betaFeedback

WorksMathematics

One-Dimensional Symmetry Reductions of the Monge-Ampère Equation on Some Semi-Homogeneous Convex Cones in R4\mathbb{R}^4

Abstract

Motivated by non-symmetric conic optimization, we construct six explicit families of semi-homogeneous proper convex cones KK in R4\mathbb R^4. For each family, a specified three-dimensional subgroup of linear automorphisms of KK has generic orbits of codimension one. This yields an orbit invariant tt and reduces the Monge-Ampère equation det⁡F′′=e2F\det F''=e^{2F}, defining the canonical barrier on KK, to an ordinary differential equation (ODE) with independent variable tt. The families include asymmetric power, rotational-power, exponential, and nonabelian-type cones. For the power and exponential cases we derive first integrals and parametric solutions of the ODE. The reductions open the way for computing the canonical barrier on these cones numerically.

Introduction

Interior-point methods for conic optimization require efficiently evaluable logarithmically homogeneous self-concordant barriers. Such barriers are explicitly known for symmetric cones [9]. Nonsymmetric examples of cones used in optimization are power and exponential cones [13, 14, 15]. Generalized power cones also have useful modeling and structural properties [12]. All these cones are semi-homogeneous, i.e., their automorphism group Aut⁡(K)\operatorname{Aut}(K) has generic orbits of at most codimension 11 [6].

For every proper convex cone K⊂RnK \subset\mathbb{R}^{n}, the canonical barrier [7, 5] is the unique solution F:K∘→RF: K^{\circ} \to\mathbb{R} of the Monge–Ampère equation

det⁡F′′=e2F,F′′≻0,F∣∂K=+∞.(1)\det F^{\prime\prime}=e^{2F}, \qquad F^{\prime\prime}\succ0, \qquad F|_{\partial K}=+\infty. \tag*{(1)}

It satisfies the logarithmic homogeneity condition

F(λx)=F(x)−nlog⁡λ,λ>0, x∈K∘.(2)F(\lambda x)=F(x)-n\log\lambda,\qquad\lambda>0,\ x\in K^{\circ}. \tag*{(2)}

Here the level sets of FF are the complete hyperbolic affine spheres asymptotic to ∂K\partial K, and existence and uniqueness follow from corresponding work[1, 2, 3, 11, 4, 10, 8].

In this paper we consider explicit families of 4-dimensional semi-homogeneous cones. We designate an explicit connected subgroup G⊂Aut⁡(K)G \subset\operatorname{Aut}(K) with generic codimension 1 orbits. All orbit statements refer to GG unless Aut⁡(K)\operatorname{Aut}(K) is explicitly mentioned. For special cases inside each family Aut⁡(K)\operatorname{Aut}(K) may be transitive on the interior K∘K^\circ of the cone, but the subgroup GG still has orbits of smaller dimension.

We construct six explicit cone families in R4\mathbb{R}^{4}, including new nonabelian-type examples, and give full set of generators for GG. The rotational-power family consists of generalized power cones already used in conic optimization. For this family, our contribution is an explicit reduction of (1) to an ODE. The asymmetric-power and exponential reductions generalize Hildebrand’s three-dimensional power and exponential cases, whereas the two nonabelian-type families and their reduced boundary-value problems are new. For each family an orbit invariant t:K∘→Rt: K^\circ\to\mathbb{R} and an explicit summand F0F_0 satisfying the relation

F(gx)=F(x)−log⁡∣det⁡g∣,g∈Aut⁡(K).(3)F(gx) = F(x) - \log\lvert\det g\rvert,\qquad g \in\operatorname{Aut}(K). \tag*{(3)}

are given such that the canonical barrier can be written as

F=F0+ϕ(t)(4)F = F_0 + \phi(t) \tag*{(4)}

with the function ϕ\phi satisfying the ODE which is obtained by reducing (1), and F0F_0 being logarithmically homogeneous and constant on each orbit of the unimodular subgroup of GG. For the power and exponential cones we construct first integrals and parametric solutions. The nonabelian-type cases (here nonabelian refers to the Lie algebra of GG) yield explicit boundary-value problems. Positivity, smooth matching, and asymptotic behaviour at the boundary select candidate solutions, but uniqueness of the integration constants is asserted only in the former cases.

Preliminaries

A cone K⊂R4K \subset\mathbb{R}^{4} is called proper if it is closed, pointed, convex, and has non-empty interior. Here pointed means K∩(−K)={0}K \cap(-K) = \{0\}. Its automorphism group and Lie algebra are defined by

Aut⁡(K)={A∈GL⁡4(R):A[K]=K},\operatorname{Aut}(K) = \{A \in\operatorname{GL}_{4}(\mathbb{R}) : A[K] = K\},
aut⁡(K)={X∈gl4(R):exp⁡(tX)[K]=K for all t∈R}.\operatorname{aut}(K) = \{X \in\mathfrak{gl}_{4}(\mathbb{R}) : \exp(tX)[K] = K \text{ for all } t \in\mathbb{R}\}.

The canonical barrier satisfies the transformation law (3) [7]. In particular, every unimodular automorphism preserves FF. For the one-parameter subgroups G⊂Aut⁡(K)G \subset\operatorname{Aut}(K) used below we have det⁡g>0\det g > 0, so the absolute value may be omitted. Throughout the paper, GG denotes the connected subgroup generated by the specified Lie algebra for each of the cone families. Statements about generic orbit dimension or codimension refer to GG unless Aut⁡(K)\operatorname{Aut}(K) is mentioned explicitly.

Explicit Cone Families

Throughout the paper, points in R4\mathbb{R}^{4} are written as column vectors (x,y,z,w)T(x,y,z,w)^{T}, and cl⁡Q\operatorname{cl} Q denotes the closure of a subset Q⊂R4Q \subset\mathbb{R}^{4}. For the two nonabelian-type families we use

Δ=2xz−y2.\Delta= 2xz-y^{2}.

We consider the following six explicit families of 4D semi-homogeneous cones. The first 4 cases lead to abelian Lie algebras, the last two are of nonabelian type.

  1. The asymmetric power cone

Kα,βp,q,r=cl⁡{(x,y,z,w)T:x,y,z>0, −αx1/py1/qz1/r<w<βx1/py1/qz1/r}.(5)K_{\alpha,\beta}^{p,q,r}=\operatorname{cl}\left\{(x,y,z,w)^{T}:x,y,z>0,\ -\alpha x^{1/p}y^{1/q}z^{1/r}<w<\beta x^{1/p}y^{1/q}z^{1/r}\right\}. \tag*{(5)}

where α,β>0\alpha,\beta>0, p,q,r>1p,q,r>1, and p−1+q−1+r−1=1p^{-1}+q^{-1}+r^{-1}=1. These assumptions exclude the homogeneous degenerations. E.g., p=q=2p=q=2, r=∞r=\infty and α=β\alpha=\beta leads to K≃L3×R+K\simeq\mathcal{L}_{3}\times\mathbb{R}_{+}, while q=r=∞q=r=\infty yields K≃R+4K\simeq\mathbb{R}_{+}^{4}.

