ScalingStacks

3.4 Legendre transform [0239]

Original official author HTML, exact retained edition. Historical TeX conversion verdicts remain unchanged. Cited-edition alignment and mathematical self-containment are not assessed.

Complete original source context Β· Original author HTML

3.4 Legendre transform

The Legendre transform of the convex function 𝔴\mathfrak{w} is given by

π”΄βˆ—β€‹(p)=sups∈[1,∞)s​pβˆ’π”΄β‘(s).\mathfrak{w}^{*}(p)=\sup_{s\in[1,\infty)}sp-\mathfrak{w}(s).

In particular, we have

π”΄βˆ—β€‹(p)=s​d​𝔴d​sβˆ’π”΄β‘(s)Β at ​p=d​𝔴d​s\mathfrak{w}^{*}(p)=s\frac{d\mathfrak{w}}{ds}-\mathfrak{w}(s)\qquad\text{ at }p=\frac{d\mathfrak{w}}{ds}

We are now going to rewrite the first integral (17) in terms of π”΄βˆ—β€‹(p)\mathfrak{w}^{*}(p). First, since limsβ†’βˆžπ”΄β‘(s)=∞\lim_{s\rightarrow\infty}\mathfrak{w}(s)=\infty we see from (17) that

d​𝔴d​s([1,∞))=[0,pβˆ—=(n2​(nβˆ’1)​w02)1n)\frac{d\mathfrak{w}}{ds}([1,\infty))=[0,p_{*}=\left(\frac{n}{2(n-1)w_{0}^{2}}\right)^{\frac{1}{n}})

and so the Legendre transform is defined on this interval. Furthermore, since d​𝔴d​s​(s=1)=0\frac{d\mathfrak{w}}{ds}(s=1)=0 we have dβ€‹π”΄βˆ—d​p​(p=0)=1\frac{d\mathfrak{w}^{*}}{dp}(p=0)=1, and π”΄βˆ—β€‹(p=0)=βˆ’π”΄β‘(s=1)=βˆ’w0\mathfrak{w}^{*}(p=0)=-\mathfrak{w}(s=1)=-w_{0} by the properties of the Legendre transform. Now from the involution property of the Legendre transform,

𝔴⁑(s⁑(p))=s⁑(p)​pβˆ’π”΄βˆ—β€‹(p)=dβ€‹π”΄βˆ—d​p​pβˆ’π”΄βˆ—β€‹(p).\mathfrak{w}(s(p))=s(p)p-\mathfrak{w}^{*}(p)=\frac{d\mathfrak{w}^{*}}{dp}p-\mathfrak{w}^{*}(p).

The first integral (17) is rewritten as

nβˆ’1n​pn+12​(dβ€‹π”΄βˆ—d​p​pβˆ’π”΄βˆ—β€‹(p))βˆ’2=12​w02,\frac{n-1}{n}p^{n}+\frac{1}{2}\left(\frac{d\mathfrak{w}^{*}}{dp}p-\mathfrak{w}^{*}(p)\right)^{-2}=\frac{1}{2w_{0}^{2}},

namely

(p​dd​pβ€‹π”΄βˆ—βˆ’π”΄βˆ—)2​(1w02βˆ’2​(nβˆ’1)​pnn)=1.(p\frac{d}{dp}\mathfrak{w}^{*}-\mathfrak{w}^{*})^{2}\left(\frac{1}{w_{0}^{2}}-\frac{2(n-1)p^{n}}{n}\right)=1.

To simplify matters we can rescale. Define

y=1pβˆ—β€‹p,g⁑(y)=1w0β€‹π”΄βˆ—.y=\frac{1}{p_{*}}p,\quad g(y)=\frac{1}{w_{0}}\mathfrak{w}^{*}.

Then the ODE is recast on the interval y∈[0,1)y\in[0,1) as

(y​d​gd​yβˆ’g⁑(y))2​(1βˆ’yn)=1,(y\frac{dg}{dy}-g(y))^{2}(1-y^{n})=1, (22)

subject to the initial conditions

g⁑(0)=βˆ’1,g′​(0)=pβˆ—w0=(n2​(nβˆ’1)​w02+n)1n.g(0)=-1,\quad g^{\prime}(0)=\frac{p_{*}}{w_{0}}=\left(\frac{n}{2(n-1)w_{0}^{2+n}}\right)^{\frac{1}{n}}.

This can be integrated explicitly:

Lemma 3.7.

The function

g(y)=βˆ’2F1[12,βˆ’1n,nβˆ’1n;yn]+pβˆ—w0y,g(y)=-\,_{2}F_{1}[\frac{1}{2},-\frac{1}{n},\frac{n-1}{n};y^{n}]+\frac{p_{*}}{w_{0}}y,

where F12​[a,b,c;z]\,{}_{2}F_{1}[a,b,c;z] is the hypergeometric function (38). In particular,

g⁑(1)=pβˆ—w0βˆ’Ξ“β‘(1βˆ’1n)​πΓ⁑(12βˆ’1n).g(1)=\frac{p_{*}}{w_{0}}-\frac{\Gamma(1-\frac{1}{n})\sqrt{\pi}}{\Gamma(\frac{1}{2}-\frac{1}{n})}.
Proof.

