1 Introduction
For
$\alpha>0$
and nonnegative
$f\in L^1(\mathbb {S}^n)$
with positive integral, we are interested in finding a weak solution to the Monge–Ampére equation

or in other words, a weak solution to Lutwak’s
$L^p$
-Minkowski problem on
$S^n$
when
$-n-1<p<1$
for
$p=1-\frac 1\alpha $
where
$\bar {\nabla }$
is the Levi-Civita connection of
$\mathbb {S}^n$
,
$\bar {g}_{ij}$
, with
$\bar {g}$
being the induced round metric on the unit sphere. By a weak (Alexandrov) solution, we mean the following: Given a nontrivial finite Borel measure
$\mu $
on
$\mathbb {S}^n$
(for example,
$d\mu =f\,d\theta $
for the Lebesgue measure
$\theta $
on
$S^n$
and the f in (1.1)), find a convex body
$\Omega \subset \Bbb R^{n+1}$
with
$o\in \Omega $
such that

where
$u(x)=\max _{z\in \Omega }\langle x,z\rangle $
is the support function and
$S_\Omega $
is the surface area measure of
$\Omega $
(see [Reference Schneider45]). If
$\partial \Omega $
is
$C^2_+$
, then

where
$K(x)$
is the Gaussian curvature at the point of
$\partial \Omega $
where
$x\in S^n$
is the exterior unit normal (see [Reference Schneider45]). Concerning the regularity of the solution of (1.1), if
$f\in C^{0,\beta }(S^n)$
and u are positive, then u is
$C^{2,\beta }$
according to Caffarelli’s regularity theory in [Reference Caffarelli15, Reference Caffarelli16]. On the other hand, even if f is positive and continuous for
$\alpha>\frac 1n$
, there might exist weak solution where
$u(x)=0$
for some
$x\in S^n$
and u is not even
$C^1$
according to Example 4.2 in [Reference Bianchi, Böröczky and Colesanti7]. Moreover, even if
$f\in C^{0,\beta }(S^n)$
is positive, it is possible that
$u(x)=0$
for some
$x\in S^n$
for
$\alpha>\frac 1n$
, but Choi, Kim, and Lee [Reference Choi, Kim and Lee19] still managed to obtain some regularity results in this case.
The case
$\alpha =\frac 1{n+2}$
of the Monge–Ampére equation (1.1) is the critical case when the left-hand side of (1.1) is invariant under linear transformations of
$\Omega $
, and the case
$\alpha =1$
is the so-called logarithmic Minkowski problem posed by Firey [Reference Firey23]. Setting
$p=1-\frac 1\alpha <1$
, the Monge–Ampére equation (1.1) is Lutwak’s
$L^p$
-Minkowski problem

In this notation, (1.2) reads as

that equation makes sense for any
$p\in \Bbb R$
. Within the rapidly developing
$L^p$
-Brunn–Minkowski theory (where
$p=1$
is the classical case originating from Minkowski’s oeuvre) initiated by Lutwak [Reference Lutwak39–Reference Lutwak41], if
$p>1$
and
$p\neq n+1$
, then Hug, Lutwak, Yang, and Zhang [Reference Hug, Lutwak, Yang and Zhang30] (improving on Chou and Wang [Reference Chou and Wang20]) prove that (1.4) has an Alexandrov solution if and only if the
$\mu $
is not concentrated onto any closed hemisphere, and the solution is unique. We note that there are examples in [Reference Guan and Lin25] (see also [Reference Hug, Lutwak, Yang and Zhang30]) and show that if
$1<p<n+1$
, then it may happen that the density function f is a positive continuous in (1.3) and
$o\in \partial K$
holds for the unique Alexandrov solution, and actually Bianchi, Böröczky, and Colesanti [Reference Bianchi, Böröczky and Colesanti7] exhibit an example that
$o\in \partial K$
even if the density function f is a positive continuous in (1.3) assuming
$-n-1<p<1$
.
In the case
$p\in (0,1)$
(or equivalently,
$\alpha>1$
), if the measure
$\mu $
is not concentrated onto any great subsphere of
$S^n$
, then Chen, Li, and Zhu [Reference Chen, Li and Zhu17] prove that there exists an Alexandrov solution
$K\in \mathcal {K}_o^n$
of (1.4) using a variational argument (see also [Reference Bianchi, Böröczky, Colesanti and Yang8]). We note that for
$p\in (0,1)$
and
$n\geq 2$
, no complete characterization of
$L^p$
-surface area measures is known (see [Reference Böröczky and Trinh12] for the case
$n=1$
, and [Reference Bianchi, Böröczky, Colesanti and Yang8, Reference Saroglou43] for partial results about the case when
$n\geq 2$
and the support of
$\mu $
is contained in a great subsphere of
$S^n$
).
Concerning the case
$p=0$
(or equivalently,
$\alpha =1$
), the still open logarithmic Minkowski problem (1.3) or (1.4) was posed by Firey [Reference Firey23] in 1974. The paper [Reference Böröczky, Lutwak, Yang and Zhang11] characterized even measures
$\mu $
such that (1.4) has an even solution for
$p=0$
by the so-called subspace concentration condition (see (a) and (b) in Theorem 1.1). In general, Chen, Li, and Zhu [Reference Chen, Li and Zhu18] proved that if a nontrivial finite Borel measure
$\mu $
on
$S^{n-1}$
satisfies the same subspace concentration condition, then (1.4) has a solution for
$p=0$
. On the other hand, Böröczky and Hegedus [Reference Böröczky and Hegedűs10] provide conditions on the restriction of the
$\mu $
in (1.4) to a pair of antipodal points.
If
$-n-1<p<0$
(or equivalently,
$\frac 1{n+2}<\alpha <1$
) and
$f\in L_{\frac {n+1}{n+1+p}}(S^{n})$
in (1.3), then (1.3) has a solution according to [Reference Bianchi, Böröczky, Colesanti and Yang8]. For a rather special discrete measure
$\mu $
satisfying that
$\mu $
is not concentrated on any closed hemisphere and any n unit vectors in the support of
$\mu $
are independent, Zhu [Reference Zhu47] solves the
$L^p$
-Minkowski problem (1.4) for
$p<0$
. The
$p=-n-1$
(or equivalently,
$\alpha =\frac 1{n+2}$
) case of the
$L^p$
-Minkowski problem is the critical case because its link with the
$\mathrm {SL}(n)$
invariant centro-affine curvature whose reciprocal is
$u^{n+2}\det (\bar {\nabla }^2_{ij} u+u\bar {g}_{ij})$
(see [Reference Hug29] or [Reference Ludwig38]). For positive results concerning the critical case
$p=-n-1$
, see, for example, [Reference Guang, Li and Wang28, Reference Jian, Lu and Zhu34], and for obstructions for a solution, see, for example, [Reference Chou and Wang20, Reference Du22].
In the super-critical case
$p<-n-1$
(or equivalently,
$\alpha <\frac 1{n+2}$
), there is a recent important work by Li, Guang, and Wang [Reference Guang, Li and Wang27] proving that for any positive
$C^2$
function f, there exists a
$C^4$
solution of (1.3). See also [Reference Du22] for non-existence examples.
The main contribution of this paper is to provide a very natural argument based on anisotropic flows developed by Andrews [Reference Andrews4] to handle the case
$-n-1<p<1$
, or equivalently, the case
$\frac 1{n+2}<\alpha <\infty $
.
Entropy functional. For any convex body
$\Omega $
, a fixed positive function f on
$\mathbb {S}^n$
and
$\alpha \in (0, \infty )$
, define