  1. The rotational-power cone

Kλ2,2=cl⁡{(x,y,z,w)T:x,y>0, z2+w2<xλy1−λ}.(6)K_{\lambda}^{2,2}=\operatorname{cl}\left\{(x,y,z,w)^{T}:x,y>0,\ \sqrt{z^{2}+w^{2}}<x^{\lambda}y^{1-\lambda}\right\}. \tag*{(6)}

where α,β>0\alpha,\beta>0 and 0<λ<10<\lambda<1. The exceptional value λ=12\lambda=\frac{1}{2} yields L4\mathcal{L}_{4}: with u=(x+y)/2u=(x+y)/2, and v=(x−y)/2v=(x-y)/2, the inequality becomes u2>v2+z2+w2u^{2}>v^{2}+z^{2}+w^{2}. Further, λ=0,1\lambda=0,1 gives L3×R+\mathcal{L}_{3}\times\mathbb{R}_{+}. In the other cases we get non-homogeneous cones [12].

  1. The exponential-power cone

Kpow,a=cl⁡{(x,y,z,w)T:z,w>0, yz>exp⁡(x/z)(wz)−a},a>0.(7)K_{\mathrm{pow},a}=\operatorname{cl}\left\{(x,y,z,w)^{T}:z,w>0,\ \frac{y}{z}>\exp(x/z)\left(\frac{w}{z}\right)^{-a}\right\},\qquad a>0. \tag*{(7)}
  1. The exponential-quadratic cone

Kpar=cl⁡{(x,y,z,w)T:z>0, yz>exp⁡(xz+(wz)2)}.(8)K_{\mathrm{par}}=\operatorname{cl}\left\{(x,y,z,w)^{T}:z>0,\ \frac{y}{z}>\exp\left(\frac{x}{z}+\left(\frac{w}{z}\right)^{2}\right)\right\}. \tag*{(8)}
  1. The asymmetric mixed-power cone

Kγ,α,βN=cl⁡{(x,y,z,w)T:x,z>0, Δ>0, −αΔγ/2z1−γ<w<βΔγ/2z1−γ}.(9)K_{\gamma,\alpha,\beta}^{N}=\operatorname{cl}\left\{(x,y,z,w)^{T}:x,z>0,\ \Delta>0,\ -\alpha\Delta^{\gamma/2}z^{1-\gamma}<w<\beta\Delta^{\gamma/2}z^{1-\gamma}\right\}. \tag*{(9)}

where α,β>0\alpha,\beta>0 and 0<γ<20<\gamma<2.

  1. The logarithmic-parabolic cone

Klog⁡N=cl⁡{(x,y,z,w)T:x,z>0, Δ>0, w<zlog⁡Δz2}.(10)K_{\log}^{N}=\operatorname{cl}\left\{(x,y,z,w)^{T}:x,z>0,\ \Delta>0,\ w<z\log\frac{\Delta}{z^{2}}\right\}. \tag*{(10)}

All inequalities inside the closure describe the interiors of the respective cones. Their convexity, pointedness, and explicit automorphisms are verified in the next section. Those coordinate variables which are explicitly designated as positive will be referred to as base coordinates further on.

Convexity and Explicit Automorphisms

We now define the subgroup G⊂Aut⁡(K)G \subset\operatorname{Aut}(K) by specifying a basis of its Lie algebra for every cone family. For a displayed Lie algebra element A∈R4×4A \in\mathbb{R}^{4\times4}, the one-parameter group exp⁡(sA)\exp(sA) acts on the column vector (x,y,z,w)T(x,y,z,w)^{T}. The scalar generator

S=(1000010000100001)S=\begin{pmatrix}1&0&0&0\\0&1&0&0\\0&0&1&0\\0&0&0&1\end{pmatrix}

preserves every cone and is always an element of the Lie algebra of GG. We now either give two more generators, or three generators whose sum is SS, for each family. We use only these explicit subgroups and do not claim that the displayed generators exhaust the full automorphism algebra.

Asymmetric power cone. Let h=x1/py1/qz1/rh=x^{1/p}y^{1/q}z^{1/r}. The infinitesimal automorphisms of KK include

Ax=(1000000000000001/p),Ay=(0000010000000001/q),Az=(0000000000100001/r).A_x=\begin{pmatrix}1&0&0&0\\0&0&0&0\\0&0&0&0\\0&0&0&1/p\end{pmatrix},\qquad A_y=\begin{pmatrix}0&0&0&0\\0&1&0&0\\0&0&0&0\\0&0&0&1/q\end{pmatrix},\qquad A_z=\begin{pmatrix}0&0&0&0\\0&0&0&0\\0&0&1&0\\0&0&0&1/r\end{pmatrix}.

Indeed, each positive base-coordinate scaling multiplies ww by the corresponding factor hh, and Ax+Ay+Az=SA_x+A_y+A_z=S. Since the exponents are positive and sum to one, hh is concave and homogeneous of degree one on R+3\mathbb{R}_+^3. Both inequalities w≤βhw\le\beta h and −w≤αh-w\le\alpha h therefore define convex hypographs, proving convexity of (5).

Rotational-power cone. The following three matrices generate base scalings and rotations of the elliptic fibers, respectively:

Ax=(1000000000λ0000λ),Ay=(00000100001−λ00001−λ),A_x=\begin{pmatrix}1&0&0&0\\0&0&0&0\\0&0&\lambda&0\\0&0&0&\lambda\end{pmatrix},\qquad A_y=\begin{pmatrix}0&0&0&0\\0&1&0&0\\0&0&1-\lambda&0\\0&0&0&1-\lambda\end{pmatrix},
R=(00000000000−10010).R=\begin{pmatrix}0&0&0&0\\0&0&0&0\\0&0&0&-1\\0&0&1&0\end{pmatrix}.

Here Ax+Ay=SA_x+A_y=S, while exp⁡(sR)\exp(sR) preserves z2+w2z^2+w^2. The function xλy1−λx^\lambda y^{1-\lambda} is concave, and xλy1−λ−z2+w2x^\lambda y^{1-\lambda}-\sqrt{z^2+w^2} is concave. Its nonnegative superlevel set is exactly (6), hence this cone is convex.

Exponential-power cone. Besides SS, define the Lie algebra generators

B=(0010010000000000),C=(00000−a0000000001).B=\begin{pmatrix}0&0&1&0\\0&1&0&0\\0&0&0&0\\0&0&0&0\end{pmatrix},\qquad C=\begin{pmatrix}0&0&0&0\\0&-a&0&0\\0&0&0&0\\0&0&0&1\end{pmatrix}.

Their actions are

exp⁡(sB)(x,y,z,w)T=(x+sz,esy,z,w)T,\exp(sB)(x,y,z,w)^T=(x+sz,e^s y,z,w)^T,
exp⁡(sC)(x,y,z,w)T=(x,e−asy,z,esw)T,\exp(sC)(x,y,z,w)^T=(x,e^{-as}y,z,e^s w)^T,

so the defining inequality in (7) is invariant. Moreover, f(x,w)=exw−af(x,w)=e^{xw^{-a}} has

∇2f=f⋅(1−a/w−a/wa(a+1)/w2),det⁡∇2f=af2/w2>0.\nabla^2 f=f\cdot\begin{pmatrix}1&-a/w\\-a/w&a(a+1)/w^2\end{pmatrix},\qquad\det\nabla^2 f=af^2/w^2>0.

Thus ff is convex, its perspective zf(x/z,w/z)zf(x/z,w/z) is convex, and (7) is its closed epigraph in the coordinate yy.

Exponential-quadratic cone. Here the basis elements are SS and

B=(0010010000000000),C=(000−2000000000010).B=\begin{pmatrix}0&0&1&0\\0&1&0&0\\0&0&0&0\\0&0&0&0\end{pmatrix},\qquad C=\begin{pmatrix}0&0&0&-2\\0&0&0&0\\0&0&0&0\\0&0&1&0\end{pmatrix}.

