Fractional Zakai Equations

• Fractional Zakai equations are stochastic partial integro-differential equations that replace traditional time derivatives with Riemann–Liouville fractional derivatives.
• They incorporate non-Markovian time-changes and jump processes to effectively model anomalous dynamics such as subdiffusion and heavy-tailed noise.
• These equations guarantee unique and regular solutions under smoothness, ellipticity, and boundedness conditions, providing robust frameworks for applications in finance, geophysics, and biology.

A fractional Zakai equation is a stochastic partial integro-differential equation (SPIDE) governing the evolution of the un-normalized conditional distribution (filter) in nonlinear filtering problems where the state, observation, or both processes are time-changed to exhibit non-Markovian, subdiffusive, or heavy-tailed phenomena. These equations arise by introducing fractional calculus—typically via the Riemann–Liouville fractional derivative—in place of the standard time derivative, or by incorporating operators related to jump (Lévy) processes, thereby generalizing the classical Zakai framework to new regimes of anomalous dynamics and jump noise (Umarov et al., 2013, Mikulevicius et al., 2010).

1. Fundamentals of Fractional Zakai Equations

The classical Zakai equation provides the evolution of the unnormalized conditional law $\pi_t(f)$ associated with a hidden Markov process observed through noise. In the classical setting, for a state process $X_t$ satisfying

$dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$

with observation

$dZ_t = h(X_t)\,dt + dW_t,$

the Zakai equation for $\pi_t(f)$ is

$$d\pi_t(f) = \pi_t(Af)\,dt + \sum_{k=1}^m \pi_t(h_k f)\,dZ_t^{(k)},$$

where $A$ is the infinitesimal generator of $X_t$. In its density form, this becomes

$$du(t,x) = A^* u(t,x)\,dt + \sum_{k=1}^m h_k(x) u(t,x)\,dZ_t^{(k)}.$$

Fractional Zakai equations arise when the temporal parameter is replaced by a random time-change process $T_t$, typically the inverse of a stable subordinator, or when stochastic dynamics incorporate nonlocal jump terms. The hallmark of the fractional Zakai equation is the replacement of the time derivative with a Riemann–Liouville fractional derivative, or via an integro-differential generator (e.g., fractional Laplacian) in the drift (Umarov et al., 2013, Mikulevicius et al., 2010).

2. Mathematical Structure and Operator Formulation

In the time-changed regime, the state process is given by $X_t$0, where $X_t$1 is the inverse (first-hitting time) of a strictly increasing $X_t$2-stable Lévy process ($X_t$3), so that

$X_t$4

with $X_t$5 the subordinator. The density of $X_t$6 in the $X_t$7 plane $X_t$8 solves

$X_t$9

where $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$0 is the Riemann–Liouville fractional derivative.

The corresponding fractional Zakai equation for the un-normalized filter $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$1 is

$dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$2

where $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$3 is the Riemann–Liouville fractional derivative of order $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$4: $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$5 For densities $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$6,

$dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$7

An alternative fractional regime leads to generators of jump processes: $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$8 yielding a Zakai SPIDE with a fractional Laplacian-type term in the drift (Mikulevicius et al., 2010).

3. Analysis of Regularity and $dX_t = b(X_t)\,dt + \sigma(X_t)\,dB_t,$9-Estimates

A distinguishing feature of fractional Zakai equations, especially in the presence of jumps, is the appearance of nonlocal singular integral operators,

$dZ_t = h(X_t)\,dt + dW_t,$0

where $dZ_t = h(X_t)\,dt + dW_t,$1 is bounded and 0-homogeneous. The symbol of this operator,

$dZ_t = h(X_t)\,dt + dW_t,$2

encodes fractional diffusion and is sectorial under ellipticity assumptions.

The core $dZ_t = h(X_t)\,dt + dW_t,$3-theory, originally established by Mikulevičius and Pragarauskas, demonstrates the mapping properties of $dZ_t = h(X_t)\,dt + dW_t,$4 on Sobolev and Besov scales: $dZ_t = h(X_t)\,dt + dW_t,$5 These results guarantee a fractional Sobolev gain and imply existence, uniqueness, and regularity for solutions $dZ_t = h(X_t)\,dt + dW_t,$6 of the Zakai SPIDE in appropriate function spaces (Mikulevicius et al., 2010).