where

Here,
$u_{z}(x):=\sup _{y\in \Omega }\left \langle y-z,x\right \rangle $
is the support function of
$\Omega $
in direction x with respect to
$z_0$
and
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n} h(x)\, d\theta (x)=\frac {1}{\omega _n} \int _{\mathbb {S}^n} h(x)$
with
$\omega _n$
being the surface area of
$\mathbb {S}^n$
and
$\theta $
is the Lebesgue measure on
$S^n$
. When
$\alpha =1$
and
$f(x)\equiv 1$
, then the above quantity agrees with the entropy in [Reference Guan and Ni26], first introduced by Firey [Reference Firey23] for the centrally symmetric
$\Omega $
. General integral quantities were studied by Andrews in [Reference Andrews2, Reference Andrews4]. Here, we shall assume that
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n} f(x)\, d\theta (x)=1$
, namely,
$\frac 1{\omega _n}\,f(x)d\theta (x)$
is a probability measure. For the special case
$f\equiv 1$
,
$\mathcal {E}_{\alpha , f}(\Omega ) $
becomes the entropy
$\mathcal {E}_\alpha (\Omega )$
in [Reference Andrews, Guan and Ni6].
For positive
$f\in C^{\infty }(\mathbb S^n)$
, consider the anisotropic flow for convex hypersurfaces
$\tilde X(\cdot , \tau ): M_{\tau }\to \mathbb {R}^{n+1}$
:

where
$\nu (x, \tau )$
is the unit exterior normal at
$\tilde X(x, \tau )$
of
$\tilde M_\tau =\tilde X(M, \tau )$
, and
$\tilde K(x,\tau )$
is the Gauss curvature of
$\tilde M_\tau $
at
$\tilde X(x,\tau )$
. Andrews [Reference Andrews4] proved that flow (1.7) contracts to a point under finite time if the initial hypersurface
$M_0$
is strictly convex. Under a proper normalization, the normalized anisotropy flow of (1.7) is

The basic observation is that a critical point for entropy
$\mathcal {E}_{\alpha , f} (\Omega )$
defined in (1.5) under volume normalization is a solution to equation (1.1). The entropy is monotone along flow (1.8). One may view (1.1) is an “optimal solution” to this variational problem as the flow (1.8) provides a natural path to reach it. This approach was devised in [Reference Andrews, Böröczky, Guan and Ni5] with the aim to obtain convergence of the normalized flow (1.8). The main arguments in [Reference Andrews, Böröczky, Guan and Ni5] follows those in [Reference Andrews, Guan and Ni6, Reference Guan and Ni26] where convergence of isotropic flows by power of Gauss curvature (i.e.,
$f=1$
) was established. Unfortunately, the entropy point estimate in [Reference Andrews, Guan and Ni6, Reference Guan and Ni26] fails for general anisotropic flows except
$\frac {1}{n+2}<\alpha \le \frac {1}n$
[Reference Andrews4]. The convergence was obtained in [Reference Andrews, Böröczky, Guan and Ni5] assuming
$M_0$
and f are invariant under a subgroup G of
$O(n+1)$
which has no fixed point. We note that an inverse Gauss curvature flow argument was considered by Bryan, Ivaki, and Scheuer [Reference Bryan, Ivaki and Scheuer14] to produce a origin-symmetric solution to (1.1).
Since we are only interested in finding a weak solution to (1.2), one only needs certain “weak” convergence of the flow (1.8). The key steps are to control diameter with entropy under appropriate conditions on measure
$\mu =f d\theta $
on
$\mathbb S^n$
and use monotonicity of entropy to produce a solution to (1.2). The following is our main result.
Theorem 1.1 For
$\alpha>\frac 1{n+2}$
and finite nontrivial Borel measure
$\mu $
on
$\mathbb {S}^n$
,
$n\geq 1$
, there exists a weak solution of (1.2) provided the following holds:
-
(i) If
$\alpha>1$ and
$\mu $ is not concentrated onto any great subsphere
$x^\bot \cap \mathbb {S}^n$ ,
$x\in \mathbb {S}^n$ .
-
(ii) If
$\alpha =1$ and
$\mu $ satisfies that for any linear
$\ell $ -subspace
$L\subset \Bbb R^{n+1}$ with
$1\leq \ell \leq n$ , we have
-
(a)
$\displaystyle \mu (L\cap \mathbb {S}^n)\leq \frac {\ell }{n+1}\cdot \mu (\mathbb {S}^n)$ ;
-
(b) equality in (a) for a linear
$\ell $ -subspace
$L\subset \Bbb R^{n+1}$ with
$1\leq d\leq n$ implies the existence of a complementary linear
$(n+1-\ell )$ -subspace
$\widetilde {L}\subset \Bbb R^{n+1}$ such that
$\mathrm {supp}\,\mu \subset L\cup \widetilde {L}$ .
-
-
(iii) If
$\frac 1{n+2}<\alpha <1$ and
$d\mu =f\,d\theta $ for nonnegative
$f\in L^{\frac {n+1}{n+2-\frac 1\alpha }}( \mathbb {S}^n)$ with
$\int _{\mathbb {S}^n}f>0$ .
Let us briefly discuss what is known about uniqueness of the solution of the
$L^p$
-Minkowski problem (1.4). If
$p>1$
and
$p\neq n$
, then Hug, Lutwak, Yang, and Zhang [Reference Hug, Lutwak, Yang and Zhang30] proved that the Alexandrov solution of the
$L^p$
-Minkowski problem (1.4) is unique. However, if
$p<1$
, then the solution of the
$L^p$
-Minkowski problem (1.3) may not be unique even if f is positive and continuous. Examples are provided by Chen, Li, and Zhu [Reference Chen, Li and Zhu17, Reference Chen, Li and Zhu18] if
$p\in [0,1)$
, and Milman [Reference Milman42] shows that for any
$C\in \mathcal {K}_{(0)}$
, one finds
$q\in (-n,1)$
such that if
$p<q$
, then there exist multiple solutions to the
$L^p$
-Minkowski problem (1.4) with
$\mu =S_{C,p}$
; or in other words, there exists
$K\in \mathcal {K}_{(0)}$
with
$K\neq C$
and
$S_{K,p}=S_{C,p}$
. In addition, Jian, Lu, and Wang [Reference Jian, Lu and Wang33] and Li, Liu, and Lu [Reference Li, Liu and Lu37] prove that for any
$p<0$
, there exists positive even
$C^\infty $
function f with rotational symmetry such that the
$L^p$
-Minkowski problem (1.3) has multiple positive even
$C^\infty $
solutions. We note that in the case of the centro-affine Minkowski problem
$p=-n$
, Li [Reference Li36] even verified the possibility of existence of infinitely many solutions without affine equivalence, and Stancu [Reference Stancu46] related unique solution in the cases
$p=0$
and
$p=-n$
.
The case when f is a constant function in the
$L^p$
-Minkowski problem (1.3) has received a special attention since [Reference Firey23]. When
$p=-(n+1)$
, (1.3) is self-similar solution of affine curvature flow. It is proved by Andrews that all solutions are centered ellipsoids. If
$n=2$
and
$p=2$
, the uniqueness was proved by Andrews [Reference Andrews3]. For general n and
$p>-(n+1)$
, through the work of Lutwak [Reference Lutwak40], Guan-Ni [Reference Guan and Ni26], and Andrews, Guan, and Ni [Reference Andrews, Guan and Ni6], Brendle, Choi, and Daskalopoulos [Reference Brendle, Choi and Daskalopoulos13] finally classified that the only solutions are centered balls. See also [Reference Crasta and Fragalá21, Reference Ivaki and Milman32, Reference Saroglou44] for other approaches. Stability versions of these results have been obtained by Ivaki [Reference Ivaki31], but still no stability version is known in the case
$p\in [0,1)$
if we allow any solutions of (1.3) not only even ones.
Concerning recent versions of the
$L^p$
-Minkowski problem, see [Reference Böröczky, Koldobsky and Volberg9].
The paper is structured as follows: The required diameter bounds are discussed in Section 2. Section 3 verifies the main properties of the Entropy, Section 4 proves our main result (Theorem 4.1) about flows, and finally Theorem 1.1 is proved in Section 5 via weak approximation.
2 Entropy and diameter estimates
For
$\delta \in [0,1)$
and linear i-subspace L of
$\Bbb R^{n+1}$
with
$1\leq \mathrm {dim}\,L\leq n$
, we consider the collar

