TheoremBase

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.

Proof

We use the notation of the statement. Throughout we use Concatenation Identifies a Product of Euclidean Spaces with a Euclidean Space, by which ι\iota is a linear bijection, and claim 3 of that lemma, by which

ι(ξ,η)⋅ι(ξ′,η′)=ξ⋅ξ′+η η′and∥ι(ξ,η)∥2=∥ξ∥2+η2\iota(\xi,\eta)\cdot\iota(\xi',\eta')=\xi\cdot\xi'+\eta\,\eta'\qquad\text{and}\qquad\lVert\iota(\xi,\eta)\rVert^{2}=\lVert\xi\rVert^{2}+\eta^{2}

for ξ,ξ′∈Rn\xi,\xi'\in\mathbb{R}^{n} and η,η′∈R\eta,\eta'\in\mathbb{R}. Algebraic manipulations of dot products use The Dot Product on Rn\mathbb{R}^n Is Symmetric, Bilinear and Positive Definite, and ∥v∥2=v⋅v\lVert v\rVert^{2}=v\cdot v is claim 1 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square.

Claim 1. Let z∈Uz\in U. If z=yz=y the asserted inequality reads f(y)≥f(y)f(y)\ge f(y), which holds. So assume z≠yz\neq y, and let θ\theta be the smaller of 11 and r/∥z−y∥r/\lVert z-y\rVert, a real number with 0<θ≤10<\theta\le1. Put w=θz+(1−θ)y=y+θ(z−y)w=\theta z+(1-\theta)y=y+\theta(z-y), which lies in UU because UU is convex. By claim 5 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square, ∥w−y∥=θ∥z−y∥≤r\lVert w-y\rVert=\theta\lVert z-y\rVert\le r, so w∈Bˉ(y,r)w\in\bar{B}(y,r) and the hypothesis gives

f(w)≥f(y)+p⋅(w−y)=f(y)+θ p⋅(z−y).f(w)\ge f(y)+p\cdot(w-y)=f(y)+\theta\,p\cdot(z-y).

On the other hand ff is convex on UU, so f(w)≤θf(z)+(1−θ)f(y)f(w)\le\theta f(z)+(1-\theta)f(y). Combining and subtracting f(y)f(y),

θ p⋅(z−y)≤θ(f(z)−f(y)),\theta\,p\cdot(z-y)\le\theta\bigl(f(z)-f(y)\bigr),

and dividing by the positive number θ\theta gives f(z)≥f(y)+p⋅(z−y)f(z)\ge f(y)+p\cdot(z-y). As z∈Uz\in U was arbitrary, p∈∂Uf(y)p\in\partial_{U}f(y) by Subdifferential of a Real-Valued Function on a Convex Subset of Rn\mathbb{R}^n (i).

Claim 2. Since UU is open, yy is an interior point of UU, so A Convex Function is Lipschitz on a Ball around an Interior Point provides r,M∈Rr,M\in\mathbb{R} with 0<r0<r, 0≤M0\le M, Bˉ(y,r)⊆U\bar{B}(y,r)\subseteq U and

∣f(z)−f(w)∣≤M ∥z−w∥for all z,w∈Bˉ(y,r).(L)|f(z)-f(w)|\le M\,\lVert z-w\rVert\qquad\text{for all }z,w\in\bar{B}(y,r). \tag{L}

Put c=f(y)+Mr+1c=f(y)+Mr+1 and

K={ ι(z,t)  :  z∈Bˉ(y,r), t∈R, f(z)≤t  and  t≤c }⊆Rn+1.K=\bigl\{\,\iota(z,t)\;:\;z\in\bar{B}(y,r),\ t\in\mathbb{R},\ f(z)\le t\ \text{ and }\ t\le c\,\bigr\}\subseteq\mathbb{R}^{n+1}.

KK is nonempty: f(y)≤f(y)f(y)\le f(y) and f(y)≤cf(y)\le c, so ι(y,f(y))∈K\iota(y,f(y))\in K.