First note that if gg solves the equation, then so does g+c​yg+cy for any constant cc. Thus, we may reduce to the initial condition g⁑(0)=βˆ’1,g′​(0)=0g(0)=-1,g^{\prime}(0)=0 below. The initial condition specifies a sign choice of the square root, whence

dd​y​(gy)=1y2​(1βˆ’yn)12,\frac{d}{dy}\left(\frac{g}{y}\right)=\frac{1}{y^{2}(1-y^{n})^{\frac{1}{2}}},

and the Lemma is reduced to Prop. 5.3. ∎

We can implement the matching condition w′​(1)=12​w​(1)w^{\prime}(1)=\frac{1}{2}w(1) as follows. From (21) and the definition of the Legendre transform,

w′​(t=1)=π”΄βˆ—β€‹(pβˆ—)=w0​g​(1)=w0​(pβˆ—w0βˆ’Ξ“β‘(1βˆ’1n)​πΓ⁑(12βˆ’1n)).w^{\prime}(t=1)=\mathfrak{w}^{*}(p_{*})=w_{0}g(1)=w_{0}\left(\frac{p_{*}}{w_{0}}-\frac{\Gamma(1-\frac{1}{n})\sqrt{\pi}}{\Gamma(\frac{1}{2}-\frac{1}{n})}\right).

On the other hand, we have

w⁑(t=1)\displaystyle w(t=1) =limsβ†’βˆžπ”΄β‘(s)s\displaystyle=\lim_{s\rightarrow\infty}\frac{\mathfrak{w}(s)}{s}
=limpβ†’pβˆ—p⁑(dβ€‹π”΄βˆ—/d​p)βˆ’π”΄βˆ—(dβ€‹π”΄βˆ—/d​p)\displaystyle=\lim_{p\rightarrow p_{*}}\frac{p(d\mathfrak{w}^{*}/dp)-\mathfrak{w}^{*}}{(d\mathfrak{w}^{*}/dp)}
=limpβ†’pβˆ—pβˆ’π”΄βˆ—(dβ€‹π”΄βˆ—/d​p)\displaystyle=\lim_{p\rightarrow p_{*}}p-\frac{\mathfrak{w}^{*}}{(d\mathfrak{w}^{*}/dp)}

but from the first integral we have (dβ€‹π”΄βˆ—/d​p)β†’βˆž(d\mathfrak{w}^{*}/dp)\rightarrow\infty as pβ†’pβˆ—p\rightarrow p_{*}, while π”΄βˆ—β†’π”΄βˆ—β€‹(1)<+∞\mathfrak{w}^{*}\rightarrow\mathfrak{w}^{*}(1)<+\infty. Thus we get

w⁑(1)=pβˆ—=(n2​(nβˆ’1)​w02)1/n.w(1)=p_{*}=\left(\frac{n}{2(n-1)w_{0}^{2}}\right)^{1/n}.

All together, the matching amounts to solving the equation

w0​(pβˆ—w0βˆ’Ξ“β‘(1βˆ’1n)​πΓ⁑(12βˆ’1n))=12​pβˆ—,w_{0}\left(\frac{p_{*}}{w_{0}}-\frac{\Gamma(1-\frac{1}{n})\sqrt{\pi}}{\Gamma(\frac{1}{2}-\frac{1}{n})}\right)=\frac{1}{2}p_{*},

Or equivalently

12​(n2​(nβˆ’1)​w02)1/n=w0​Γ⁑(1βˆ’1n)​πΓ⁑(12βˆ’1n)\frac{1}{2}\left(\frac{n}{2(n-1)w_{0}^{2}}\right)^{1/n}=w_{0}\frac{\Gamma(1-\frac{1}{n})\sqrt{\pi}}{\Gamma(\frac{1}{2}-\frac{1}{n})}

which yields

w0=(12)n+1n+2​(n(nβˆ’1))1n+2​(1π​Γ⁑(12βˆ’1n)Γ⁑(1βˆ’1n))nn+2\displaystyle w_{0}=\left(\frac{1}{2}\right)^{\frac{n+1}{n+2}}\left(\frac{n}{(n-1)}\right)^{\frac{1}{n+2}}\left(\frac{1}{\sqrt{\pi}}\frac{\Gamma(\frac{1}{2}-\frac{1}{n})}{\Gamma(1-\frac{1}{n})}\right)^{\frac{n}{n+2}} (23)

which tends to 12\frac{1}{2} as nβ†’+∞n\rightarrow+\infty. For our purpose the important fact is that w0>0w_{0}>0 for nβ‰₯3n\geq 3, which means we have found the initial condition that ensures matching.

Remark 3.8.

If n=2n=2, then g⁑(1)=pβˆ—w0g(1)=\frac{p_{*}}{w_{0}}, so w′​(t=1)=pβˆ—w^{\prime}(t=1)=p_{*}, and w⁑(t=1)=pβˆ—w(t=1)=p_{*}. The matching condition has no positive solution. Correspondingly, our generalization of the Calabi ansatz only yield nontrivial examples in dimension at least three.

Original mathematics by the credited authors. Source-backed reader collection; mathematical self-containment is not assessed.