Markovian Lifts in Stochastic Processes

• Markovian lifts are state-space augmentation techniques that replace hidden memory with an explicit enlarged state, restoring the Markov property.
• They are applied to stochastic Volterra equations, nonreversible MCMC, and quantum or algebraic settings to enhance model tractability and analysis.
• Methodologies use projection maps, semigroup representations, and infinite-dimensional frameworks to derive limit theorems, ergodic properties, and efficient approximations.

Markovian lifts are state-space augmentation procedures that embed a given object into a larger Markovian one. In stochastic Volterra theory, the lift restores the Markov property of memory-dependent dynamics by encoding the past in an infinite-dimensional state variable; in Markov chain Monte Carlo, lifting enlarges the chain with auxiliary variables to create nonreversible motion while preserving a prescribed stationary law; and in adjacent literatures the term also denotes constructions for classical and quantum composite states, abstractions of parametric Markov chains, and generalized Lawrence liftings in algebraic statistics. This suggests that the unifying principle is not a single canonical model but a recurrent operation: replace hidden memory, constraints, or parameter dependence by an explicit enlarged state on which a Markovian or Markov-compatible dynamics acts (Hamaguchi, 2023, Vucelja, 2014, Accardi et al., 2011).

1. Core idea and formal patterns

A Markovian lift replaces a process whose future depends on more than its current visible state by a process whose enlarged state is sufficient for future evolution. In the stochastic Volterra setting, the original dynamics has convolution kernels and is typically neither Markovian nor a semimartingale; the lifted object is a stochastic evolution equation on a Hilbert, Banach, or measure-valued state space, with the original process recovered by a projection such as $X_t=\Xi \mathcal{X}_t$ or $V_t=(g,A_t)$ (Friesen et al., 2024, Cuchiero et al., 2018). In nonreversible MCMC, the base chain is already Markovian, but the lift adds an auxiliary variable—often a direction, replica, or activity label—to create persistence and net stochastic flow while preserving the target stationary distribution under projection (Vucelja, 2014).

Two recurrent mathematical templates appear across the literature. The first is a projection-based reconstruction,

$X_t=\Xi \mathcal{X}_t,$

or, more generally,

$V_t=(g,A_t),$

where the lifted state contains the information lost in the observable coordinates (Friesen et al., 2024, Cuchiero et al., 2018). The second is a semigroup representation of memory kernels. For completely monotone kernels, Bernstein-type Laplace representations such as

$K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$

or

$K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$

allow the lift to be driven by multiplication or shift semigroups on weighted function spaces (Friesen et al., 2024, Hamaguchi, 2023).

A common misconception is that a Markovian lift is automatically finite-dimensional. In the principal stochastic-analysis applications, the canonical lift is generally infinite-dimensional: weighted $L^2$-type spaces, weighted Sobolev spaces, Filipović spaces, dual Banach spaces, or spaces of matrix-valued measures are used precisely because the memory structure cannot, in general, be captured by finitely many coordinates (Huber, 2024, Bianchi et al., 10 Sep 2025, Cuchiero et al., 2019).

2. Stochastic Volterra equations and infinite-dimensional state augmentation

For stochastic Volterra integral equations, the basic problem is that convolution with a kernel $K$ introduces path dependence. A representative equation studied in the recent small-time theory is

$$X_t=g(t)+\int_0^t K(t-s)b(X_s)\,ds+\int_0^t K(t-s)\sigma(X_s)\,dB_s,\qquad t\ge 0.$$

When $K$ is completely monotone, one can define a Hilbert space $V_t=(g,A_t)$0 with inner product

$V_t=(g,A_t)$1

use the shift semigroup $V_t=(g,A_t)$2, and define the projection

$V_t=(g,A_t)$3

The lifted process $V_t=(g,A_t)$4 then solves

$V_t=(g,A_t)$5

with $V_t=(g,A_t)$6 (Friesen et al., 2024).

Related frameworks formulate the lift as a stochastic evolution equation on weighted Sobolev spaces or on Hilbert spaces tailored to the kernel singularity. One construction uses a measure-valued process $V_t=(g,A_t)$7 satisfying

$V_t=(g,A_t)$8

with the original SVE recovered by $V_t=(g,A_t)$9. Another uses the abstract SEE

$X_t=\Xi \mathcal{X}_t,$0

where $X_t=\Xi \mathcal{X}_t,$1 and $X_t=\Xi \mathcal{X}_t,$2 encode operator-valued Volterra kernels (Huber, 2024, Bianchi et al., 10 Sep 2025).

The functional-analytic realization depends on the kernel class. Laplace-transform lifts use weighted $X_t=\Xi \mathcal{X}_t,$3-spaces with semigroup $X_t=\Xi \mathcal{X}_t,$4 and projection $X_t=\Xi \mathcal{X}_t,$5. Shift-based lifts on Filipović spaces use $X_t=\Xi \mathcal{X}_t,$6 and $X_t=\Xi \mathcal{X}_t,$7, thereby accommodating kernels beyond the completely monotone class, including singular and non-monotonic fractional kernels for $X_t=\Xi \mathcal{X}_t,$8 (Bianchi et al., 10 Sep 2025). In the affine-jump case, the lift may instead be a measure-valued mild solution of

$X_t=\Xi \mathcal{X}_t,$9

with

$V_t=(g,A_t),$0

(Cuchiero et al., 2018).

