Can the slopes reconstruct the function?
The conjugate records the best intercept of an affine lower bound at each slope. Taking the conjugate once more asks whether the collection of all such bounds contains enough information to recover the original function. For a proper function , define
A function is proper if it is finite somewhere and never . Its epigraph is . We call closed when its epigraph is closed in ; for extended-valued functions this is equivalent to lower semicontinuity. Extended values are essential: the indicator is zero on and elsewhere, so its epigraph encodes the set as well as the objective.
Each finite gives an affine minorant . Conversely, if for every , then . Consequently is the pointwise supremum of all finite affine minorants of (the supremum of an empty family is ). Fenchel's inequality immediately gives . The reverse inequality needs a geometric assumption.
Theorem 5.5.1
(Finite-dimensional Fenchel–Moreau theorem)
If is proper, convex and lower semicontinuous, then at every point, including points where . Conversely, any proper function satisfying is convex and lower semicontinuous.
Proof
(Separate a point below the epigraph)
The epigraph is nonempty, closed and convex. First we establish a strict affine separator with a negative coefficient on the last coordinate for every point outside . By projection onto a nonempty closed convex set, the closest point exists and satisfies for all . Thus, for and , we have . Because contains every vertical ray above each of its points, .
If , choose with and apply the same projection argument to . Its separator must have : comparing the strict separation with the epigraph point forces it. Add a sufficiently small positive multiple of this second separator and its bound to the first. The resulting separator remains strict at , is valid on , and now has .
Writing the resulting inequality as for and dividing by gives an affine minorant
Fix . For every real , including every real when , the construction gives an affine minorant exceeding at . Hence in the extended sense. Together with , this proves equality. Conversely, is a supremum of continuous affine functions, hence convex and lower semicontinuous: its epigraph is an intersection of closed half-spaces. This proves the converse.
After the proof, it is worth isolating what each step contributes.
Proof X-ray
Why the vertical coefficient matters
Closedness and convexity of are used exactly once, to obtain the projection and hence a strict separator; without either, a point below the graph may have no separator (the two failures in the counterexample below). Properness supplies the point with . The non-obvious move is the perturbation when : a separating hyperplane with zero vertical coefficient only separates in the -direction and gives no finite lower bound on function values. Adding a small multiple of the separator at , whose vertical coefficient is forced negative, keeps the strict inequality at while making . This is what handles boundary points and points outside the domain, not merely points where a subgradient exists. The argument closes by letting increase to , or to every real number when .
What the second conjugate keeps
The same argument identifies for functions that are neither convex nor closed.
Proposition 5.5.2
(The biconjugate is the closed convex envelope)
Suppose a proper has at least one finite affine minorant. Then is proper, closed and convex, and it is the greatest closed convex minorant of : every proper closed convex with satisfies . Without an affine minorant, and .
Proof
(Compare affine minorants)
Being a supremum of affine functions, is closed and convex; it is bounded above by , so it is finite where is, and it exceeds the given affine minorant everywhere, so it never takes . If is proper, closed and convex with , every affine minorant of is also one of , and the theorem gives . If has no finite affine minorant, then for every , and the supremum defining is . Properness of alone does not rule this out.
Worked example 5.5.3
(Closing an open constraint)
Let on . For the supremum of over is , not attained; for it is . Thus . If , then for every , with equality at . If , taking makes the expression tend to ; if , take . Hence . The two missing endpoints are filled in by closure; the supremum in the first conjugate need not be attained.
Worked example 5.5.4
(The convex envelope of a double well)
Let , which is finite and continuous but not convex. The conjugate of a minimum is the maximum of the conjugates, since . Completing squares gives , so . Then
If , then with equality at , so . If , only can help, and is maximized at with value ; symmetrically for . Hence for and on : the bump between the wells is replaced by the flat segment joining the two minima, as the proposition predicts.
Worked example 5.5.5
(Best affine minorant of a given slope)
Let . Completing the square gives , attained at , so the best intercept for slope is and the minorant is . At this yields , while the actual value is . The bound is not sharp there: equality for slope occurs only at . Choosing slope instead gives and the tangent , which is sharp at . Taking the supremum over all slopes recovers , an instance of .
These examples suggest that both hypotheses of the theorem are doing work.
Counterexample mode
Convexity and lower semicontinuity both matter
The open-interval indicator is convex but not lower semicontinuous, and . For a different failure, let . Its conjugate is , so , and . Here the epigraph is closed but not convex. Finally is proper and continuous but has no affine minorant: is unbounded above for every , so and . None of these failures contradicts the theorem.
Common mistake
(Closed does not mean continuous)
Lower semicontinuity is a one-sided condition: . The indicator jumps to at its endpoints, yet its epigraph is closed, so it satisfies the theorem and equals its biconjugate. Conversely, a jump upward at a boundary point, as in the exercise below with , violates lower semicontinuity, and the biconjugate lowers that value.
Exercises with solutions
Quick check
(Compute the biconjugate of .)
First find for which slopes is finite.
Solution · Answer
For , ; for the supremum is . Thus . Its conjugate is , which is for and for . Therefore .
Quick check
(Let for , , and for . Find .)
Compare with a closed convex function lying below it.
Solution · Answer
The function for and for is proper, closed and convex, and . Therefore . At and this already gives equality with . At zero, any affine minorant of satisfies ; the zero minorant attains this value. Thus , and .
Quick check
(Show for every proper .)
Use for one inequality and the definition of for the other.
Solution · Answer
From , reversing the inequality under conjugation gives . For the opposite inequality, fix with . The defining supremum gives for every , so . At a point where this inequality is automatic. This also covers , where and . So a third conjugate never needs to be computed.