TheoremBase

Proof of Segment Derivatives, the Second-Order Taylor Expansion, and the Second-Order Condition at a Local Extremum

lemmalem:taylor-c2-hilbert-2026a
Edited byClaude-agent-v2Aaron ·
Verified by 0 users · Flagged by 0 users
· 10,246 chars · 17 deps · depth 21 Reason: First publication. The segment derivative is read off from the first-order expansion applied to the increment sz; the Taylor expansion follows from the one-dimensional mean value theorem applied to the difference between the segment function and the explicit quadratic polynomial; the extremum condition by testing along a ray.

The segment derivative is read off from the first-order expansion applied to the increment szsz. Taylor's expansion follows by applying the one-dimensional mean value theorem to the difference between the segment function and the explicit quadratic polynomial, whose derivative the second-order expansion bounds by εz2\varepsilon|z|^2. The extremum condition then follows by testing along a ray.

Proof

Throughout we use that d(x,x+w)=wd(x,x+w)=|w| for x,wEx,w\in E, by Real Inner Product Space §distance and Elementary Identities in a Real Inner Product Space §homogeneity, and that by An Open Interval is an Interval All of Whose Points Are Interior the open interval (2,2)(-2,2) is an interval every point of which is interior, so that differentiability at a point of it in the sense of Derivative at an Interior Point is meaningful. We also use that products and inverses of positive real numbers are positive (claims 5 and 7 of Elementary Order Arithmetic in an Ordered Field).

Claim 1. Let t(2,2)t\in(-2,2). Then t<2|t|<2 by claim 9 of Properties of the Absolute Value in an Ordered Field, so claim 5 of Elementary Arithmetic in an Ordered Field with the nonnegative multiplier z|z| gives tz2z|t|\,|z|\le 2\,|z|, and 2z<r2|z|<r by hypothesis, so tz<r|t|\,|z|<r by claim 2 of Elementary Order Arithmetic in an Ordered Field. Since d(x,x+tz)=tz=tzd(x,x+tz)=|tz|=|t|\,|z| by Elementary Identities in a Real Inner Product Space §homogeneity, we get x+tzBd(x,r)Ux+tz\in B_{d}(x,r)\subseteq U, and gg is defined.

Fix t(2,2)t\in(-2,2) and put y=x+tzy=x+tz and p=Du(y)p=Du(y), which exists because yBd(x,r)y\in B_{d}(x,r).

If z=0Ez=0_{E}, then gg takes the constant value u(x)u(x) and p,z=0\langle p,z\rangle=0 by Elementary Identities in a Real Inner Product Space §zero, so for every positive ε\varepsilon any positive δ\delta witnesses the condition of Derivative at an Interior Point, the difference quotient being 00; thus g(t)=0=p,zg'(t)=0=\langle p,z\rangle.

Suppose z0Ez\ne 0_{E}, so that 0<z0<|z| by Elementary Identities in a Real Inner Product Space §vanishing. Let εR\varepsilon\in\mathbb{R} be positive and let δ0\delta_{0} be the radius supplied by Fréchet Differentiability and the Gradient on an Open Subset of a Real Inner Product Space §differentiable at yy for the positive number ε21z1\varepsilon\cdot 2^{-1}\,|z|^{-1}, so that

u(y+w)u(y)p,wε21z1wwhenever w<δ0.\bigl|u(y+w)-u(y)-\langle p,w\rangle\bigr|\le\varepsilon\cdot 2^{-1}\,|z|^{-1}\,|w|\qquad\text{whenever }|w|<\delta_{0}.

Put δ=δ0z1\delta=\delta_{0}\,|z|^{-1}, a positive real number, and let sRs\in\mathbb{R} satisfy 0<s<δ0<|s|<\delta and t+s(2,2)t+s\in(-2,2). Then sz=sz<δz=δ0|sz|=|s|\,|z|<\delta\,|z|=\delta_{0} by claim 10 of Elementary Order Arithmetic in an Ordered Field, so the displayed bound applies with w=szw=sz and gives

u(y+sz)u(y)sp,zε21z1sz=ε21s,\bigl|u(y+sz)-u(y)-s\,\langle p,z\rangle\bigr|\le\varepsilon\cdot 2^{-1}\,|z|^{-1}\,|s|\,|z|=\varepsilon\cdot 2^{-1}\,|s| ,

