Level-3 Large Deviation Principle

• Level-3 Large Deviation Principle is a process-level framework establishing LDPs for empirical process measures, capturing the full temporal dynamics of complex systems.
• It employs entropy rate functions, convex duality, and variational formulations, with applications in models like nonlinear Hawkes processes and discrete Markov chains.
• The principle enables recovery of lower-level large deviation results via contractions, bridging stationary and nonequilibrium analyses with superexponential estimates.

A Level-3 Large Deviation Principle ordinarily denotes a process-level large deviation principle for empirical process measures on path space. In the standard hierarchy, Level 1 concerns scalar observables or finite-dimensional marginals, Level 2 concerns empirical measures or occupation measures, and Level 3 concerns the empirical process or full stationary path-measure LDP in the Donsker–Varadhan sense (Zhu, 2011). The classification is determined by the random object and topology: several works explicitly distinguish genuine Level-3 results from finite-time path-space Freidlin–Wentzell LDPs for stochastic evolution equations, occupation-profile LDPs, and other measure-valued statements that are stronger than scalar LDPs but are not Level 3 in the classical Donsker–Varadhan sense (Wang et al., 2010, Li et al., 2013, Johnston et al., 2021).

1. Standard meaning and canonical random objects

In the modern literature, a genuine Level-3 theorem is typically formulated for an empirical process obtained by averaging shifts of a trajectory. For nonlinear Hawkes processes, the empirical process is

$$R_{t,\omega}(A)=\frac1t\int_0^t \mathbf 1_A(\theta_s\omega_t)\,ds,$$

a probability measure on the path space $\Omega$ of countable, locally finite subsets of $\mathbb R$, with $\omega_t$ the periodized path segment and $\theta_s$ the shift (Zhu, 2011). For a Markov chain on a discrete countable state space $S$, the analogous object is

$L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$

a probability measure on $S^{\mathbb N}$, where $T$ is the left shift (Daures, 15 Jul 2025). On a one-sided full shift $\Omega=A^{\mathbb N}$, the empirical process is

$\Omega$0

with $\Omega$1 the left shift (Cuneo et al., 2017).

In interacting-particle systems, the same level is realized by empirical measures of full particle trajectories rather than one-time empirical densities. For locally interacting Brownian motions on the one-dimensional torus, the empirical process is

$\Omega$2

a random probability measure on $\Omega$3 (Seo, 2015). For the white-forced 2D Navier–Stokes system in a bounded domain, the path-space empirical measure is

$\Omega$4

viewed as a random element of $\Omega$5 with $\Omega$6 (Zhao, 17 Jun 2025).

These formulations make clear why Level 3 is not merely “an LDP on a path space.” A small-noise LDP for a single random path $\Omega$7 is a path-space LDP, but it is not process-level in the Donsker–Varadhan sense unless the random object itself is an empirical process or stationary path measure (Wang et al., 2010).

2. Rate functions and Donsker–Varadhan structure

The characteristic Level-3 rate function is an entropy rate or its convex-dual representation. For stationary nonlinear Hawkes processes, the rate is the process-level entropy

$\Omega$8

and, when $\Omega$9, it has the explicit Girsanov form

$\mathbb R$0

where $\mathbb R$1 is the compensator density under $\mathbb R$2 and $\mathbb R$3 is the Hawkes intensity (Zhu, 2011).

For selectively decoupled measures on shift spaces, the Level-3 rate function is first defined by convex duality,

$\mathbb R$4

and under Upper Decoupling it becomes the specific relative entropy density

$\mathbb R$5

while $\mathbb R$6 off the invariant measures (Cuneo et al., 2017).

For general Markov chains on discrete state spaces, the rate function retains the classical Donsker–Varadhan formulas only on an admissible shift-invariant set: $\mathbb R$7 Under $\mathbb R$8, it also admits the entropy-rate representation

$\mathbb R$9

This admissibility cutoff is the decisive reducible-chain modification (Daures, 15 Jul 2025).

For the white-forced 2D Navier–Stokes system, the Level-3 rate function is a continuous-time Donsker–Varadhan entropy on two-sided path space: $\omega_t$0 The corresponding Level-2 occupation-measure rate is given exactly by the generator formula

$\omega_t$1

This is an exact Donsker–Varadhan identification in a noncompact infinite-dimensional setting (Zhao, 17 Jun 2025).

Not every genuine Level-3 theorem has a purely stationary entropy form. For interacting Brownian motions in nonequilibrium dynamics, the good rate function is

$\omega_t$2

with

$\omega_t$3

and

$\omega_t$4

where $\omega_t$5 is the marginal density of $\omega_t$6 and $\omega_t$7 is the minimizing admissible drift (Seo, 2015). The entropy term remains central, but the formula is adapted to a finite-time nonequilibrium diffusion system.