KK is bounded: let ι(z,t)∈K\iota(z,t)\in K. Then ∥z∥≤∥y∥+r\lVert z\rVert\le\lVert y\rVert+r by claim 6 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square, and by (L) f(y)−Mr≤f(z)≤t≤cf(y)-Mr\le f(z)\le t\le c, so ∣t∣≤∣f(y)−Mr∣+∣c∣|t|\le|f(y)-Mr|+|c|. Hence ∥ι(z,t)∥2=∥z∥2+t2\lVert\iota(z,t)\rVert^{2}=\lVert z\rVert^{2}+t^{2} is bounded above by a number independent of the point, and KK is bounded.

KK is closed: we use Sequential Characterization of Closed Subsets of a Metric Space. Let (ξm)m∈N(\xi_{m})_{m\in\mathbb{N}} be a sequence in KK converging to ξ∈Rn+1\xi\in\mathbb{R}^{n+1}. Since ι\iota is a bijection we may write ξm=ι(zm,tm)\xi_{m}=\iota(z_{m},t_{m}) and ξ=ι(z,t)\xi=\iota(z,t), and by linearity of ι\iota and the norm identity,

∥ξm−ξ∥2=∥ι(zm−z,tm−t)∥2=∥zm−z∥2+(tm−t)2,\lVert\xi_{m}-\xi\rVert^{2}=\lVert\iota(z_{m}-z,t_{m}-t)\rVert^{2}=\lVert z_{m}-z\rVert^{2}+(t_{m}-t)^{2},

so ∥zm−z∥≤∥ξm−ξ∥\lVert z_{m}-z\rVert\le\lVert\xi_{m}-\xi\rVert and ∣tm−t∣≤∥ξm−ξ∥|t_{m}-t|\le\lVert\xi_{m}-\xi\rVert; hence (zm)(z_{m}) converges to zz and (tm)(t_{m}) to tt. The closed ball Bˉ(y,r)\bar{B}(y,r) is closed by claim 3 of Elementary Properties of the Closed Ball in a Metric Space, so z∈Bˉ(y,r)z\in\bar{B}(y,r) by Sequential Characterization of Closed Subsets of a Metric Space again. From tm≤ct_{m}\le c and Order Properties of Limits of Real Sequences we get t≤ct\le c. Finally (L) gives ∣f(zm)−f(z)∣≤M∥zm−z∥|f(z_{m})-f(z)|\le M\lVert z_{m}-z\rVert, so (f(zm))(f(z_{m})) converges to f(z)f(z), and from f(zm)≤tmf(z_{m})\le t_{m} and Order Properties of Limits of Real Sequences we get f(z)≤tf(z)\le t. Hence ξ∈K\xi\in K, and KK is closed.

KK is convex: let ι(z1,t1),ι(z2,t2)∈K\iota(z_{1},t_{1}),\iota(z_{2},t_{2})\in K and let θ∈R\theta\in\mathbb{R} with 0≤θ≤10\le\theta\le1. By linearity of ι\iota,

θ ι(z1,t1)+(1−θ) ι(z2,t2)=ι(θz1+(1−θ)z2, θt1+(1−θ)t2).\theta\,\iota(z_{1},t_{1})+(1-\theta)\,\iota(z_{2},t_{2})=\iota\bigl(\theta z_{1}+(1-\theta)z_{2},\ \theta t_{1}+(1-\theta)t_{2}\bigr).

The point θz1+(1−θ)z2\theta z_{1}+(1-\theta)z_{2} lies in Bˉ(y,r)\bar{B}(y,r) by claim 2 of Euclidean Balls are Convex; the number θt1+(1−θ)t2\theta t_{1}+(1-\theta)t_{2} is at most cc; and since ff is convex on U⊇Bˉ(y,r)U\supseteq\bar{B}(y,r),

f(θz1+(1−θ)z2)≤θf(z1)+(1−θ)f(z2)≤θt1+(1−θ)t2.f\bigl(\theta z_{1}+(1-\theta)z_{2}\bigr)\le\theta f(z_{1})+(1-\theta)f(z_{2})\le\theta t_{1}+(1-\theta)t_{2}.

Hence the combination lies in KK.