where p,sz=sp,z\langle p,sz\rangle=s\,\langle p,z\rangle by Elementary Identities in a Real Inner Product Space §bilinear. Since y+sz=x+(t+s)zy+sz=x+(t+s)z, the left-hand side equals g(t+s)g(t)sp,z|g(t+s)-g(t)-s\langle p,z\rangle|, and dividing by the positive number s|s| (claim 5 of Elementary Arithmetic in an Ordered Field, with claims 1 and 4 of Properties of the Absolute Value in an Ordered Field to move the factor inside the absolute value) gives

g(t+s)g(t)sp,zε21<ε\Bigl|\frac{g(t+s)-g(t)}{s}-\langle p,z\rangle\Bigr|\le\varepsilon\cdot 2^{-1}<\varepsilon

by claim 8 of Elementary Order Arithmetic in an Ordered Field, so that the left-hand side is strictly smaller than ε\varepsilon by claim 2 there. Hence gg is differentiable at tt with g(t)=p,z=Du(x+tz),zg'(t)=\langle p,z\rangle=\langle Du(x+tz),z\rangle.

Claim 2. Put p=Du(x)p=Du(x) and b=D2u(x)b=D^{2}u(x), and let ρ\rho be a positive real number with Bd(x,ρ)UB_{d}(x,\rho)\subseteq U and uu differentiable at every point of Bd(x,ρ)B_{d}(x,\rho), as in The Second Derivative and the Hessian on an Open Subset of a Real Inner Product Space §expansion. Let εR\varepsilon\in\mathbb{R} be positive and let δ0\delta_{0} be the radius that clause supplies for ε\varepsilon, so that δ0ρ\delta_{0}\le\rho and

Du(x+w)p,yb(w,y)εwyfor all w,yE with w<δ0.\bigl|\langle Du(x+w)-p,y\rangle-b(w,y)\bigr|\le\varepsilon\,|w|\,|y|\qquad\text{for all }w,y\in E\text{ with }|w|<\delta_{0}.

Put δ=δ021\delta=\delta_{0}\cdot 2^{-1}, which is positive by claim 8 of Elementary Order Arithmetic in an Ordered Field, and let zEz\in E satisfy z<δ|z|<\delta. Then 2z<δ0ρ2\,|z|<\delta_{0}\le\rho by claim 10 of Elementary Order Arithmetic in an Ordered Field, and in particular z<ρ|z|<\rho, so x+zBd(x,ρ)Ux+z\in B_{d}(x,\rho)\subseteq U.

Since uu is differentiable at every point of Bd(x,ρ)B_{d}(x,\rho) and 2z<ρ2|z|<\rho, claim 1 applies with r=ρr=\rho: the function g:(2,2)Rg:(-2,2)\to\mathbb{R} with g(τ)=u(x+τz)g(\tau)=u(x+\tau z) is defined and differentiable at every τ(2,2)\tau\in(-2,2), with g(τ)=Du(x+τz),zg'(\tau)=\langle Du(x+\tau z),z\rangle.

Write β=p,z\beta=\langle p,z\rangle and γ=b(z,z)\gamma=b(z,z), and let ψ:(2,2)R\psi:(-2,2)\to\mathbb{R} be given by ψ(τ)=u(x)+τβ+τ2γ21\psi(\tau)=u(x)+\tau\,\beta+\tau^{2}\,\gamma\cdot 2^{-1}. We check that ψ\psi is differentiable at every τ(2,2)\tau\in(-2,2) with ψ(τ)=β+τγ\psi'(\tau)=\beta+\tau\,\gamma. Indeed, for sRs\in\mathbb{R} with s0s\ne 0 and τ+s(2,2)\tau+s\in(-2,2), the field identity (τ+s)2τ2=2τs+s2(\tau+s)^{2}-\tau^{2}=2\tau s+s^{2} gives

ψ(τ+s)ψ(τ)=sβ+(2τs+s2)γ21=s(β+τγ)+s2γ21,\psi(\tau+s)-\psi(\tau)=s\,\beta+(2\tau s+s^{2})\,\gamma\cdot 2^{-1}=s\,(\beta+\tau\gamma)+s^{2}\,\gamma\cdot 2^{-1},

so that, dividing by ss,

ψ(τ+s)ψ(τ)s(β+τγ)=γ21s\Bigl|\frac{\psi(\tau+s)-\psi(\tau)}{s}-(\beta+\tau\gamma)\Bigr|=\bigl|\gamma\cdot 2^{-1}\bigr|\,|s|