The action of BB is the same as above, whereas

exp⁡(sC)(x,y,z,w)T=(x−2sw−s2z,y,z,w+sz)T.\exp(sC)(x,y,z,w)^{T}=(x-2sw-s^{2}z,y,z,w+sz)^{T}.

It leaves x/z+(w/z)2x/z+(w/z)^2 unchanged. The Hessian of f(x,w)=ex+w2f(x,w)=e^{x+w^2} is

∇2f=f(12w2w4w2+2),det⁡∇2f=2f2>0.\nabla^2f=f\begin{pmatrix}1&2w\\2w&4w^2+2\end{pmatrix},\qquad\det\nabla^2f=2f^2>0.

Consequently the perspective zf(x/z,w/z)zf(x/z,w/z) is convex and (8) is a convex epigraph.

Asymmetric mixed-power cone. Besides SS, explicit generators are

D=(200001000000000γ),N=(0100001000000000).D=\begin{pmatrix}2&0&0&0\\0&1&0&0\\0&0&0&0\\0&0&0&\gamma\end{pmatrix},\qquad N=\begin{pmatrix}0&1&0&0\\0&0&1&0\\0&0&0&0\\0&0&0&0\end{pmatrix}.

The action of NN is

(x,y,z,w)T↦(x+sy+12s2z,y+sz,z,w)T,(x,y,z,w)^{T}\mapsto(x+sy+\tfrac{1}{2}s^2z,y+sz,z,w)^{T},

which fixes Δ\Delta and zz. The action of DD multiplies Δ\Delta by e2se^{2s} and ww by eγse^{\gamma s}, and therefore preserves w/(Δγ/2z1−γ)w/(\Delta^{\gamma/2}z^{1-\gamma}). The function Δ\sqrt{\Delta} is concave for positive base coordinates. Since the weighted geometric mean is concave and coordinate-wise increasing on R+2\mathbb{R}^{2}_{+}, Δγ/2z1−γ\Delta^{\gamma/2}z^{1-\gamma} is concave for 0≤γ≤20\le\gamma\le2. The two hypograph inequalities in (9) are consequently convex.

Logarithmic-parabolic cone. Together with SS, use

N=(0100001000000000),L=(2000010000000020).N=\begin{pmatrix}0&1&0&0\\0&0&1&0\\0&0&0&0\\0&0&0&0\end{pmatrix},\qquad L=\begin{pmatrix}2&0&0&0\\0&1&0&0\\0&0&0&0\\0&0&2&0\end{pmatrix}.

The first basis element again fixes Δ\Delta, zz, ww. The second one gives

(x,y,z,w)T↦(e2sx,esy,z,w+2sz)T,(x,y,z,w)^{T}\mapsto(e^{2s}x,e^sy,z,w+2sz)^{T},

which preserves w−zlog⁡(Δ/z2)w-z\log(\Delta/z^2). Finally, the defining inequality is equivalent on z>0z>0 to

x>y22z+z2exp⁡(w/z).x>\frac{y^2}{2z}+\frac{z}{2}\exp(w/z).

The right-hand side is a sum of two convex perspective functions, so its epigraph is convex.

The closures above have nonempty interiors. Pointedness is immediate for the power and rotational-power cones from the nonnegativity of their base coordinates. For the exponential-power cone, the face z=0z=0 has x≤0x\le0, y≥0y\ge0, and w≥0w\ge0; for the exponential-quadratic cone it also has w=0w=0. Hence neither cone contains a line. The base of the asymmetric mixed-power cone is pointed, so this cone is pointed. In the logarithmic case, a vector and its negative can both lie in the closure only if x=y=z=0x=y=z=0 and simultaneously w≤0w\le0 and −w≤0-w\le0, hence w=0w=0. Thus all six sets are proper convex cones.

Lemma 1 (Generic orbit rank). For every family, the infinitesimal action of GG has rank three on an open dense subset of K∘K^\circ, so its generic orbits have codimension one. If t:K∘→Rt: K^\circ\to\mathbb{R} is an orbit invariant and dtx≠0dt_x \ne0, then Tx(G⋅x)=ker⁡dtxT_x(G\cdot x) = \ker dt_x. Hence every smooth GG-invariant function locally factors through tt.

Proof. For each cone family, construct a 4×34 \times3 matrix whose columns are AkxA_kx, where AkA_k is a Lie algebra generator and xx runs through K∘K^\circ. We shall show that this matrix is generically of full rank. Selected 3×33 \times3 minors for the six families are, respectively, xyzxyz, −xyw-xyw or xyzxyz, −wz2-wz^2, −z3-z^3, zΔz\Delta, and −zΔ-z\Delta. They are nonzero in the interior, except on the proper closed rotation axis {z=w=0}\{z=w=0\} in the rotational-power case. Thus the generic rank is three and so is the generic dimension of the GG-orbits. Invariance gives Tx(G⋅x)⊆ker⁡dtxT_x(G\cdot x) \subseteq\ker dt_x. If dtx≠0dt_x \ne0, then both spaces have dimension three. The local factorization follows from the submersion theorem.

Reduction of the Monge–Ampère Equation

Let tt be a complete local invariant of the generic GG-orbits, and let F0:K∘→RF_0: K^\circ\to\mathbb{R} satisfy (3). Then the difference ϕ=F−F0\phi=F-F_0 is GG-invariant and hence F=F0+ϕ(t)F=F_0+\phi(t) locally by Lemma 1. The chain rule gives

F′′=F0′′+ϕ¨ ∇t∇tT+ϕ˙ ∇2t.(11)F''=F_0''+\ddot{\phi}\,\nabla t\nabla t^{\mathsf T}+\dot{\phi}\,\nabla^2t. \tag*{(11)}

Equivariance then reduces PDE (1) to an ODE on ϕ(t)\phi(t) by the determinant calculations displayed below for each case. Dots denote tt-derivatives, and the endpoints of the domain of tt provide the boundary conditions. Below we shall specify in each case an explicit summand F0F_0 and deduce the reduction for the corresponding ϕ\phi. Below, cc and CC both denote real integration constants.

Asymmetric power cone

Set

h=x1/py1/qz1/r,t=wh.h=x^{1/p}y^{1/q}z^{1/r},\qquad t=\frac{w}{h}.

so that −α<t<β-\alpha<t<\beta. Invariance and logarithmic homogeneity give

F=−p+1plog⁡x−q+1qlog⁡y−r+1rlog⁡z+ϕ(t).(12)F=-\frac{p+1}{p}\log x-\frac{q+1}{q}\log y-\frac{r+1}{r}\log z+\phi(t). \tag*{(12)}

For t>0t>0 put

τ=log⁡t,φ=ϕ+τ,ξ=dφdτ=1+tϕ˙,\tau=\log t,\qquad\varphi=\phi+\tau,\qquad\xi=\frac{d\varphi}{d\tau}=1+t\dot{\phi},

and define

A=pqr,P(ξ)=(ξ−1)(ξ+p)(ξ+q)(ξ+r).A=pqr,\qquad P(\xi)=(\xi-1)(\xi+p)(\xi+q)(\xi+r).

