TheoremBase

An integral-free construction: h is an explicit convergent series of continuously differentiable quadratic-then-linear ramps switched on at a sequence of radii tending to zero, and each clause is checked directly from elementary estimates for the ramp, the mean value theorem, and facts about series of real numbers.

Proof

Each result cited is universally quantified over the data in its own statement. For a real tt write t+=max⁡(t,0)t^{+}=\max(t,0). Series of real numbers, their convergence and their sums are those of Series of Real Numbers §convergent; linearity of sums is Elementary Properties of Series of Real Numbers §linearity and comparison of sums is Elementary Properties of Series of Real Numbers §order; the comparison test is Series of Nonnegative Real Numbers, Comparison, and the Geometric Series §comparison, the geometric series with ratio 12\tfrac12 (sum 11, and nn-th tail (12)n(\tfrac12)^{n}, with (12)n→0(\tfrac12)^{n}\to0) is Series of Nonnegative Real Numbers, Comparison, and the Geometric Series §geometric, and the tail bound for a dominated series of nonnegative terms is Series of Nonnegative Real Numbers, Comparison, and the Geometric Series §tail-bound. Differentiability at a point of the interval I=RI=\mathbb{R}, every point of which is an interior point, is that of Single-Variable Calculus on an Interval §derivative.

Step 1: the ramp. Define P,Q,σ:R→RP,Q,\sigma:\mathbb{R}\to\mathbb{R} by

P(t)=12 (t+)2,Q(t)=P(t)−P(t−1),σ(t)=t+−(t−1)+.P(t)=\tfrac12\,(t^{+})^{2},\qquad Q(t)=P(t)-P(t-1),\qquad \sigma(t)=t^{+}-(t-1)^{+}.

Sorting tt into the three ranges t≤0t\le0, 0≤t≤10\le t\le1 and t≥1t\ge1 gives Q(t)=0Q(t)=0 and σ(t)=0\sigma(t)=0 for t≤0t\le0; Q(t)=t2/2Q(t)=t^{2}/2 and σ(t)=t\sigma(t)=t for 0≤t≤10\le t\le1; and Q(t)=t−12Q(t)=t-\tfrac12 and σ(t)=1\sigma(t)=1 for t≥1t\ge1. In particular 0≤σ(t)≤10\le\sigma(t)\le1 for every tt.

We claim that ∣P(t+u)−P(t)−t+u∣≤u2/2|P(t+u)-P(t)-t^{+}u|\le u^{2}/2 for all real t,ut,u. If t≥0t\ge0 and t+u≥0t+u\ge0 the left side is ∣(t+u)2/2−t2/2−tu∣=u2/2|(t+u)^{2}/2-t^{2}/2-tu|=u^{2}/2. If t≥0t\ge0 and t+u<0t+u<0, then 0≤t<−u0\le t<-u and the quantity inside the absolute value is −t2/2−tu=t(−u−t/2)-t^{2}/2-tu=t(-u-t/2), which is nonnegative, and u2/2−t(−u−t/2)=(u+t)2/2≥0u^{2}/2-t(-u-t/2)=(u+t)^{2}/2\ge0. If t<0t<0 and t+u≥0t+u\ge0, the quantity is (t+u)2/2(t+u)^{2}/2 and 0≤t+u<u0\le t+u<u, so it lies between 00 and u2/2u^{2}/2. If t<0t<0 and t+u<0t+u<0 it is 00. Applying the claim at tt and at t−1t-1 (note (t−1)+(t-1)^{+} is the coefficient there) and using the triangle inequality gives

∣Q(t+u)−Q(t)−σ(t)u∣≤u2for all real t,u.(1)|Q(t+u)-Q(t)-\sigma(t)u|\le u^{2}\qquad\text{for all real }t,u. \tag{1}