Let
$B(1)\subset \Bbb R^{n+1}$
be the unit ball centered at the origin.
Theorem 2.1 Let
$\alpha>\frac 1{n+2}$
, let
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f=1$
for a bounded measurable function f on
$\mathbb {S}^n$
with
$\inf f>0$
, and let
$\Omega \subset \Bbb R^{n+1}$
be a convex body such that
$|\Omega |=|B(1)|$
and
$\mathrm {diam}\, \Omega = D$
. For any
$\delta ,\tau \in (0,1)$
, we have
-
(i) if
$\alpha>1$ , and
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\Psi (z^\bot \cap \mathbb {S}^n,\delta )}f\leq 1-\tau $ for any
$z\in S^n$ , then
$$ \begin{align*}\exp\left(\frac{\alpha-1}{\alpha}\,\mathcal{E}_{\alpha, f} (\Omega)\right) \geq \gamma_1 \tau\delta^{1-\frac1\alpha}D^{1-\frac1\alpha}, \end{align*} $$
$\gamma _1>0$ depends on n and
$\alpha $ ;
-
(ii) if
$\alpha =1$ , and
$$ \begin{align*}\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\Psi(L\cap \mathbb{S}^n,\delta)}f< \frac{(1-\tau)i}{n+1}, \end{align*} $$
$\Bbb R^{n+1}$ ,
$i=1,\ldots ,n$ , then
$$ \begin{align*}\mathcal{E}_{1, f} (\Omega)\geq\tau\log D +\log\delta-4\log(n+1); \end{align*} $$
-
(iii) if
$\frac 1{n+2}<\alpha <1$ ,
$p=1-\frac 1\alpha $ (where
$-n-1<p<0$ ),
$\tau \leq \frac 12\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f\cdot u^{1-\frac 1\alpha }$ and
(2.1)for any$$ \begin{align} \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\Psi(z^\bot\cap \mathbb{S}^n,\delta)}f^{\frac{n+1}{n+1+p}}\leq \tau^{\frac{n+1}{n+1+p}}, \end{align} $$
$z\in S^{n-1}$ , then
$$ \begin{align*}\mbox{either}\ D\leq 16n^2/\delta^2, \ \ \mbox{or } D\leq \left(\frac12\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n}f\cdot u^{1-\frac1\alpha}\right)^{\frac2{p}}. \end{align*} $$
$\tau \leq \frac 12\exp \left (\frac {\alpha -1}{\alpha }\,\mathcal {E}_{\alpha , f} (\Omega )\right )$ , then
$$ \begin{align*}\mbox{either}\ D\leq 16n^2/\delta^2, \ \ \mbox{or } D\leq \left(\frac12\exp\left(\frac{\alpha-1}{\alpha}\,\mathcal{E}_{\alpha, f} (\Omega)\right)\right)^{\frac2{p}}. \end{align*} $$
Remark 2.2 We note that for any
$\alpha \ge 1$
, bounded f with
$\inf f>0$
and
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f=1$
, and
$\tau \in (0,1)$
, there exists
$\delta \in (0,1)$
such that conditions in (i) and (ii) hold. In the case of
$1>\alpha >\frac 1{n+2}$
, (iii) holds if in addition that
$\tau \leq \frac 12\exp \left (\frac {1-\alpha }{\alpha }\,\mathcal {E}_{\alpha , f} (\Omega )\right )$
for the convex body
$\Omega \subset \Bbb R^{n+1}$
.
Proof Given
$\alpha>\frac 1{n+2}$
, bounded f with
$\inf f>0$
and
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f=1$
, and
$\tau \in (0,1)$
, the existence of suitable
$\delta \in (0,1)$
follows from the fact that the Lebesgue measure is a Borel measure.
Now, we assume that the conditions in (i)–(iii) hold. We may assume that the centroid of
$\Omega $
is the origin; thus, Kannan, Lovász, and Simonovics [Reference Kannan, Lovász and Simonovits35] yield the existence of an o-symmetric ellipsoid such that