Differentiation of (12) by virtue of (11), followed by the matrix determinant lemma, reduces (1) to

P′(ξ)ξdξdτ−P(ξ)=Ae2φ.(13)\frac{P'(\xi)}{\xi}\frac{d\xi}{d\tau}-P(\xi)=Ae^{2\varphi}. \tag*{(13)}

Since dφ/dτ=ξd\varphi/d\tau=\xi, integration gives

e−φP(ξ)=Aeφ+c,(14)e^{-\varphi}P(\xi)=Ae^\varphi+c, \tag*{(14)}

where c∈Rc\in\mathbb{R}. Thus, with

D(ξ)=c2+4AP(ξ),H(ξ)=1ξ−1−1p(ξ+p)−1q(ξ+q)−1r(ξ+r),D(\xi)=c^2+4AP(\xi),\qquad H(\xi)=\frac{1}{\xi-1}-\frac{1}{p(\xi+p)}-\frac{1}{q(\xi+q)}-\frac{1}{r(\xi+r)},

every local nonconstant branch satisfying the algebraic admissibility conditions has the parametrization

t(ξ)=t0exp⁡{∫ξ0ξH(s)2(1+cσD(s)) ds},t(\xi)=t_0\exp\left\{\int_{\xi_0}^{\xi}\frac{H(s)}{2}\left(1+\frac{c\sigma}{\sqrt{D(s)}}\right)\,ds\right\},
φ(ξ)=log⁡(−c+σD(ξ)2At(ξ)),σ∈{−1,1}.(15)\varphi(\xi)=\log\left(\frac{-c+\sigma\sqrt{D(\xi)}}{2At(\xi)}\right),\qquad\sigma\in\{-1,1\}. \tag*{(15)}

Here t0>0t_0>0 and ξ0\xi_0 fix a reference point of the parametrization. One takes components on which D>0D>0, the logarithm is defined, and the Hessian of (12) is positive definite. The formula on t<0t<0 is obtained by reflection. The solution defining the canonical barrier must also satisfy smooth matching at t=0t=0, Hessian positivity, and φ(t)→+∞\varphi(t)\to+\infty as t↓−αt\downarrow-\alpha and t↑βt\uparrow\beta. These conditions constrain the constants, but a general existence-and-uniqueness theorem for the tuple (c,t0,ξ0,σ)(c,t_0,\xi_0,\sigma) is not claimed here. After normalizing β=1\beta=1 and, if necessary, reflecting tt, we may without loss of generality restrict to −2A<c≤0-2A<c\leq0. The limit case c=0c=0 gives the symmetric power cones with α=β\alpha=\beta.

For completeness, in the symmetric case α=β=1\alpha=\beta=1 the final expression is elementary. With ζ∈R\zeta\in\mathbb{R},

t(ζ)=ζ(ζ2+p+1)−1/(2p)(ζ2+q+1)−1/(2q)(ζ2+r+1)−1/(2r),t(\zeta)=\zeta(\zeta^2+p+1)^{-1/(2p)}(\zeta^2+q+1)^{-1/(2q)}(\zeta^2+r+1)^{-1/(2r)},
φ(ζ)=−12log⁡(pqr)+p+12plog⁡(ζ2+p+1)+q+12qlog⁡(ζ2+q+1)+r+12rlog⁡(ζ2+r+1).(16)\varphi(\zeta)=-\frac{1}{2}\log(pqr)+\frac{p+1}{2p}\log(\zeta^2+p+1)+\frac{q+1}{2q}\log(\zeta^2+q+1)+\frac{r+1}{2r}\log(\zeta^2+r+1). \tag*{(16)}

Rotational-power cone

A generalization of this case to arbitrary dimension was considered in [7 Section 6] with a different but isomorphic definition of the cones. Set

ρ=z2+w2,t=ρxλy1−λ∈[0,1),\rho=\sqrt{z^2+w^2},\qquad t=\frac{\rho}{x^\lambda y^{1-\lambda}}\in[0,1),
F=−(1+2λ)log⁡x−(3−2λ)log⁡y+ϕ(t).(17)F = -(1+2\lambda)\log x-(3-2\lambda)\log y+\phi(t). \tag*{(17)}

At the normal section x=y=1x=y=1, z=tz=t, w=0w=0, define

ax=1+2λ+λtϕ˙,ay=3−2λ+(1−λ)tϕ˙,Q=λ2ax+(1−λ)2ay.a_x=1+2\lambda+\lambda t\dot{\phi},\qquad a_y=3-2\lambda+(1-\lambda)t\dot{\phi},\qquad Q=\frac{\lambda^2}{a_x}+\frac{(1-\lambda)^2}{a_y}.

The determinant of the Hessian gives the scalar ODE

ϕ˙taxay[ϕ¨−ϕ˙(ϕ˙+tϕ¨)Q]=e2ϕ.(18)\frac{\dot{\phi}}{t}a_xa_y\left[\ddot{\phi}-\dot{\phi}(\dot{\phi}+t\ddot{\phi})Q\right]=e^{2\phi}. \tag*{(18)}

The formula at t=0t=0 is understood by its smooth limit. To integrate it, put

τ=log⁡t,ψ=ϕ+2τ,ξ=2+tϕ˙,\tau=\log t,\qquad\psi=\phi+2\tau,\qquad\xi=2+t\dot{\phi},
Pλ(ξ)=(ξ−2)2(1+λξ)(1+(1−λ)ξ).P_\lambda(\xi)=(\xi-2)^2(1+\lambda\xi)(1+(1-\lambda)\xi).

Equation (18) has the first integral

eψ−Pλ(ξ)e−ψ=C.(19)e^\psi-P_\lambda(\xi)e^{-\psi}=C. \tag*{(19)}

Here C∈RC\in\mathbb{R}. Let

Sλ(ξ)=C2+4Pλ(ξ),Uλ,σ(ξ)=C+σSλ(ξ)2,σ∈{−1,1}.S_\lambda(\xi)=\sqrt{C^2+4P_\lambda(\xi)},\qquad U_{\lambda,\sigma}(\xi)=\frac{C+\sigma S_\lambda(\xi)}{2},\qquad\sigma\in\{-1,1\}.

On every local branch with Uλ,σ>0U_{\lambda,\sigma}>0 and positive-definite Hessian, a parametric candidate is

τ(ξ)=τ0+∫ξ0ξPλ′(s)sσSλ(s)Uλ,σ(s) ds,\tau(\xi)=\tau_0+\int_{\xi_0}^{\xi}\frac{P'_\lambda(s)}{s\sigma S_\lambda(s)U_{\lambda,\sigma}(s)}\,\mathrm{d}s,
t(ξ)=eτ(ξ),ϕ(ξ)=log⁡Uλ,σ(ξ)−2τ(ξ).(20)t(\xi)=e^{\tau(\xi)},\qquad\phi(\xi)=\log U_{\lambda,\sigma}(\xi)-2\tau(\xi). \tag*{(20)}

The constants τ0∈R\tau_0\in\mathbb{R} and ξ0\xi_0 fix the reference point. A canonical branch must satisfy F′′≻0F''\succ0, smoothness at the rotation axis (ϕ(0)\phi(0) finite and ϕ˙(0)=0\dot{\phi}(0)=0), and ϕ(t)→+∞\phi(t)\to+\infty as t↑1t\uparrow1. We do not prove here that these conditions determine a unique tuple (C,τ0,ξ0,σ)(C,\tau_0,\xi_0,\sigma).

Exponential-quadratic cone

The cone is given by

K={(x,y,0,0)∣x≤0,y≥0}∪{(x,y,z,w)∣z>0,yz≥exp⁡(xz+w2z2)}.K=\{(x,y,0,0)\mid x\leq0,y\geq0\}\cup\left\{(x,y,z,w)\mid z>0,\frac{y}{z}\geq\exp\left(\frac{x}{z}+\frac{w^2}{z^2}\right)\right\}.