3. Representative Level-3 theorems

The nonlinear Hawkes theorem is a full Level-3 LDP on $\omega_t$8, the stationary simple point-process laws with finite first moment, equipped with the strengthened weak topology requiring weak convergence together with convergence of $\omega_t$9. The empirical process $\theta_s$0 under the empty-history law $\theta_s$1 satisfies the lower and upper bounds

$\theta_s$2

for open $\theta_s$3, and

$\theta_s$4

for closed $\theta_s$5 (Zhu, 2011). The upper bound is upgraded from compact to closed sets by superexponential estimates adapted to the strengthened topology.

For general discrete-state Markov chains, the theorem is a Level-3 weak LDP on $\theta_s$6 with the weak topology and speed $\theta_s$7, with lower bounds on open sets and upper bounds on compact sets. The weakness is essential: the paper assumes only that the state space is discrete and does not assume irreducibility, exponential tightness, goodness of the rate function, or any uniformity condition such as the usual assumption $\theta_s$8 (Daures, 15 Jul 2025). In the reducible case, the rate function may be nonconvex.

For invariant measures on shift spaces over finite alphabets satisfying Selective Lower Decoupling, the empirical process $\theta_s$9 satisfies a full LDP on $S$0 with speed $S$1 and a good convex rate function (Cuneo et al., 2017). Under Upper Decoupling, the rate becomes the specific relative entropy density. The theorem is formulated without assuming Gibbs structure or thermodynamic formalism.

For locally interacting Brownian motions, the final theorem is a Level-3 LDP on

$S$2

with the topology of weak convergence and speed $S$3 (Seo, 2015). The result is annealed, nonequilibrium, and finite-time; it concerns empirical laws of full particle trajectories rather than stationary infinite-volume processes. The paper states that only two lattice systems had previously been treated in this nonequilibrium process-level context, so this result is the third in that context and the first for a diffusion-type interacting particle system.

For the white-forced 2D Navier–Stokes system in a bounded domain, the Level-3 theorem is uniform over initial laws in

$S$4

and gives a good rate function on $S$5 with $S$6 (Zhao, 17 Jun 2025). The theorem is formulated under nondegenerate additive noise $S$7 for all Stokes modes.

4. Proof architectures

The Hawkes proof is direct at process level and avoids Markovian approximations. The lower bound is obtained by an entropy/tilting argument based on conditioning on the past, controlling inherited memory through the decay of the kernel $S$8, and using good-past sets $S$9. The upper bound starts from exponential martingale estimates for a class of functionals

$L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$0

followed by periodization-error estimates and superexponential controls of local multiplicities and large local counts (Zhu, 2011).

For reducible discrete-state Markov chains, the central methodological novelty is a subadditive proof at the pair-measure level followed by block and projective-limit lifts. The core construction is “slicing and stitching”: finite trajectories are sliced into subwords associated with irreducible classes, reordered according to the total order of reachable classes in an admissible empirical measure, and stitched together with short connecting words respecting reachability (Daures, 15 Jul 2025). Level 3 is then obtained by a weak-LDP version of Dawson–Gärtner applied to $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$1-block empirical measures.

On shift spaces, the proof is based on Ruelle–Lanford functions and a selective gluing map $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$2 built from the Selective Lower Decoupling hypothesis (Cuneo et al., 2017). The argument first proves admissibility for scalar Birkhoff sums and then lifts that admissibility to empirical-process neighborhoods in $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$3. Exponential tightness is immediate because $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$4 and $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$5 is compact in the weak topology.

For interacting Brownian motions, the path-space theorem is not proved directly from the original diffusion system. The decisive intermediate object is the empirical density of colors, whose LDP supplies the finite-dimensional projections of the empirical process (Seo, 2015). The proof combines super-exponential replacement estimates converting collision local times into local densities, exponential tightness, a color-density LDP with rate

$L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$6

Girsanov tilts with color-dependent drifts and modified label-switching rates, and Dawson–Gärtner projective limits.

For the white-forced 2D Navier–Stokes system, the proof combines an improved Kifer criterion, a lift argument inspired by Donsker–Varadhan, an improved abstract result on generalized Markov semigroups, and a resolvent approximation scheme utilizing the resolvent operators of the Markov semigroup (Zhao, 17 Jun 2025). The improved Kifer criterion replaces the exponential growth condition used in earlier work by exponential tightness along subsequences. The lift argument passes from level-2 control of occupation measures to path-space control of periodized empirical trajectory measures. The resolvent approximation is what yields the exact generator-based Donsker–Varadhan formula in the noncompact infinite-dimensional phase space.