Consequently QQ is differentiable at every t∈Rt\in\mathbb{R} with Q′(t)=σ(t)Q'(t)=\sigma(t), in the sense of Single-Variable Calculus on an Interval §derivative: given ε>0\varepsilon>0 take δ=ε\delta=\varepsilon; for 0<∣u∣<δ0<|u|<\delta, dividing (1) by ∣u∣|u| gives ∣(Q(t+u)−Q(t))/u−σ(t)∣≤∣u∣<ε|(Q(t+u)-Q(t))/u-\sigma(t)|\le|u|<\varepsilon. By (1) and ∣σ(t)∣≤1|\sigma(t)|\le1 we also have ∣Q(t+u)−Q(t)∣≤∣u∣+u2|Q(t+u)-Q(t)|\le|u|+u^{2}, so QQ is continuous at every point (given ε>0\varepsilon>0, δ=min⁡(1,ε/2)\delta=\min(1,\varepsilon/2) works: for ∣u∣<δ|u|<\delta we have u2≤∣u∣u^{2}\le|u|, so ∣u∣+u2≤2∣u∣<ε|u|+u^{2}\le2|u|<\varepsilon), and hence continuous on every closed interval [a,b][a,b] in the sense of The Real Line: Standing Notation and Background for Calculus §continuity. Let a<ba<b. The restriction of QQ to [a,b][a,b] is continuous on [a,b][a,b], and it is differentiable at every point xx of (a,b)(a,b) with derivative σ(x)\sigma(x), because passing to a subinterval changes neither differentiability nor the derivative, as recorded in Single-Variable Calculus on an Interval §derivative. The mean value theorem Mean Value Theorem on a Closed Real Interval gives c∈(a,b)c\in(a,b) with Q(b)−Q(a)=σ(c)(b−a)Q(b)-Q(a)=\sigma(c)(b-a), and since 0≤σ(c)≤10\le\sigma(c)\le1,

0≤Q(b)−Q(a)≤b−awhenever a≤b,(2)0\le Q(b)-Q(a)\le b-a\qquad\text{whenever }a\le b, \tag{2}

the case a=ba=b being trivial.

Step 2: rescaled ramps. For a real ρ>0\rho>0 define qρ,σρ:R→Rq_{\rho},\sigma_{\rho}:\mathbb{R}\to\mathbb{R} by qρ(s)=ρ Q(s/ρ−1)q_{\rho}(s)=\rho\,Q(s/\rho-1) and σρ(s)=σ(s/ρ−1)\sigma_{\rho}(s)=\sigma(s/\rho-1). From Step 1 we read off the following, for all real s,t,us,t,u.

(a) If s≤ρs\le\rho then s/ρ−1≤0s/\rho-1\le0, so qρ(s)=0q_{\rho}(s)=0 and σρ(s)=0\sigma_{\rho}(s)=0; and always 0≤σρ(s)≤10\le\sigma_{\rho}(s)\le1.

(b) If s≤ts\le t then 0≤qρ(t)−qρ(s)≤t−s0\le q_{\rho}(t)-q_{\rho}(s)\le t-s, by (2) applied to s/ρ−1≤t/ρ−1s/\rho-1\le t/\rho-1 and multiplied by ρ\rho.

(c) 0≤qρ(s)≤s+0\le q_{\rho}(s)\le s^{+}. Indeed, if s≤ρs\le\rho this is (a); if s>ρs>\rho then (b) and qρ(ρ)=0q_{\rho}(\rho)=0 give 0≤qρ(s)≤s−ρ≤s0\le q_{\rho}(s)\le s-\rho\le s.

(d) qρ(s)≥s−32ρq_{\rho}(s)\ge s-\tfrac32\rho. Indeed, if s≥2ρs\ge2\rho then s/ρ−1≥1s/\rho-1\ge1 and qρ(s)=ρ(s/ρ−1−12)=s−32ρq_{\rho}(s)=\rho(s/\rho-1-\tfrac12)=s-\tfrac32\rho; if s<2ρs<2\rho then (b) gives qρ(2ρ)−qρ(s)≤2ρ−sq_{\rho}(2\rho)-q_{\rho}(s)\le2\rho-s, and qρ(2ρ)=ρ/2q_{\rho}(2\rho)=\rho/2, so qρ(s)≥s−32ρq_{\rho}(s)\ge s-\tfrac32\rho.

