Write for the infimal convolution. It is proper by the permitted hypothesis. If and are finite, choose within of their respective infima. For , put and . Convexity of and gives
Let . If either endpoint value is infinite, the desired inequality is automatic. Thus the infimal convolution is convex, without assuming the infimum is attained.
For its convex conjugate, replace a negative infimum by a supremum and then change variables :
The two suprema separate because and are independent. Properness of and makes each supremum strictly greater than , so this separation remains valid when one or both are . Therefore . This is the conjugate of an infimal convolution. The bounded-domain assumption is unnecessary for these two calculations; the assumed properness and lower semicontinuity of will be used when applying subgradient inversion under convex conjugacy.
Apply the local bound from condition (1) separately to and . For any and , there are such that
With , both inequalities hold when , so
This is exactly continuity at . Since was arbitrary, is continuous on : upper semicontinuity of and of is equivalent to simultaneous upper semicontinuity and lower semicontinuity of .
Variational regularization Created 2026-09-24 Updated 2026-11-05
Variational regularization reconstructs an unknown by balancing data fit with a penalty expressing desired structure. A standard model is
The linear operator predicts data, denotes observations, and the extended-real functional can impose smoothness, sparsity or hard constraints. This construction includes nonlinear reconstruction maps even when itself is linear.[ref]
For , the model is quadratic Tikhonov regularization. For , the penalty controls the total variation seminorm on a domain and can retain jumps without demanding a square-integrable classical gradient. Adding the indicator functional of a constraint set enforces constraints such as nonnegativity.
Existence, uniqueness and stability are separate conclusions. A direct method in the calculus of variations requires suitable compactness of bounded-energy sets and lower semicontinuity; convexity alone does not supply compactness. A unique minimizer at fixed also does not prove convergence to the true object as noise vanishes: the regularization parameter choice must balance bias and noise.