4. Existence, Uniqueness, and Solution Properties

For fractional Zakai equations, existence and uniqueness of mild (square-integrable) solutions are established under standard smoothness and growth conditions on coefficients and the Lévy kernel:

• Coefficients $dZ_t = h(X_t)\,dt + dW_t,$7 are $dZ_t = h(X_t)\,dt + dW_t,$8 smooth, bounded, with all derivatives bounded.
• The inverse subordinator $dZ_t = h(X_t)\,dt + dW_t,$9 is independent of the driving Brownian motion(s).
• The initial condition $\pi_t(f)$0 is independent and possesses a rapidly decaying $\pi_t(f)$1 density.

Under these conditions, the Zakai problem in the "inner time" variable $\pi_t(f)$2 has a unique solution, which is then transferred to the fractional regime via integration against the subordinator density $\pi_t(f)$3, yielding uniqueness in the fractional case as well (Umarov et al., 2013). For jump-driven models, additional ellipticity and regularity assumptions on $\pi_t(f)$4 are required (Mikulevicius et al., 2010).

The solution operator exhibits a fractional gain in Sobolev regularity, and the conditional density $\pi_t(f)$5 is well-defined as a distribution in dual Sobolev spaces.

5. Special Cases, Classical Limit, and Connections

Special cases clarify the relation to traditional filtering:

• Fractional Brownian filtering: With $\pi_t(f)$6 and constant coefficients (except $\pi_t(f)$7), the fractional Zakai equation reduces to

$\pi_t(f)$8

directly connecting the fractional derivative to subdiffusive dynamics.

• Classical limit: As $\pi_t(f)$9, the fractional derivative $$d\pi_t(f) = \pi_t(Af)\,dt + \sum_{k=1}^m \pi_t(h_k f)\,dZ_t^{(k)},$$0, and the equation becomes the standard Zakai equation.
• Mixed time-changes: If only the state or only the observation process is time-changed, the resulting Zakai equation retains the fractional derivative on the drift, with the appropriate martingale integrals (Umarov et al., 2013).
• Jump-driven signal and observation: When the underlying process generator is non-local with stable-type jumps, the Zakai equation features fractional Laplacian terms and corresponding singular integral structure (Mikulevicius et al., 2010).

6. Applications, Interpretation, and Practical Implications

Fractional Zakai equations model filtering scenarios with anomalous dynamics such as subdiffusion, trapping, or heavy-tailed waiting times, introducing memory into the filtering evolution via the nonlocal fractional derivative. They provide non-Markovian filtering for hidden states or observations with long-range dependence.

Applications arise in geophysics, finance, and biology, particularly where sensors or systems display power-law waiting times or pronounced jumps. The non-local evolution encoded in $$d\pi_t(f) = \pi_t(Af)\,dt + \sum_{k=1}^m \pi_t(h_k f)\,dZ_t^{(k)},$$1 and integro-differential drift terms allows for the accommodation of phenomena not captured by classical filtering theory (Umarov et al., 2013, Mikulevicius et al., 2010).

7. Numerical Considerations and Approximation Schemes

Numerical treatment of fractional Zakai equations requires discretization of the fractional operators. Practical strategies include:

• Discretizing the Riemann–Liouville derivative using Grünwald–Letnikov or convolution-based quadrature.
• Combining finite-difference or finite-element schemes for fractional diffusion with standard SPDE integrators for martingale drivers.
• Implementing Monte Carlo simulation for the inverse subordinator $$d\pi_t(f) = \pi_t(Af)\,dt + \sum_{k=1}^m \pi_t(h_k f)\,dZ_t^{(k)},$$2 within particle filtering algorithms.

Analogously, jump-driven Zakai SPIDEs necessitate numerical schemes that robustly approximate both the drift and nonlocal fractional operator, with stability measured in Sobolev/Besov norms, and complexity $$d\pi_t(f) = \pi_t(Af)\,dt + \sum_{k=1}^m \pi_t(h_k f)\,dZ_t^{(k)},$$3 or exploiting FFT-based techniques (Mikulevicius et al., 2010, Umarov et al., 2013).

---

For foundational results and further developments, see [Mikulevičius–Pragarauskas, (Mikulevicius et al., 2010)] and [Umarov–Daum–Nelson, (Umarov et al., 2013)].