---
title: Local Large Deviation Principle
url: https://www.emergentmind.com/topics/local-large-deviation-principle
type: topic
---

# Local Large Deviation Principle

Searching arXiv for recent and foundational papers on local large deviation principles and closely related formulations.
Search 1: recent papers explicitly on "local large deviation principle".
Search 2: general random-process LLDP framework and necessity/sufficiency criteria.
Search 3: representative applications across stochastic processes, SPDEs, graphs, and dynamics.
The local large deviation principle (LLDP) is a localized form of large deviation theory that identifies exponential asymptotics for probabilities of shrinking neighborhoods around fixed points, increments, paths, or empirical objects, rather than for arbitrary open or closed sets in a whole state space. In a basic formulation for random vectors \(\{\zeta_T\}_{T\ge 0}\subset\mathbb R^d\), the LLDP requires that
\[
\lim_{T\to\infty}T^{-1}\ln \mathbf P(|\zeta_T-\alpha|<\varepsilon_T)=-D(\alpha)
\]
for every \(\alpha\in\mathbb R^d\) and for \(\varepsilon_T\to 0\) slowly enough, where \(D\) is the local rate function [2604.22257]. Recent work extends this pointwise formulation to conditional increments, finite-dimensional distributions, and full trajectories of general stochastic processes, using conditional logarithmic moment generating function asymptotics and Legendre transforms in a Gärtner–Ellis-type framework [2604.27485].

## 1. Definition and formal structure

For a family \(\{\zeta_T\}_{T\ge 0}\) of random vectors in \(\mathbb R^d\), the LLDP is defined by the pair of limits
\[
\lim_{\varepsilon \searrow 0}\liminf_{T\to\infty}\frac1{T}\ln \mathbf P(\zeta_T \in (\alpha)_\varepsilon) = - D(\alpha),
\]
\[
\lim_{\varepsilon \searrow 0}\limsup_{T\to\infty}\frac1{T}\ln \mathbf P(\zeta_T \in (\alpha)_\varepsilon) = - D(\alpha),
\]
for every \(\alpha\in \mathbb R^d\), where \((\alpha)_\varepsilon=\{\beta:|\beta-\alpha|<\varepsilon\}\) and \(D:\mathbb R^d\to[0,\infty]\), \(D\not\equiv\infty\) [2604.22257]. An equivalent formulation uses any \(\varepsilon_T\to 0\) “slowly enough,” meaning that there exists some \(\widetilde\varepsilon_T\to 0\) such that the statement holds for \(\widetilde\varepsilon_T\), and then holds for every \(\varepsilon_T\to 0\) with \(\varepsilon_T\ge c\,\widetilde\varepsilon_T\) for some \(c>0\) [2604.22257].

This is a local statement in the precise sense that it concerns shrinking neighborhoods of a prescribed point. It is therefore weaker than a full LDP over arbitrary Borel sets unless additional tail control is available. Once the LLDP holds, the rate function \(D\) is automatically lower semicontinuous [2604.22257].

The same localization idea appears in path space. For a rescaled process \(z_T(s)=Z(sT)/T\) on \([0,1]\), one studies probabilities of the form \(\mathbf P(z_T\in (f)_{\varepsilon_T})\), where \((f)_\varepsilon\) is the \(\varepsilon\)-neighborhood of \(f\) in a chosen metric, often the uniform metric on \(\mathbb D[0,1]\) [2604.27485]. In that setting, the LLDP describes the exponential cost of following a prescribed trajectory rather than merely reaching a prescribed terminal value.

## 2. Truncated cumulant functions and the relaxed Gärtner–Ellis paradigm

A central structural result is that the correct local analogue of the scaled cumulant generating function is not the full moment generating function, but a truncated one. If the LLDP holds for \(\{\zeta_T\}\) with rate function \(D\), then for \(M_T\to\infty\) slowly enough the limit
\[
A(\mu):=\lim_{T\to\infty}\frac1T\ln \mathbf E\!\left(e^{T\langle \mu,\zeta_T\rangle};\ |\zeta_T|\le M_T\right)
\]
exists for every \(\mu\in\mathbb R^d\), and equals the Legendre–Fenchel transform
\[
A(\mu)=\mathcal L_D(\mu):=\sup_{\alpha\in\mathbb R^d}\bigl(\langle \mu,\alpha\rangle-D(\alpha)\bigr)
\]
[2604.22257]. Conversely, if this truncated limit exists and \(A\) is essentially smooth, then the LLDP holds with rate function
\[
D=\mathcal L_A.
\]
The result is described as a “relaxed version” of the Gärtner–Ellis theorem because it avoids the restrictive exponential integrability assumptions required for the full cumulant generating function [2604.22257].

