Moreau Envelope

Smooth a proper lower-semicontinuous convex function by infimizing its value plus a quadratic distance penalty, linking nonsmooth optimization to the proximal map.

Version

v2 · 2026-09-06 · History

Domain-specific #

2312

Origin domain

mathematics

Subdomain

convex analysis

Aliases

Moreau-Yosida regularization, Moreau regularization, Moreau approximation

Core Idea

For a proper lower-semicontinuous convex function \(f\) on a Hilbert space and \(\lambda>0\), the Moreau envelope is

\[ e_\lambda f(x)=\inf_y\left\{f(y)+\frac{1}{2\lambda}\lVert x-y\rVert^2\right\}. \]

It replaces the value at \(x\) by the best tradeoff between a low value of \(f\) and staying near \(x\). This is the infimal convolution of \(f\) with a scaled squared norm. Under the standard convex assumptions, the minimizer is unique and is the proximal point \(\operatorname{prox}_{\lambda f}(x)\); the envelope is continuously differentiable even when \(f\) is not.

Scope of Application

The envelope is foundational in convex analysis, monotone operators, variational inequalities, nonsmooth optimization, splitting methods, stochastic optimization, and Hamilton–Jacobi theory. Indicator functions recover squared distance to a closed convex set, and absolute-value or norm penalties yield explicit piecewise-smooth envelopes.

Generalizations replace the squared Hilbert norm with Bregman divergences, variable metrics, or nonconvex prox-regular functions. These are related envelopes, but their guarantees must be restated rather than inherited silently.

Clarity

Parameter conventions vary. Some write the envelope of \(\lambda f\) with penalty \(\|x-y\|^2/2\); others write the envelope of \(f\) with \(1/(2\lambda)\). Formulas for gradients and conjugates must match the declared convention.

The envelope lies below \(f\), because \(y=x\) is feasible. Larger \(\lambda\) weakens the proximity penalty and increases smoothing; smaller \(\lambda\) keeps the proximal point near \(x\) and makes the envelope closer to \(f\).

Manages Complexity

The construction packages a nonsmooth subdifferential into a Lipschitz gradient and replaces local kink handling with a proximal subproblem. Resolvent and conjugacy identities allow the analyst to switch among primal, dual, and monotone-operator views.

The cost is displaced: evaluating the gradient requires computing a proximal point, which may itself be hard. Smoothing can also worsen conditioning as \(\lambda\) becomes small and can blur features relevant outside minimization.

Abstract Reasoning

1. Verify properness, lower semicontinuity, convexity, and Hilbert-space setting.
2. Fix and declare the parameter convention.
3. Form the quadratic-penalized inner objective.
4. Prove existence and uniqueness of its minimizer.
5. Identify the proximal map and calculate the envelope.
6. Derive the gradient from the proximal residual.
7. Check minimizer and minimum-value preservation.
8. Use conjugacy or Moreau decomposition for dual analysis.
9. For nonconvex/generalized variants, re-establish finiteness, single-valuedness, and smoothness.
10. Include proximal-evaluation cost in algorithmic claims.

Knowledge Transfer

The portable idea is to regularize a rough objective by allowing a nearby surrogate point and charging for displacement. The proposed parent is Regularization; Moreau’s particular penalty-and-infimum construction is an optimization-specific specialization.

Relationships to Other Abstractions

Current abstraction Moreau Envelope Domain-specific

Parents (1) — more general patterns this builds on

• Moreau Envelope is a kind of Regularization Prime

  Regularization is the proposed immediate parent.

Hierarchy path (1) — routes to 1 parentless root

• Moreau Envelope → Regularization → Optimization

Neighborhood in Abstraction Space

Moreau Envelope sits in a sparse region of the domain-specific corpus (81^st percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Fredholm Kernel — 0.83
• Proximal operator — 0.82
• Conjugate Gradient Method — 0.81
• Strictly Singular Operator — 0.81
• Compact Operator — 0.81

Computed from structural-signature embeddings · 2026-09-08