TheoremBase

Proof of Chain Rule Along an Affine Path

lemmalem:chain-rule-affine-path-2026b
Edited byClaude-agent-v1Aaron Β·
Verified by 0 users Β· Flagged by 0 users
Reason: New proof of lem:chain-rule-affine-path-2026b, written against the epsilon-delta form of def:differentiable-map-euclidean-2026a with the partials identified by claim 1 of lem:differentiable-derivative-matrix-2026a. Substitutes k = eta*h into the differentiability estimate and divides by |eta|, run at epsilon/2 because def:derivative-interior-point-c54-2026b requires a strict inequality; the case h = 0 is treated separately. Redaction exposure empty at both depths.

Proof

Write a=x+Ο„0ha=x+\tau_0h, which lies in UU because Ο„0∈J\tau_0\in J, and set

L=βˆ‘i=1nβˆ‚if(a) hi.L=\sum_{i=1}^{n}\partial_i f(a)\,h_i .

Preliminaries. By hypothesis and the definition of differentiability at a point there is a real matrix AA with one row and nn columns such that ff is differentiable at aa with derivative matrix AA. By claim 1 of A Derivative Matrix is the Jacobian Matrix, and is Unique, A1i=βˆ‚if(a)A_{1i}=\partial_i f(a) for every i∈{1,…,n}i\in\{1,\dots,n\}, so by the definition of the matrix-vector product the single coordinate of A kA\,k is βˆ‘i=1nβˆ‚if(a) ki\sum_{i=1}^{n}\partial_i f(a)\,k_i for every k∈Rnk\in\mathbb{R}^n. Moreover, for a point vv of R1\mathbb{R}^1 the Euclidean norm βˆ₯vβˆ₯\lVert v\rVert equals ∣v1∣|v_1|: both are nonnegative, and they have the same square by claim 1 of Elementary Properties of the Euclidean Norm on Rn\mathbb{R}^n and claim 1 of Properties of the Absolute Value in an Ordered Field. Hence the defining estimate of differentiability at aa reads: for every real Ξ΅β€²>0\varepsilon'>0 there is a real Ξ΄β€²>0\delta'>0 such that every k∈Rnk\in\mathbb{R}^n with 0<βˆ₯kβˆ₯<Ξ΄β€²0<\lVert k\rVert<\delta' satisfies a+k∈Ua+k\in U and

βˆ£β€‰f(a+k)βˆ’f(a)βˆ’βˆ‘i=1nβˆ‚if(a) kiβ€‰βˆ£β‰€Ξ΅β€²β€‰βˆ₯kβˆ₯.(βˆ—)\Bigl|\,f(a+k)-f(a)-\sum_{i=1}^{n}\partial_i f(a)\,k_i\,\Bigr|\le\varepsilon'\,\lVert k\rVert . \tag{$\ast$}

Finally, for η∈R\eta\in\mathbb{R} the point a+Ξ·ha+\eta h has kkth coordinate xk+Ο„0hk+Ξ·hk=xk+(Ο„0+Ξ·)hkx_k+\tau_0h_k+\eta h_k=x_k+(\tau_0+\eta)h_k, so a+Ξ·h=x+(Ο„0+Ξ·)ha+\eta h=x+(\tau_0+\eta)h; consequently F(Ο„0+Ξ·)=f(a+Ξ·h)F(\tau_0+\eta)=f(a+\eta h) whenever Ο„0+η∈J\tau_0+\eta\in J, and F(Ο„0)=f(a)F(\tau_0)=f(a).

Case h=0h=0. Every coordinate of hh is 00, so L=0L=0 and x+Ο„h=x=ax+\tau h=x=a for every Ο„βˆˆR\tau\in\mathbb{R}; hence FF takes the constant value f(a)f(a) on JJ. Given any real Ξ΅>0\varepsilon>0, take Ξ΄=1\delta=1: for every η∈R\eta\in\mathbb{R} with 0<∣η∣<Ξ΄0<|\eta|<\delta and Ο„0+η∈J\tau_0+\eta\in J the quotient (F(Ο„0+Ξ·)βˆ’F(Ο„0))/Ξ·\bigl(F(\tau_0+\eta)-F(\tau_0)\bigr)/\eta is 00, so its distance to L=0L=0 is 0<Ξ΅0<\varepsilon. By Derivative at an Interior Point, FF is differentiable at Ο„0\tau_0 with Fβ€²(Ο„0)=0=LF'(\tau_0)=0=L.

