TheoremBase

The three-slope inequality follows from convexity at the point t written as a weighted average of s and u; it makes the difference quotients at a fixed point increasing, which gives the infimum, the supporting lines and a local Lipschitz bound, and density of the rationals with continuity gives the supremum formula.

Proof

Throughout, for x∈Ix\in I and y∈Iy\in I with y≠xy\neq x write

Qx(y)=f(y)−f(x)y−x,so that Qx(y)=Qy(x);Q_{x}(y)=\frac{f(y)-f(x)}{y-x},\qquad\text{so that } Q_{x}(y)=Q_{y}(x);

and x↦cxx\mapsto c_{x}, once cxc_{x} exists, is a map I→RI\to\mathbb{R} given by an infimum (Routine Facts Used Without Citation: Arithmetic, Maps, Predicativity and Indices (iv)), so the set in (v) is formed as written. Every x∈Ix\in I lies strictly between two points of II: since II is open, there is a real r>0r>0 with BdR(x,r)⊆IB_{d_{\mathbb{R}}}(x,r)\subseteq I, and x−r/2x-r/2 and x+r/2x+r/2 lie in this ball by Open Ball in a Metric Space (ii), as their distance to xx is r/2<rr/2<r.

Step 1: increasing slopes. Let s,t,u∈Is,t,u\in I with s<t<us<t<u, and put λ=(u−t)/(u−s)\lambda=(u-t)/(u-s). Then 0<λ<10<\lambda<1, 1−λ=(t−s)/(u−s)1-\lambda=(t-s)/(u-s) and λs+(1−λ)u=t\lambda s+(1-\lambda)u=t, so convexity (Convex Functions on an Interval of ℝ (i), with ss, uu and the weight λ\lambda) gives f(t)≤λf(s)+(1−λ)f(u)f(t)\le\lambda f(s)+(1-\lambda)f(u). Subtracting f(s)f(s), respectively f(u)f(u), from both sides and rearranging,

f(t)−f(s)≤t−su−s(f(u)−f(s)),f(u)−f(t)≥u−tu−s(f(u)−f(s)).f(t)-f(s)\le\frac{t-s}{u-s}\bigl(f(u)-f(s)\bigr),\qquad f(u)-f(t)\ge\frac{u-t}{u-s}\bigl(f(u)-f(s)\bigr).

Dividing the first inequality by t−s>0t-s>0 and the second by u−t>0u-t>0 gives the two inequalities of (i).

Step 2: the quotients at a point increase. Let x∈Ix\in I and y,z∈I∖{x}y,z\in I\setminus\{x\} with y<zy<z. We show Qx(y)≤Qx(z)Q_{x}(y)\le Q_{x}(z). If x<y<zx<y<z, this is the left inequality of (i) for (s,t,u)=(x,y,z)(s,t,u)=(x,y,z). If y<z<xy<z<x, it is the right inequality for (y,z,x)(y,z,x), since Qx(y)=(f(x)−f(y))/(x−y)Q_{x}(y)=(f(x)-f(y))/(x-y) and Qx(z)=(f(x)−f(z))/(x−z)Q_{x}(z)=(f(x)-f(z))/(x-z). If y<x<zy<x<z, it is the outer inequality for (y,x,z)(y,x,z):

Qx(y)=f(x)−f(y)x−y≤f(z)−f(x)z−x=Qx(z).Q_{x}(y)=\frac{f(x)-f(y)}{x-y}\le\frac{f(z)-f(x)}{z-x}=Q_{x}(z).

Step 3: the infimum cxc_{x}. Let x∈Ix\in I, and choose s,t∈Is,t\in I with s<x<ts<x<t, as above; so the set Rx={Qx(y):y∈I, y>x}R_{x}=\{Q_{x}(y):y\in I,\ y>x\} contains Qx(t)Q_{x}(t), and by Step 2 every s′∈Is'\in I with s′<xs'<x gives the lower bound Qx(s′)Q_{x}(s') of RxR_{x}. Hence RxR_{x} has an infimum cx∈Rc_{x}\in\mathbb{R}, by The Real Numbers, with the Natural Numbers, Integers and Rationals Identified with Subsets of the Reals, and Completeness (viii); as a lower bound of RxR_{x} it satisfies cx≤Qx(t′)c_{x}\le Q_{x}(t') for every t′∈It'\in I with t′>xt'>x, and as the greatest lower bound it satisfies Qx(s′)≤cxQ_{x}(s')\le c_{x} for every s′∈Is'\in I with s′<xs'<x. Since Qx(s′)=(f(x)−f(s′))/(x−s′)Q_{x}(s')=(f(x)-f(s'))/(x-s'), this is (ii).

Step 4: supporting lines. Let x,y∈Ix,y\in I. For y=xy=x there is nothing to prove. If y>xy>x, then cx≤Qx(y)c_{x}\le Q_{x}(y) by (ii), and multiplying by y−x>0y-x>0 gives f(y)−f(x)≥cx(y−x)f(y)-f(x)\ge c_{x}(y-x). If y<xy<x, then Qx(y)≤cxQ_{x}(y)\le c_{x} by (ii), and multiplying by y−x<0y-x<0 gives f(y)−f(x)≥cx(y−x)f(y)-f(x)\ge c_{x}(y-x) again. This is (iii).