(e) Writing eρ(s,u)=qρ(s+u)−qρ(s)−σρ(s)ue_{\rho}(s,u)=q_{\rho}(s+u)-q_{\rho}(s)-\sigma_{\rho}(s)u, we have ∣eρ(s,u)∣≤u2/ρ|e_{\rho}(s,u)|\le u^{2}/\rho and ∣eρ(s,u)∣≤2∣u∣|e_{\rho}(s,u)|\le2|u|. The first bound is (1) at the point s/ρ−1s/\rho-1 with increment u/ρu/\rho, multiplied by ρ\rho; the second follows from ∣qρ(s+u)−qρ(s)∣≤∣u∣|q_{\rho}(s+u)-q_{\rho}(s)|\le|u|, which is (b), and ∣σρ(s)u∣≤∣u∣|\sigma_{\rho}(s)u|\le|u|, which is (a).

Step 3: choice of constants. The constants are chosen in the following order. First, for each k∈Nk\in\mathbb{N} the hypothesis on mm, applied with η=2−k−3\eta=2^{-k-3}, gives a real rk′>0r'_{k}>0 with m(s)≤2−k−3sm(s)\le2^{-k-3}s for every s∈[0,r]s\in[0,r] with s<rk′s<r'_{k}; fix such a choice for every kk. Such a sequence (rk′)k∈N(r'_{k})_{k\in\mathbb{N}} exists by Axiom of Dependent Choice, applied to the set SS of pairs (k,ρ)(k,\rho) with k∈Nk\in\mathbb{N} and ρ>0\rho>0 real such that m(s)≤2−k−3sm(s)\le2^{-k-3}s for every s∈[0,r]s\in[0,r] with s<ρs<\rho, to the relation consisting of the pairs ((k,ρ),(k+1,ρ′))\bigl((k,\rho),(k+1,\rho')\bigr) of elements of SS, and to a starting element (1,ρ1)∈S(1,\rho_{1})\in S: the hypothesis on mm (with η=2−4\eta=2^{-4}) shows that SS has such an element, and (with η=2−k−4\eta=2^{-k-4}) that every (k,ρ)∈S(k,\rho)\in S is related to some element of SS; the sequence (ak)k∈N(a_{k})_{k\in\mathbb{N}} given by the axiom has first coordinate kk at every index kk, by induction on kk, and we let rk′r'_{k} be the second coordinate of aka_{k}. Second, define r1=min⁡(r,r1′)r_{1}=\min(r,r'_{1}) and recursively rk+1=min⁡(rk/2,rk+1′)r_{k+1}=\min(r_{k}/2,r'_{k+1}) for k∈Nk\in\mathbb{N}. Then for every kk: rk>0r_{k}>0, rk+1≤rk/2r_{k+1}\le r_{k}/2, rk≤rr_{k}\le r, rk≤rk′r_{k}\le r'_{k}, the sequence (rk)(r_{k}) is nonincreasing, and by induction rk≤(12)k−1r1r_{k}\le(\tfrac12)^{k-1}r_{1}. Hence

m(s)≤2−k−3sfor every k∈N and every s∈[0,r] with s<rk.(3)m(s)\le2^{-k-3}s\qquad\text{for every }k\in\mathbb{N}\text{ and every }s\in[0,r]\text{ with }s<r_{k}. \tag{3}

Third, since B≥m(0)=0B\ge m(0)=0, the number A=4B/r1A=4B/r_{1} satisfies A≥0A\ge0.