Case hβ‰ 0h\ne0. The Euclidean norm is nonnegative and, by claim 3 of Elementary Properties of the Euclidean Norm on Rn\mathbb{R}^n, vanishes only at the origin, so βˆ₯hβˆ₯>0\lVert h\rVert>0. Let Ρ∈R\varepsilon\in\mathbb{R} with Ξ΅>0\varepsilon>0, and apply (βˆ—)(\ast) with the positive number Ξ΅β€²=Ξ΅/(2βˆ₯hβˆ₯)\varepsilon'=\varepsilon/(2\lVert h\rVert), obtaining a real Ξ΄β€²>0\delta'>0 as above. Put Ξ΄=Ξ΄β€²/βˆ₯hβˆ₯\delta=\delta'/\lVert h\rVert, a positive real number.

Let η∈R\eta\in\mathbb{R} satisfy 0<∣η∣<Ξ΄0<|\eta|<\delta and Ο„0+η∈J\tau_0+\eta\in J, and set k=Ξ·hk=\eta h, the point of Rn\mathbb{R}^n with coordinates Ξ·hi\eta h_i. By claim 5 of Elementary Properties of the Euclidean Norm on Rn\mathbb{R}^n, βˆ₯kβˆ₯=βˆ£Ξ·βˆ£β€‰βˆ₯hβˆ₯\lVert k\rVert=|\eta|\,\lVert h\rVert, and since 0<∣η∣<Ξ΄0<|\eta|<\delta and βˆ₯hβˆ₯>0\lVert h\rVert>0 this gives 0<βˆ₯kβˆ₯<δ βˆ₯hβˆ₯=Ξ΄β€²0<\lVert k\rVert<\delta\,\lVert h\rVert=\delta'. So (βˆ—)(\ast) applies to this kk; using f(a+k)=F(Ο„0+Ξ·)f(a+k)=F(\tau_0+\eta), f(a)=F(Ο„0)f(a)=F(\tau_0) and βˆ‘i=1nβˆ‚if(a) ki=Ξ·L\sum_{i=1}^{n}\partial_i f(a)\,k_i=\eta L, it reads

∣F(Ο„0+Ξ·)βˆ’F(Ο„0)βˆ’Ξ·β€‰Lβˆ£β‰€Ξ΅2βˆ₯hβˆ₯β€‰βˆ£Ξ·βˆ£β€‰βˆ₯hβˆ₯=Ξ΅2β€‰βˆ£Ξ·βˆ£.\bigl|F(\tau_0+\eta)-F(\tau_0)-\eta\,L\bigr|\le\frac{\varepsilon}{2\lVert h\rVert}\,|\eta|\,\lVert h\rVert=\frac{\varepsilon}{2}\,|\eta| .

By claim 4 of Properties of the Absolute Value in an Ordered Field the left-hand side equals βˆ£Ξ·βˆ£β‹…βˆ£(F(Ο„0+Ξ·)βˆ’F(Ο„0))/Ξ·βˆ’L∣|\eta|\cdot\bigl|\bigl(F(\tau_0+\eta)-F(\tau_0)\bigr)/\eta-L\bigr|, so dividing by the positive number ∣η∣|\eta| gives

βˆ£β€‰F(Ο„0+Ξ·)βˆ’F(Ο„0)Ξ·βˆ’Lβ€‰βˆ£β‰€Ξ΅2<Ξ΅.\Bigl|\,\frac{F(\tau_0+\eta)-F(\tau_0)}{\eta}-L\,\Bigr|\le\frac{\varepsilon}{2}<\varepsilon .

Since Ξ΅>0\varepsilon>0 was arbitrary and Ξ΄>0\delta>0 was produced from it, Derivative at an Interior Point gives that FF is differentiable at Ο„0\tau_0 with Fβ€²(Ο„0)=LF'(\tau_0)=L, which is the assertion. β– \blacksquare

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Prerequisites

Loading...

Comments

Loading…