· 8,007 chars · 15 deps · depth 12 Reason: First publication of the proof: convexity along a segment for the local-to-global claim, and a supporting hyperplane to a bounded closed convex piece of the epigraph for nonemptiness.
The local-to-global claim follows by convexity along the segment from the point. For nonemptiness, a bounded closed convex piece of the epigraph is projected from points just below the graph; the normalised normals subconverge to a normal whose last coordinate is negative, and dividing by it produces a subgradient on a ball.
Claim 1. Let z∈U. If z=y the asserted inequality reads f(y)≥f(y), which holds. So assume z=y, and let θ be the smaller of 1 and r/∥z−y∥, a real number with 0<θ≤1. Put w=θz+(1−θ)y=y+θ(z−y), which lies in U because U is convex. By claim 5 of Elementary Properties of the Euclidean Norm on Rn, ∥w−y∥=θ∥z−y∥≤r, so w∈Bˉ(y,r) and the hypothesis gives
f(w)≥f(y)+p⋅(w−y)=f(y)+θp⋅(z−y).
On the other hand f is convex on U, so f(w)≤θf(z)+(1−θ)f(y). Combining and subtracting f(y),
K is nonempty:f(y)≤f(y) and f(y)≤c, so ι(y,f(y))∈K.
K is bounded: let ι(z,t)∈K. Then ∥z∥≤∥y∥+r by claim 6 of Elementary Properties of the Euclidean Norm on Rn, and by (L) f(y)−Mr≤f(z)≤t≤c, so ∣t∣≤∣f(y)−Mr∣+∣c∣. Hence ∥ι(z,t)∥2=∥z∥2+t2 is bounded above by a number independent of the point, and K is bounded.
The point θz1+(1−θ)z2 lies in Bˉ(y,r) by claim 2 of Euclidean Balls are Convex; the number θt1+(1−θ)t2 is at most c; and since f is convex on U⊇Bˉ(y,r),
Projections from below the graph. For k∈N put ξk=ι(y,f(y)−1/k). Since ι is injective, ξk∈K would force f(y)≤f(y)−1/k, which is false; so ξk∈/K. By claim 1 of Nearest-Point Projection onto a Nonempty Closed Convex Subset of Euclidean Space, applied to the nonempty closed convex set K, the nearest point πK(ξk)∈K is defined; write πK(ξk)=ι(zk,tk) and put vk=ξk−πK(ξk), which is nonzero because ξk∈/K while πK(ξk)∈K. By claim 2 of that lemma, vk⋅(w−πK(ξk))≤0 for every w∈K; multiplying by the positive number 1/∥vk∥ and putting uk=vk/∥vk∥, a point with ∥uk∥=1 by claim 5 of Elementary Properties of the Euclidean Norm on Rn, we get
The points uk all lie in Bˉ(0,1), which is bounded, so Bolzano-Weierstrass Theorem in Euclidean Space provides u∈Rn+1 and a strictly increasing sequence (pl)l∈N in N such that (upl)l∈N converges to u. Choose l0 with ∥upl0−u∥<1/2; then by claim 6 of Elementary Properties of the Euclidean Norm on Rn, 1=∥upl0∥≤∥u∥+∥upl0−u∥, so ∥u∥≥1/2 and in particular u=0. Write u=ι(a,b) with a∈Rn and b∈R.
Fix w∈K. For every l, by (N) and the dot product identity,
The last coordinate is negative. Taking z=y and t=f(y)+Mr+1=c in (H), which is legitimate since f(y)≤c, gives b(Mr+1)≤0, so b≤0 because 0<Mr+1. Suppose b=0. Then ∥u∥2=∥a∥2, so a=0. For z∈Bˉ(y,r) we have, by (L), f(z)≤f(y)+M∥z−y∥≤f(y)+Mr<c, so ι(z,f(z))∈K and (H) gives a⋅(z−y)≤0. Taking z=y+ra/∥a∥, which lies in Bˉ(y,r) because ∥z−y∥=r, we get a⋅(z−y)=r∥a∥2/∥a∥=r∥a∥>0, a contradiction. Hence b<0.
Conclusion. Put p=(1/(−b))a. For z∈Bˉ(y,r) we have ι(z,f(z))∈K as just shown, so (H) gives a⋅(z−y)+b(f(z)−f(y))≤0, that is a⋅(z−y)≤(−b)(f(z)−f(y)). Dividing by the positive number −b and using Bilinearity and Symmetry of the Dot Product on Rn,