Invariance under the unimodular part of GG gives the ansatz F(x,y,z,w)=−log⁡y−3log⁡z+ϕ(t)F(x,y,z,w)=-\log y-3\log z+\phi(t), where t=log⁡yz−xz−(wz)2t=\log\frac{y}{z}-\frac{x}{z}-\left(\frac{w}{z}\right)^2.

The derivatives of FF are given by

F′=(−ϕ˙zϕ˙−1y−3+ϕ˙z+xϕ˙z2+2w2ϕ˙z3−2wϕ˙z2)F'=\begin{pmatrix} -\frac{\dot{\phi}}{z}\\ \frac{\dot{\phi}-1}{y}\\ -\frac{3+\dot{\phi}}{z}+\frac{x\dot{\phi}}{z^{2}}+\frac{2w^{2}\dot{\phi}}{z^{3}}\\ -\frac{2w\dot{\phi}}{z^{2}} \end{pmatrix}

The full Hessian, with no omitted symmetric entries, is

F′′=(ϕ˙z2−ϕ¨yzϕ˙+ϕ¨z2−xϕ¨z3−2w2ϕ¨z42wϕ˙z3−ϕ¨yz−ϕ˙−ϕ¨+1y2−ϕ˙yz+xϕ¨y2z+2w2ϕ¨yz3−2wϕ˙yz2ϕ˙+ϕ¨z2−xϕ¨z3−2w2ϕ¨z4−ϕ˙yz+xϕ¨y2z+2w2ϕ¨yz32z+ϕ˙(1z−2xz2−6w2z4)+ϕ¨(−1z+xz2+2w2z3)−2w(ϕ˙+2ϕ˙)z3−2xϕ˙z4−4w3ϕ˙z52wϕ˙z3−2ϕ˙z22wϕ˙z3−2wϕ˙yz22w(ϕ˙+2ϕ˙)z3−2xwϕ˙z4−4w3ϕ˙z54w2ϕ˙z4−2ϕ˙z2)F''=\begin{pmatrix} \frac{\dot{\phi}}{z^{2}} & -\frac{\ddot{\phi}}{yz} & \frac{\dot{\phi}+\ddot{\phi}}{z^{2}}-\frac{x\ddot{\phi}}{z^{3}}-\frac{2w^{2}\ddot{\phi}}{z^{4}} & \frac{2w\dot{\phi}}{z^{3}}\\ -\frac{\ddot{\phi}}{yz} & -\frac{\dot{\phi}-\ddot{\phi}+1}{y^{2}} & -\frac{\dot{\phi}}{yz}+\frac{x\ddot{\phi}}{y^{2}z}+\frac{2w^{2}\ddot{\phi}}{y z^{3}} & -\frac{2w\dot{\phi}}{y z^{2}}\\ \frac{\dot{\phi}+\ddot{\phi}}{z^{2}}-\frac{x\ddot{\phi}}{z^{3}}-\frac{2w^{2}\ddot{\phi}}{z^{4}} & -\frac{\dot{\phi}}{yz}+\frac{x\ddot{\phi}}{y^{2}z}+\frac{2w^{2}\ddot{\phi}}{y z^{3}} & \frac{2}{z}+\dot{\phi}\left(\frac{1}{z}-\frac{2x}{z^{2}}-\frac{6w^{2}}{z^{4}}\right)+\ddot{\phi}\left(-\frac{1}{z}+\frac{x}{z^{2}}+\frac{2w^{2}}{z^{3}}\right)-\frac{2w(\dot{\phi}+2\dot{\phi})}{z^{3}}-\frac{2x\dot{\phi}}{z^{4}}-\frac{4w^{3}\dot{\phi}}{z^{5}} & \frac{2w\dot{\phi}}{z^{3}}-\frac{2\dot{\phi}}{z^{2}}\\ \frac{2w\dot{\phi}}{z^{3}} & -\frac{2w\dot{\phi}}{y z^{2}} & \frac{2w(\dot{\phi}+2\dot{\phi})}{z^{3}}-\frac{2xw\dot{\phi}}{z^{4}}-\frac{4w^{3}\dot{\phi}}{z^{5}} & \frac{4w^{2}\dot{\phi}}{z^{4}}-\frac{2\dot{\phi}}{z^{2}} \end{pmatrix}

Direct determinant computation gives

det⁡F′′=2ϕ˙(−ϕ˙3+ϕ˙2+4ϕ˙ϕ¨−3ϕ˙)y2z6\det F''=\frac{2\dot{\phi}\left(-\dot{\phi}^{3}+\dot{\phi}^{2}+4\dot{\phi}\ddot{\phi}-3\dot{\phi}\right)}{y^{2}z^{6}}

Consequently, the reduced ODE is

2ϕ˙(−ϕ˙3+ϕ˙2+4ϕ˙ϕ¨−3ϕ˙)=e2ϕ(21)2\dot{\phi}\left(-\dot{\phi}^{3}+\dot{\phi}^{2}+4\dot{\phi}\ddot{\phi}-3\dot{\phi}\right)=e^{2\phi} \tag*{(21)}

By Sylvester’s criterion (using the principal minor indexed by the first, second, and fourth coordinates), positive definiteness is equivalent to the following system:

{ϕ¨>0,ϕ˙<1,2ϕ˙ϕ¨(ϕ˙−1)>0,ϕ¨>ϕ˙2(1−ϕ˙)3−4ϕ˙.(22)\begin{cases} \ddot{\phi}>0,\\ \dot{\phi}<1,\\ 2\dot{\phi}\ddot{\phi}(\dot{\phi}-1)>0,\\ \ddot{\phi}>\dfrac{\dot{\phi}^{2}(1-\dot{\phi})}{3-4\dot{\phi}}. \tag*{(22)} \end{cases}

Let p=ϕ˙p=\dot{\phi} and P(p)=2p3(1−p)P(p)=2p^{3}(1-p). Then the ODE takes the form

ddϕ(e−ϕP(p))=−eϕ\frac{d}{d\phi}\left(e^{-\phi}P(p)\right)=-e^{\phi}

As in [6 Sec. 4.1], this yields −P+c24=(eϕ−c2)2-P+\frac{c^{2}}{4}=\left(e^{\phi}-\frac{c}{2}\right)^{2}. The inequalities in (22) imply P<0P<0. Also, eϕ>0e^{\phi}>0, so

eϕ=c2+−P+c24(23)e^{\phi}=\frac{c}{2}+\sqrt{-P+\frac{c^{2}}{4}} \tag*{(23)}

Let κ=−1ϕ˙\kappa=-\frac{1}{\dot{\phi}}. From (21) we have

dtdϕ˙=1ϕ˙=8ϕ˙2−6ϕ˙2−P+c24(c2+−P+c24)\frac{dt}{d\dot{\phi}}=\frac{1}{\dot{\phi}}=\frac{8\dot{\phi}^{2}-6\dot{\phi}}{2\sqrt{-P+\frac{c^{2}}{4}}\left(\frac{c}{2}+\sqrt{-P+\frac{c^{2}}{4}}\right)}

This integrates to

t=∫0κ8+6κ22+2κ+c24κ4(c2κ2+2+2κ+c24κ4) dκt=\int_{0}^{\kappa}\frac{8+6\kappa}{2\sqrt{2+2\kappa+\frac{c^{2}}{4}\kappa^{4}}\left(\frac{c}{2}\kappa^{2}+\sqrt{2+2\kappa+\frac{c^{2}}{4}\kappa^{4}}\right)}\,d\kappa

As in [6 Sec. 4.1], the affine half-line {(x,1,1,0)∣x≤0}⊂K\{(x,1,1,0)\mid x\leq0\}\subset K intersects all level surfaces of FF. Hence F(x,1,1,0)→−∞F(x,1,1,0)\to-\infty as x→−∞x\to-\infty, which gives ϕ(t)→−∞\phi(t)\to-\infty and eϕ(t)→0e^{\phi(t)}\to0 as t→+∞t\to+\infty. It follows from (23) that c≤0c\leq0.

