TheoremBase

Following Li and Shi, the centres are chosen recursively as near-maximisers of the partially perturbed function over a nested sequence of closed superlevel sets whose diameters shrink by the gauge hypothesis, and the limit of the centres is the strict maximiser.

Proof

Each result cited is universally quantified over the data in its own statement. Natural numbers are read in R\mathbb{R} through the canonical map as in The Real Numbers: Standing Notation and Background §numbers, so 1/n1/n is a positive real for every n∈Nn\in\mathbb{N}; the successor of nn is written n+1n+1, as in claim 1 of Natural Numbers. Since X≠∅X\ne\varnothing and ff is bounded above, σ=sup⁡x∈Xf(x)\sigma=\sup_{x\in X}f(x) exists by The Real Numbers: Standing Notation and Background §bounds, and f(x)≤σf(x)\le\sigma for every x∈Xx\in X. Since each ckc_{k} is positive and 0≤g(x,y)0\le g(x,y), claim 5 of Elementary Arithmetic in an Ordered Field gives

0≤ck g(x,y)for all k∈N and x,y∈X.(0)0\le c_{k}\,g(x,y)\qquad\text{for all }k\in\mathbb{N}\text{ and }x,y\in X.\tag{0}

Step 1 (Gauge radii). For n∈Nn\in\mathbb{N} let BnB_{n} be the set of reals β>0\beta>0 such that d(x,y)<1/nd(x,y)<1/n for all x,y∈Xx,y\in X with g(x,y)≤βg(x,y)\le\beta. The hypothesis on gg, applied with η=1/n\eta=1/n, shows that BnB_{n} is nonempty. Thus (Bn)n∈N(B_{n})_{n\in\mathbb{N}} is a family of nonempty subsets of R\mathbb{R}, and Axiom of Countable Choice gives a sequence (βn)n∈N(\beta_{n})_{n\in\mathbb{N}} with βn∈Bn\beta_{n}\in B_{n} for every nn, that is,

0<βn,andg(x,y)≤βn ⟹ d(x,y)<1n(x,y∈X).(1)0<\beta_{n},\qquad\text{and}\qquad g(x,y)\le\beta_{n}\ \Longrightarrow\ d(x,y)<\tfrac{1}{n}\quad(x,y\in X).\tag{1}

Step 2 (Recursive construction). Let T\mathcal{T} be the set of quadruples (n,z,H,A)(n,z,H,A) with n∈Nn\in\mathbb{N}, A⊆XA\subseteq X, z∈Az\in A, and H:X→RH:X\to\mathbb{R} bounded above. For (n,z,H,A)∈T(n,z,H,A)\in\mathcal{T} put A+={x∈A: H(z)≤H(x)}A^{+}=\{x\in A:\ H(z)\le H(x)\}. Then z∈A+z\in A^{+}, so H(A+)H(A^{+}) is nonempty, and it is bounded above because HH is; let M(n,z,H,A)=sup⁡x∈A+H(x)M(n,z,H,A)=\sup_{x\in A^{+}}H(x) (The Real Numbers: Standing Notation and Background §bounds).

Let RR be the relation on T\mathcal{T} consisting of the pairs ((n,z,H,A),(n+1,y,H′,A+))\bigl((n,z,H,A),(n+1,y,H',A^{+})\bigr) with (n,z,H,A)∈T(n,z,H,A)\in\mathcal{T}, y∈A+y\in A^{+}, H′(x)=H(x)−cn+1 g(x,y)H'(x)=H(x)-c_{n+1}\,g(x,y) for every x∈Xx\in X, and

H(y)>M(n,z,H,A)−cn+1βn.H(y)>M(n,z,H,A)-c_{n+1}\beta_{n}.

Every (n,z,H,A)∈T(n,z,H,A)\in\mathcal{T} has an RR-successor in T\mathcal{T}: cn+1βnc_{n+1}\beta_{n} is positive by (1) and claim 5 of Elementary Order Arithmetic in an Ordered Field, so claim 3 of Approximation Property of the Supremum and the Infimum in R\mathbb{R}, applied to H(A+)H(A^{+}) with ε=cn+1βn\varepsilon=c_{n+1}\beta_{n}, gives y∈A+y\in A^{+} with H(y)>M(n,z,H,A)−cn+1βnH(y)>M(n,z,H,A)-c_{n+1}\beta_{n}; the function H′H' so defined satisfies H′(x)≤H(x)H'(x)\le H(x) for every xx by (0) and claim 3 of Elementary Arithmetic in an Ordered Field, so it is bounded above, and y∈A+y\in A^{+}, whence (n+1,y,H′,A+)∈T(n+1,y,H',A^{+})\in\mathcal{T}.

