Space-Time Gibbs Measures

• Space-Time Gibbs Measures are probability measures on trajectories that integrate spatial and temporal locality via conditional specifications or dynamical equilibrium methods.
• They are constructed using different formulations such as Dobrushin–Lanford–Ruelle conditions, Gibbs tilts of Wiener measures, and Kubo–Martin–Schwinger equilibrium criteria.
• These measures have practical importance across disciplines including statistical physics, neuroscience, mathematical finance, and the study of nonlinear Hamiltonian PDEs.

Space-time Gibbs measures are probability measures on trajectories rather than on static configurations. In the Dobrushin–Lanford–Ruelle setting, they are path-space laws whose conditional distributions on finite space-time regions are quasilocal and compatible with a family of specifications; in increment-based constructions, they are Gibbs tilts of Wiener path measures by a space-time Hamiltonian; in Hamiltonian PDEs, the equilibrium law can be characterized dynamically by the classical Kubo–Martin–Schwinger condition. Across these formulations, locality is expressed simultaneously in space and in time, and the resulting objects appear in statistical physics, neuroscience, engineering, mathematical finance, nonlinear Hamiltonian PDEs, and the stochastic heat equation (Lacker et al., 2019, Mukherjee, 2017, Ammari et al., 2021).

1. Conceptual domain and principal formulations

The most explicit DLR-type formulation in the supplied literature is the second-order path-space theory for locally interacting diffusions on a locally finite graph. There the relevant random field is the collection of histories $(X_v[t])_{v\in V}\in \mathcal{C}_t^V$, with $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$, and locality is governed not by the ordinary graph boundary but by the “double boundary” $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$. A law is a second-order Markov random field when, for every finite $A\subset V$, $Y_A$ is conditionally independent of $Y_{(A\cup\partial^2 A)^c}$ given $Y_{\partial^2 A}$ (Lacker et al., 2019).

A distinct formulation arises on Brownian path space through translation-invariant interactions depending only on increments. In that setting the finite-volume Gibbs measure is defined by tilting Wiener increments with a Hamiltonian

$$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$

so the interaction is intrinsically space-time but does not proceed through local specifications in the explicit DLR sense (Mukherjee, 2017).

A third formulation replaces DLR locality by dynamical equilibrium. For nonlinear Hamiltonian PDEs, the classical KMS condition is used as an alternative to DLR: for suitable observables $F,G$ and Hamiltonian vector field $X$,

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$0

Within the admissible class considered, this characterizes Gibbs measures as the unique KMS equilibrium states (Ammari et al., 2021).

A related but more macroscopic perspective studies time-evolved Gibbs measures through cumulant-generating functions and push-forward dynamics. That framework does not define space-time Gibbs measures explicitly nor DLR specifications in space-time, but it isolates “macroscopic mixing” and “transitive mixing” as large-scale stability properties of time evolution (Lefevere et al., 2021).

Formulation	State space	Structural statement
Second-order DLR	$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$1 or $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$2	Path histories form a $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$3MRF with specifications $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$4
Increment Gibbs tilt	Brownian increments on $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$5	Infinite-volume Gibbs measure exists under Assumption A or B
Classical KMS	Gaussian/Wiener phase spaces for PDEs	Gibbs measures are the unique KMS equilibrium states
Macroscopic stability	Push-forward trajectory laws of observables	Gibbs-like cumulants and densities are stable at large $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$6

2. Second-order path-space Gibbs measures for locally interacting diffusions

The locally interacting diffusion model is indexed by the vertices of a simple, locally finite graph $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$7 with finite or countably infinite vertex set. At each vertex $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$8, the process satisfies

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$9

or, in the path-history notation emphasized in the source,

$$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$0

The drift is progressively measurable in the past histories of the vertex and its neighbors, while the diffusion coefficient depends only on the history of the vertex itself; the framework therefore allows non-Markovian, history-dependent drifts and also permits path-dependent $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$1 (Lacker et al., 2019).