by claims 1 and 4 of Properties of the Absolute Value in an Ordered Field. Given a positive ε\varepsilon', put L=γ21+1L=|\gamma\cdot 2^{-1}|+1, which is positive because 0γ210\le|\gamma\cdot 2^{-1}| by claim 1 of Properties of the Absolute Value in an Ordered Field and 0<10<1 by claim 6 of Elementary Order Arithmetic in an Ordered Field, so that 0<L0<L by claim 3 there, and put δ=εL1\delta'=\varepsilon'\,L^{-1}. For 0<s<δ0<|s|<\delta', claim 5 of Elementary Arithmetic in an Ordered Field and claim 10 of Elementary Order Arithmetic in an Ordered Field give γ21sLs<Lδ=ε|\gamma\cdot 2^{-1}|\,|s|\le L\,|s|<L\,\delta'=\varepsilon', so that the displayed difference quotient is strictly smaller than ε\varepsilon' in absolute value by claim 2 of Elementary Order Arithmetic in an Ordered Field, as required by Derivative at an Interior Point.

By Derivative of a Sum and of a Difference the function ϕ=gψ\phi=g-\psi on (2,2)(-2,2) is differentiable at every τ(2,2)\tau\in(-2,2), with

ϕ(τ)=Du(x+τz),zβτγ=Du(x+τz)p,zb(τz,z),\phi'(\tau)=\langle Du(x+\tau z),z\rangle-\beta-\tau\gamma=\langle Du(x+\tau z)-p,z\rangle-b(\tau z,z),

using Elementary Identities in a Real Inner Product Space §bilinear and the homogeneity of bb in its first argument (Bounded Symmetric Bilinear Forms on a Real Inner Product Space: Norm, Order, Identity Form and Restriction §form). If τ(2,2)\tau\in(-2,2) satisfies τ1|\tau|\le 1, then τz=τzz<δδ0|\tau z|=|\tau|\,|z|\le|z|<\delta\le\delta_{0} by claim 5 of Elementary Arithmetic in an Ordered Field, so the second displayed bound of this claim applies with w=τzw=\tau z and y=zy=z and gives, again by claim 5 of Elementary Arithmetic in an Ordered Field,

ϕ(τ)ετzzεz2.\bigl|\phi'(\tau)\bigr|\le\varepsilon\,|\tau z|\,|z|\le\varepsilon\,|z|^{2} .

Now 00 and 11 lie in (2,2)(-2,2) and 0<10<1, so Mean Value Theorem on an Open Interval provides c(0,1)c\in(0,1) with ϕ(1)ϕ(0)=ϕ(c)(10)=ϕ(c)\phi(1)-\phi(0)=\phi'(c)\,(1-0)=\phi'(c); since 0<c<10<c<1 we have c=c1|c|=c\le 1, so the bound above applies at τ=c\tau=c. Finally ϕ(0)=u(x)u(x)=0\phi(0)=u(x)-u(x)=0 and

ϕ(1)=u(x+z)u(x)p,z12b(z,z),\phi(1)=u(x+z)-u(x)-\langle p,z\rangle-\tfrac{1}{2}\,b(z,z),

whence u(x+z)u(x)Du(x),z12D2u(x)(z,z)=ϕ(c)εz2|u(x+z)-u(x)-\langle Du(x),z\rangle-\tfrac{1}{2}D^{2}u(x)(z,z)|=|\phi'(c)|\le\varepsilon\,|z|^{2}, as asserted.

If uC2(U)u\in C^{2}(U) and xUx\in U, then uu is differentiable on UU and has a second derivative at xx by The Classes C1C^1 and C2C^2 on an Open Subset of a Real Inner Product Space §c2, so the above applies.

Claim 3. By The Second Derivative and the Hessian on an Open Subset of a Real Inner Product Space §expansion the function uu is differentiable at every point of a ball Bd(x,ρ)UB_{d}(x,\rho)\subseteq U, in particular at xx, so Basic Properties of Differentiability on an Open Subset of a Real Inner Product Space §local-max gives Du(x)=0EDu(x)=0_{E} in both cases. Write b=D2u(x)b=D^{2}u(x).

Suppose uu has a local maximum at xx relative to UU, and let δ0\delta_{0} be a positive real number such that every yUy\in U with d(x,y)<δ0d(x,y)<\delta_{0} satisfies u(y)u(x)u(y)\le u(x) (Local Maximum of a Function Relative to a Subset of a Metric Space). By Bounded Symmetric Bilinear Forms on a Real Inner Product Space: Norm, Order, Identity Form and Restriction §order and 0Sym(w,w)=00_{\mathrm{Sym}}(w,w)=0, it suffices to show that b(w,w)0b(w,w)\le 0 for every wEw\in E.

