---
title: Lasry–Lions Double Envelope
url: https://www.emergentmind.com/topics/lasry-lions-double-envelope
type: topic
---

# Lasry–Lions Double Envelope

The **Lasry–Lions double envelope** denotes a two-sided regularization mechanism built from alternating infimal and supremal envelope operations. In its classical form on a Hilbert space, it packages the Lasry–Lions inf/sup-convolution regularization into a symmetric operator
\[
R_t := \check T_t \circ T_{2t} \circ \check T_t,
\]
where
\[
T_t u(x) := \inf_{y\in H}\left(u(y)+\frac{|y-x|^2}{t}\right), \qquad \check T_t u(x) := \sup_{y\in H}\left(u(y)-\frac{|y-x|^2}{t}\right).
\]
This operator yields explicit \(C^{1,1}\) approximants, preserves order through a pinching principle, and provides a direct proof of Ilmanen’s insertion lemma [1001.3129]. The expression “Lasry–Lions double envelope” is also used more broadly for structurally analogous two-sided envelope constructions in price-formation free boundary models, Hamilton–Jacobi theory, Riemannian and Lorentzian geometry, viscosity methods for fully nonlinear PDE, and nonconvex optimization [1105.1121, 1509.03609, 1401.5053, 2606.23976, 1903.06644, 2103.08533].

## 1. Classical operator and \(C^{1,1}\) regularization

In the Hilbert-space setting of bounded functions \(u:H\to\mathbb R\), the basic Lasry–Lions envelopes are the inf-convolution \(T_t\) and sup-convolution \(\check T_t\). They satisfy
\[
T_t(-u) = -\check T_t(u),
\]
and form semigroups:
\[
T_t\circ T_s = T_{t+s}, \qquad \check T_t\circ \check T_s = \check T_{t+s}.
\]
The symmetric double envelope emphasized in "Lasry-Lions regularization and a Lemma of Ilmanen" [1001.3129] is
\[
R_t f := \check T_t\bigl(T_{2t}(\check T_t f)\bigr).
\]

The classical Lasry–Lions theorem recorded there states that if \(u\) is bounded, then for \(0<s<t\), the function
\[
\check T_s\circ T_t u
\]
is \(C^{1,1}\); if \(u\) is uniformly continuous, then \(\check T_s\circ T_t u\to u\) uniformly as \(t\to 0\) [1001.3129]. The paper’s reformulation replaces the asymmetric two-parameter expression \(\check T_s\circ T_t\) by the symmetric operator \(R_t\), whose principal advantage is a transparent comparison principle.

The regularity mechanism is expressed in terms of semiconcavity and semiconvexity. A function \(u:H\to\mathbb R\) is \(k\)-semiconcave if
\[
x\mapsto u(x)-\frac{|x|^2}{k}
\]
is concave, and \(k\)-semiconvex if \(-u\) is \(k\)-semiconcave. The relevant characterization is that a bounded function \(u\) is \(t\)-semiconcave iff it lies in the image of \(T_t\), while a continuous function is \(C^{1,1}\) iff it is both semiconcave and semiconvex [1001.3129]. A key lemma states that if \(u\) is continuous and both \(k\)-semiconcave and \(k\)-semiconvex, then \(u\in C^{1,1}\), and the gradient is Lipschitz with constant \(6/k\) [1001.3129]. Accordingly, the double envelope regularizes by forcing both one-sided second-order controls simultaneously.

The fundamental theorem is that for every bounded \(f\) and every \(t>0\),
\[
R_t f := \check T_t\circ T_{2t}\circ \check T_t f
\]
is \(C^{1,1}\) [1001.3129]. This makes the double envelope an explicit approximation-to-identity with second-order regularity.

## 2. Algebraic structure, approximation, and pinching

The elementary envelopes already encode one-sided curvature. The Hilbert-space theory proves that \(T_tu\) is \(t\)-semiconcave and \(\check T_tu\) is \(t\)-semiconvex. More sharply, if \(u\) is \(k\)-semiconcave, then for each \(t<k\),
\[
\check T_t u \text{ is } (k-t)\text{-semiconcave},
\]
and if \(u\) is \(k\)-semiconvex, then for each \(t<k\),
\[
T_t u \text{ is } (k-t)\text{-semiconvex}
\]
[1001.3129]. These facts are proved by direct convexity arguments on the quadratic penalization.