Let u be the support function of
$\Omega $
, and let
$R=\max \{\|y\|:\,y\in \Omega \}\geq D/2$
and
$z_0\in \mathbb {S}^n$
such that
$Rz_0\in \partial \Omega $
. We observe that the definition of the entropy yields

Case 1:
$\alpha>1$
.
According to the condition in (i), we may choose
$\zeta \in \{+1,-1\}$
such that

and hence
$\frac {R\zeta z_0}{n+1}\in \Omega $
by (2.2). Since
$u_\sigma (x)\geq \langle \frac {R\zeta z_0}{n+1},x\rangle \geq \frac {R\delta }{n+1}$
for
$x\in \Phi $
, we have

Case 2:
$\alpha =1$
.
To simplify notation, we consider the Borel probability measure
$\mu (A)=\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _Af$
on
$S^n$
. Let
$e_1,\ldots ,e_{n+1}\in \mathbb {S}^n$
be the principal directions associated with the ellipsoid E in (2.2), and let
$r_1,\ldots ,r_{n+1}>0$
be the half axes of E with
$r_ie_i\in \partial E$
where we may assume that
$r_1\leq \cdots \leq r_{n+1}$
. In particular, (2.2) yields that

We observe that for any
$v\in \mathbb {S}^n$
, there exists
$e_i$
such that
$|\langle v,e_i\rangle |\geq \frac {1}{\sqrt {n+1}}> \frac {\delta }{n+1}$
. For
$i=1,\ldots ,n+1$
, we define

In particular,
$B_i\subset \Psi (L_i\cap \mathbb {S}^n,\delta )$
for
$i=1,\ldots ,n$
and
$L_i=\mathrm {lin}\{e_1,\ldots ,e_i\}$
.
It follows that
$\mathbb {S}^n$
is partitioned into the Borel sets
$B_1,\ldots ,B_{n+1}$
, and as
$B_i\subset \Psi (L_i\cap \mathbb {S}^n,\delta )$
for
$i=1,\ldots ,n$
, we have


For
$\zeta =\frac {1-\tau }{n+1}$
, we have
$0< \zeta <\frac 1{n+1}$
, and define




As
$r_ie_i\in \Omega $
, it follows from the definition of
$B_i$
that
$u(x)\geq \langle x,r_ie_i\rangle \geq r_i\cdot \frac {\delta }{n+1}$
for
$x\in B_i$
,
$i=1,\ldots ,n+1$
. We deduce from applying (2.3), (2.5)–(2.9),
$r_1\leq \cdots \leq r_{n+1}$
, and
$\zeta <\frac 1{n+1}$
that

Now,
$D\leq (n+1)\mathrm {diam}\,E=2(n+1)r_{n+1}\leq (n+1)^2r_{n+1}$
and
$\tau <1$
, and hence

In particular, we conclude that

Case 3:
$\frac 1{n+2}<\alpha <1$
.
In this case,
$-(n+1)<1-\frac 1\alpha <0$
. We may assume that

and we consider

Concerning
$\Phi _0$
, we have

On the other hand, we have
$\pm \frac {R}{(n+1)}\,z_0\in \Omega $
by (2.2), thus any
$x\in \Phi _1$
satisfies

and hence
$|\langle x,z_0\rangle |\leq (n+1)\sqrt {\frac {2}{R}}\leq \frac {4n}{\sqrt {D}}\leq \delta $
; or in other words,

It follows from
$|\Omega |=|B(1)|$
and the Blaschke–Santaló inequality (cf. [Reference Schneider45]) that

For
$p=1-\frac 1\alpha \in (-n-1,0)$
, Hölder’s inequality and
$\int _{\Phi _1}f^{\frac {n+1}{n+1+p}}< \tau ^{\frac {n+1}{n+1+p}}$
yield

Finally, adding the last estimate to (2.10) yields

and hence the conditions either
$\tau \leq \frac 12\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f\cdot u^{1-\frac 1\alpha }$
or
$\tau \leq \frac 12\exp \left (\frac {1-\alpha }{\alpha }\,\mathcal {E}_{\alpha , f} (\Omega )\right )$
on
$\tau $
implies (iii).
3 Anisotropic flows and monotonicity of entropies
The following theorem was proved by Andrews in [Reference Andrews4] (see also for a discussion of contracting of non-homogeneous fully nonlinear anisotropic curvature flows in [Reference Guan, Huang and Liu24]).
Theorem 3.1 [Reference Andrews4]
For any
$\alpha>0$
and positive
$f\in C^{\infty }(\mathbb S^n)$
and any initial smooth, strictly convex hypersurface
$\tilde M_0\subset \mathbb R^{n+1}$
, the hypersurfaces
$\tilde M_{\tau }$
given by the solution of (1.7) exist for a finite time T and converge in Hausdorff distance to a point
$p \in \mathbb R^{n+1}$
as
$\tau $
approaches T.
Assuming

solution (1.7) yields a smooth convex solution to the normalized flow (1.8) with volume preserved.
Set