Let us consider two cases: in the first, we arrive at a contradiction, while in the second, we obtain a parameterized expression for the function.

Case 1: c<0c<0. The integrand above has the asymptotic expansion

8+6κ22+2κ+c24κ4(c2κ2+2+2κ+c24κ4)=3+1κ+O(κ−2),κ→+∞.\frac{8+6\kappa}{2\sqrt{2+2\kappa+\frac{c^{2}}{4}\kappa^{4}}\left(\frac{c}{2}\kappa^{2}+\sqrt{2+2\kappa+\frac{c^{2}}{4}\kappa^{4}}\right)}=3+\frac{1}{\kappa}+O(\kappa^{-2}),\qquad\kappa\to+\infty.

It follows that

t=3κ+log⁡κ+O(1),κ=t3−13log⁡t+O(1).t=3\kappa+\log\kappa+O(1),\qquad\kappa=\frac{t}{3}-\frac{1}{3}\log t+O(1).

Using (23),

ϕ=log⁡(c2+c24+2κ−3+2κ−4)=log⁡(−2c)−3log⁡κ+O(κ−1)=log⁡(−54c)−3log⁡t+O(log⁡tt).\begin{aligned} \phi=\log\left(\frac{c}{2}+\sqrt{\frac{c^{2}}{4}+2\kappa^{-3}+2\kappa^{-4}}\right) \\ &=\log\left(-\frac{2}{c}\right)-3\log\kappa+O(\kappa^{-1}) \\ &=\log\left(-\frac{54}{c}\right)-3\log t+O\left(\frac{\log t}{t}\right). \end{aligned}

Therefore

lim⁡z→0+F(−1,1,z,0)=lim⁡z→0+(−3log⁡z+ϕ(1z−log⁡z))=lim⁡z→0+(−3log⁡z+log⁡(−54c)−3log⁡(z−1−log⁡z))=log⁡(−54c).\begin{aligned} \lim_{z\to0^{+}}F(-1,1,z,0)=\lim_{z\to0^{+}}\left(-3\log z+\phi\left(\frac{1}{z}-\log z\right)\right) \\ &=\lim_{z\to0^{+}}\left(-3\log z+\log\left(-\frac{54}{c}\right)-3\log\left(z^{-1}-\log z\right)\right) \\ &=\log\left(-\frac{54}{c}\right). \end{aligned}

This contradicts the boundary condition F∣∂K=+∞F|_{\partial K}=+\infty. Hence the complete solution cannot correspond to c<0c<0.

Case 2: c=0c=0. We get

t=∫−∞ϕ˙4s2−3s2(s4−s3) ds=∫−∞ϕ˙(32s2−12s+12(s−1)) ds=−32ϕ˙+12log⁡ϕ˙−1ϕ˙.\begin{aligned} t=\int_{-\infty}^{\dot{\phi}}\frac{4s^{2}-3s}{2(s^{4}-s^{3})}\,\mathrm{d}s=\int_{-\infty}^{\dot{\phi}}\left(\frac{3}{2s^{2}}-\frac{1}{2s}+\frac{1}{2(s-1)}\right)\,\mathrm{d}s \\ &=-\frac{3}{2\dot{\phi}}+\frac{1}{2}\log\frac{\dot{\phi}-1}{\dot{\phi}}. \end{aligned}

And now from (23) we obtain

ϕ=12log⁡(2(ϕ˙4−ϕ˙3)).\phi=\frac{1}{2}\log\left(2(\dot{\phi}^{4}-\dot{\phi}^{3})\right).

It yields the parametric representation for the solution ϕ(t)\phi(t), where κ\kappa runs through all positive reals:

(tϕ)=12(log⁡(1+κ)+3κlog⁡(1+κ)−4log⁡κ+log⁡2).\begin{pmatrix}t\\ \phi\end{pmatrix}=\frac{1}{2}\begin{pmatrix}\log(1+\kappa)+3\kappa\\ \log(1+\kappa)-4\log\kappa+\log2\end{pmatrix}.

Exponential-power cones

The exponential-power family is defined as follows:

K={(x,y,0,w)∣x≤0,y≥0,w≥0}K=\{(x,y,0,w)\mid x\leq0,y\geq0,w\geq0\}
∪{(x,y,z,w)∣z>0,w>0,yz≥exp⁡(xz)(wz)−a},a>0\cup\left\{(x,y,z,w)\mid z>0,w>0,\frac{y}{z}\geq\exp\left(\frac{x}{z}\right)\left(\frac{w}{z}\right)^{-a}\right\},\qquad a>0

The equivariant ansatz and orbit invariant are

F=−log⁡y−log⁡w−2log⁡z+ϕ(t),t=log⁡yz−xz+alog⁡wz,F=-\log y-\log w-2\log z+\phi(t),\qquad t=\log\frac{y}{z}-\frac{x}{z}+a\log\frac{w}{z},

The Hessian is

1y2w2z4∣ϕ¨−ϕ¨zyϕ˙−ϕ¨(xz−(a+1))−aϕ¨zw−ϕ¨zyϕ¨+1−ϕ˙ϕ¨(yz−(a+1)yz)aϕ˙ywϕ˙−ϕ¨(xz−(a+1))ϕ¨(yz−(a+1)yz)ϕ¨(xz−(a+1))2+2+ϕ˙(a+1−2xz)aϕ¨(xw−(a+1)zw)−aϕ¨zwaϕ˙ywaϕ¨(xwz2−(a+1)wz)a2ϕ˙+1−aϕ˙∣.\frac{1}{y^{2}w^{2}z^{4}} \left| \begin{array}{cccc} \ddot{\phi} & -\frac{\ddot{\phi}z}{y} & \dot{\phi}-\ddot{\phi}\left(\frac{x}{z}-(a+1)\right) & -a\ddot{\phi}\frac{z}{w}\\ -\frac{\ddot{\phi}z}{y} & \ddot{\phi}+1-\dot{\phi} & \ddot{\phi}\left(\frac{y}{z}-\frac{(a+1)y}{z}\right) & a\dot{\phi}\frac{y}{w}\\ \dot{\phi}-\ddot{\phi}\left(\frac{x}{z}-(a+1)\right) & \ddot{\phi}\left(\frac{y}{z}-\frac{(a+1)y}{z}\right) & \ddot{\phi}\left(\frac{x}{z}-(a+1)\right)^{2}+2+\dot{\phi}(a+1-2\frac{x}{z}) & a\ddot{\phi}\left(\frac{x}{w}-(a+1)\frac{z}{w}\right)\\ -a\ddot{\phi}\frac{z}{w} & a\dot{\phi}\frac{y}{w} & a\ddot{\phi}\left(\frac{xw}{z^{2}}-(a+1)\frac{w}{z}\right) & a^{2}\dot{\phi}+1-a\dot{\phi} \end{array} \right|.