5. Terminological boundaries and recurrent misclassifications

A recurrent misclassification is to treat every LDP on a trajectory space as Level 3. The slow–fast SPDE paper “Large deviations for slow-fast stochastic partial differential equations” proves a Freidlin–Wentzell path-space LDP for the slow component

$L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$7

on $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$8 with speed $L_n^\infty=\frac1n\sum_{i=1}^n \delta_{T^i(X)},$9 and rate function determined by the skeleton equation

$S^{\mathbb N}$0

but it is explicit that this is not a Donsker–Varadhan Level-3 LDP: the random object is the path of the slow component over a fixed finite interval, not an empirical process, occupation law, or stationary path measure (Wang et al., 2010).

The random interlacements paper proves an LDP for the occupation-time density profile

$S^{\mathbb N}$1

on $S^{\mathbb N}$2 with speed $S^{\mathbb N}$3 and rate function $S^{\mathbb N}$4, and explicitly classifies it as an occupation-measure or density-profile LDP, closer to Level 2 than to a true Level-3 theorem (Li et al., 2013).

The SPDE paper on super-Brownian motion and Fleming–Viot processes proves a path-space LDP for the trajectories of the measure-valued process itself, first on $S^{\mathbb N}$5 and then by contraction on $S^{\mathbb N}$6 or $S^{\mathbb N}$7. It describes this as a process-level/path-space LDP for measure-valued trajectories, not as a classical empirical-process Level-3 theorem (Fatheddin et al., 2012).

The heavy-tailed Markov renewal paper proves a full LDP for the pair $S^{\mathbb N}$8, where $S^{\mathbb N}$9 is an empirical measure on $T$0 and $T$1 is the empirical jump flow. It is described as a Level-2.5 analogue, or a generalized Level-3-type result for semi-Markov dynamics, rather than a path-space Donsker–Varadhan Level-3 LDP (Mariani et al., 2012).

The cube-projection paper proves an LDP for the sequence of random projected laws $T$2 on $T$3 with speed $T$4 and explicit good rate function

$T$5

and describes the theorem as “Level-3-type” because the random variable is itself a probability measure, while simultaneously noting that it is not Level 3 in the classical Sanov/Donsker–Varadhan sense since the randomness comes from a random geometric environment rather than from an empirical process (Johnston et al., 2021). This suggests a structural extension of Level-3 language to measure-valued deviations, but only by analogy.

A different source of ambiguity is unrelated terminology: in the rough-path paper “Large and moderate deviation for rough slow-fast system with level 3 geometric rough path,” “level 3” refers to the rough-path lift of the driver, not to the Donsker–Varadhan hierarchy (Yang et al., 23 Sep 2025). The LDP there is Freidlin–Wentzell-type for rough differential equations and slow–fast averaging, with rate function

$T$6

so the phrase “level 3” belongs to rough-path regularity rather than to the level of the large deviation statement.

6. Structural themes, contractions, and scope

A genuine Level-3 theorem controls enough information to recover lower-level statements by contraction. For Hawkes processes, continuity of

$T$7

in the strengthened weak topology yields the Level-1 LDP for $T$8 with

$T$9

(Zhu, 2011). For shift spaces, the contraction

$\Omega=A^{\mathbb N}$0

recovers Level-1 LDPs for Birkhoff averages $\Omega=A^{\mathbb N}$1 (Cuneo et al., 2017). For interacting Brownian motions, the process-level object projects to empirical density and empirical density of colors, and the whole proof architecture is built around these finite-dimensional projections (Seo, 2015).

The effective domain of a Level-3 rate function is often a structural invariant of the dynamics. In shift spaces, $\Omega=A^{\mathbb N}$2 off $\Omega=A^{\mathbb N}$3 (Cuneo et al., 2017). In the general discrete-state Markov-chain theorem, finiteness is restricted to the admissible shift-invariant set $\Omega=A^{\mathbb N}$4, and this admissibility constraint is exactly what can make the true rate function nonconvex (Daures, 15 Jul 2025). In the white-forced 2D Navier–Stokes theorem, the rate is finite only on shift-invariant path laws and is uniform over exponentially square-integrable initial distributions (Zhao, 17 Jun 2025).

The major dividing lines in the subject are therefore not only between Level 1, 2, and 3, but also between stationary and nonequilibrium settings, irreducible and reducible dynamics, full and weak LDPs, and classical entropy-rate formulas versus admissibility-constrained or control-theoretic actions. The literature surveyed here shows that the term “Level-3 Large Deviation Principle” remains technically precise when reserved for empirical-process or path-measure LDPs, while adjacent path-space and measure-valued results require explicit qualification to avoid conflating Donsker–Varadhan Level 3 with other large-deviation regimes (Wang et al., 2010, Daures, 15 Jul 2025, Zhao, 17 Jun 2025).