For finite graphs, the initial law $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$2 is assumed absolutely continuous with respect to a product reference measure $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$3, with density factorizing over $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$4-cliques: $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$5 The second-order Hammersley–Clifford statement identifies this factorization with the $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$6MRF property, under strict positivity, by passing to the square graph $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$7 and applying the usual theorem for $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$8MRFs on $$\partial_G^2 A:=\partial_G A\cup\partial_G(A\cup\partial_G A)$$9. In the infinite-graph setting, the assumption is instead that $A\subset V$0 is a $A\subset V$1MRF with finite second moments and local absolute continuity relative to a product law (Lacker et al., 2019).

The path-space Gibbs structure becomes explicit through Girsanov’s theorem. If $A\subset V$2 denotes the law on $A\subset V$3 induced by the interacting SDE and $A\subset V$4 the law of the driftless system with product initial law, then in finite volume

$A\subset V$5

with

$A\subset V$6

This factorization exhibits the locality mechanism: the initial factor depends on $A\subset V$7-cliques, and each local martingale $A\subset V$8 depends only on $A\subset V$9. The corresponding conditional kernel on a finite $Y_A$0 is proportional to the product of the clique factors intersecting $Y_A$1 and the stochastic exponentials for $Y_A$2, with normalization given by a finite-volume conditional expectation under the reference law (Lacker et al., 2019).

This construction shows why the relevant Gibbs object is naturally a law on histories. The source emphasizes that even for gradient-type drifts and independent initial states, the static time-$Y_A$3 marginals need not be $Y_A$4MRF or $Y_A$5MRF. On a finite path of five vertices, the interacting Ornstein–Uhlenbeck–type system

$Y_A$6

provides such an example. The preserved structure is therefore the space-time $Y_A$7MRF on path space, not an automatic Gibbs property of one-time marginals (Lacker et al., 2019).

3. Propagation, bijection, and uniqueness for path-space Gibbs laws

Under the finite-graph assumptions, Theorem 3.4 states that for each $Y_A$8, $Y_A$9 is a $Y_{(A\cup\partial^2 A)^c}$0MRF on $Y_{(A\cup\partial^2 A)^c}$1, and moreover $Y_{(A\cup\partial^2 A)^c}$2 is a $Y_{(A\cup\partial^2 A)^c}$3MRF on $Y_{(A\cup\partial^2 A)^c}$4. Under the infinite-graph assumptions, the same conclusion holds by Theorem 3.8. In both cases, the conditional laws $Y_{(A\cup\partial^2 A)^c}$5 supply the DLR kernels for a second-order Gibbs measure on path space (Lacker et al., 2019).

A central structural result is the bijection between configuration Gibbs measures and space-time Gibbs measures. For a given initial $Y_{(A\cup\partial^2 A)^c}$6MRF $Y_{(A\cup\partial^2 A)^c}$7, define

$Y_{(A\cup\partial^2 A)^c}$8

$Y_{(A\cup\partial^2 A)^c}$9

Then

$Y_{\partial^2 A}$0

and the map $Y_{\partial^2 A}$1 is a bijection with inverse $Y_{\partial^2 A}$2. In particular, if $Y_{\partial^2 A}$3 for all $Y_{\partial^2 A}$4 and the initial projection equals $Y_{\partial^2 A}$5, then $Y_{\partial^2 A}$6 (Lacker et al., 2019).

This induces a Gibbs uniqueness principle on infinite graphs. If the initial Gibbs set $Y_{\partial^2 A}$7 is a singleton, then the associated path-space Gibbs set is also a singleton: $Y_{\partial^2 A}$8 Hence the joint path law is completely determined by the initial condition and the space-time second-order specifications $Y_{\partial^2 A}$9 (Lacker et al., 2019).

The infinite-volume construction relies on projection and approximation results that preserve second-order locality. A finite-volume graph $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$0 is obtained by augmenting the induced subgraph so that the outer ring is fully connected, which ensures $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$1 for sets $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$2 and preserves $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$3-clique factorizations on interior sets. The resulting finite-volume kernels are consistent on interiors, and uniform second-moment bounds together with relative entropy control yield tightness and convergence of the approximations. A plausible implication is that second-order conditioning is not a technical artifact of the proof but the minimal locality class in which both approximation and quasilocal consistency can be maintained on general locally finite graphs (Lacker et al., 2019).