For uniformly continuous data, both \(T_tu\) and \(\check T_tu\) converge uniformly to \(u\) as \(t\to0\). More generally, if \(u\) is \(p\)-continuous with modulus
\[
p(r):=\sup\{\,u(x+re)-u(x): x\in H,\ \|e\|\le 1\,\},
\]
then there exists a nondecreasing error function \(\varepsilon(t)\to0\) such that
\[
u-\varepsilon(t)\le T_tu\le u, \qquad u\le \check T_tu\le u+\varepsilon(t),
\]
with explicit estimate
\[
\varepsilon(t)=\sup_{r>0}\left(p(r)-\frac{r^2}{t}\right)
\]
[1001.3129]. The envelopes preserve the same modulus of continuity.

A central algebraic fact is the fixed-point criterion
\[
\check T_t\circ T_t(u)\le u,
\]
with equality iff \(u\) is \(t\)-semiconvex. Similarly,
\[
T_t\circ \check T_t(v)\ge v,
\]
with equality iff \(v\) is \(t\)-semiconvex [1001.3129]. The proof uses the Legendre–Fenchel biconjugate viewpoint: \(\check T_t\circ T_t(u)+|x|^2/t\) is the convex biconjugate of \(u(x)+|x|^2/t\). This identifies the double envelope not merely as a smoothing map but as a convex-analytic projector onto functions having controlled second-order geometry.

The comparison principle most characteristic of the symmetric operator is the **pinching property**. If \(u\) and \(-v\) are \(k\)-semiconcave and
\[
u \ge f \ge v,
\]
then for every \(t<k\),
\[
u \ge R_t f \ge v.
\]
Thus a function squeezed between semiconcave barriers remains squeezed after regularization [1001.3129]. An immediate consequence is stationarity on sufficiently regular data: if \(f\) is already \(C^{1,1}\), then for \(t\) small enough,
\[
R_t f = f.
\]
This pinching statement is the mechanism by which the paper derives Ilmanen’s insertion lemma: if \(u\) and \(v\) are bounded functions on \(H\) with \(u>v\), and both \(u\) and \(-v\) are semiconcave, then there exists a \(C^{1,1}\) function \(w\) such that
\[
u>w>v
\]
[1001.3129].

## 3. Intrinsic and geometric variants

In Hamilton–Jacobi theory, the quadratic kernel is replaced by an action functional. For a Tonelli Lagrangian, the fundamental solution
\[
A_t(x,y)=\inf_{\gamma\in\Gamma_{x,y}^t}\int_0^t L(\gamma(s),\dot\gamma(s))\,ds
\]
induces the Lax–Oleinik operators
\[
\check T_t u(x):=\sup_{y\in\mathbb R^n}\{u(y)-A_t(x,y)\}, \qquad
T_t u(x):=\inf_{y\in\mathbb R^n}\{u(y)+A_t(y,x)\}.
\]
In the quadratic case, these coincide with the classical Lasry–Lions sup/inf convolutions. "Lasry-Lions, Lax-Oleinik and Generalized characteristics" [1509.03609] treats these operators as the Hamiltonian realization of the Lasry–Lions regularization. For semiconcave \(u\), \(\check T_t u\) is \(C^{1,1}\) for small \(t\), converges to \(u\) as \(t\to0^+\), and its local maximizers generate generalized characteristics. The derivative of the regularized function selects the unique element \(p_0\in D^+u(x)\) minimizing \(H(x,p)\) over the superdifferential [1509.03609].

For discounted Hamilton–Jacobi equations,
\[
\lambda u^\lambda(x) + H(x,Du^\lambda(x)) = 0,
\]
"Lasry-Lions approximations for discounted Hamilton-Jacobi equations" [1709.08676] replaces the Euclidean quadratic kernel by the fundamental solution \(\hat A_{0,t}\) of the time-dependent Lagrangian \(\hat L(t,x,v)=e^{\lambda t}L(x,v)\), and defines the intrinsic regularization
\[
\hat T_t u^\lambda(x)=\sup_{y\in \mathbb{R}^n}\{u^\lambda(y)-\hat A_{0,t}(x,y)\}.
\]
If \(u^\lambda\) is a viscosity solution, then \(\hat T_tu^\lambda\) is of class \(C^{1,1}\), converges uniformly to \(u^\lambda\) as \(t\to0^+\), and its gradients converge to the unique \(q_x\in D^+u^\lambda(x)\) satisfying
\[
H(x,q_x)=\min_{p\in D^+u^\lambda(x)} H(x,p)
\]
[1709.08676]. This preserves the selection principle familiar from the quadratic theory while encoding the Hamiltonian dynamics directly into the kernel.