The truncation is the decisive feature. It ignores remote tails, which are irrelevant for local probabilities but can make \(\mathbf E(e^{T\langle\mu,\zeta_T\rangle})\) infinite. One example given in the literature has
\[
\mathbf E(e^{T\mu \zeta_T})=\infty \quad \text{for all } \mu\neq 0,
\]
yet the LLDP still holds, and even the LDP holds because the tail is exponentially tight [2604.22257]. The local theory is therefore compatible with distributions that lie outside the reach of the classical Gärtner–Ellis theorem.

The regularity condition is the standard convex-analysis notion of essential smoothness. A convex function \(f\) is essentially smooth if \(\operatorname{int}(\operatorname{dom}f)\neq\varnothing\), \(f\) is differentiable on \(\operatorname{int}(\operatorname{dom}f)\), and \(|f(\mu_k)|\to\infty\) whenever \(\mu_k\in \operatorname{int}(\operatorname{dom}f)\) converges to a boundary point of \(\operatorname{dom}f\) [2604.22257]. Unlike the classical Gärtner–Ellis theorem, the sufficient condition for the LLDP does not require \(0\in\operatorname{int}(\operatorname{dom}A)\), because the local problem does not require a global upper bound over unbounded sets [2604.22257].

A further distinction from classical convex large deviation theory is that an LLDP rate function need not be convex. When \(D\) is non-convex, \(A\) can still exist, but without essential smoothness one may have \(D\neq \mathcal L_A\); in that case \(\mathcal L_A\) is the largest convex lower semicontinuous minorant of \(D\) [2604.22257].

## 3. Conditional, finite-dimensional, and functional LLDPs for processes

A general process-level framework is developed for a real-valued stochastic process \(Z=\{Z(t):t\in\mathbb R\}\) with trajectories in \(\mathbb D(\mathbb R)\) [2604.27485]. The basic object is the rescaled variable \(z(T)=Z(T)/T\), and later the rescaled path
\[
z_T(s):=\frac{Z(sT)}{T}, \qquad s\in[0,1].
\]
The key hypothesis is a uniform asymptotic conditional logarithmic moment generating function for increments. In one formulation, for any fixed \(0\le s_1<s_2<\infty\), any \(\alpha\in\mathbb R\), and any \(\mu\in\operatorname{dom}A\),
\[
\frac1{T_2-T_1}\ln \mathbf E\!\left(e^{\mu(Z(T_2)-\alpha T)}\mid \mathcal F_{T_1}\right) = A(\mu)+o(1),
\]
uniformly over \(\omega\in C_{s_1}(T)\), where \(T_i=s_iT\) and
\[
C_{s_1}(T)=\left\{\left|\frac{Z(s_1T)}{T}-\alpha\right|<\eta_T\right\},
\qquad \eta_T\to 0
\]
[2604.27485]. The function \(A\) is assumed convex and essentially smooth, and its Legendre transform
\[
D(\alpha)=\mathcal L_A(\alpha):=\sup_{\mu\in\mathbb R}\{\alpha\mu-A(\mu)\}
\]
is the local rate function [2604.27485].

Under this assumption, the uniform conditional LLDP for increments states that there exists a sequence \(\varepsilon_T=\overline{o}(1)\) such that
\[
\lim_{T\to\infty}\frac1T \ln \mathbf P\!\left( \frac{Z(T_2)-Z(T_1)}{T_2-T_1}\in (\beta)_{\varepsilon_T} \ \bigg|\ \mathcal F_{T_1} \right) =-(s_2-s_1)D(\beta),
\]
uniformly on \(C_{s_1}(T)\), for any fixed \(0\le s_1<s_2<\infty\) and \(\alpha,\beta\in\mathbb R\) [2604.27485]. This is a genuinely conditional and local statement: the neighborhood around \(\beta\) shrinks, and the estimate is uniform over histories for which the current macroscopic state is close to \(\alpha\).

The increment LLDP extends to finite-dimensional distributions of \(z_T\). For a partition \(0=s_0<s_1<\cdots<s_K=1\) with \(h_k=s_k-s_{k-1}\), one obtains
\[
\lim_{T\to\infty}\frac1T \ln \mathbf P\!\left(\bigcap_{k=1}^K G_{k,T}(\varepsilon_T)\mid \mathcal F_0\right) = -\sum_{k=1}^K h_k D(\beta_k),
\]
uniformly on
\[
C_0(T)=\left\{\left|\frac{Z(0)}{T}-\alpha_0\right|<\eta_T\right\},
\]
where \(G_{k,T}(\varepsilon)\) is the event that the \(k\)-th increment average lies in an \(\varepsilon\)-neighborhood of \(\beta_k\) [2604.27485]. Each subinterval contributes additively to the exponent.