4. Propagation of Gibbsianness under finite space-time interactions

A complementary line of work studies infinite-dimensional diffusions on $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$4 whose interaction has finite extent both in space and in time and is not supposed to be smooth or Markov. The initial measure is Gibbs for an interaction $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$5 satisfying strong summability,

$$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$6

and the dynamics are given by

$$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$7

The drift is adapted, local in a finite spatial neighborhood, and depends only on a finite time window in the past; it is also bounded uniformly in space, time, and paths (Roelly et al., 2013).

The main result asserts that if the strongness of the initial interaction is below a suitable level and the dynamical interaction is bounded from above in the required way, then for all $$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$8 the time-evolved single-time marginal

$$\mathscr H_T(\omega)=\int_0^T\int_0^T H(t-s,\omega(t)-\omega(s))\,ds\,dt,$$9

is again a Gibbs measure on $F,G$0 with some absolutely summable interaction $F,G$1, in the sense that

$F,G$2

The smallness assumptions are expressed by a Dobrushin-type condition on the initial interaction,

$F,G$3

and a Kotecký–Preiss condition for the dynamical polymer expansion (Roelly et al., 2013).

The proof proceeds through a space-time cluster expansion of the Girsanov factor. After passing to free bridge measures, the time-$F,G$4 density has the form

$F,G$5

with local functionals $F,G$6 built from the drift. Time is partitioned into blocks of size $F,G$7, polymers are defined from compatible space and time clusters, and the density is expanded as a gas of nonintersecting space-time clusters. Uniform bounds on polymer activities follow from boundedness of the drift and exponential ergodicity of the free dynamics, the latter encoded in decay estimates for $F,G$8 under the invariant measure of the free single-site dynamics (Roelly et al., 2013).

The logarithm of the finite-volume density then acquires a convergent polymer expansion,

$F,G$9

producing an effective double-layer interaction on the time slices $X$0. After Dobrushin uniqueness on the bi-layer system and a disintegration step, Kozlov’s representation theorem yields a single-time specification at time $X$1 and hence a time-dependent interaction $X$2. In contrast with the path-space $X$3MRF theory, the preserved Gibbsian object here is the one-time marginal. The two results are therefore complementary rather than equivalent: one preserves a full space-time Gibbs law on histories, while the other propagates Gibbsianness of single-time marginals under small, finite-range space-time perturbations (Roelly et al., 2013).

5. Increment Gibbs measures on Brownian path space

A different notion of space-time Gibbs measure is defined on the $X$4-algebra generated by Brownian increments $X$5 on $X$6. The interaction is translation invariant in the sense that the Hamiltonian depends only on differences $X$7: $X$8 The corresponding finite-volume Gibbs measure is

$X$9

with coupling parameter $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$00 (Mukherjee, 2017).

Two interaction regimes are treated. Assumption A allows long-range dependence in time with bounded spatial interaction and decay up to

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$01

Assumption B allows compactly supported time dependence and spatial singularities through a factorization $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$02 with $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$03 either bounded, Coulomb-like $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$04 in the specified range for $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$05, or $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$06 in $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$07. Under either assumption, for any $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$08 the infinite-volume Gibbs measure exists,

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$09

and the rescaled increment process satisfies a central limit theorem: $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$10 with $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$11 (Mukherjee, 2017).

The proof replaces the original long memory by a “Markovianized” chain of block increments. Interactions within a block are absorbed into a diagonal single-block measure, interactions between neighboring blocks define a kernel $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$12, and the associated integral operator

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$13

has a principal eigenvalue $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$14 and positive eigenfunction $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$15 by Krein–Rutman theory. The resulting tilted transition kernel

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$16

satisfies a Doeblin minorization, which yields geometric ergodicity. Solving the Poisson equation $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$17 for the increment observable $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$18 produces a Green–Kubo or Dirichlet-form representation of the limiting variance and ensures strict positivity (Mukherjee, 2017).