Let F1:X→RF_{1}:X\to\mathbb{R} be given by F1(x)=f(x)−c1 g(x,x1)F_{1}(x)=f(x)-c_{1}\,g(x,x_{1}); as above F1(x)≤f(x)≤σF_{1}(x)\le f(x)\le\sigma, so (1,x1,F1,X)∈T(1,x_{1},F_{1},X)\in\mathcal{T}. By Axiom of Dependent Choice there is a sequence (am)m∈N(a_{m})_{m\in\mathbb{N}} in T\mathcal{T} with a1=(1,x1,F1,X)a_{1}=(1,x_{1},F_{1},X) and (am,am+1)∈R(a_{m},a_{m+1})\in R for every mm. By the form of RR and induction on mm, am=(m,xm,Fm,Tm)a_{m}=(m,x_{m},F_{m},T_{m}), where (xm)m∈N(x_{m})_{m\in\mathbb{N}} is a sequence in XX whose first term is the given point x1x_{1}, each Fm:X→RF_{m}:X\to\mathbb{R} is a function and each Tm⊆XT_{m}\subseteq X, such that T1=XT_{1}=X and, for every m∈Nm\in\mathbb{N},

xm∈Tm,Tm+1={x∈Tm: Fm(xm)≤Fm(x)},xm+1∈Tm+1,(2)x_{m}\in T_{m},\qquad T_{m+1}=\{x\in T_{m}:\ F_{m}(x_{m})\le F_{m}(x)\},\qquad x_{m+1}\in T_{m+1},\tag{2} Fm+1(x)=Fm(x)−cm+1 g(x,xm+1)(x∈X),(3)F_{m+1}(x)=F_{m}(x)-c_{m+1}\,g(x,x_{m+1})\quad(x\in X),\tag{3} Fm(xm+1)>sup⁡x∈Tm+1Fm(x)−cm+1βm.(4)F_{m}(x_{m+1})>\sup_{x\in T_{m+1}}F_{m}(x)-c_{m+1}\beta_{m}.\tag{4}

Step 3 (Partial sums, the series, and semicontinuity). For x∈Xx\in X and n∈Nn\in\mathbb{N} let pn(x)=∑k=1nck g(x,xk)p_{n}(x)=\sum_{k=1}^{n}c_{k}\,g(x,x_{k}), the nn-th partial sum of (ck g(x,xk))k∈N(c_{k}\,g(x,x_{k}))_{k\in\mathbb{N}}. Then

Fn(x)=f(x)−pn(x)(x∈X, n∈N),(5)F_{n}(x)=f(x)-p_{n}(x)\qquad(x\in X,\ n\in\mathbb{N}),\tag{5}

by induction on nn: for n=1n=1 this is the definition of F1F_{1} together with ∑k=11ak=a1\sum_{k=1}^{1}a_{k}=a_{1} from claim 1 of Properties of Finite Sums, and the step from nn to n+1n+1 is (3) together with the recursion pn+1(x)=pn(x)+cn+1 g(x,xn+1)p_{n+1}(x)=p_{n}(x)+c_{n+1}\,g(x,x_{n+1}) of the same claim. In particular F1(x1)=f(x1)−c1 g(x1,x1)=f(x1)F_{1}(x_{1})=f(x_{1})-c_{1}\,g(x_{1},x_{1})=f(x_{1}), since g(x1,x1)=0g(x_{1},x_{1})=0.

Fix x∈Xx\in X. By (0) and the hypotheses, 0≤ck0\le c_{k}, 0≤g(x,xk)≤G0\le g(x,x_{k})\le G for every kk, and ∑k=1∞ck\sum_{k=1}^{\infty}c_{k} converges, so the tail bound for dominated series (with μk=ck\mu_{k}=c_{k}, wk=g(x,xk)w_{k}=g(x,x_{k}) and M=GM=G) shows that the series P(x)=∑k=1∞ck g(x,xk)P(x)=\sum_{k=1}^{\infty}c_{k}\,g(x,x_{k}) converges and 0≤P(x)−pn(x)0\le P(x)-p_{n}(x) for every nn. Hence Φ(x)=f(x)−P(x)\Phi(x)=f(x)-P(x) is defined and, by claim 3 of Elementary Arithmetic in an Ordered Field and (5), since Fn(x)−Φ(x)=P(x)−pn(x)F_{n}(x)-\Phi(x)=P(x)-p_{n}(x),

Φ(x)≤Fn(x)(x∈X, n∈N).(6)\Phi(x)\le F_{n}(x)\qquad(x\in X,\ n\in\mathbb{N}).\tag{6}