The Riemannian manifold extension replaces squared norms by squared geodesic distances. For a manifold \(M\), the intrinsic double envelope is
\[
(f_{\lambda})^{\mu}(x)=\sup_{z\in M}\inf_{y\in M}\left\{f(y)+\frac{1}{2\lambda}d(z,y)^2-\frac{1}{2\mu}d(x,z)^2\right\}.
\]
Under bounded sectional curvature, positive injectivity radius, and positive convexity radius, every bounded uniformly continuous \(f:M\to\mathbb R\) can be uniformly approximated by globally \(C^{1,1}\) functions \((f_\lambda)^\mu\) for \(0<\mu<\lambda/2\) [1401.5053]. The paper formulates \(C^{1,1}(M)\) intrinsically through Lipschitz control of \(\nabla f\) under parallel transport, and proves preservation of ordering, infima, minimizers, invariance under isometries, and Lipschitz continuity; convexity is also preserved when \(K\le0\) [1401.5053]. Counterexamples show that global \(C^{1,1}\) regularization can fail without boundedness of \(f\) or bounded sectional curvature.

A Lorentzian analogue appears in "A Lorentzian Lasry-Lions regularization theorem" [2606.23976]. On a globally hyperbolic spacetime, with causal action
\[
c_t(x,y)= \begin{cases} -t^{1-p}\,\ell(x,y)^p, & y\in J^+(x),\ +\infty, & \text{otherwise}, \end{cases}
\]
the forward and backward Lax–Oleinik operators are
\[
T_tu(y)=\inf_{x\in J^-(y)}\bigl(u(x)+c_t(x,y)\bigr), \qquad
\hat T_tu(x)=\sup_{y\in J^+(x)}\bigl(u(y)-c_t(x,y)\bigr).
\]
If the forward evolution \(T_tu\) is locally semiconcave and has future-directed timelike superdifferentials near a point, then for \(t\) near \(t_0\) and sufficiently small \(s>0\), the double evolution \(\hat T_s\circ T_tu\) is \(C^{1,1}_{\mathrm{loc}}\) near that point [2606.23976]. This is the Lorentzian counterpart of the classical result \(\hat T_s\circ T_tu\in C^{1,1}\) for \(0<s<t\).

## 4. Free-boundary and price-formation interpretation

A different but related use of the term occurs in the Lasry–Lions price-formation model. In "On a price formation free boundary model by Lasry & Lions" [1105.1121], the unknown is a sign-changing density \(f(x,t)\) and a free boundary \(p(t)\), interpreted as the transaction price or equilibrium price. The model is
\[
{f_t} - {f_{xx} = \lambda(t)\bigl(\delta(x-p(t)+a)-\delta(x-p(t)-a)\bigr),
\]
with
\[
\lambda(t) = -f_x(p(t),t), \qquad f(p(t),t)=0.
\]

The paper constructs a transformed function
\[
F(x,t)= \begin{cases} \displaystyle \sum_{n=0}^\infty f^+(x+na,t), & x<p(t),\\[1.2ex]
\displaystyle -\sum_{n=0}^\infty f^-(x-na,t), & x>p(t), \end{cases}
\]
and describes this as a “double envelope” construction: on the left of the interface one stacks translated positive parts, while on the right one stacks translated negative parts [1105.1121]. The transformed function satisfies the heat equation
\[
F_t - F_{xx}=0
\]
in the distributional sense, and the free boundary becomes the zero level set of \(F\). Conversely, if the zero level set of a suitably prepared heat solution \(F\) is a smooth graph \(x=p(t)\), one reconstructs the original \(f\) from discrete differences of \(F^\pm\) [1105.1121].

This usage differs from the Hilbert-space operator \(R_t=\check T_t\circ T_{2t}\circ \check T_t\), but the structural idea is analogous: a nonlinear interface problem is encoded by a two-sided envelope transformation whose output evolves linearly. The paper proves global existence of a unique smooth solution and continuity of the price path \(p(t)\), and derives asymptotics: if the total positive and negative masses are unequal, then
\[
p(t)\sim \sqrt{t}\,q_\infty \qquad \text{as } t\to\infty,
\]
where \(q_\infty\) is determined by
\[
\operatorname{erf}(q_\infty)=\frac{M^-}{M^+};
\]
if the masses are balanced, then \(p(t)\) remains bounded and converges to a constant [1105.1121].