Step 5: continuity. We use the ε\varepsilon-δ\delta form of continuity on II of The Real Line: Standing Notation and Background for Calculus on Intervals (vi). Let x∈Ix\in I and let ε>0\varepsilon>0 be real. Choose s,t∈Is,t\in I with s<x<ts<x<t, as above, and put M=∣Qx(s)∣+∣Qx(t)∣M=|Q_{x}(s)|+|Q_{x}(t)|. For y∈Iy\in I with s<y<ts<y<t and y≠xy\neq x, Step 2 gives Qx(s)≤Qx(y)≤Qx(t)Q_{x}(s)\le Q_{x}(y)\le Q_{x}(t), hence ∣Qx(y)∣≤M|Q_{x}(y)|\le M and

∣f(y)−f(x)∣=∣Qx(y)∣ ∣y−x∣≤M ∣y−x∣,|f(y)-f(x)|=|Q_{x}(y)|\,|y-x|\le M\,|y-x|,

which also holds for y=xy=x. Put δ=min⁡{x−s, t−x, ε/(M+1)}>0\delta=\min\{x-s,\ t-x,\ \varepsilon/(M+1)\}>0. Every y∈Iy\in I with ∣y−x∣<δ|y-x|<\delta satisfies s<y<ts<y<t, so

∣f(y)−f(x)∣≤(M+1) ∣y−x∣<(M+1) δ≤ε.|f(y)-f(x)|\le(M+1)\,|y-x|<(M+1)\,\delta\le\varepsilon .

Hence ff is continuous on II, which is (iv).

Step 6: the rational supporting lines. Let y∈Iy\in I and let Sy={f(q)+cq(y−q):q∈I∩Q}S_{y}=\{f(q)+c_{q}(y-q):q\in I\cap\mathbb{Q}\}. By (iii), applied at qq, every element of SyS_{y} is at most f(y)f(y). Choose s∈Is\in I with s<ys<y, as above; by The Real Numbers, with the Natural Numbers, Integers and Rationals Identified with Subsets of the Reals, and Completeness (x) there is q∈Qq\in\mathbb{Q} with s<q<ys<q<y, and q∈Iq\in I by Intervals of the Real Line (i). So SyS_{y} is nonempty and bounded above by f(y)f(y); it has a supremum by The Real Numbers, with the Natural Numbers, Integers and Rationals Identified with Subsets of the Reals, and Completeness (vii), and sup⁡Sy≤f(y)\sup S_{y}\le f(y).

For the reverse inequality put r=(s+y)/2r=(s+y)/2, so s<r<ys<r<y and r∈Ir\in I, and

L=f(r)−f(s)r−s.L=\frac{f(r)-f(s)}{r-s}.

For q∈Iq\in I with r<q<yr<q<y, (ii) (with s<qs<q) and then (i) (with s<r<qs<r<q) give

cq≥f(q)−f(s)q−s≥L,hencef(q)+cq(y−q)≥f(q)+L(y−q)≥f(q)−∣L∣ (y−q).c_{q}\ge\frac{f(q)-f(s)}{q-s}\ge L,\qquad\text{hence}\qquad f(q)+c_{q}(y-q)\ge f(q)+L(y-q)\ge f(q)-|L|\,(y-q).

Now let ε>0\varepsilon>0 be real. By (iv) there is a real δ>0\delta>0 such that ∣f(q)−f(y)∣<ε/2|f(q)-f(y)|<\varepsilon/2 for every q∈Iq\in I with ∣q−y∣<δ|q-y|<\delta. Put a=max⁡{r, y−δ, y−ε/(2∣L∣+2)}a=\max\{r,\ y-\delta,\ y-\varepsilon/(2|L|+2)\}, so a<ya<y; by The Real Numbers, with the Natural Numbers, Integers and Rationals Identified with Subsets of the Reals, and Completeness (x) there is q∈Qq\in\mathbb{Q} with a<q<ya<q<y, and q∈Iq\in I since r<q<yr<q<y. Then (∣L∣+1)(y−q)<ε/2(|L|+1)(y-q)<\varepsilon/2, and

f(q)+cq(y−q)≥f(q)−∣L∣ (y−q)>(f(y)−ε2)−ε2=f(y)−ε.f(q)+c_{q}(y-q)\ge f(q)-|L|\,(y-q)>\Bigl(f(y)-\frac{\varepsilon}{2}\Bigr)-\frac{\varepsilon}{2}=f(y)-\varepsilon .

So sup⁡Sy>f(y)−ε\sup S_{y}>f(y)-\varepsilon for every real ε>0\varepsilon>0, hence sup⁡Sy≥f(y)\sup S_{y}\ge f(y), and with the first part f(y)=sup⁡Syf(y)=\sup S_{y}. This is (v).

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…