We record a location fact. Let 0<s<r10<s<r_{1}. Since (12)n→0(\tfrac12)^{n}\to0 by the geometric series clause and s/r1>0s/r_{1}>0, there is n′∈Nn'\in\mathbb{N} with (12)n′<s/r1(\tfrac12)^{n'}<s/r_{1}; with n=n′+1n=n'+1 we get (12)n−1<s/r1(\tfrac12)^{n-1}<s/r_{1}, that is (12)n−1r1<s(\tfrac12)^{n-1}r_{1}<s, hence rn≤(12)n−1r1<sr_{n}\le(\tfrac12)^{n-1}r_{1}<s; so the set of n∈Nn\in\mathbb{N} with rn≤sr_{n}\le s is nonempty, and its least element n0n_{0} satisfies n0≥2n_{0}\ge2 because r1>sr_{1}>s. With k=n0−1∈Nk=n_{0}-1\in\mathbb{N} we get

rk+1≤s<rk.(4)r_{k+1}\le s<r_{k}. \tag{4}

Step 4: definition of hh. Write qj=qrjq_{j}=q_{r_{j}}, σj=σrj\sigma_{j}=\sigma_{r_{j}} and ej=erje_{j}=e_{r_{j}} for j∈Nj\in\mathbb{N}. For s∈Rs\in\mathbb{R} we have 0≤2−jqj(s)≤s+2−j0\le2^{-j}q_{j}(s)\le s^{+}2^{-j} by (c), so by the comparison test the series ∑j=1∞2−jqj(s)\sum_{j=1}^{\infty}2^{-j}q_{j}(s) converges, the dominating series ∑js+2−j\sum_{j}s^{+}2^{-j} converging (with sum s+s^{+}) by the geometric series and linearity of sums, with sum in [0,s+][0,s^{+}] by comparison of sums. Likewise 0≤2−jσj(s)≤2−j0\le2^{-j}\sigma_{j}(s)\le2^{-j} by (a), so ∑j=1∞2−jσj(s)\sum_{j=1}^{\infty}2^{-j}\sigma_{j}(s) converges with sum in [0,1][0,1]. Define H,D:R→RH,D:\mathbb{R}\to\mathbb{R} by

H(s)=A q2(s)+∑j=1∞2−jqj(s),D(s)=A σ2(s)+∑j=1∞2−jσj(s).H(s)=A\,q_{2}(s)+\sum_{j=1}^{\infty}2^{-j}q_{j}(s),\qquad D(s)=A\,\sigma_{2}(s)+\sum_{j=1}^{\infty}2^{-j}\sigma_{j}(s).

Then H(s)≥0H(s)\ge0 and 0≤D(s)≤A+10\le D(s)\le A+1 for every ss. If s<0s<0 then s<rjs<r_{j} for every jj, so every qj(s)q_{j}(s) vanishes by (a) and H(s)=0H(s)=0. Let h:[0,∞)→[0,∞)h:[0,\infty)\to[0,\infty) be the restriction of HH to [0,∞)[0,\infty). Then the function hˉ\bar h of the statement coincides with HH on all of R\mathbb{R}.

Clause (Monotone). Let 0≤s≤t0\le s\le t. By (b), Aq2(s)≤Aq2(t)A q_{2}(s)\le Aq_{2}(t) (as A≥0A\ge0) and 2−jqj(s)≤2−jqj(t)2^{-j}q_{j}(s)\le2^{-j}q_{j}(t) for every jj, so comparison of sums gives h(s)≤h(t)h(s)\le h(t); thus hh is nondecreasing on [0,∞)[0,\infty) in the sense of Monotone Real Function §nondecreasing. Since 0≤rj0\le r_{j} for every jj, (a) gives qj(0)=0q_{j}(0)=0 for every jj, so h(0)=0h(0)=0.

