurst index. Magnen and Unterberger describe how, for $F(t,x)=F(t,x_{\wedge t})$7, a weak non-Gaussian perturbation of the Gaussian measure can desingularize iterated integrals, producing a rough-path lift and hence a “pathwise Stratonovich-like” stochastic calculus for a process that is not a semimartingale when $F(t,x)=F(t,x_{\wedge t})$8 (Magnen et al., 2010). This suggests that rough-path enhancement is not merely a regularity device; it can also serve as a renormalized carrier of second-order information when classical iterated integrals diverge.

3. Functional Itô calculus and Föllmer-type pathwise integration

Functional Itô calculus begins with two directional derivatives on stopped paths. The vertical derivative perturbs the path by a jump at the present time, while the horizontal derivative extends the path flat over a short future interval. Under continuity and regularity assumptions, $F(t,x)=F(t,x_{\wedge t})$9 admits a pathwise change-of-variable formula along any càdlàg path $t$0 of finite quadratic variation: $t$1 The stochastic integral here is the Föllmer integral, defined as a limit of non-anticipative Riemann sums (Cont et al., 2010).

For continuous semimartingales, Cont and Fournié extend this structure to square-integrable martingales. Their vertical-gradient operator $t$2 is closable and extends to an isometry from the martingale space $t$3 to $t$4, becoming the inverse of the Itô integral. The resulting representation

$t$5

is a constructive martingale representation, and in the Brownian case $t$6 matches the conditional Malliavin derivative: $t$7 for $t$8 (Cont et al., 2010).

Föllmer’s calculus for càdlàg paths with quadratic variation along a fixed sequence of partitions provides a complementary pathwise route. The integral

$t$9

is defined by convergence of non-anticipative Riemann sums; admissible integrands include $t$0 for $t$1. The pathwise Itô formula, integration by parts,

$t$2

and associativity of iterated Föllmer integrals all hold under the stated hypotheses (Hirai, 2017). This framework also solves explicit integral equations, including the homogeneous linear equation

$t$3

whose unique solution is the Doléans exponential $t$4, and related inhomogeneous or path-dependent equations, including Azéma–Yor type constructions and drawdown-constrained equations (Hirai, 2017).

A common misconception is that “pathwise” automatically means “probability-free but otherwise unique.” The Föllmer framework shows that uniqueness can depend on the partition sequence, whereas the rough-path framework relocates the ambiguity into the choice of enhancement. The two viewpoints are compatible in spirit but not identical in how second-order information is specified.

4. Local times, pathwise differentiation, and singular observables

For continuous paths, Davis, Obłój, and Siorpaes define pathwise local time as the limit of discrete local times along partitions whose oscillation on compacts vanishes. This yields a pathwise Itô formula for $t$5 functions and a pathwise Itô–Tanaka formula for $t$6 functions or differences of convex functions: $t$7 They also prove change-of-variables and change-of-time results for quadratic variation and local time (Davis et al., 2015).

Their analysis makes the partition issue explicit. Existence of pathwise local time is equivalent, for $t$8, to boundedness and weak convergence of the discrete local times in $t$9, and to a dual occupation-time formula. At the same time, the paper shows pathological dependence on partitions: for sufficiently oscillatory paths, any prescribed increasing process can be realized as pathwise quadratic variation along a suitable refining partition. For semimartingales, however, this degeneracy disappears when partitions are optional, and the limits agree with classical quadratic variation and local time (Davis et al., 2015).

A different notion of differentiation is developed by Allouba. For a continuous semimartingale $w$0 and Brownian motion $w$1, the pathwise stochastic derivative is

$w$2

whenever the derivative exists. If

$w$3

then $w$4 almost surely, giving a fundamental theorem of stochastic calculus. The same derivative satisfies a chain rule, mean-value and Rolle-type theorems, and algebraic product and quotient rules (Allouba, 2010). This is a differentiation theory dual to Itô integration, with quadratic covariation replacing ordinary time differentiation.

Houdré and Vázquez enlarge the functional calculus further by introducing derivatives along a curve $w$5. The derivative

$w$6

recovers Dupire’s horizontal derivative when $w$7 and vertical derivatives when $w$8. Under suitable regularity, they derive a coinvariant Itô formula for continuous paths of bounded quadratic variation along partitions (Houdré et al., 12 Apr 2026). Their max-functional example shows that singular observables may fail to be Dupire differentiable while remaining accessible through carefully chosen directional curves. This suggests that pathwise observables with geometric singularities often require directional, rather than purely coordinatewise, differentiation.