Note that
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n} d\sigma _t(x) =\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n} d\theta (x) =1$
.
Since the un-normalized flow (1.7) shrinks to a point in finite time, we may assume that it is the origin. Then the support function
$u(x,t)$
is positive for the normalized flow (1.8).
Lemma 3.2
-
(a) The entropy
$ \mathcal {E}_{\alpha , f}(\Omega _{t})$ defined in (1.5) is monotonically decreasing,
(3.2)$$ \begin{align} \mathcal{E}_{\alpha, f}(\Omega_{t_2})\le \mathcal{E}_{\alpha, f}(\Omega_{t_1}), \quad \forall t_1\le t_2 \in [0, \infty).\end{align} $$
-
(b) There is
$D>0$ depending only on
$\inf f, \sup f, \alpha , \Omega _0$ such that
(3.3)$$ \begin{align} \mathrm{diam}\,\Omega_t=D(t)\le D, \ \forall t\ge 0.\end{align} $$
-
(c)
$\forall t_0\in [0, \infty )$ ,
(3.4)$$ \begin{align} \mathcal{E}_{\alpha, f}(\Omega_{t_0}, 0)\ge \mathcal{E}_{\alpha, f, \infty} +\int_{t_0}^{\infty}\left(\frac{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h^{\alpha+1}(x,t)\, d\sigma_t} {\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h(x,t) \, d\sigma_t \cdot \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h^{\alpha}(x,t)\, d\sigma_t}-1\right)\, dt,\end{align} $$
where
$$\begin{align*}h(x,t)=h_0(x, t), \ \mathcal{E}_{\alpha, f, \infty}\doteqdot\lim_{t\to \infty} \mathcal{E}_{\alpha, f}(\Omega_{t}).\end{align*}$$
Proof
-
(a) We follow argument in [Reference Guan and Ni26]. For each
$T_0>$ fixed, pick
$T> T_0$ . Let
$a^{T}=(a^{T}_1,\ldots , a^{T}_{n+1})$ be an interior point of
$\Omega _T$ . Set
$u^T=u- e^{t-T}\sum _{i=1}^{n+1} a^T_ix_i$ ; it satisfies equation
(3.5)$$ \begin{align} \frac{\partial}{\partial t}u^T(x, t)= -\frac{f^\alpha(x) K^{\alpha}(x, t)}{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n}f^\alpha K^{\alpha-1}} +u^T(x,t).\end{align} $$
Note that since
$a^T$ is an interior point of
$\Omega _T$ and
$u(x,T)$ is the support function of
$\Omega _T$ with respect to
$a^T$ ,
$u^T(x, T)> 0, \forall x\in \mathbb S^n$ . We claim
$$\begin{align*}u^T(x, t)>0, \ \forall t\in [0, T).\end{align*}$$
$u^T(x_0, t')\le 0$ for some
$0<t'<T, x_0\in \mathbb S^n$ , and equation (3.5) implies
$u^T(x_0, t)<0$ for all
$t>t'$ , which contradicts to
$u^T(x, T)> 0$ .
Set
$a^T(t)=e^{t-T}a^T$ . By the claim,
$a^T(t)$ is in the interior of
$\Omega _t, \ \forall t\le T$ . Denote
$$\begin{align*}d\sigma_{T,t}=u^T(x,t)K^{-1}(x,t)d\theta,\end{align*}$$
(3.6)We have$$ \begin{align} \frac{\partial}{\partial t}u_{a^T(t)}(x,t)= -\frac{f^\alpha(x) K^{\alpha}(x, t)}{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n}h_{a^T(t)}^{\alpha}(x,t)\, d\sigma_{T,t}} +u_{a^T(t)}(x,t).\end{align} $$
$$\begin{align*}\frac{\partial}{\partial t} \mathcal{E}_{\alpha, f}(\Omega_{t}, a^T(t))=\frac{-\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha+1}(x,t)\, d\sigma_{T,t}} {\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}(x,t) \, d\sigma_{T,t} \cdot \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha}(x,t)\, d\sigma_{T,t}}+1.\end{align*}$$
Thus,
$\forall t<T$ ,
(3.7)Therefore,$$ \begin{align} &\mathcal{E}_{\alpha, f}(\Omega_{t}, a^T(t))-\mathcal{E}_{\alpha, f}(\Omega_T, a^T)\\ \nonumber &= \int_{t}^{T}\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb S^n} \left(\frac{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha+1}(x,t)\, d\sigma_{T,t}} {\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}(x,t) \, d\sigma_{T,t} \cdot \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha}(x,t)\, d\sigma_{T,t}}-1\right)\, dt\ge 0. \end{align} $$
$$\begin{align*}\mathcal{E}_{\alpha, f}(\Omega_{t})\ge\mathcal{E}_{\alpha, f}(\Omega_T, a^T), \ \forall t<T.\end{align*}$$
$a^T$ is arbitrary, (3.2) is proved.
-
(b) The boundedness of
$D(t)$ follows from Theorem 2.1 combined with the estimate
$\mathcal {E}_{\alpha , 1}(\Omega _{t})\leq \mathcal {E}_{\alpha , 1}(B(1))$ from (a) (see also [Reference Andrews, Guan and Ni6, Reference Guan and Ni26]). The only nontrivial case is when
$\frac 1{n+2}<\alpha <1$ because we have to choose a
$\tau $ independent of t. However, we may choose any
$\tau \in (0,1)$ with
$\tau \leq \frac 12\exp \left (\frac {1-\alpha }{\alpha }\,\mathcal {E}_{\alpha , f} (B(1))\right )$ according to
$\mathcal {E}_{\alpha , 1}(\Omega _{t})\leq \mathcal {E}_{\alpha , 1}(B(1))$ .
-
(c)
$\forall \epsilon>0, \ \forall t_0$ fixed, pick
$T>T_0>t_0$ . As
$ \mathcal {E}_{\alpha , f}(\Omega _{T})$ is bounded by (a),
$\exists a^T$ inside
$\Omega _T$ such that
$ \mathcal {E}_{\alpha , f}(\Omega _{T})\le \mathcal {E}_{\alpha , f}(\Omega _{T}, a^T)+\epsilon $ . By (3.7),
$$ \begin{align*} &\mathcal{E}_{\alpha, f}(\Omega_{t_0}, a^T(t_0))-\mathcal{E}_{\alpha, f}(\Omega_{T})\\ &\ge \int_{t_0}^{T_0}\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb S^n} \left(\frac{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha+1}(x,t)\, d\sigma_{T,t}} {\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}(x,t) \, d\sigma_{T,t} \cdot \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h_{a^T(t)}^{\alpha}(x,t)\, d\sigma_{T,t}}-1\right)\, dt-\epsilon. \end{align*} $$
As
$|a^T|\le D, \ \forall T$ , let
$T\to \infty $ ,
$$\begin{align*}a^T(t)\to 0, \ u^T(x,t)\to u(x,t), \ \ \mbox{ uniformly for}\ 0\le t\le T_0, x\in \mathbb S^n. \end{align*}$$
We obtain
$\forall t_0<T_0$ ,
$$ \begin{align*} \mathcal{E}_{\alpha, f}(\Omega_{t_0}, 0)-\mathcal{E}_{\alpha, f, \infty}\ge \int_{t_0}^{T_0}\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb S^n} \left(\frac{\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h^{\alpha+1}(x,t)\, d\sigma_t} {\frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h(x,t) \, d\sigma_t \cdot \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\mathbb{S}^n} h^{\alpha}(x,t)\, d\sigma_t}-1\right)\, dt-\epsilon.\end{align*} $$
Then let
$T_0\to \infty $ , as
$\epsilon>0$ is arbitrary, we obtain (3.4).
4 Weak convergence
The goal of this section is to prove the following statement.
Theorem 4.1 For a
$C^\infty $
function
$f:\mathbb {S}^n\to (0,\infty )$
and
$\alpha>\frac 1{n+2}$
with
$ \frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}f=1$
, there exist
$\lambda>0$
and a convex body
$\Omega \subset \Bbb R^{n+1}$
with
$o\in \Omega $
whose support function u is a (possibly weak) solution of the Monge–Ampère equation

and
$\Omega $
satisfies that

where
$C^{-1}<\lambda <C$
for a
$C>1$
depending only on the
$\alpha , \tau , \delta $
in Theorem 2.1 such that f satisfies the conditions in Theorem 2.1.
From now on, we will assume that the f in Theorem 4.1 satisfies the corresponding condition in Theorem 2.1 and
$\Omega _0=B(1)$
in (1.8). We note that for any
$z\in B(1)$
,
$v_z\leq 2$
for the support function
$v_z$
of
$B(1)$
at z, and hence if
$\alpha>\frac 1{n+2}$
, then