3. Limit theorems, invariant measures, and ergodic structure

Once the lift is Markovian, small-time and long-time asymptotics become accessible by semigroup and coupling methods. For Hilbert space-valued lifts of SVIEs with completely monotone kernels, finite-dimensional projections satisfy a small-time CLT: for $V_t=(g,A_t),$1,

$V_t=(g,A_t),$2

with

$V_t=(g,A_t),$3

The same framework yields a functional CLT and limit theorems for smooth transformations, and covers kernels including Riemann–Liouville kernels with short- and long-range dependencies (Friesen et al., 2024).

For time-asymptotic questions, several complementary results are now available. Weighted Sobolev-space lifts provide existence conditions for invariant measures for both the lifted process and the projected SVE, together with an Ito-type formula adapted to the nonlocal setting (Huber, 2024). Abstract Hilbert-space lifts yield existence and characterization of possibly multiple limit distributions and stationary processes, a law of large numbers with rate,

$V_t=(g,A_t),$4

and a CLT for time averages in the Gaussian domain of attraction (Bianchi et al., 10 Sep 2025). A notable feature is that the limit distribution may depend on persistent memory components through $V_t=(g,A_t),$5, so non-uniqueness of stationary distributions is part of the theory rather than an anomaly (Bianchi et al., 10 Sep 2025).

Further regularity and ergodicity results use asymptotic coupling and generalized Harris theory. Markovian lifts of SVIEs with completely monotone kernels satisfy an asymptotic log-Harnack inequality, which yields asymptotic strong Feller-type consequences and uniqueness results for invariant measures under the stated nondegeneracy and decay assumptions (Hamaguchi, 2023). Exponential ergodicity has subsequently been established for a class of SVE lifts on a Gelfand triplet by constructing a generalized coupling and a distance-like function adapted to the highly degenerate infinite-dimensional dynamics; the same work proves that invariant measures and stationary laws on path space can be weakly approximated by those of finite-dimensional SDEs, thereby giving a rigorous approximation result for stationary solutions of the original SVE (Hamaguchi, 3 Mar 2026).

These results also correct a frequent misunderstanding. Restoring the Markov property does not, by itself, imply classical finite-dimensional strong Feller theory. In several Volterra lift frameworks the noise is finite-dimensional while the state space is infinite-dimensional, and strong Feller may fail; the appropriate substitutes are generalized Feller semigroups, asymptotic log-Harnack inequalities, Wasserstein contraction, and distance-like couplings (Cuchiero et al., 2019, Hamaguchi, 2023, Hamaguchi, 3 Mar 2026).

4. Affine Volterra processes, rough volatility, and Heston-type models

In mathematical finance, Markovian lifts provide a tractable representation of rough and affine Volterra models. The generalized Feller framework for affine stochastic Volterra processes with jumps treats lifted SPDEs on weighted Banach spaces and shows existence, uniqueness, approximation, and affine transform formulas. This includes rough volatility models of general jump diffusion type, and was explicitly motivated by the observation that “in this Markovian light the theory of stochastic Volterra processes becomes almost classical” (Cuchiero et al., 2018).

Cuchiero and Teichmann developed Markovian lifts of positive semidefinite affine Volterra type processes by considering SPDEs on cones of $V_t=(g,A_t),$6-valued measures and proving the generalized Feller property (Cuchiero et al., 2019). Their framework introduces Volterra Wishart processes with fractional kernels, constructed from matrix products of infinite-dimensional Ornstein–Uhlenbeck processes, as well as positive definite Volterra pure jump processes yielding multivariate Hawkes type processes. The projected covariance process takes the form

$V_t=(g,A_t),$7

with affine transform formulas governed by matrix Riccati–Volterra equations (Cuchiero et al., 2019).

The lifted Heston model is the best-known finite-dimensional approximation paradigm. It replaces the rough Heston variance by

$V_t=(g,A_t),$8

so that $V_t=(g,A_t),$9 recovers classical Heston and $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$0 approximates rough Heston through exponential-sum kernels $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$1 (Jaber, 2018). The model is finite-dimensional and Markovian, retains tractable Fourier-Laplace pricing formulas via a system of $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$2 ODEs, and was presented as reconciling the classical Heston model with its rough counterpart (Jaber, 2018).

Subsequent numerical work treats the lifted Heston model as a Markovian lift of the rough Heston model with $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$3 square-root state processes and proposes an implicit integrated variance scheme based on an $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$4-optimal linear projection and inverse Gaussian sampling. The stated aim is efficient Monte Carlo simulation with large time steps, while preserving positivity through constrained projection (Zaugg et al., 9 Oct 2025). This computational line is consistent with the broader observation that finite-dimensional Markovian approximation methods, including multi-factor approximations, become mathematically rigorous when viewed from the lifted-space perspective (Friesen et al., 2024).