5. Jump processes, quasi-sure constructions, and diffusion–jump parallelism

A pathwise stochastic integral can also be constructed simultaneously under a non-dominated family of semimartingale laws. Nutz begins with an $w$9-predictable integrand $W$0 and a càdlàg integrator $W$1 that is a semimartingale under each $W$2. Averaging $W$3 against a dominating predictable increasing process $W$4 produces finite-variation approximations $W$5, so that the Lebesgue–Stieltjes integrals

$W$6

are defined pathwise. A medial limit then aggregates the measurewise limits into a single càdlàg process $W$7 that coincides $W$8-almost surely with the usual Itô integral for every $W$9 (Nutz, 2011). Bartl, Kupper, and Neufeld pursue a different quasi-sure route on continuous Hilbert-space-valued paths: an outer measure defined through pathwise superhedging in $W$0 and $W$1 yields model-free Itô integrals, weak BDG-type bounds, and well-posed SDEs on “typical paths” (Bartl et al., 2018).

For pure-jump dynamics, Foxall considers hybrid jump processes driven by an adapted point process of locally finite intensity. If $W$2 is the jump rate, $W$3 the continuous drift, and $W$4 the jump kernel, then

$W$5

The predictable drift and diffusivity are

$W$6

and the compensated process is a local martingale with predictable quadratic variation $W$7 (Foxall, 2016). The paper derives exponential martingales, sample-path estimates, ODE approximation, and first-passage bounds within this jump calculus.

A full stochastic calculus for pathwise observables of Markov-jump processes is developed in exact parallel with overdamped diffusion. For transition rates $W$8, the jump counts satisfy the Langevin-type equation

$W$9

with centered noise increments $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$0. This leads to explicit covariation formulas for current- and density-type observables, as well as thermodynamic uncertainty relations, correlation bounds, transport bounds, and linear-response formulas. The same framework admits a continuum limit under lattice refinement, recovering the Fokker–Planck and Itô–Langevin calculus term by term (Stutzer et al., 6 Aug 2025).

The diffusion–jump correspondence is therefore not merely heuristic. In the cited jump framework, the discrete drift, noise, entropy production, and current covariations converge to their diffusion analogues. A plausible implication is that pathwise stochastic thermodynamics can be organized by observable type rather than by whether the microscopic dynamics are continuous or jump-like.

6. Ambiguities, applications, and current extensions

One disputed issue is the relation between pathwise solution concepts and the classical Itô–Stratonovich choice. Ryter proposes a pathwise construction based on short-time successive approximation of the integral equation

$$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$1

leading to a non-Gaussian basic increment, a modified anti-Itô integral, and a Fokker–Planck equation in which the noise-induced drift cancels exactly. In this formulation the pair $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$2, with $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$3, completely specifies both the SDE and its path-wise solution, and no further choice between Itô or Stratonovich is required (Ryter, 2020). This is a specific proposal rather than a universally adopted replacement for existing stochastic integrals.

A separate controversy concerns coherent-state path integrals. Altland and collaborators argue that ambiguities in continuous-time coherent-state path integrals arise from illegitimate use of standard calculus rules in a setting analogous to stochastic calculus. Treating the Hubbard–Stratonovich field as white noise yields an Itô-style substitution rule, resolves the chemical-potential shift, and establishes a one-to-one correspondence between path-integral discretization parameter $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$4 and Cahill–Glauber ordering index $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$5 (Rançon, 2019). Here stochastic-calculus ideas function as a consistency principle for path integrals rather than as a probabilistic limit theorem.

Path-wise observables also have direct experimental and computational roles. In precisely controlled colloidal systems, two equivalent Itô forms of the overdamped Langevin equation are used to extract local velocities, first-passage and transition-path statistics, and to assess data quality. The experiments verify, among other relations, the equivalence between $$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$6 and the long-time mean velocity, and confirm the Berezhkovskiĭ–Hummer symmetry for transition-path times in equilibrium (Dieball et al., 2024). In goal-oriented learning of stochastic dynamical systems, an error bound

$$J_t=\int_0^t\sum_{i\ne j}\kappa_{ij}(s)\,dn_{ij}(s)$$7

is used as a variational loss for path-space observables, including mean first hitting times on unbounded time domains. The accompanying Fréchet derivative of expected path functionals, obtained by a Girsanov argument, supplies an implementable stochastic-gradient formula (Zou et al., 20 Mar 2026).

Across these frameworks, the common principle is continuity or differentiability of observables on an appropriately enriched path space: rough paths use iterated integrals, functional Itô calculus uses horizontal and vertical derivatives, Föllmer calculus uses quadratic variation along partitions, local-time theories add occupation densities, and jump calculi use counting-noise decompositions. The main limitation, documented repeatedly in the literature, is that the relevant enrichment is not unique a priori. Partition choice, enhancement choice, and discretization convention may matter unless structural hypotheses—such as semimartingale sample paths with optional partitions, or a fixed rough-path lift—remove the ambiguity.