The following is a consequence of Theorem 2.1 and Lemma 3.2.
Lemma 4.2 There exist
$C_{\alpha , \tau , \delta }>0, D_{\alpha , \tau , \delta }>0$
, and
$c_{\alpha , \tau , \delta }\in \mathbb R$
depending only on constants
$\alpha , \tau , \delta $
in Theorem 2.1 such that, along (1.8), we have

Proof For each
$\alpha>\frac 1{n+2}$
fixed with condition on f as in Theorem 2.1,
$\mathcal {E}_{\alpha , f}(\Omega _{t})$
is bounded from below in terms of the diameter
$D(t)$
. Since
$|\Omega _t|=|B(1)|$
, we have
$D(t)\ge 2$
by the Isodiametric Inequality (cf. [Reference Schneider45]). By Theorem 2.1,
$\mathcal {E}_{\alpha , f}(\Omega _{t})$
is bounded from below by a constant
$c_{\alpha , \tau , \delta }>0$
, and hence
$\mathcal {E}_{\alpha , f, \infty } \ge c_{\alpha , \tau , \delta }$
. It follows from Lemma 3.2 that
$\mathcal {E}_{\alpha , f}(\Omega _{t})\le \mathcal {E}_{\alpha , f}(B(1))$
, and this estimate combined with (4.3) and Theorem 2.1 yields
$D(t)\le D_{\alpha , \tau , \delta }$
where
$D_{\alpha , \tau , \delta }$
depends only on constants in condition on f in Theorem 2.1. Finally, the inequalities follow from Lemma 3.2.
Set

We note that
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n} h(x,t) \, d\sigma _{t} $
is monotone and bounded from below and above by Lemma 4.2, and hence we have

By Lemma 3.2 and Corollary 4.2,

Since the integrand is nonnegative,
$\exists t_k\to \infty $
such that

This implies

After considering a subsequence, we may assume that

where u is the support function of
$\Omega $
. In view of (4.9) and (4.6),

The following lemma is crucial for the weak convergence, which is a refined form of the classical Hölder inequality.Footnote 1
Lemma 4.3 Let
$p,\ q\in \mathbb R^+$
with
$\frac 1p+\frac 1q=1$
, and set
$\beta =\min \{\frac 1p, \frac 1q\}$
. Let
$(M,\mu )$
be a measurable space;
$\forall F\in L^p, \ G\in L^q$
,

Proof We first prove the following Claim.
$\forall s, t\in \mathbb R$
,

We may assume
$t\ge s$
, set
$\tau =t-s$
, and (4.13) is equivalent to

Set

We have
$\xi (0)=0$
,

If
$\beta =\frac 1q$
, then
$\frac 1q\le \frac 12$
; since
$\tau \ge 0$
,

If
$\beta =\frac 1p$
, then
$\frac 1q\ge \frac 12$
; we have

We conclude that

In turn,

This yields (4.14) and (4.13). The Claim is verified.
Back to the proof of the lemma. We may assume

Set

Put them into (4.13) and integrate, as
$\frac 1p+\frac 1q=1$
,

We prove weak convergence.
Proposition 4.4
$\forall \alpha>\frac {1}{n+2}$
, suppose that (4.10) and (4.11) hold. Denote

Then

where
$\eta $
is defined in (4.5) which is bounded from below and above in (4.6). As a consequence, there is a convex body
$\Omega \subset \mathbb R^{n+1}$
with
$o\in \Omega $
,

and its support function u satisfies

Proof We only need to verify (4.15). By (4.11), it is equivalent to prove

Since
$D(t_k)$
is bounded,


By (4.8), (4.11), and Lemma 4.3, with
$p=\alpha +1$
,
$F^{\frac {1}{1+\alpha }}=h(x,t_k)$
,
$G=1$
,

For
$t_k$
fixed, let

and set

It is straightforward to check that
$\exists A_{\alpha }\ge 1$
depending only on
$\alpha $
such that

Since
$ |\gamma ^{\frac {1+\alpha }2}_k(x)-1|\le 2^{1+\alpha }, \ \forall x\in \Sigma _k$
, let
$\delta =\min \{1+\alpha , 2\}$
,