At path level, the state space is \(\mathbb D[0,1]\) with the uniform metric
\[
\|f\|:=\sup_{s\in[0,1]}|f(s)|.
\]
For absolutely continuous \(f\in\mathbb C_a\), the action is
\[
I(f)=\int_0^1 D(f'(s))\,ds.
\]
For general \(f\in\mathbb D\), the deviation integral is
\[
J(f):=\lim_{K\to\infty}\sum_{k=1}^K h_k\,D\!\left(\frac{f(s_k)-f(s_{k-1})}{h_k}\right),
\]
and the paper notes that it can be represented as \(J(f)=\sup I(f^{\boldsymbol s^K})\), where \(f^{\boldsymbol s^K}\) is the piecewise linear interpolation along a partition [2604.27485]. Under the conditional mgf assumption one obtains the upper bound
\[
\limsup_{T\to\infty}\frac1T \ln \mathbf P\!\left(z_T\in (f)_{\varepsilon_T}\mid \mathcal F_0\right) \le -J(f),
\]
uniformly on \(\{|Z(0)/T-f(0)|<\eta_T\}\). If an additional oscillation condition is imposed,
\[
\sup\bigl\{|Z(uT)-Z(vT)|:0\le u<v\le \min(u+\delta,1)\bigr\} \le V(T)+W(\delta)T
\]
with \(V(t)=o(t)\) and \(W(h)=o(1)\), then the matching lower bound holds and hence
\[
\lim_{T\to\infty}\frac1T \ln \mathbf P\!\left(z_T\in (f)_{\varepsilon_T}\mid \mathcal F_0\right) = -J(f)
\]
[2604.27485]. The same scheme extends to triangular arrays \(Z=Z^{(T)}\) [2604.27485].

## 4. Representative models and explicit local rate functions

The LLDP has been established in markedly different settings, and the form of the local event depends on the model.

| Setting | Local event | Rate statement |
|---|---|---|
| Inhomogeneous birth-death process | \(\xi_T(\cdot)\in U_\varepsilon(f)\) | \(v(T)=T^{l+1}\) or \(T^{m+1}\); \(I(f)=P_l\int_0^1 f(t)^l\,dt\), \((\sqrt{P_l}-\sqrt{Q_m})^2\int_0^1 f(t)^l\,dt\), or \(Q_m\int_0^1 f(t)^m\,dt\), according to the cases \(l>m\), \(l=m\), or \(l<m\) [1806.08956] |
| Wiener process with random resetting | \(\mathbf E_T\in U_\varepsilon(f)\) | speed \(T\); \(I(f)=\frac12\int_0^1 \dot f(t)^2\,dt\) on \(AC_0[0,1]\) or \(AC_0'[0,1]\) [1911.06751] |
| Typed random graph | \(P\in B_p\) | \(\mathbb P_{(\varpi_n,\omega_n)}\{P\in B_p\}=\exp\{-nJ(\varpi,\omega)(p)+o(n)\}\), with \(J(\varpi,\omega)(p)=H(p\|q)\) if \(p_1=\varpi\) and \((p,\omega)\) are sub-consistent [1708.03895] |
| Mixing Smale space with conditional Gibbs measure on \(W_\delta^u(x)\) | \(\mu_{x,G}^u\{y:\zeta_y^n\in J\}\) or \(\mu_{x,G}^u\{y:\zeta_y^n\in K\}\) | speed \(n\); \(I(\nu)=P(G)-\int G\,d\nu-h_\nu(f)\) on \(\mathcal M_f(X)\), \(I(\nu)=\infty\) otherwise [2510.00972] |

These examples illustrate that the “locality” may refer to a uniform path neighborhood, a neighborhood of an empirical profile, or a leafwise conditional measure in a hyperbolic system. In the birth-death model, the result gives rough exponential asymptotics for excursions of a rescaled trajectory near a prescribed nonnegative continuous path, with the dominant polynomial growth of the birth or death rate fixing both speed and action [1806.08956]. In the resetting model, the reset mechanism changes the proof but not the leading action: for positive or negative excursions the rate remains the Brownian action \(\frac12\int_0^1 \dot f^2\) [1911.06751].

Graph models supply two different local structures. For typed random graphs, the empirical locality measure has an LLDP with relative-entropy rate \(H(p\|q)\), where \(q(a,e)\) is an explicit product-Poisson reference law induced by the prescribed empirical type and link measures [1708.03895]. For marked SINR graphs, a spectral potential
\[
U_Q(g,\omega)=(g, e^{-R_D}\,\omega\otimes\omega)
\]
is used to derive an LLDP and then a conditional LDP for the empirical connectivity measure given the empirical marked measure, at speed \(\lambda\) [1909.04529].