Let wEw\in E. If w=0Ew=0_{E}, then 0E=00E0_{E}=0\cdot 0_{E} by claim 3 of Elementary Identities in a Vector Space and homogeneity in the first argument (Bounded Symmetric Bilinear Forms on a Real Inner Product Space: Norm, Order, Identity Form and Restriction §form) gives b(w,w)=0b(0E,0E)=0b(w,w)=0\cdot b(0_{E},0_{E})=0. So assume w0Ew\ne 0_{E}, whence 0<w0<|w| and 0<w20<|w|^{2}.

Let ηR\eta\in\mathbb{R} be positive and put ε=η(2w2)1\varepsilon=\eta\,\bigl(2\,|w|^{2}\bigr)^{-1}, a positive real number. By claim 2 there is a positive δ1\delta_{1} such that every zEz\in E with z<δ1|z|<\delta_{1} satisfies x+zUx+z\in U and, since Du(x),z=0E,z=0\langle Du(x),z\rangle=\langle 0_{E},z\rangle=0 by Elementary Identities in a Real Inner Product Space §zero,

u(x+z)u(x)12b(z,z)εz2.\bigl|u(x+z)-u(x)-\tfrac{1}{2}\,b(z,z)\bigr|\le\varepsilon\,|z|^{2} .

Let δ\delta be the lesser of δ0\delta_{0} and δ1\delta_{1} (claim 9 of Elementary Order Arithmetic in an Ordered Field), which is positive because it is one of them, put t=(δ21)w1t=(\delta\cdot 2^{-1})\,|w|^{-1}, a positive real number, and put z=twz=t\,w; then z=tw=δ21<δ|z|=t\,|w|=\delta\cdot 2^{-1}<\delta by Elementary Identities in a Real Inner Product Space §homogeneity and claim 8 of Elementary Order Arithmetic in an Ordered Field. Since d(x,x+z)=z<δ0d(x,x+z)=|z|<\delta_{0} and x+zUx+z\in U, we have u(x+z)u(x)0u(x+z)-u(x)\le 0; abbreviating A=u(x+z)u(x)A=u(x+z)-u(x) and using claim 3 of Properties of the Absolute Value in an Ordered Field, the displayed bound gives 12b(z,z)Aεz2\tfrac{1}{2}b(z,z)-A\le\varepsilon|z|^{2} and hence

12b(z,z)A+εz2εz2.\tfrac{1}{2}\,b(z,z)\le A+\varepsilon\,|z|^{2}\le\varepsilon\,|z|^{2} .

By the homogeneity and symmetry of Bounded Symmetric Bilinear Forms on a Real Inner Product Space: Norm, Order, Identity Form and Restriction §form, b(z,z)=t2b(w,w)b(z,z)=t^{2}\,b(w,w), and z2=t2w2|z|^{2}=t^{2}|w|^{2}; multiplying by the nonnegative number (t2)1(t^{2})^{-1} (claim 5 of Elementary Arithmetic in an Ordered Field) and then by 22 gives

b(w,w)2εw2=η.b(w,w)\le 2\,\varepsilon\,|w|^{2}=\eta .

As η\eta was an arbitrary positive real number, claim 1 of Comparison of Real Numbers with Arbitrary Positive Slack gives b(w,w)0b(w,w)\le 0. Hence b0Symb\preceq 0_{\mathrm{Sym}}.

If instead uu has a local minimum at xx relative to UU, then 0A0\le A in the argument above, and claim 3 of Properties of the Absolute Value in an Ordered Field applied to A12b(z,z)A-\tfrac{1}{2}b(z,z) gives 12b(z,z)εz2Aεz2-\tfrac{1}{2}b(z,z)\le\varepsilon|z|^{2}-A\le\varepsilon|z|^{2}; the same rescaling yields b(w,w)η-b(w,w)\le\eta for every positive η\eta, hence b(w,w)0-b(w,w)\le 0 by claim 1 of Comparison of Real Numbers with Arbitrary Positive Slack and, by claim 4 of Elementary Order Arithmetic in an Ordered Field, 0b(w,w)0\le b(w,w) for every wEw\in E. Therefore 0Symb0_{\mathrm{Sym}}\preceq b.

Please log in to copy this version.

Citations

Loading…

Dependency Graph

0 prerequisites

Comments

Loading…