Each FnF_{n} is upper semicontinuous on XX. Indeed, for every kk the function ck g(⋅,xk)c_{k}\,g(\cdot,x_{k}) is lower semicontinuous on XX by claim 3 of Sums and Nonnegative Multiples of Semicontinuous Functions, since g(⋅,xk)g(\cdot,x_{k}) is and 0≤ck0\le c_{k}. As ff is upper semicontinuous on XX, claim 3 of Negation, Restriction, and Separated Differences of Semicontinuous Functions shows that F1=f−c1 g(⋅,x1)F_{1}=f-c_{1}\,g(\cdot,x_{1}) is upper semicontinuous on XX; and if FnF_{n} is, then so is Fn+1=Fn−cn+1 g(⋅,xn+1)F_{n+1}=F_{n}-c_{n+1}\,g(\cdot,x_{n+1}), by (3) and the same claim.

Step 4 (Nesting and monotone values). We claim that, whenever n≤mn\le m,

Tm⊆Tn,xm∈Tn,Fn(xn)≤Fm(xm).(7)T_{m}\subseteq T_{n},\qquad x_{m}\in T_{n},\qquad F_{n}(x_{n})\le F_{m}(x_{m}).\tag{7}

First, for every mm we have Tm+1⊆TmT_{m+1}\subseteq T_{m} by (2), and

Fm+1(xm+1)=Fm(xm+1)−cm+1 g(xm+1,xm+1)=Fm(xm+1)≥Fm(xm)F_{m+1}(x_{m+1})=F_{m}(x_{m+1})-c_{m+1}\,g(x_{m+1},x_{m+1})=F_{m}(x_{m+1})\ge F_{m}(x_{m})

by (3), g(xm+1,xm+1)=0g(x_{m+1},x_{m+1})=0, and xm+1∈Tm+1x_{m+1}\in T_{m+1} in (2). Now argue by induction on mm. For m=1m=1, n≤1n\le1 and 1≤n1\le n (claim 4 of Properties of the Order on the Natural Numbers) force n=1n=1 by claim 2 there, and the first and third assertions are trivial. If they hold for mm and n≤m+1n\le m+1, then either n=m+1n=m+1, where they are trivial, or n≤mn\le m by claim 5 of Properties of the Order on the Natural Numbers, in which case Tm+1⊆Tm⊆TnT_{m+1}\subseteq T_{m}\subseteq T_{n} and Fn(xn)≤Fm(xm)≤Fm+1(xm+1)F_{n}(x_{n})\le F_{m}(x_{m})\le F_{m+1}(x_{m+1}). The second assertion follows from the first because xm∈Tmx_{m}\in T_{m} by (2).

Step 5 (The sets TnT_{n} are closed). We show by induction on nn: if y∈Xy\in X is such that for every real δ>0\delta>0 some w∈Tnw\in T_{n} satisfies d(y,w)<δd(y,w)<\delta, then y∈Tny\in T_{n}. For n=1n=1 this is clear since T1=XT_{1}=X. Assume it for nn, and let yy have this property relative to Tn+1T_{n+1}. Since Tn+1⊆TnT_{n+1}\subseteq T_{n}, yy has it relative to TnT_{n}, so y∈Tny\in T_{n}. Suppose Fn(xn)≤Fn(y)F_{n}(x_{n})\le F_{n}(y) fails; then Fn(y)<Fn(xn)F_{n}(y)<F_{n}(x_{n}) since the order is total, and γ=Fn(xn)−Fn(y)\gamma=F_{n}(x_{n})-F_{n}(y) is positive by claim 1 of Elementary Order Arithmetic in an Ordered Field. As FnF_{n} is upper semicontinuous at yy relative to XX (Step 3), there is δ>0\delta>0 with Fn(w)<Fn(y)+γ=Fn(xn)F_{n}(w)<F_{n}(y)+\gamma=F_{n}(x_{n}) for every w∈Xw\in X with d(y,w)<δd(y,w)<\delta. Taking such a ww in Tn+1T_{n+1} contradicts Fn(xn)≤Fn(w)F_{n}(x_{n})\le F_{n}(w), which holds by (2). Hence Fn(xn)≤Fn(y)F_{n}(x_{n})\le F_{n}(y), and y∈Tn+1y\in T_{n+1} by (2).