By (4.19),

Hence,

5 The general Monge–Ampère equations – proof of Theorem 1.1
In order to prove Theorem 1.1, we need weak approximation in the following sense.
Lemma 5.1 For
$\delta ,\varepsilon \in (0,\frac 12)$
and a Borel probability measure
$\mu $
on
$\mathbb {S}^n$
,
$n\geq 1$
, there exists a sequence
$d\mu _k=\frac 1{\omega _n}\,f_k\,d\theta $
of Borel probability measures whose weak limit is
$\mu $
and
$f_k\in C^\infty ( \mathbb {S}^n)$
satisfies
$f_k>0$
and the following properties:
-
(i) If
$\mu \left (\Psi (z^\bot \cap \mathbb {S}^n,2\delta )\right )\leq 1-\varepsilon $ for any
$z\in S^{n-1}$ , then
(5.1)$$ \begin{align} \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\Psi(z^\bot\cap \mathbb{S}^n,\delta)}f_k\leq 1-\varepsilon \mbox{ for any}\ z\in S^{n-1}. \end{align} $$
-
(ii) If
$\mu (\Psi (L\cap \mathbb {S}^n,2\delta ))<(1-2\delta )\cdot \frac {\ell }{n+1}$ for any linear
$\ell $ -subspace L of
$\Bbb R^{n+1}$ ,
$\ell =1,\ldots ,n$ , then
(5.2)$$ \begin{align} \mu_k\left(\Psi\left(L\cap \mathbb{S}^n,\delta\right)\right)<(1-\delta)\cdot \frac{\ell}{n+1}. \end{align} $$
-
(iii) If
$d\mu =\frac 1{\omega _n}\,f\,d\theta $ for
$f\in L^{r}(\mathbb {S}^n)$ where
$r>1$ , and
(5.3)$$ \begin{align} \frac{\ \ }{\ \ }{\hskip -0.4cm}\int_{\Psi(z^\bot\cap \mathbb{S}^n,2\delta)}f^r\leq \varepsilon \end{align} $$
for any
$z\in S^{n-1}$ , then
(5.4)$$ \begin{align} \int_{\Psi(z^\bot\cap \mathbb{S}^n,\delta)}f_k^r\leq 2^r\varepsilon \mbox{ for any}\ z\in S^{n-1}. \end{align} $$
Proof For
$k\geq 1$
, let
$\{B_{k,i}\}_{i=1,\ldots ,m(k)}$
be a partition of
$S^n$
into spherically convex Borel measurable sets
$B_{k,i}$
with
$\mathrm {diam}B_{k,i}\leq \frac 1k$
and
$\theta (B_{k,i})>0$
. For each
$B_{k,i}$
, we choose a
$C^\infty $
function
$h_{k,i}:\mathbb {S}^n\to [0,\infty )$
such that for
$M_{k,i}=\max h_{k,i}$
and the probability measure
$d\tilde {\theta }=\frac 1{\omega _n}\,d\theta $
, we have:
-
•
$h_{k,i}=0$ if
$x\not \in B_{k,i}$ ;
-
•
$M_{k,i}\leq (1+\frac 1k)\cdot \frac {\mu (B_{k,i})}{\tilde {\theta }(B_{k,i})}$ ;
-
•
$\theta \left (\left \{x\in B_{k,i}:h_{k,i}(x)<M_{k,i}\right \}\right )<\frac 1k\,\theta (B_{k,i})$ ;
-
•
$\int _{B_{k,i}}h_{k,i}\,d\tilde {\theta }=\mu (B_{k,i})$ .
We consider the positive
$C^\infty $
function
$\tilde {f}_k\hspace{-0.5pt}=\hspace{-0.5pt}\frac 1k\hspace{-0.5pt}+\hspace{-0.5pt}\sum _{i=1}^{m(k)}h_{k,i}$
, and hence
$f_k\hspace{-0.5pt}=\hspace{-0.5pt}\left (\hspace{2pt} \frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}\tilde {f}_k\right )^{-1}\tilde {f}$
satisfies that the probability measure
$d\mu _k=f_k\,d\tilde {\theta }$
tends weakly to
$\mu $
, and for large
$k\geq 1/\delta $
,
$\mu _k$
satisfies (i), and if (ii) holds, then
$\mu _k$
also satisfies (5.2).
Turning to (iii), we assume that
$d\mu =f\,d\tilde {\theta }$
for
$f\in L^{r}(\mathbb {S}^n)$
where
$r>1$
, and f satisfies (5.3). For any large k and
$i=1,\ldots ,m(k)$
, we deduce from the Hölder inequality that

Summing this estimate up for large k and all
$B_{k,i}$
with
$B_{k,i}\cap \Psi (z^\bot \cap \mathbb {S}^n,\delta )\neq \emptyset $
, and using that
$\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}\tilde {f}_k\geq 2^{-1/2}$
for large k, we deduce that

For
$\alpha>0$
and
$p=1-\frac 1\alpha $
, the
$L^p$
-surface area
$dS_{\Omega ,p}=u^{1-p}dS_\Omega $
was introduced in the seminal works [Reference Lutwak39–Reference Lutwak41] for a convex body
$\Omega \subset \Bbb R^{n+1}$
with
$o\in \Omega $
and support function u. Since the surface area measure is weakly continuous for
$p<1$
, and if
$K\subset \Bbb R^{n+1}$
is an at most n-dimensional compact convex set, then
$S_{K,p}\equiv 0$
for
$p<1$
, we have the following statement.
Lemma 5.2 If convex bodies
$\Omega _m\subset \Bbb R^{n+1}$
tend to a compact convex set
$K\subset \Bbb R^{n+1}$
where
$o\in \Omega _m,K$
, and
$\liminf _{m\to \infty }S_{\Omega _m,p}>0$
, then
$\mathrm {int}K\neq \emptyset $
and
$S_{\Omega _m,p}$
tends weakly to
$S_{K,p}$
.
For the reader’s sake, let us recall Theorem 1.1.
Theorem 5.3 For
$\alpha>\frac 1{n+2}$
and finite nontrivial Borel measure
$\mu $
on
$\mathbb {S}^n$
,
$n\geq 1$
, there exists a weak solution of (1.2) provided the following holds:
-
(i) If
$\alpha>1$ and
$\mu $ is not concentrated onto any great subsphere
$x^\bot \cap \mathbb {S}^n$ ,
$x\in \mathbb {S}^n$ .
-
(ii) If
$\alpha =1$ and
$\mu $ satisfies that for any linear
$\ell $ -subspace
$L\subset \Bbb R^{n+1}$ with
$1\leq \ell \leq n$ , we have:
-
(a)
$\displaystyle \mu (L\cap \mathbb {S}^n)\leq \frac {\ell }{n+1}\cdot \mu (\mathbb {S}^n)$ ;
-
(b) equality in (a) for a linear
$\ell $ -subspace
$L\subset \Bbb R^{n+1}$ with
$1\leq d\leq n$ implies the existence of a complementary linear
$(n+1-\ell )$ -subspace
$\widetilde {L}\subset \Bbb R^{n+1}$ such that
$\mathrm {supp}\,\mu \subset L\cup \widetilde {L}$ .
-
-
(iii) If
$\frac 1{n+2}<\alpha <1$ , assume
$d\mu =fd\theta $ for nonnegative
$f\in L^{\frac {n+1}{n+2-\frac 1\alpha }}( \mathbb {S}^n)$ with
$\int _{\mathbb {S}^n}f>0$ .
Proof Let
$\alpha>\frac 1{n+2}$
. After rescaling, we may assume that the
$\mu $
in (1.2) is a probability measure. We consider the sequence
$d\mu _k=\frac 1{\omega _n}f_k\,d\theta $
of Lemma 5.1 of Borel probability measures whose weak limit is
$\mu $
and
$f_k\in C^\infty ( \mathbb {S}^n)$
satisfies
$f_k>0$
. For each
$f_k$
, let
$\Omega _k\subset \Bbb R^{n+1}$
be the convex body with
$o\in \Omega _k$
provided by Theorem 4.1 whose support function
$u_k$
is the solution of the Monge–Ampère equation