This framework covers the Nelson model, massless bosons, polaron models with ultraviolet cutoff, and self-interacting polymers with $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$19 in dimension $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$20. It also applies to the multiplicative-noise stochastic heat equation in dimensions $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$21 after time and space mollification of the noise. Through a Feynman–Kac representation, the annealed path measure becomes exactly of the Gibbs form above, and the averaged diffusively rescaled solution converges pointwise to a deterministic diffusion equation with homogenized coefficient $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$22 (Mukherjee, 2017).

A recurring threshold in this setting is the time-decay exponent. The source states that decay exponent $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$23 is critical: for slower decays, including kernels with effective decay $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$24 for $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$25, multiple infinite-volume Gibbs measures can occur, which obstructs mixing and invalidates the CLT under convex combinations. This is one of the clearest phase-transition mechanisms presently identified in the supplied material for path-space Gibbs measures (Mukherjee, 2017).

6. Alternative equilibrium characterizations, macroscopic stability, and conceptual boundaries

The classical KMS condition provides a dynamical alternative to DLR. In the setting of nonlinear Hamiltonian PDEs on Gaussian probability spaces, Gibbs measures are shown to be the unique KMS equilibrium states for nonlinear Schrödinger, Hartree, and nonlinear wave or Klein–Gordon equations. The Gibbs law is written relative to a Gaussian reference measure $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$26 as

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$27

and the KMS identity with respect to the Hamiltonian vector field $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$28 implies the density equation

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$29

Using Malliavin calculus, Gaussian integration by parts, and the principle that vanishing Malliavin gradient forces almost sure constancy, one obtains $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$30. In this sense, Gibbs equilibrium and dynamical equilibrium coincide within the class of measures absolutely continuous with respect to the Gaussian reference and with density in $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$31 (Ammari et al., 2021).

The KMS viewpoint is intrinsically space-time because it is formulated through time-translation invariance and correlation identities rather than through spatial conditional specifications. The source further derives stationary and equilibrium hierarchies for $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$32-point densities, describing these as space-time equilibrium constraints. This suggests that “space-time Gibbs measure” can refer either to a DLR path-space object or to an equilibrium law characterized by temporal analyticity and Poisson-bracket identities, depending on context (Ammari et al., 2021).

At a still different level, macroscopic stability of time-evolved Gibbs measures can be studied without introducing DLR specifications. In the lattice mechanical system with random mirrors, “macroscopic mixing” is expressed through the cumulant-generating function

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$33

and requires

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$34

“Transitive mixing” is a two-stage analogue after a parameter switch at time $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$35, and it implies the entropy monotonicity

$\mathcal{C}_t=C([0,t];\mathbb{R}^d)$36

The paper explicitly notes that it does not define space-time Gibbs measures nor DLR specifications in space-time; the result is instead that macroscopic observables over space-time windows are governed by Gibbs-like cumulants and densities, even though the microscopic dynamics are non-ergodic and not mixing (Lefevere et al., 2021).

Several conceptual boundaries follow from the supplied literature. First, space-time Gibbsianity of histories does not imply that one-time marginals remain Gibbs in the same locality class, as shown by the interacting Ornstein–Uhlenbeck example on a five-vertex path (Lacker et al., 2019). Second, uniqueness is model-dependent: it can follow from a bijection with the initial Gibbs law for locally interacting diffusions, from Doeblin minorization in the Markovianized increment chain, or from Malliavin regularity in KMS theory, but it can also fail beyond critical time-decay thresholds or outside the admissible absolute continuity class (Lacker et al., 2019, Mukherjee, 2017, Ammari et al., 2021). Third, microscopic non-mixing does not exclude macroscopic Gibbs-like stability, which separates dynamical equilibration of observables from full path-space quasilocality (Lefevere et al., 2021).

Open directions stated in the supplied material include extending Markovianization to interactions that are simultaneously long-range in time and singular in space, developing DLR formulations and phase-transition theory near the critical decay threshold, characterizing $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$37 more explicitly in quantum models, strengthening KMS uniqueness beyond $\mathcal{C}_t=C([0,t];\mathbb{R}^d)$38 densities, and extending the Hamiltonian-PDE theory to stochastic or quantum settings (Mukherjee, 2017, Ammari et al., 2021).