Clause (Differentiable). Fix s∈Rs\in\mathbb{R} and ε>0\varepsilon>0. Choose, in this order, J∈NJ\in\mathbb{N} with (12)J<ε/8(\tfrac12)^{J}<\varepsilon/8, which exists since (12)n→0(\tfrac12)^{n}\to0; then the positive number CJ=A/r2+∑j=1J2−j/rj+1C_{J}=A/r_{2}+\sum_{j=1}^{J}2^{-j}/r_{j}+1; then δ=ε/(2CJ)\delta=\varepsilon/(2C_{J}). Let uu be real with 0<∣u∣<δ0<|u|<\delta. The three series defining H(s+u)H(s+u), H(s)H(s) and D(s)D(s) converge, so by linearity of sums

H(s+u)−H(s)−D(s)u=A e2(s,u)+∑j=1∞2−jej(s,u).H(s+u)-H(s)-D(s)u=A\,e_{2}(s,u)+\sum_{j=1}^{\infty}2^{-j}e_{j}(s,u).

By (e), 0≤2−j∣ej(s,u)∣≤2∣u∣ 2−j0\le2^{-j}|e_{j}(s,u)|\le2|u|\,2^{-j}, so by the comparison test the series ∑j2−j∣ej(s,u)∣\sum_{j}2^{-j}|e_{j}(s,u)| converges, the dominating series ∑j2∣u∣ 2−j\sum_{j}2|u|\,2^{-j} converging by the geometric series and linearity of sums, and comparison of sums applied to −2−j∣ej(s,u)∣≤2−jej(s,u)≤2−j∣ej(s,u)∣-2^{-j}|e_{j}(s,u)|\le2^{-j}e_{j}(s,u)\le2^{-j}|e_{j}(s,u)| (the left series converging by linearity) gives ∣∑j2−jej(s,u)∣≤∑j2−j∣ej(s,u)∣\bigl|\sum_{j}2^{-j}e_{j}(s,u)\bigr|\le\sum_{j}2^{-j}|e_{j}(s,u)|. The tail bound, applied with μj=2−j\mu_{j}=2^{-j}, wj=∣ej(s,u)∣w_{j}=|e_{j}(s,u)|, M=2∣u∣M=2|u| and n=Jn=J, together with the geometric series tail (12)J(\tfrac12)^{J}, gives

∑j=1∞2−j∣ej(s,u)∣≤∑j=1J2−j∣ej(s,u)∣+2∣u∣ (12)J≤u2∑j=1J2−jrj+2∣u∣ (12)J,\sum_{j=1}^{\infty}2^{-j}|e_{j}(s,u)|\le\sum_{j=1}^{J}2^{-j}|e_{j}(s,u)|+2|u|\,(\tfrac12)^{J}\le u^{2}\sum_{j=1}^{J}\frac{2^{-j}}{r_{j}}+2|u|\,(\tfrac12)^{J},

where the second inequality is the first bound of (e). Using the first bound of (e) also for e2e_{2},

∣H(s+u)−H(s)−D(s)u∣≤u2CJ+2∣u∣ (12)J<∣u∣ ε2+∣u∣ ε4<ε∣u∣,|H(s+u)-H(s)-D(s)u|\le u^{2}C_{J}+2|u|\,(\tfrac12)^{J}<|u|\,\frac{\varepsilon}{2}+|u|\,\frac{\varepsilon}{4}<\varepsilon|u|,

since ∣u∣CJ<δCJ=ε/2|u|C_{J}<\delta C_{J}=\varepsilon/2. Dividing by ∣u∣|u|, ∣(H(s+u)−H(s))/u−D(s)∣<ε|(H(s+u)-H(s))/u-D(s)|<\varepsilon. As ss is an interior point of I=RI=\mathbb{R} and ε\varepsilon was arbitrary, hˉ=H\bar h=H is differentiable at ss in the sense of Single-Variable Calculus on an Interval §derivative, with hˉ′(s)=D(s)\bar h'(s)=D(s), the derivative being unique as recorded there. Hence h′(s)=D(s)h'(s)=D(s) for every s≥0s\ge0.