$\exists \lambda _k>0$
under control, with
$|\lambda _k\Omega |=|B(1)|$
,
$\Omega _k$
satisfies that

We also need the observations that

and if
$p=1-\frac 1\alpha $
, then

We claim that if there exists
$\Delta>0$
depending on n,
$\alpha $
, and
$\mu $
such that

To prove this claim, we note that (5.9) yields the existence of a subsequence of
$\{\Omega _k\}$
tending to a compact convex set
$\Omega $
with
$o\in \Omega $
, which is a convex body by (5.8) and Lemma 5.2. Moreover, Lemma 5.2 also yields that
$\Omega $
is an Alexandrov solution of (1.2), verifying the claim (5.9).
We divide the rest of the argument verifying Theorem 5.3 into three cases.
Case 1:
$\alpha>1$
.
Since
$\mu $
is not concentrated to any great subsphere, there exist
$\delta \in (0,\frac 12)$
depending on
$\mu $
such that
$\mu \left (\Psi (z^\bot \cap \mathbb {S}^n,2\delta )\right )\leq 1-2\delta $
for any
$z\in S^{n-1}$
. It follows from Lemma 5.1 that we may assume that

Now, Theorem 4.1 implies that
$\lambda _k\geq c$
for a constant
$c>0$
depending on n,
$\delta $
, and
$\alpha $
, and in turn Theorem 4.1, (4.3), and
$\frac 1{\alpha }-1<0$
yield that

for a constant
$C>0$
depending on n,
$\delta $
, and
$\alpha $
. Therefore, Theorem 2.1 and (5.10) imply that the sequence
$\{\Omega _k\}$
is bounded, and in turn the claim (5.9) implies Theorem 5.3 if
$\alpha>1$
.
Case 2:
$\alpha =1$
.
The argument is by induction on
$n\geq 0$
where we do not put any restriction on the probability measure
$\mu $
in the case
$n=0$
. For the case
$n=0$
, we observe that any finite measure
$\mu $
on
$S^0$
can be represented in the form
$d\mu =u\,dS_\Omega $
for a suitable segment
$\Omega \subset \Bbb R^1$
.
For the case
$n\geq 1$
, assuming that we have verified Theorem 5.3(ii) in smaller dimensions, we consider a Borel measure probability
$\mu $
on
$S^n$
satisfying (a) and (b).
Case 2.1: There exists a linear
$\ell $
-subspace
$L\subset \Bbb R^{n+1}$
with
$1\leq \ell \leq n$
and
$\mu (L\cap \mathbb {S}^n)= \frac {\ell }{n+1}\cdot \mu (\mathbb {S}^n)$
.
Let
$\widetilde {L}\subset \Bbb R^{n+1}$
be the complementary linear
$(n+1-\ell )$
-subspace with
$\mathrm {supp}\,\mu \subset L\cup \widetilde {L}$
, and hence
$\mu (\widetilde {L}\cap \mathbb {S}^n)= \frac {n+1-\ell }{n+1}\cdot \mu (\mathbb {S}^n)$
. It follows by induction that there exist an
$\ell $
-dimensional compact convex set
$K'\subset L$
and an
$(n+1-\ell )$
-dimensional compact convex set
$\widetilde {K}'\subset \widetilde {L}$
such that
and
. Finally, for
$K=\widetilde {L}^\bot \cap (K'+L^\bot )$
and
$\widetilde {K}=L^\bot \cap (\widetilde {K}'+\widetilde {L}^\bot )$
, there exist
$\alpha ,\tilde {\alpha }>0$
such that

Case 2.2:
$\mu (L\cap \mathbb {S}^n)< \frac {\ell }{n+1}\cdot \mu (\mathbb {S}^n)$
for any linear
$\ell $
-subspace
$L\subset \Bbb R^{n+1}$
with
$1\leq \ell \leq n$
.
It follows by a compactness argument that there exists
$\delta \in (0,\frac 12)$
depending on
$\mu $
such that
$\mu (\Psi (L\cap \mathbb {S}^n,2\delta ))<(1-2\delta )\cdot \frac {\ell }{n+1}$
for any linear
$\ell $
-subspace L of
$\Bbb R^{n+1}$
,
$\ell =1,\ldots ,n$
. We consider the sequence of probability measures
$d\mu _k=\frac 1{\omega _n}f_k\,d\theta $
of Lemma 5.1 tending weakly to
$\mu $
such that
$f_k>0$
,
$f_k\in C^\infty (\mathbb {S}^n)$
, and

for any linear
$\ell $
-subspace L of
$\Bbb R^{n+1}$
,
$\ell =1,\ldots ,n$
.
For each
$f_k$
, let
$\Omega _k\subset \Bbb R^{n+1}$
with
$o\in \Omega _k$
be the convex body provided by Theorem 4.1 whose support function
$u_k$
is the solution of the Monge–Ampère equation (4.1) and satisfies (4.2) with
$f=f_k$
and
$\lambda =\lambda _k$
where
$|B(1)|=|\lambda _k\Omega _k|$
for
$\lambda _k>0$
, and

and hence
$\lambda _k=1$
. In particular, (4.3) yields

Since
$\mathcal {E}_{1, f_k} (\Omega _k)$
is bounded, (5.11) and Theorem 2.1 imply that the sequence
$\Omega _k$
stays bounded, as well. Therefore, the claim (5.9) yields Theorem 5.3 if
$\alpha =1$
.
Case 3:
$\frac 1{n+2}<\alpha <1$
.
We set
$p=1-\frac 1\alpha \in (-n-1,0)$
and
$r=\frac {n+1}{n+1+p}>1$
, and

and choose
$\delta \in (0,\frac 12)$
such that

for any
$z\in S^{n-1}$
. We deduce from Lemma 5.1 that if
$z\in S^{n-1}$
, then

We deduce from (5.5), (5.7), and
$|\lambda _k\Omega _k|=|B(1)|=\frac {\omega _n}{n+1}$
that

In particular, (4.3) and the upper bound on the entropy yield that

It follows from (5.15) that
$\lambda _k\leq 2^{\frac {|p|}{|p|+n}}$
, and in turn (5.14) yields that

Therefore,
$\tau \leq \frac 12\frac {\ \ }{\ \ }{\hskip -0.4cm}\int _{\mathbb {S}^n}u_k^{p}f_k$
(cf. (5.12)), (5.13), and Theorem 2.1 yield that the sequence
$\{\Omega _k\}$
is bounded, and in turn the claim (5.9) implies Theorem 5.3 if
$\frac 1{n+2}<\alpha <1$
.