5. Nonreversible lifted chains and acceleration of mixing

In Markov chain Monte Carlo, lifting has a different but related meaning. The base chain is already Markovian; the lift enlarges the state space to break detailed balance while preserving stationarity. A standard construction duplicates the state space,

$K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$5

and imposes skew detailed balance,

$K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$6

instead of ordinary detailed balance (Vucelja, 2014). The motivation is to reduce diffusive exploration by introducing directional persistence.

The literature is explicit that detailed balance is sufficient, not necessary, for convergence. Nonreversible lifted chains satisfy global balance and can mix faster than reversible ones, but the speedup is structurally limited. The review of lifting as a nonreversible MCMC algorithm reports square-root improvements in some cases, such as $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$7 to $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$8 for a ring or torus, and $K(t)=K(\infty)+\int_0^\infty e^{-xt}\mu(dx)$9 to $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$0 for the mean-field Ising model, while also noting cases such as the two-dimensional Ising model where only prefactors, not asymptotic scaling, improve (Vucelja, 2014). For continuous-state chains, evolving-set methods establish lower bounds showing that any lift can at best improve mixing time by roughly a square-root factor, with the bound expressed through a structural quantity $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$1 rather than $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$2 (Ramanan et al., 2016). A complementary analysis of lifted chains on graphs shows that the admissible speedup depends crucially on initialization and invariance constraints: with suitable initialization and without invariance one can reach diameter-time mixing, whereas imposing both no controlled initialization and invariance eliminates acceleration over the best local chain (Apers et al., 2017).

Recent work has sharpened the convergence theory through “second-order lifts.” Given a nonreversible Markov process $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$3 projecting to a reversible diffusion $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$4, the lift is second-order if

$K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$5

This framework supports a flow Poincaré inequality and variational hypocoercivity estimates for PDMPs and run-and-tumble systems, again with relaxation rates controlled at square-root scale by the spectral gap of the collapsed reversible process (Eberle et al., 6 Mar 2025).

Concrete lifted chains have also become model systems in their own right. The lifted directed-worm algorithm introduces a mode variable $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$6 and uses geometric allocation to minimize backscattering and maximize net stochastic flow, with reported efficiency gains of approximately $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$7, $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$8, and $K(t)=\int_{[0,\infty)} e^{-\theta t}\mu(d\theta)$9 relative to the standard worm algorithm, the Wolff cluster algorithm, and the previous lifted worm algorithm for the four-dimensional hypercubic lattice Ising model (Suwa, 2022). The lifted TASEP is presented as a “second-generation lifting” of a reversible Metropolis algorithm, with a non-local activity variable and an exactly solvable transition matrix amenable to Bethe ansatz analysis (Essler et al., 2023).

6. Adjacent meanings: classical and quantum liftings, parametric abstractions, and algebraic liftings

The term “lifting” also appears in settings where the enlarged object is not primarily introduced to restore Markovianity of a memory equation. In classical and quantum information theory, lifting means embedding a state of a subsystem into a state of a composite system. A classical lifting is a map

$L^2$0

and a Markovian lifting is specified by conditional probabilities through

$L^2$1

The classical Ohya lifting,

$L^2$2

is nondemolishing and reproduces the marginal on both subsystems. In the quantum case, product liftings preserve completely positive reduced dynamics, whereas more general correlated liftings may lead to merely positive, not completely positive, reduced maps (Accardi et al., 2011).

In verification of parametric Markov chains, “parameter lifting” denotes an abstraction-refinement methodology rather than a state-process embedding. Generalized parameter lifting abstracts a pMC over a region $L^2$3 to an interval Markov chain via interval substitution,

$L^2$4

and refines the abstraction by a big-step transformation that reduces parameter dependencies before analysis (Heck et al., 8 Apr 2025). The paper explicitly presents GPL as lifting restrictive assumptions of standard parameter lifting and reports accelerations “by up to orders of magnitude” (Heck et al., 8 Apr 2025).

In algebraic statistics, generalized Lawrence liftings are matrices

$L^2$5

used to study Markov bases, universal Gröbner bases, and Graver complexity (Charalambous et al., 2013). Although this usage is structurally different from stochastic-process lifts, it preserves the same general motif: an enlargement is introduced so that hidden combinatorial structure becomes explicit.

Across these literatures, a consistent distinction remains essential. A Markovian lift is not synonymous with a finite-factor approximation, a nonreversible chain, or a product extension of a state. Rather, each is a specialized realization of an enlargement principle whose mathematical content is dictated by the problem class. A plausible implication is that the most substantive commonality lies in the projection map: the lifted object is valuable precisely because the original one is recovered by an explicit collapse, while the enlarged object admits tools unavailable at the base level.