Step 6 (Small sets). Let n∈Nn\in\mathbb{N} and x∈Tn+2x\in T_{n+2}. Then x∈Tn+1x\in T_{n+1} by (7), so Fn(x)≤sup⁡w∈Tn+1Fn(w)F_{n}(x)\le\sup_{w\in T_{n+1}}F_{n}(w), and by (4) and claims 1 and 2 of Elementary Order Arithmetic in an Ordered Field, Fn(x)<Fn(xn+1)+cn+1βnF_{n}(x)<F_{n}(x_{n+1})+c_{n+1}\beta_{n}. On the other hand Fn+1(xn+1)≤Fn+1(x)F_{n+1}(x_{n+1})\le F_{n+1}(x) by (2) applied with m=n+1m=n+1, which by (3) and g(xn+1,xn+1)=0g(x_{n+1},x_{n+1})=0 reads Fn(xn+1)≤Fn(x)−cn+1 g(x,xn+1)F_{n}(x_{n+1})\le F_{n}(x)-c_{n+1}\,g(x,x_{n+1}). Combining the two,

cn+1 g(x,xn+1)≤Fn(x)−Fn(xn+1)<cn+1βn,c_{n+1}\,g(x,x_{n+1})\le F_{n}(x)-F_{n}(x_{n+1})<c_{n+1}\beta_{n},

and multiplying by cn+1−1>0c_{n+1}^{-1}>0 (claims 7 and 10 of Elementary Order Arithmetic in an Ordered Field) gives g(x,xn+1)<βng(x,x_{n+1})<\beta_{n}. By (1),

d(x,xn+1)<1nfor every n∈N and every x∈Tn+2.(8)d(x,x_{n+1})<\tfrac{1}{n}\qquad\text{for every }n\in\mathbb{N}\text{ and every }x\in T_{n+2}.\tag{8}

Step 7 (Centres). Let a real ε>0\varepsilon>0 be given. By claim 8 of Elementary Order Arithmetic in an Ordered Field and claim 3 of The Archimedean Property of the Real Numbers there is n∈Nn\in\mathbb{N} with 1/n<ε/21/n<\varepsilon/2. For m,ℓ∈Nm,\ell\in\mathbb{N} with m≥n+2m\ge n+2 and ℓ≥n+2\ell\ge n+2, (7) gives xm,xℓ∈Tn+2x_{m},x_{\ell}\in T_{n+2}, so (8) and the symmetry and triangle inequality of Metric Space give

d(xm,xℓ)≤d(xm,xn+1)+d(xℓ,xn+1)<1n+1n<ε,d(x_{m},x_{\ell})\le d(x_{m},x_{n+1})+d(x_{\ell},x_{n+1})<\tfrac{1}{n}+\tfrac{1}{n}<\varepsilon,

using claims 3 and 8 of Elementary Order Arithmetic in an Ordered Field. Thus (xm)(x_{m}) is a Cauchy sequence, and since (X,d)(X,d) is complete it converges to some xˉ∈X\bar{x}\in X. This is the centres assertion.

Moreover xˉ∈Tn\bar{x}\in T_{n} for every n∈Nn\in\mathbb{N}. Given a real δ>0\delta>0, convergence gives N∈NN\in\mathbb{N} with d(xm,xˉ)<δd(x_{m},\bar{x})<\delta for all m≥Nm\ge N; let mm be the larger of NN and nn (trichotomy, claim 3 of Properties of the Order on the Natural Numbers). Then xm∈Tnx_{m}\in T_{n} by (7) and d(xˉ,xm)<δd(\bar{x},x_{m})<\delta, so Step 5 gives xˉ∈Tn\bar{x}\in T_{n}.

Step 8 (The sets TnT_{n} shrink to xˉ\bar{x}). Let y∈Xy\in X lie in TnT_{n} for every n∈Nn\in\mathbb{N}. For each nn, both yy and xˉ\bar{x} lie in Tn+2T_{n+2}, so as in Step 7, (8) gives d(y,xˉ)≤d(y,xn+1)+d(xˉ,xn+1)<2/nd(y,\bar{x})\le d(y,x_{n+1})+d(\bar{x},x_{n+1})<2/n. Given a real ε>0\varepsilon>0, choosing nn with 1/n<ε/21/n<\varepsilon/2 as in Step 7 yields d(y,xˉ)<εd(y,\bar{x})<\varepsilon; since 0≤d(y,xˉ)0\le d(y,\bar{x}), vanishing gives d(y,xˉ)=0d(y,\bar{x})=0, so y=xˉy=\bar{x} by Metric Space.