Clause (Derivative). Since 0<rj0<r_{j} for every jj, (a) gives σj(0)=0\sigma_{j}(0)=0 for every jj, hence h′(0)=D(0)=0h'(0)=D(0)=0. With K=A+1K=A+1 we have ∣h′(s)∣=D(s)≤K|h'(s)|=D(s)\le K for every s≥0s\ge0, by Step 4. Let η>0\eta>0. Choose k∈Nk\in\mathbb{N} with k≥2k\ge2 and (12)k<η(\tfrac12)^{k}<\eta, and put r′′=rk>0r''=r_{k}>0. Let 0≤s<r′′0\le s<r''. For j≤kj\le k we have s<rk≤rjs<r_{k}\le r_{j}, so σj(s)=0\sigma_{j}(s)=0 by (a); in particular σ2(s)=0\sigma_{2}(s)=0 as k≥2k\ge2. The tail bound with μj=2−j\mu_{j}=2^{-j}, wj=σj(s)∈[0,1]w_{j}=\sigma_{j}(s)\in[0,1], M=1M=1 and n=kn=k, whose partial sum up to kk vanishes, then gives 0≤h′(s)=∑j=1∞2−jσj(s)≤(12)k<η0\le h'(s)=\sum_{j=1}^{\infty}2^{-j}\sigma_{j}(s)\le(\tfrac12)^{k}<\eta.

Clause (Majorant). Let s∈[0,r]s\in[0,r]. If s=0s=0 then h(0)=0=m(0)h(0)=0=m(0). If s≥r1s\ge r_{1}, then, as the series part of H(s)H(s) is nonnegative, (d) and r2≤r1/2r_{2}\le r_{1}/2 give

h(s)≥A q2(s)≥A(s−32r2)≥A(r1−34r1)=Ar14=B≥m(s).h(s)\ge A\,q_{2}(s)\ge A\bigl(s-\tfrac32r_{2}\bigr)\ge A\bigl(r_{1}-\tfrac34r_{1}\bigr)=\frac{Ar_{1}}{4}=B\ge m(s).

If 0<s<r10<s<r_{1}, let kk be as in (4). For j≥k+3j\ge k+3 we have rj≤rk+3≤rk+1/4≤s/4r_{j}\le r_{k+3}\le r_{k+1}/4\le s/4, so (d) gives qj(s)≥s−38s≥s/2q_{j}(s)\ge s-\tfrac38s\ge s/2. Define bj=0b_{j}=0 for j≤k+2j\le k+2 and bj=2−js/2b_{j}=2^{-j}s/2 for j≥k+3j\ge k+3. Then 0≤bj≤2−jqj(s)0\le b_{j}\le2^{-j}q_{j}(s) for every jj, by (c) for j≤k+2j\le k+2. The sequence (bj)(b_{j}) differs from (s22−j)(\tfrac{s}{2}2^{-j}) by the finitely supported sequence equal to s22−j\tfrac{s}{2}2^{-j} for j≤k+2j\le k+2 and 00 afterwards, so by linearity of sums and the geometric series tail,

∑j=1∞bj=s2(∑j=1∞2−j−∑j=1k+22−j)=s2(12)k+2=2−k−3s.\sum_{j=1}^{\infty}b_{j}=\frac{s}{2}\Bigl(\sum_{j=1}^{\infty}2^{-j}-\sum_{j=1}^{k+2}2^{-j}\Bigr)=\frac{s}{2}\Bigl(\frac12\Bigr)^{k+2}=2^{-k-3}s.

Comparison of sums and Aq2(s)≥0Aq_{2}(s)\ge0 then give h(s)≥∑j2−jqj(s)≥2−k−3sh(s)\ge\sum_{j}2^{-j}q_{j}(s)\ge2^{-k-3}s. Since s∈[0,r]s\in[0,r] and s<rks<r_{k}, estimate (3) gives m(s)≤2−k−3s≤h(s)m(s)\le2^{-k-3}s\le h(s). This proves h(s)≥m(s)h(s)\ge m(s) for every s∈[0,r]s\in[0,r] and completes the proof.

Citations

Loading…

Dependencies

Uses0

Loading…

Comments

Log in to comment.

Loading…