The same free-boundary structure appears as the large-transaction-rate limit of a Boltzmann-type kinetic model in "On a Boltzmann type price formation model" [1302.5650]. There, the unfolded quantities
\[
F^k(x,t)=\sum_{l=0}^{\infty} f^k(x+al,t),\qquad G^k(x,t)=\sum_{l=0}^{\infty} g^k(x-al,t)
\]
lead to a heat-equation potential \(\Phi=F^k-G^k\), and in the limit \(k\to\infty\) the transaction measure concentrates to
\[
\mu^\infty(x,t)=\lambda(t)\delta(x-p(t))
\]
under a compatibility condition [1302.5650]. The limiting densities are recovered by
\[
f^\infty(x,t)=\Phi^+(x,t)-\Phi^+(x+a,t), \qquad g^\infty(x,t)=\Phi^-(x,t)-\Phi^-(x-a,t),
\]
which matches the envelope viewpoint of the earlier price-formation analysis [1302.5650].

## 5. Envelope methods in nonlinear PDE and viscosity theory

The Lasry–Lions and Alvarez–Lasry–Lions envelope methodology also appears as a substitute for unavailable variational structure. In "The Brunn-Minkowski inequality for the principal eigenvalue of fully nonlinear homogeneous elliptic operators" [1903.06644], the problem is the Dirichlet eigenvalue equation
\[
F(Du,D^2u)=\lambda |u|^\alpha u \quad \text{in }\Omega, \qquad u=0 \quad \text{on }\partial\Omega,
\]
for fully nonlinear homogeneous elliptic \(F\). Because there is no variational characterization of \(\lambda(\Omega)\) in general, the proof uses the transformation \(v=-\log u\), the operator
\[
G(\xi,X):=-F\big(\xi,\ \xi\otimes \xi - X\big),
\]
and the weighted infimal convolution
\[
(v_0\#\cdots\# v_k)_t(x) := \inf\left\{ \sum_{i=0}^k t_i v_i(x_i) :\ x=\sum_{i=0}^k t_i x_i \right\}
\]
on Minkowski sums of domains [1903.06644].

In that paper, “double-envelope” refers to the Lasry–Lions / Alvarez–Lasry–Lions envelope mechanism in two related senses: the infimal convolution envelope used for the Brunn–Minkowski inequality, and the convex envelope \(v^{**}\) used for log-concavity of eigenfunctions [1903.06644]. The crucial technical input is viscosity stability under the envelope, encoded by second-order subjet transfer inequalities. This construction yields a viscosity supersolution on the Minkowski sum domain, from which the maximum principle gives the eigenvalue inequality
\[
\lambda\big((1-t)\Omega_0+t\Omega_1\big)^{-1/(\alpha+2)} \ge (1-t)\lambda(\Omega_0)^{-1/(\alpha+2)} +t\lambda(\Omega_1)^{-1/(\alpha+2)}
\]
[1903.06644]. Under stronger assumptions, the convex envelope argument shows that positive eigenfunctions are log-concave [1903.06644].

This usage broadens the notion of Lasry–Lions double envelope from explicit smoothing of a single function to a PDE mechanism that propagates differential inequalities through envelope constructions. A plausible implication is that the essential content of the method is not tied to a particular formula such as \(R_t\), but to the interaction between one-sided regularization, convexification, and comparison principles.

A common source of terminological confusion is the appearance of **Lasry–Lions monotonicity** in mean field game theory. "Mean Field Game Master Equations with Anti-monotonicity Conditions" [2201.10762] concerns the sign of monotonicity inequalities in master equations and explicitly states that it does **not** use, develop, or rely on the Lasry–Lions double envelope construction. The relation is therefore nominal rather than methodological.

## 6. Nonconvex optimization and homotopy formulations

In optimization, the double envelope is a smooth surrogate for nonsmooth, nonconvex terms. For a proper lsc function \(h\), the Moreau envelope is
\[
h^\lambda(x)\coloneqq \inf_w\left\{h(w)+\frac{1}{2\lambda}\|w-x\|^2\right\},
\]
and the Lasry–Lions double envelope is
\[
h^{\lambda,\mu}(x)\coloneqq -(-h^\lambda)^\mu(x)
=\sup_w\left\{h^\lambda(w)-\frac{1}{2\mu}\|w-x\|^2\right\},
\qquad 0<\mu\le \lambda.
\]
For smoothness, one typically requires \(0<\mu<\lambda<\gamma_h\), where \(\gamma_h\) is the prox-boundedness threshold of \(h\) [2103.08533].