Not all localized principles are full upper-and-lower LLDPs. For generalized multiple intersection local times of multidimensional Brownian motion, the estimate is a local upper bound on closed cylindrical sets:
\[
\limsup_{t\to\infty}\frac{1}{t^2}\log \theta_{t u_1,\dots,t u_{k-1}}\bigl(tC_{T,A}\bigr) \le -\inf_{\phi\in C_{T,A}\cap E_{u_1,\dots,u_{k-1}}} I(\phi),
\]
with \(I(\phi)=\frac12\int_0^1|\phi'(t)|^2\,dt\) on the Cameron–Martin space [2406.07173]. For randomly forced nonlinear wave equations with localized damping, the level-2 LDP for empirical measures has a full upper bound but a lower bound only on open sets intersected with a distinguished subset \(\mathcal W\) of equilibrium states, which is why the result is described as having a lower bound “of a local type” [2409.11717].

## 5. Relation to full large deviation principles, topology, and tightness

The LLDP is weaker than a full LDP unless supplemented by tail control. A precise relationship is available: if the LDP holds, then the LLDP holds with the same rate function; conversely, LLDP plus exponential tightness implies the LDP [2604.22257]. This equivalence explains why local asymptotics sometimes suffice to recover a global theory and sometimes do not.

In some models the LLDP is explicitly upgraded to a full LDP by standard compactness and covering arguments. For typed random graphs, the local entropy asymptotics for neighborhoods of the empirical locality law yield the full LDP for the same object, and in the Erdős–Rényi specialization the rate reduces to the entropy relative to the Poisson law \(q_c(k)=e^{-c}c^k/k!\) under the mean constraint \(\langle p\rangle=c\) [1708.03895].

In other settings the local form is intrinsic. For inhomogeneous birth-death processes, a standard full LDP in the Skorokhod space is generally unavailable except in the homogeneous case \(l=m=0\), because the family is not exponentially dense [1806.08956]. For Wiener process with random resetting, no full LDP in the Skorokhod \(J_1\) topology is claimed, because the family is not exponentially tight there; the result is instead local and based on the uniform metric [1911.06751]. For the locally damped wave equation, the lack of smoothing effect leads to the introduction of asymptotic exponential tightness, which is weaker than classical exponential tightness and supports only a local lower bound [2409.11717].

The topology of the local neighborhood is model-dependent. Euclidean neighborhoods \((\alpha)_\varepsilon\) are natural for random vectors [2604.22257]. Uniform neighborhoods in \(\mathbb D[0,1]\) are used for pathwise results in general stochastic processes, birth-death processes, and resetting models [2604.27485, 1806.08956, 1911.06751]. Weak neighborhoods of empirical measures are used in graph models [1708.03895]. In hyperbolic dynamics, the same rate function as in the global theory can appear, but with probabilities computed under conditional Gibbs measures supported on local unstable leaves rather than under the global equilibrium state [2510.00972].

## 6. Terminology, adjacent usages, and common sources of confusion

The phrase “local large deviation principle” is not used uniformly across the literature. In the canonical probabilistic sense, “local” refers to shrinking neighborhoods around a point, increment, path, or empirical measure, as in the random-vector and random-process formulations above [2604.22257, 2604.27485]. In Smale spaces, the deviation statement is local because it is taken with respect to conditional Gibbs measures on local unstable manifolds, even though the rate function remains the standard entropy-pressure functional [2510.00972].

In nearby areas, “local” may refer to the observable rather than the deviation principle. Branching Brownian motion studies the local mass \(Z_t(B_t)\), meaning the number of particles in a bounded region or moving ball, and derives large deviation asymptotics for atypically small local mass [1811.09037]. Random walk among random conductances studies local times and proves an annealed LDP for normalized occupation measures in a finite domain [1104.1548]. These are closely related subjects, but they are not LLDPs in the shrinking-neighborhood sense.

A separate source of ambiguity appears in SPDE and SDE titles involving “locally monotone,” “fully local monotone,” or “localized conditions.” In those works the established deviation result is a Freidlin–Wentzell or Wentzell–Freidlin LDP, while “local” describes the coefficient hypothesis rather than the large deviation principle. One paper states explicitly that it does **not** develop a separate “local large deviation principle” and that the word “local” refers to the monotonicity structure of the SPDE [1605.06618]. Related results cover fully local monotone coefficients, multiplicative noise, Lévy noise, and gradient-dependent noise in Gelfand-triple frameworks [2212.05257, 2402.18108, 2212.10282].

This suggests that the LLDP is best understood not as a single theorem with a universal formalism, but as a localization scheme inside large deviation theory. In one branch, it sharpens pointwise or pathwise asymptotics through shrinking neighborhoods and truncated cumulant functions [2604.22257]. In another, it supplies conditional or leafwise analogues of classical global principles [2604.27485, 2510.00972]. Across applications, its distinguishing feature is always the same: it resolves the exponential cost of being near a prescribed local configuration, rather than the cost of belonging to a broad set.

Source: https://www.emergentmind.com/topics/local-large-deviation-principle