Equation (1) reduces to

−aϕ˙4+aϕ˙3+4aϕ˙2ϕ¨−3aϕ˙ϕ¨+ϕ˙3−ϕ˙2−3ϕ˙ϕ¨+2ϕ¨=e2ϕ.(24)-a\dot{\phi}^{4}+a\dot{\phi}^{3}+4a\dot{\phi}^{2}\ddot{\phi}-3a\dot{\phi}\ddot{\phi}+\dot{\phi}^{3}-\dot{\phi}^{2}-3\dot{\phi}\ddot{\phi}+2\ddot{\phi}=e^{2\phi}. \tag*{(24)}

Let p=ϕ˙p=\dot{\phi}. Equivalently,

M(p)p˙+N(p)=e2ϕ,M(p)\dot{p}+N(p)=e^{2\phi},

where

M(p)=4ap2−3(a+1)p+2,N(p)=−ap4+(a+1)p3−p2.M(p)=4ap^{2}-3(a+1)p+2,\qquad N(p)=-ap^{4}+(a+1)p^{3}-p^{2}.

The cancellation

N′(p)+pM(p)=0N'(p) + pM(p) = 0

gives

ddt(eϕ+N(p)e−ϕ)=e−ϕ(pe2ϕ+N′(p)p˙−pN(p))=e−ϕ(p(Mp˙+N)−pMp˙−pN)=0.\begin{aligned} \frac{d}{dt}\left(e^{\phi} + N(p)e^{-\phi}\right) &= e^{-\phi}\left(pe^{2\phi} + N'(p)\dot{p} - pN(p)\right) \\ &= e^{-\phi}\left(p(M\dot{p}+N)-pM\dot{p}-pN\right)=0. \end{aligned}
eϕ+N(p)e−ϕ=C.(25)e^{\phi}+N(p)e^{-\phi}=C. \tag*{(25)}

The condition for positive definiteness of the Hessian is written as follows:

{ϕ¨>0,ϕ˙<1,aϕ˙<1,Mϕ¨+N>0.\left\{ \begin{aligned} \ddot{\phi} &> 0,\\ \dot{\phi} &< 1,\\ a\dot{\phi} &< 1,\\ M\ddot{\phi}+N &> 0. \end{aligned} \right.

Then −N=ϕ˙2(ϕ˙−1)(aϕ˙−1)>0-N=\dot{\phi}^{2}(\dot{\phi}-1)(a\dot{\phi}-1)>0 and

eϕ=c2+−N+c24(26)e^{\phi}=\frac{c}{2}+\sqrt{-N+\frac{c^{2}}{4}} \tag*{(26)}

From (24) we obtain

dtdϕ=1ϕ˙=Me2ϕ−N=M2−N+c24(c2+−N+c24).\frac{dt}{d\phi}=\frac{1}{\dot{\phi}}=\frac{M}{e^{2\phi}-N} =\frac{M}{2\sqrt{-N+\frac{c^{2}}{4}}\left(\frac{c}{2}+\sqrt{-N+\frac{c^{2}}{4}}\right)}.

This integrates to

t=∫0κ4a+3(a+1)κ+2κ22a+(a+1)κ+κ2+c24κ4(c2κ2+a+(a+1)κ+κ2+c24κ4) dκt=\int_{0}^{\kappa} \frac{4a+3(a+1)\kappa+2\kappa^{2}} {2\sqrt{a+(a+1)\kappa+\kappa^{2}+\frac{c^{2}}{4}\kappa^{4}} \left(\frac{c}{2}\kappa^{2}+\sqrt{a+(a+1)\kappa+\kappa^{2}+\frac{c^{2}}{4}\kappa^{4}}\right)} \,d\kappa

Using (26), we can prove that c≤0c\leq0: affine half-line {(x,1,1,1)∣x≤0}⊂K\{(x,1,1,1)\mid x\leq0\}\subset K intersects all level surfaces of FF. That yields:

lim⁡x→−∞F(x,1,1,1)=−∞⟹lim⁡t→+∞ϕ(t)=−∞⟹lim⁡t→+∞eϕ=0⟹c≤0\lim_{x\to-\infty}F(x,1,1,1)=-\infty \Longrightarrow \lim_{t\to+\infty}\phi(t)=-\infty \Longrightarrow \lim_{t\to+\infty}e^{\phi}=0 \Longrightarrow c\leq0

As in the previous subsection, let us consider two cases: in the first, we arrive at a contradiction, while in the second, we obtain a parameterized expression for the function.

Case 1: c<0c<0. The integrand above has the asymptotic expansion

2+a+1κ+O(κ−2),κ→+∞.2+\frac{a+1}{\kappa}+O(\kappa^{-2}),\qquad\kappa\to+\infty.

This integrates to

t=2κ+(a+1)log⁡κ+O(1)⟹κ=t2−a+12log⁡t+O(1)t = 2\kappa+ (a+1)\log\kappa+ O(1) \Longrightarrow\kappa= \frac{t}{2} - \frac{a+1}{2}\log t + O(1)

We obtain the asymptotics of ϕ\phi using (26):

ϕ=log⁡(−4c)−2log⁡t+O(log⁡tt)\phi= \log\left(-\frac{4}{c}\right) - 2\log t + O\left(\frac{\log t}{t}\right)

Therefore

lim⁡z→0+F(−1,1,z,1)=lim⁡z→0+(−2log⁡z+ϕ(1z−(a+1)log⁡z))=log⁡(−4c).\lim_{z\to0^+} F(-1,1,z,1) = \lim_{z\to0^+}\left(-2\log z + \phi\left(\frac{1}{z}-(a+1)\log z\right)\right) = \log\left(-\frac{4}{c}\right).

Again, this contradicts the boundary condition F∣∂K=+∞F|_{\partial K}=+\infty. Hence the complete solution cannot correspond to c<0c<0.

Case 2: c=0c=0. The integrand simplifies, and it follows that

t=κ+a2log⁡(a+κ)+12log⁡(1+κ)−a2log⁡at = \kappa+ \frac{a}{2}\log(a+\kappa) + \frac{1}{2}\log(1+\kappa) - \frac{a}{2}\log a

Finally, (26) yields the parametric representation for the solution ϕ(t)\phi(t), where κ\kappa runs through all positive reals:

(tϕ)=12(log⁡(1+κ)+alog⁡(a+κ)+2κ−alog⁡alog⁡(1+κ)+log⁡(a+κ)−4log⁡κ).\begin{pmatrix} t \\ \phi\end{pmatrix} = \frac{1}{2} \begin{pmatrix} \log(1+\kappa) + a\log(a+\kappa) + 2\kappa- a\log a \\ \log(1+\kappa) + \log(a+\kappa) - 4\log\kappa \end{pmatrix}.

Nonabelian Models

For the asymmetric mixed-power cones, let

t=wΔγ/2z1−γ,F=−3+γ2log⁡Δ−(1−γ)log⁡z+ϕ(t).t = \frac{w}{\Delta^{\gamma/2}z^{1-\gamma}}, \qquad F = -\frac{3+\gamma}{2}\log\Delta- (1-\gamma)\log z + \phi(t).