The envelope satisfies the bridging inequalities
\[
h^\lambda \le h^{\lambda,\mu}\le h^{\lambda-\mu}\le h.
\]
Accordingly, it interpolates between a Moreau-type lower relaxation and the original nonconvex function [2103.08533]. For \(0<\mu<\lambda<\gamma_h\), one has
\[
h^{\lambda,\mu}\in C^{1,1},
\]
with Lipschitz gradient. The paper also gives the stationarity relation
\[
d=\nabla h^{\lambda,\mu}(x) \quad\Longleftrightarrow\quad x-(\lambda-\mu)d\in \conv\prox_{\lambda h}(x+\mu d)
\]
[2103.08533]. These properties underlie a homotopy strategy for composite minimization
\[
\min_{x\in\mathbb{R}^n}\ \varphi(x)=g(x)+h(x),
\]
where \(g\) is \(C^1\) and \(h\) is proper, lsc, prox-bounded, and possibly nonsmooth and nonconvex. One solves instead
\[
\min_x\ \varphi_{\lambda,\mu}(x)\coloneqq g(x)+h^{\lambda,\mu}(x),
\]
along a continuation path \((\lambda_k,\mu_k)\to(0,0)\), with epi-convergence guaranteeing convergence of minima and cluster points of approximate minimizers [2103.08533]. The paper reports applications to signal decoding and spectral unmixing, using L-BFGS as the inner smooth solver [2103.08533].

"A Lasry-Lions envelope approach for mathematical programs with complementarity constraints" [2507.04694] applies the same principle to the indicator \(\delta_D\) of the complementarity set
\[
D:=\{z\in\mathbb{R}_+^2 \mid z_1 z_2=0\}.
\]
The Moreau envelope is
\[
\env_{\lambda}\varphi(x) :=\inf_{x'\in\mathbb{R}^n}\left\{\varphi(x')+\frac{1}{2\lambda}\|x'-x\|^2\right\},
\]
and the Lasry–Lions double envelope is
\[
\env_{\lambda,\mu}\varphi(x) := -\env_\mu\bigl(-\env_\lambda\varphi\bigr)(x)
= \sup_{x'\in\mathbb{R}^n}\left\{ \env_\lambda\varphi(x') -\frac{1}{2\mu}\|x'-x\|^2 \right\},
\qquad \lambda>\mu>0.
\]
For prox-bounded \(\varphi\), \(\env_{\lambda,\mu}\varphi\) is smooth with Lipschitz gradient, and
\[
\operatorname{Lip}(\nabla \env_{\lambda,\mu}\varphi) \le \max\Bigl\{\frac1\mu,\frac1{\lambda-\mu}\Bigr\}
\]
[2507.04694]. For the complementarity indicator, the resulting surrogate is everywhere smooth, unlike the standard Moreau envelope of \(\delta_D\), which is only piecewise smooth and nonsmooth along the diagonal in the positive orthant [2507.04694].

The smoothed MPCC subproblem is
\[
\min_{x\in C}\; f(x)+\sum_{i=1}^p \env_{\lambda,\mu}\delta_D(F_i(x)),
\]
or equivalently
\[
\min_{x\in C} s_{\lambda,\beta}(x), \qquad
s_{\lambda,\beta}(x)=f(x)+\frac1\lambda\sum_{i=1}^p r_\beta(F_i(x)),
\]
with \(\beta=\mu/\lambda\in(0,1)\) [2507.04694]. The method solves a sequence of smooth subproblems while decreasing \(\lambda^\nu\searrow0\). Under BCCQ, cluster points are C-stationary; under well-behaved complementarity constraints and BCQ\(^+\), they are M-stationary [2507.04694]. The paper also provides a worst-case gradient-evaluation complexity of
\[
\mathcal O\!\left(\epsilon^{-4}\log\!\left(\epsilon^{-2}\right)\right)
\]
for computing an \(\epsilon\)-approximate C-stationary point [2507.04694].

Across these optimization papers, the double envelope is no longer mainly a proof device for semiconcavity theory. It becomes a tunable smooth approximation that preserves minimizers, epi-converges to the original objective, and supports continuation methods on nonconvex landscapes. This suggests a modern reinterpretation of the Lasry–Lions double envelope as a general-purpose bridge between exact nonsmooth models and tractable smooth surrogates.

Source: https://www.emergentmind.com/topics/lasry-lions-double-envelope