Projections from below the graph. For k∈Nk\in\mathbb{N} put ξk=ι(y,f(y)−1/k)\xi_{k}=\iota(y,f(y)-1/k). Since ι\iota is injective, ξk∈K\xi_{k}\in K would force f(y)≤f(y)−1/kf(y)\le f(y)-1/k, which is false; so ξk∉K\xi_{k}\notin K. By claim 1 of Nearest-Point Projection onto a Nonempty Closed Convex Subset of Euclidean Space, applied to the nonempty closed convex set KK, the nearest point πK(ξk)∈K\pi_{K}(\xi_{k})\in K is defined; write πK(ξk)=ι(zk,tk)\pi_{K}(\xi_{k})=\iota(z_{k},t_{k}) and put vk=ξk−πK(ξk)v_{k}=\xi_{k}-\pi_{K}(\xi_{k}), which is nonzero because ξk∉K\xi_{k}\notin K while πK(ξk)∈K\pi_{K}(\xi_{k})\in K. By claim 2 of that lemma, vk⋅(w−πK(ξk))≤0v_{k}\cdot(w-\pi_{K}(\xi_{k}))\le0 for every w∈Kw\in K; multiplying by the positive number 1/∥vk∥1/\lVert v_{k}\rVert and putting uk=vk/∥vk∥u_{k}=v_{k}/\lVert v_{k}\rVert, a point with ∥uk∥=1\lVert u_{k}\rVert=1 by claim 5 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square, we get

uk⋅(w−πK(ξk))≤0for every w∈K.(N)u_{k}\cdot\bigl(w-\pi_{K}(\xi_{k})\bigr)\le0\qquad\text{for every }w\in K. \tag{N}

Since ι(y,f(y))∈K\iota(y,f(y))\in K, claims 3 and 4 of Nearest-Point Projection onto a Nonempty Closed Convex Subset of Euclidean Space give

∥πK(ξk)−ι(y,f(y))∥=∥πK(ξk)−πK(ι(y,f(y)))∥≤∥ξk−ι(y,f(y))∥=∥ι(0,−1/k)∥=1/k,\bigl\lVert\pi_{K}(\xi_{k})-\iota(y,f(y))\bigr\rVert=\bigl\lVert\pi_{K}(\xi_{k})-\pi_{K}(\iota(y,f(y)))\bigr\rVert\le\bigl\lVert\xi_{k}-\iota(y,f(y))\bigr\rVert=\lVert\iota(0,-1/k)\rVert=1/k,

and since by The Archimedean Property of the Real Numbers the numbers 1/k1/k eventually fall below any prescribed positive real, the sequence (πK(ξk))k∈N(\pi_{K}(\xi_{k}))_{k\in\mathbb{N}} converges to ι(y,f(y))\iota(y,f(y)).

The points uku_{k} all lie in Bˉ(0,1)\bar{B}(0,1), which is bounded, so The Bolzano-Weierstrass Theorem in Rn\mathbb{R}^n provides u∈Rn+1u\in\mathbb{R}^{n+1} and a strictly increasing sequence (pl)l∈N(p_{l})_{l\in\mathbb{N}} in N\mathbb{N} such that (upl)l∈N(u_{p_{l}})_{l\in\mathbb{N}} converges to uu. Choose l0l_{0} with ∥upl0−u∥<1/2\lVert u_{p_{l_{0}}}-u\rVert<1/2; then by claim 6 of Properties of the Euclidean Norm on Rn\mathbb{R}^n: Square, Distance, Vanishing, Coordinate Bounds, Homogeneity, Triangle Inequalities and Expansion of a Square, 1=∥upl0∥≤∥u∥+∥upl0−u∥1=\lVert u_{p_{l_{0}}}\rVert\le\lVert u\rVert+\lVert u_{p_{l_{0}}}-u\rVert, so ∥u∥≥1/2\lVert u\rVert\ge1/2 and in particular u≠0u\neq0. Write u=ι(a,b)u=\iota(a,b) with a∈Rna\in\mathbb{R}^{n} and b∈Rb\in\mathbb{R}.

Fix w∈Kw\in K. For every ll, by (N) and the dot product identity,

upl⋅(w−πK(ξpl))≤0.u_{p_{l}}\cdot\bigl(w-\pi_{K}(\xi_{p_{l}})\bigr)\le0 .

Both factors converge, the first to uu and the second to w−ι(y,f(y))w-\iota(y,f(y)); since for any points α,α′,β,β′\alpha,\alpha',\beta,\beta' one has, by The Dot Product on Rn\mathbb{R}^n Is Symmetric, Bilinear and Positive Definite and The Cauchy-Schwarz Inequality on Rn\mathbb{R}^n,

∣α′⋅β′−α⋅β∣≤∣(α′−α)⋅β′∣+∣α⋅(β′−β)∣≤∥α′−α∥ ∥β′∥+∥α∥ ∥β′−β∥,|\alpha'\cdot\beta'-\alpha\cdot\beta|\le|(\alpha'-\alpha)\cdot\beta'|+|\alpha\cdot(\beta'-\beta)|\le\lVert\alpha'-\alpha\rVert\,\lVert\beta'\rVert+\lVert\alpha\rVert\,\lVert\beta'-\beta\rVert ,

and the norms ∥w−πK(ξpl)∥\lVert w-\pi_{K}(\xi_{p_{l}})\rVert are bounded, the left-hand sides converge to u⋅(w−ι(y,f(y)))u\cdot(w-\iota(y,f(y))). By Order Properties of Limits of Real Sequences, writing w=ι(z,t)w=\iota(z,t),

a⋅(z−y)+b (t−f(y))≤0for every ι(z,t)∈K.(H)a\cdot(z-y)+b\,\bigl(t-f(y)\bigr)\le0\qquad\text{for every }\iota(z,t)\in K. \tag{H}

The last coordinate is negative. Taking z=yz=y and t=f(y)+Mr+1=ct=f(y)+Mr+1=c in (H), which is legitimate since f(y)≤cf(y)\le c, gives b (Mr+1)≤0b\,(Mr+1)\le0, so b≤0b\le0 because 0<Mr+10<Mr+1. Suppose b=0b=0. Then ∥u∥2=∥a∥2\lVert u\rVert^{2}=\lVert a\rVert^{2}, so a≠0a\neq0. For z∈Bˉ(y,r)z\in\bar{B}(y,r) we have, by (L), f(z)≤f(y)+M∥z−y∥≤f(y)+Mr<cf(z)\le f(y)+M\lVert z-y\rVert\le f(y)+Mr<c, so ι(z,f(z))∈K\iota(z,f(z))\in K and (H) gives a⋅(z−y)≤0a\cdot(z-y)\le0. Taking z=y+r a/∥a∥z=y+r\,a/\lVert a\rVert, which lies in Bˉ(y,r)\bar{B}(y,r) because ∥z−y∥=r\lVert z-y\rVert=r, we get a⋅(z−y)=r ∥a∥2/∥a∥=r∥a∥>0a\cdot(z-y)=r\,\lVert a\rVert^{2}/\lVert a\rVert=r\lVert a\rVert>0, a contradiction. Hence b<0b<0.

Conclusion. Put p=(1/(−b)) ap=(1/(-b))\,a. For z∈Bˉ(y,r)z\in\bar{B}(y,r) we have ι(z,f(z))∈K\iota(z,f(z))\in K as just shown, so (H) gives a⋅(z−y)+b(f(z)−f(y))≤0a\cdot(z-y)+b(f(z)-f(y))\le0, that is a⋅(z−y)≤(−b)(f(z)−f(y))a\cdot(z-y)\le(-b)\bigl(f(z)-f(y)\bigr). Dividing by the positive number −b-b and using The Dot Product on Rn\mathbb{R}^n Is Symmetric, Bilinear and Positive Definite,

f(z)≥f(y)+p⋅(z−y)for every z∈Bˉ(y,r).f(z)\ge f(y)+p\cdot(z-y)\qquad\text{for every }z\in\bar{B}(y,r).

By claim 1, p∈∂Uf(y)p\in\partial_{U}f(y), so ∂Uf(y)\partial_{U}f(y) is nonempty.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…