Step 9 (Value). We show

Fn(xn)≤Φ(xˉ)for every n∈N.(9)F_{n}(x_{n})\le\Phi(\bar{x})\qquad\text{for every }n\in\mathbb{N}.\tag{9}

Fix nn and suppose instead that Φ(xˉ)<Fn(xn)\Phi(\bar{x})<F_{n}(x_{n}), so that γ=Fn(xn)−Φ(xˉ)\gamma=F_{n}(x_{n})-\Phi(\bar{x}) is positive by claim 1 of Elementary Order Arithmetic in an Ordered Field. By Step 3 and the definition of the sum, (pm(xˉ))m∈N(p_{m}(\bar{x}))_{m\in\mathbb{N}} converges to P(xˉ)P(\bar{x}), so there is N∈NN\in\mathbb{N} with ∣pm(xˉ)−P(xˉ)∣<γ|p_{m}(\bar{x})-P(\bar{x})|<\gamma, hence P(xˉ)−pm(xˉ)<γP(\bar{x})-p_{m}(\bar{x})<\gamma by claim 9 of Properties of the Absolute Value in an Ordered Field, for all m≥Nm\ge N. Let mm be the larger of NN and nn. Since xˉ∈Tm+1\bar{x}\in T_{m+1} by Step 7, (2) gives Fm(xm)≤Fm(xˉ)F_{m}(x_{m})\le F_{m}(\bar{x}), and Fn(xn)≤Fm(xm)F_{n}(x_{n})\le F_{m}(x_{m}) by (7). Using (5) and claim 1 of Elementary Order Arithmetic in an Ordered Field,

Fn(xn)≤Fm(xˉ)=f(xˉ)−pm(xˉ)<f(xˉ)−P(xˉ)+γ=Φ(xˉ)+γ=Fn(xn),F_{n}(x_{n})\le F_{m}(\bar{x})=f(\bar{x})-p_{m}(\bar{x})<f(\bar{x})-P(\bar{x})+\gamma=\Phi(\bar{x})+\gamma=F_{n}(x_{n}),

so claim 2 of Elementary Order Arithmetic in an Ordered Field would give Fn(xn)<Fn(xn)F_{n}(x_{n})<F_{n}(x_{n}), which is impossible because a<ba<b requires a≠ba\ne b. Hence (9) holds, since the order is total. Taking n=1n=1 and recalling F1(x1)=f(x1)F_{1}(x_{1})=f(x_{1}) from Step 3 gives Φ(xˉ)≥f(x1)\Phi(\bar{x})\ge f(x_{1}), the value assertion.

Step 10 (Strict maximum). Let x∈Xx\in X with x≠xˉx\ne\bar{x}. By Step 8, xx does not lie in every TjT_{j}. Hence there is n∈Nn\in\mathbb{N} with x∈Tnx\in T_{n} and x∉Tn+1x\notin T_{n+1}: otherwise, since x∈T1=Xx\in T_{1}=X, induction on nn would give x∈Tnx\in T_{n} for every nn. By (2), x∉Tn+1x\notin T_{n+1} with x∈Tnx\in T_{n} means that Fn(xn)≤Fn(x)F_{n}(x_{n})\le F_{n}(x) fails, so Fn(x)<Fn(xn)F_{n}(x)<F_{n}(x_{n}). By (6) and (9),

Φ(x)≤Fn(x)<Fn(xn)≤Φ(xˉ),\Phi(x)\le F_{n}(x)<F_{n}(x_{n})\le\Phi(\bar{x}),

and claim 2 of Elementary Order Arithmetic in an Ordered Field gives Φ(x)<Φ(xˉ)\Phi(x)<\Phi(\bar{x}), the strict maximum assertion.

Step 11 (Localisation). By (6) with n=1n=1, the value assertion, the hypothesis on x1x_{1} and f(xˉ)≤σf(\bar{x})\le\sigma,

f(xˉ)−c1 g(xˉ,x1)=F1(xˉ)≥Φ(xˉ)≥f(x1)≥σ−ε≥f(xˉ)−ε.f(\bar{x})-c_{1}\,g(\bar{x},x_{1})=F_{1}(\bar{x})\ge\Phi(\bar{x})\ge f(x_{1})\ge\sigma-\varepsilon\ge f(\bar{x})-\varepsilon .

By claim 3 of Elementary Arithmetic in an Ordered Field, 0≤(f(xˉ)−c1 g(xˉ,x1))−(f(xˉ)−ε)=ε−c1 g(xˉ,x1)0\le\bigl(f(\bar{x})-c_{1}\,g(\bar{x},x_{1})\bigr)-\bigl(f(\bar{x})-\varepsilon\bigr)=\varepsilon-c_{1}\,g(\bar{x},x_{1}), and by the same claim c1 g(xˉ,x1)≤εc_{1}\,g(\bar{x},x_{1})\le\varepsilon, the localisation assertion.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…