Evaluation at the normal section (x,y,z,w)T=(1/2,0,1,t)T(x,y,z,w)^{T}=(1/2,0,1,t)^{T} and direct determinant computation give

(γtϕ˙+γ+3)[γ(γ−2)tϕ˙3−2(γ2−2γ+3)ϕ˙2−4γ(γ−2)tϕ˙ϕ¨+(15+2γ−γ2)ϕ¨]=e2ϕ.(27)(\gamma t\dot{\phi}+\gamma+3)\left[\gamma(\gamma-2)t\dot{\phi}^{3}-2(\gamma^{2}-2\gamma+3)\dot{\phi}^{2}-4\gamma(\gamma-2)t\dot{\phi}\ddot{\phi}+(15+2\gamma-\gamma^{2})\ddot{\phi}\right]=e^{2\phi}. \tag*{(27)}

Thus the canonical barrier is reduced to the following two-point nonlinear boundary-value problem:

t∈(−α,β),xref‘(27)‘ holds,F′′≻0,ϕ(t)⟶+∞as t↓−αand as t↑β.(28)\begin{aligned} t &\in(-\alpha,\beta), \qquad\text{{xref}`(27)` holds}, \qquad F'' \succ0,\\ \phi(t) &\longrightarrow+\infty\quad\text{as } t\downarrow-\alpha\quad\text{and as } t\uparrow\beta. \tag*{(28)} \end{aligned}

Here F′′≻0F''\succ0 refers to the Hessian of the displayed reconstructed four-variable function FF. On the normal section, t=−αt=-\alpha and t=βt=\beta correspond to the two lateral boundary hypersurfaces, along which F0F_0 remains finite. The remaining boundary strata, including Δ=0\Delta= 0 and z=0z = 0, are controlled by the logarithmic terms F0=−3+γ2log⁡Δ−(1−γ)log⁡zF_0 = -\frac{3+\gamma}{2}\log\Delta-(1-\gamma)\log z. Hence the lateral boundary condition is equivalent to the two stated blow-up conditions for ϕ\phi.

For the logarithmic-parabolic cone, use

t=wz−log⁡Δz2<0,F=−32log⁡Δ−log⁡z+ϕ(t).t=\frac{w}{z}-\log\frac{\Delta}{z^2}<0,\qquad F=-\frac{3}{2}\log\Delta-\log z+\phi(t).

At the same normal section the equation becomes

(2ϕ˙+3)[(16ϕ˙+15)ϕ¨−4ϕ˙3−6ϕ˙2]=e2ϕ(29)(2\dot{\phi}+3)\left[(16\dot{\phi}+15)\ddot{\phi}-4\dot{\phi}^{3}-6\dot{\phi}^{2}\right]=e^{2\phi} \tag*{(29)}

The corresponding boundary-value problem is

t∈(−∞,0),xref‘(29)‘ holds,F′′≻0,ϕ(t)→+∞(t↑0),ϕ(t)+4log⁡(−t)→+∞(t→−∞).(30)\begin{aligned} &t\in(-\infty,0),\qquad\text{{{xref}}`(29)` holds},\qquad F''\succ0,\\ &\phi(t)\to+\infty\quad(t\uparrow0),\qquad\phi(t)+4\log(-t)\to+\infty\quad(t\to-\infty). \tag*{(30)} \end{aligned}

Indeed, the second condition follows from logarithmic homogeneity because (−t)−1(1/2,0,1,t)T(-t)^{-1}(1/2,0,1,t)^{T} tends to the boundary ray (0,0,0,−1)T(0,0,0,-1)^{T}. Equations (28) and (30) are the final outputs of the two nonabelian reductions.

This form is also convenient computationally. One may solve the equations on truncated intervals, impose the asymptotic boundary data at their endpoints, and monitor F′′≻0F''\succ0 along the normal section. Logarithmic homogeneity and the explicit automorphisms then reconstruct the barrier on the full cone from the one-dimensional numerical solution.

Conclusion

We constructed six explicit cone families in R4\mathbb{R}^{4} and a complete set of Lie algebra generators for a three-dimensional subgroup G⊂Aut⁡(K)G\subset\operatorname{Aut}(K) in each case. Knowledge of the behaviour of the canonical barrier on generic GG-orbits permits to reduce the Monge–Ampère equation (1) to a scalar ODE. The homogeneous exceptions are explicit: λ=12\lambda=\frac{1}{2} gives L4L_{4}, while zero-weight power degenerations give products L3×R+L_{3}\times\mathbb{R}_{+} or R+4\mathbb{R}_{+}^{4}.

For the abelian power models we obtained first integrals and parametric candidate solutions of the ODE, together with the required positivity, matching, and boundary conditions. General uniqueness of the integration constants is not claimed. The exponential models remain explicit, and the nonabelian models give nonlinear boundary-value problems. The Hessian-based reductions, obtained by direct determinant computation, provide practical starting points for a numerical barrier construction.

References

  1. [1]Calabi, E.: Complete affine hyperspheres. I. In: Symposia Mathematica, vol. 10, pp. 19–38. Academic Press (1972)
  2. [2]Cheng, S.-Y., Yau, S.-T.: On the existence of a complete Kähler metric on noncompact complex manifolds and the regularity of Fefferman’s equation. Communications on Pure and Applied Mathematics 33(4), 507–544 (1980)DOI
  3. [3]Cheng, S.-Y., Yau, S.-T.: The real Monge–Ampère equation and affine flat structures. In: Proceedings of the 1980 Beijing Symposium on Differential Geometry and Differential Equations, vol. 1, pp. 339–370. Science Press, Beijing (1982)
  4. [4]Cheng, S.-Y., Yau, S.-T.: Complete affine hypersurfaces. Part I. The completeness of affine metrics. Communications on Pure and Applied Mathematics 39(6), 839–866 (1986)DOI
  5. [5]Fox, D.J.F.: A Schwarz lemma for Kähler affine metrics and the canonical potential of a proper convex cone. Annali di Matematica Pura ed Applicata 194(1), 1–42 (2015); doi:10.1007/s10231-013-0362-6
  6. [6]Hildebrand, R.: Analytic formulas for complete hyperbolic affine spheres. Beiträge zur Algebra und Geometrie / Contributions to Algebra and Geometry 55(2), 497–520 (2014); arXiv:1305.4814arxiv.org/abs/1305.4814
  7. [7]Hildebrand, R.: Canonical barriers on convex cones. Mathematics of Operations Research 39(3), 841–850 (2014)DOI
  8. [8]Li, A.-M., Simon, U., Zhao, G.: Global Affine Differential Geometry of Hypersurfaces. De Gruyter, Berlin (1993)
  9. [9]Nesterov, Y., Nemirovski, A.: Interior-Point Polynomial Algorithms in Convex Programming. SIAM, Philadelphia (1994)DOI
  10. [10]Nomizu, K., Sasaki, T.: Affine Differential Geometry: Geometry of Affine Immersions. Cambridge University Press, Cambridge (1994)
  11. [11]Sasaki, T.: A note on characteristic functions and projectively invariant metrics on a bounded convex domain. Tokyo Journal of Mathematics 8(1), 49–79 (1985)DOI
  12. [12]Lin, Y., Lindstrom, S.B., Lourenço, B.F., Pong, T.K.: Generalized power cones: optimal error bounds and automorphisms. SIAM Journal on Optimization 34(2), 1369–1404 (2024)
  13. [13]Nesterov, Y.: Towards non-symmetric conic optimization. Optimization Methods and Software 27(4–5), 893–917 (2012)DOI
  14. [14]Papp, D., Yildiz, S.: Matlab package for nonsymmetric conic optimization. INFORMS Journal on Computing 34(1), 11–19 (2022)DOI
  15. [15]Coey, C., Kapelovich, L., Vielma, J.P.: Solving natural conic formulations with Hypatia.jl. INFORMS Journal on Computing 34(5), 2686–2699 (2022); doi:10.1287/ijoc.2022.1202DOI

Paper details

vlnc:W_bft8e5zn
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 2

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents