Stability for the Affine Sobolev Inequality and its Critical Points for $p\ge 2$

Extracted source text · arXiv 2607.06415

\documentclass[11pt, twoside]{article}

%%%%%%%%%%%%%%%%%%%%%%%%%
\newcommand\version{July 7, 2026}
%%%%%%%%%%%%%%%%%%%%%%%%%

\usepackage{bbm}
\usepackage{amssymb}
\usepackage{amsfonts}
\usepackage{amsmath}
\usepackage{amsthm}
\usepackage{color}
\usepackage{mathrsfs}
\usepackage{txfonts}
\usepackage{bbm}
\usepackage{enumerate}
\usepackage{anysize}
\usepackage{indentfirst}

\usepackage{latexsym}

%\usepackage[notref,notcite]{showkeys}
%\usepackage{showkeys}
%\usepackage{citeref}
%\UseRawInputEncoding

\usepackage[colorlinks=true,
linkcolor=blue,
citecolor=red,
urlcolor=magenta,
backref=page
]{hyperref}

%\usepackage{refcheck}

%\usepackage{tikz}
%\usetikzlibrary{positioning}

\allowdisplaybreaks

\pagestyle{myheadings}\markboth{\footnotesize\rm\sc
Rupert L. Frank, Yinqin Li and Dachun Yang}
{\footnotesize\rm\sc Stability for the Affine Sobolev Inequality and its Critical Points for $p\ge 2$}

\textwidth=15cm
\textheight=23.77cm
\oddsidemargin 0.46cm
\evensidemargin 0.46cm

\parindent=13pt

%%%%% Main spaces definition%%%%%


\def\loc{\mathop\mathrm{\,loc\,}}

\newcommand{\Sph}{\mathbb S^{n-1}}
\newcommand{\Rn}{\mathbb R^n}
\newcommand{\dd}{\,d}
\newcommand{\avgint}{\fint}
\newcommand{\Ep}{E_p}
\newcommand{\Var}{\operatorname{Var}}
\newcommand{\supp}{\operatorname{supp}}

\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{definition}
\newtheorem{remark}[theorem]{Remark}
\newtheorem{definition}[theorem]{Definition}

\newcounter{assum}
\renewcommand{\theassum}{\Roman{assum}}
\newtheorem{assumption}[assum]{Assumption}

\newtheorem{question}[theorem]{Question}
\renewcommand{\appendix}{\par
\setcounter{section}{0}%
\setcounter{subsection}{0}%
\setcounter{subsubsection}{0}%
\gdef\thesection{\@Alph\c@section}%
\gdef\thesubsection{\@Alph\c@section.\@arabic\c@subsection}%
\gdef\theHsection{\@Alph\c@section.}%
\gdef\theHsubsection{\@Alph\c@section.\@arabic\c@subsection}%
\csname appendixmore\endcsname
}

\numberwithin{equation}{section}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\def\Xint#1{\mathchoice
{\XXint\displaystyle\textstyle{#1}}%
{\XXint\textstyle\scriptstyle{#1}}%
{\XXint\scriptstyle\scriptscriptstyle{#1}}%
{\XXint\scriptscriptstyle\scriptscriptstyle{#1}}%
\!\int}
\def\XXint#1#2#3{{\setbox0=\hbox{$#1{#2#3}{\int}$ }
\vcenter{\hbox{$#2#3$ }}\kern-.6\wd0}}
\def\ddashint{\Xint=}
\def\dashint{\Xint-}


%\newcommand{\eq}[1]{\begin{align*}#1\end{align*}}

\def\Xint#1{\mathchoice
{\XXint\displaystyle\textstyle{#1}}%
{\XXint\textstyle\scriptstyle{#1}}%
{\XXint\scriptstyle\scriptscriptstyle{#1}}%
{\XXint\scriptscriptstyle\scriptscriptstyle{#1}}%
\!\int}
\def\XXint#1#2#3{{\setbox0=\hbox{$#1{#2#3}{\int}$ }
\vcenter{\hbox{$#2#3$ }}\kern-.55\wd0}}

\begin{document}

\arraycolsep=1pt

\title{\bf\Large Stability for the Affine Sobolev Inequality\\ and its Critical Points for $p \ge 2$
\footnotetext{\hspace{-0.35cm} \emph{Date}: \version
\endgraf
2020 \emph{Mathematics Subject Classification}. Primary 46E35;
Secondary 26D10.
\endgraf \emph{Key words and phrases}  Affine Sobolev inequality, gradient stability, quantitative inequalities,
critical points.
\endgraf
R.L.F. acknowledges partial support through US National Science Foundation grant DMS-1954995, as well as through the German Research Foundation through EXC-2111-390814868, TRR 352–Project-ID 470903074, and
FR 2664/3-1. 
Y.L. and D.Y. are supported by the National
Natural Science Foundation of China
(Grant Nos.\ 12431006, 12371093, and 124B2004),
the Beijing Natural Science Foundation (Grant No.\ 1262011),
and the Fundamental Research Funds
for the Central Universities (Grant No.\ 2253200028).}}
\author{Rupert L. Frank, Yinqin Li and Dachun Yang}
\date{\version}
\maketitle

\vspace{-0.6cm}

\begin{center}
\begin{minipage}{13cm}
{\small {\bf Abstract}\quad 
We prove stability for the affine Sobolev inequality for exponents $p\geq 2$ with best possible norm and best possible stability exponent. We prove a corresponding results for critical points of the functional in the absence of bubbling.
}
\end{minipage}
\end{center}

%\vspace{0.2cm}

\tableofcontents

%\vspace{0.2cm}

\section{Introduction and main results}

Sharp Sobolev inequalities are central objects in nonlinear partial differential equations and geometric analysis. As we will describe later in some detail, in the last two decades there has been a lot of activity concerning the stability question for these sharp inequalities and their critical points.

The aim of the present paper is to study the stability problem for the
\emph{affine Sobolev inequality}. This is a strengthening of the
classical Sobolev inequality that concerns the \emph{affine Sobolev energy}, defined for $f\in\dot W^{1,p}(\mathbb R^n)$, the homogeneous Sobolev space, by
\begin{align*}
  \mathcal E_p(f)
  :=
  \left[
    \int_{\mathbb S^{n-1}}
    \|\nabla_\xi f\|_p^{-n}\,d\xi
  \right]^{-\frac1n}.
\end{align*}
Here $\nabla_\xi f = \xi\cdot\nabla f$. The affine Sobolev inequality states that for $p\in[1,n)$ there is a constant $S_{\rm aff}>0$, depending only on $n$ and $p$, such that, for all $f\in \dot W^{1,p}$,
\begin{align}\label{eq-affSob}
\mathcal{E}_p(f)\ge S_{\rm aff}\|f\|_{p^*}.
\end{align}
Here $p^*=np/(n-p)$ and $\|\cdot\|_{p^*}$ denotes the $L^{p^*}$-norm on $\mathbb R^n$.

An important feature of the affine Sobolev inequality is its large symmetry group and, in particular, its affine invariance, which is \emph{not} shared by the classical Sobolev inequality. More specifically, if
$$
\tilde f(x) := (\det A)^\frac{n-p}{np} f(Ax+b)
$$
for some $A\in GL(n)$ and $b\in\mathbb R^n$, then
$$
\frac{\mathcal{E}_p(\tilde f)}{\|\tilde f\|_{p^*}} = \frac{\mathcal{E}_p(f)}{\|f\|_{p^*}}.
$$
In part due to this invariance, the affine Sobolev inequality is deeply related to the Brunn--Minkowski theory in convex geometry; see the survey \cite{g02} as well as the monographs \cite{bfr26,s14}.

In what follows we are interested in the sharp form of the affine Sobolev inequality \eqref{eq-affSob} and we agree to let $S_{\rm aff}$ denote the best possible constant in this inequality. It was found by Zhang \cite{z99} for $p=1$ and
by Lutwak et al.\ \cite{lyz02} for $p\in(1,n)$, who also characterized the optimizers in this inequality, that is, the set
\begin{equation}
    \label{eq:defmaff}
    \mathcal{M}_{\rm aff}:=\left\{ f\in \dot W^{1,p}\setminus\{{\bf 0}\} :\ \mathcal{E}_p(f)= S_{\rm aff}\|f\|_{p^*} \right\}.
\end{equation}

Our first main result quantifies the fact that, when $p\geq 2$, functions $f\in \dot W^{1,p}$ for which $\mathcal{E}_p(f)/\|f\|_{p^*}$ is close to $S_{\rm aff}$ are close to elements in $\mathcal M_{\rm aff}$.

\begin{theorem}\label{thm-stab}
If $p\in[2,n)$, then there is a positive constant $c_{n,p}$, 
depending only on $n$ and $p$,
such that, for any $f\in \dot{W}^{1,p}$,
\begin{align}\label{eq-stab}
&{\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p\notag\\
&\quad\ge c_{n,p}
\inf_{v\in \mathcal M_{\rm aff},\, B\in SL(n)}\left\{\|\nabla(f\circ B -v)\|_{p}^p
+\int_{\mathbb{R}^n}
|\nabla v|^{p-2}|\nabla(f\circ B- v)|^2
\right\}.
\end{align}
\end{theorem}

\begin{remark}
    (a) Indeed, we obtain a stronger conclusion in our proof;
    see \eqref{eq-stab-e}.
    This improves upon a result in a concurrent and independent work \cite{flz26b}, where only the first term on the right side of \eqref{eq-stab} appears.\\
    (b) The first term on the right side is the $p$-th power of a distance and is reminiscent of work by Figalli and Zhang on the classical Sobolev inequality \cite{fz22}. In Section \ref{Sec-sharp} we will show that this $p$-th power of the distance is best possible, again in analogy with the classical Sobolev inequality. The second term on the right side of \eqref{eq-stab}, in contrast, is quadratic. This is reminiscent of \cite{fp24} and the quadratic, rather than $p$-th order vanishing of the distance is crucial in some applications \cite{iz25}. In Section \ref{Sec-sharp} we will show that this power $2$ of the distance is best possible. In this sense our result is best possible.\\
    (c) The infimum over $B\in SL(n)$ on the right side of \eqref{eq-stab} makes the right side affine invariant. Thus, our stability inequality inherits the same symmetries as the underlying sharp affine Sobolev inequality.\\
    (d) We should stress that the thrust of Theorem \ref{thm-stab} is the case $p>2$, as the result for $p=2$ follows easily from well-known results; see \eqref{eq-p=2} below. The restriction to $p\geq 2$ is for technical reasons and it is an open problem to remove this assumption.
\end{remark}


Our second main result concerns stability for critical points of the functional $\mathcal E_p(f)/\|f\|_{p_*}$ in the absence of bubbling. The Euler--Lagrange equation involves the \emph{affine $p$-Laplacian}, defined by
\begin{align*}
\Delta_{p}^{\rm aff}(h) :=
\mathcal{E}_p(h)^{n+p}\operatorname{div}
\left(\int_{\mathbb{S}^{n-1}}
\|\nabla_\xi h\|_p^{-n-p}|\nabla_\xi h|^{p-2}[\nabla_\xi h]\xi\,d\xi\right).
\end{align*}
For $h\in \dot W^{1,p}$ we have $\Delta_p^{\rm aff}(h)\in\dot W^{-1,p'}$, where $p'$ is the H\"older dual exponent of $p$, satisfying $\frac1p+\frac1{p'}=1$. Note that $\Delta_p^{\rm aff}$ is a nonlocal, quasilinear differential operator. It arises naturally in the context of the affine Sobolev inequality, since
$$
\frac{d}{dt}\Big|_{t=0} \mathcal E_p(h+tg) = - \mathcal E_p(h)^{1-p} \int_{\mathbb R^n} g \Delta_{p}^{\rm aff}(h).
$$
The affine invariance of the affine Sobolev energy shows that, for every $A\in GL(n)$,
\begin{align}\label{eq-T}
\Delta_p^{\rm aff}(f\circ A)=|\det A|^{\frac pn}
(\Delta_p^{\rm aff}f)\circ A.
\end{align}

Let us introduce
\begin{equation}
    \label{eq:defmaff1}
    \mathcal M_{\rm aff}^{(1)} := \left\{ f\in \mathcal M_{\rm aff} :\ \|f\|_{p_*}=1 \right\}.
\end{equation}
Then the Euler--Lagrange equation for optimizers together with the normalization show that, for all $f\in\mathcal M_{\rm aff}^{(1)}$,
\begin{equation}
    \label{eq:el}
    -\Delta_{p}^{\rm aff}(f) = S_{\rm aff}^p|f|^{p^*-2}f.
\end{equation}

Our second main result will say that if $f$ `almost' satisfies \eqref{eq:el} and it is not too far from $\mathcal M_{\rm aff}^{(1)}$, then it is close to $\mathcal M_{\rm aff}^{(1)}$ in a quantitative sense. The quantity that measures that \eqref{eq:el} is almost satisfied is the following \emph{affine Euler--Lagrange residual}
\begin{align*}
\delta_{\rm aff}(f)
:=\sup_{B\in SL(n)} \left\|\Delta_p^{\rm aff}\left(f\circ B\right) + S_{\rm aff}^p |f\circ B|^{p^*-2} (f\circ B) \right\|_{
\dot{W}^{-1,p'}}.
\end{align*}


\begin{theorem}\label{thm-bubble}
Let $p\in[2,n)$. There are positive constants 
$c_{n,p}$ and $\varepsilon_{n,p}$, depending only on $n$ and $p$,
such that, for every positive $f\in\dot{W}^{1,p}$ with
$\inf_{B\in \operatorname{SL}(n),v\in\mathcal{M}_{\rm aff}^{(1)}}
\|\nabla(f\circ B-v)\|_p<\varepsilon_{n,p}$,
\begin{align}\label{thm-bubble-e1}
\delta_{\rm aff}(f)\ge c_{n,p}
\inf_{B\in \operatorname{SL}(n),v\in\mathcal{M}_{\rm aff}^{(1)}}
\|\nabla(f\circ B-v)\|_p^{p-1}.
\end{align}
\end{theorem}

\begin{remark}
    (a) The power $p-1$ on the right side of \eqref{thm-bubble-e1} is reminiscent of work by Liu and Zhang on the classical Sobolev inequality \cite{lz25}. In Section \ref{Sec-sharp} we will show that this $(p-1)$-st power of the distance is best possible, again in analogy with critical points of the classical Sobolev inequality.\\
    (b) The infimum over $B\in SL(n)$ on the right side of \eqref{thm-bubble-e1} makes the right side affine invariant, while the left side is affine invariant be definition of the residual $\delta_{\rm aff}(f)$.\\
    (c) The term $\delta_{\rm aff}$ is a natural affine invariant quantity to measure the affine Euler--Lagrange residual.
Indeed, in \eqref{eq-p} below we will recall that $\mathcal E_p(f)$ is equivalent to $\min_{T\in SL(n)} \|\nabla(f\circ T)\|_p$, from which it follows that, with $\langle\cdot,\cdot\rangle$ denoting the $\dot W^{-1,p'}\times \dot W^{1,p}$ duality,
\begin{align*}
\delta_{\rm aff}(f)
& =\sup_{B\in\operatorname{SL}(n)}\sup_{\varphi\ne {\bf 0}}
\frac{|\langle \Delta_p^{\rm aff}\left(f\circ B\right) + S_{\rm aff}^p |f\circ B|^{p^*-2} (f\circ B),\varphi\rangle|}{\|\nabla \varphi\|_p}\\
&=\sup_{\psi\ne {\bf 0}}\sup_{T\in\operatorname{SL}(n)}
\frac{|\langle \Delta_p^{\rm aff}\left(f\right) + S_{\rm aff}^p |f|^{p^*-2} f,\psi\rangle|}{\|\nabla(\psi\circ T)\|_p} \\
& =\sup_{\psi\ne {\bf 0}}\frac{|\langle \Delta_p^{\rm aff}\left(f\right) + S_{\rm aff}^p |f|^{p^*-2} f,\psi\rangle|}{\inf_{T\in\operatorname{SL}(n)}\|\nabla(\psi\circ T)\|_p} \\
& 
\sim \sup_{\varphi\ne {\bf 0}}\frac{|\langle \Delta_p^{\rm aff}\left(f\right) + S_{\rm aff}^p |f|^{p^*-2} f,\psi\rangle|}{\mathcal{E}_p(\psi)}.
\end{align*}
Here we used \eqref{eq-T}. This shows that $\delta_{\rm aff}$
is the natural dual norm associated with the affine Sobolev energy.
\end{remark} 

%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Background}

\subsubsection*{The affine vs.~the classical Sobolev inequality}

It is natural to compare the affine Sobolev inequality with its classical variant, which states that for $p\in[1,n)$ there is a constant $S_{\rm Sob}$, depending only on $n$ and $p$, such that,
for all $f\in \dot W^{1,p}$,
\begin{align}\label{eq-Sob}
  \|\nabla f\|_p
  \ge
  S_{\rm Sob}\|f\|_{p^*}.
\end{align}
We recall that the homogeneous Sobolev space $\dot W^{1,p}$ consists of all weakly differentiable functions $f$ on $\mathbb R^n$ whose gradient belows to $L^p$ and which vanish at infinity in the sense that $|\{ |f|>\lambda\}|<\infty$ for any $\lambda>0$.

We agree to let $S_{\rm Sob}$ denote the \emph{sharp constant} in inequality \eqref{eq-Sob}. It was found by Rodemich, Aubin and Talenti \cite{r66,a76,t76}, who also characterized the optimizers in this inequality, that is, they determined the set
\begin{equation}
    \label{eq:defmsob}
    \mathcal{M}_{\rm Sob}:=\left\{ f\in \dot W^{1,p}\setminus\{{\bf 0}\} :\ \|\nabla f\|_p = S_{\rm Sob}\|f\|_{p^*} \right\}.
\end{equation}

An important feature of the sharp affine Sobolev inequality is that it strengthens the classical sharp Sobolev inequality. Indeed, by H\"older's inequality we find
\begin{align}\label{eq-holder}
  \mathcal E_p(f)\le
    |\mathbb{S}^{n-1}|^{-\frac1n-\frac1p} 
  \left[
    \int_{\mathbb S^{n-1}}
    \|\nabla_\xi f\|_p^p\,d\xi
  \right]^{\frac1p}
  =
  |\mathbb{S}^{n-1}|^{-\frac1n-\frac1p}
\alpha_{n,p}
  \|\nabla f\|_p
\end{align}
with
\begin{align}\label{eq-alpha}
\alpha_{n,q}:=\left[\int_{\mathbb{S}^{n-1}}|e\cdot\xi|^q\,d\xi\right]^{\frac1q}
= \left[ \frac{2\pi^{\frac{n-1}{2}}\Gamma(\frac{q+1}{2})}{\Gamma(\frac{q+n}{2})} \right]^\frac1q,
\quad\forall\,e\in\mathbb{S}^{n-1}.
\end{align}
By comparing the explicit values of the constants, one finds that
\begin{equation}
    \label{eq-relaSob}
    S_{\rm Sob} = \left( |\mathbb{S}^{n-1}|^{-\frac1n-\frac1p}
\alpha_{n,p} \right)^{-1} S_{\rm aff},    
\end{equation}
so the sharp affine Sobolev inequality implies the classical Sobolev inequality.

The affine Sobolev inequality is mostly of interest when $p\neq 2$, because for $p=2$ one finds, for any $f\in\dot{W}^{1,2}$,
\begin{align}\label{eq-p=2}
  \mathcal E_2(f)
  =
   |\mathbb{S}^{n-1}|^{-\frac1n-\frac12}
\alpha_{n,2}
  \min_{T\in \operatorname{SL}(n)}
  \|\nabla(f\circ T)\|_2,
\end{align}
see, for instance, \cite[Lemma 2.1 and Corollary 2.2]{st18}. The identity \eqref{eq-p=2} is special to the Hilbertian case and 
does not hold for general $p$. For general $p\in(1,n)$, one only has the equivalence 
\begin{align}\label{eq-p} 
\mathcal E_p(f) \sim \min_{T\in \operatorname{SL}(n)} \|\nabla(f\circ T)\|_p, 
\end{align} 
where the positive equivalence constants depend only on $n$ and $p$; see \cite[Theorem 1.2]{hl16}. Moreover, these equivalence constants are not the constant appearing in \eqref{eq-relaSob}.
Thus, the above linear reduction is not available
in the non-Hilbertian case.

In passing we mention some other developments of affine Sobolev inequalities,
for example, the affine Sobolev--Zhang inequality on $BV(\mathbb R^n)$ \cite{w12}, the asymmetric affine Sobolev inequalities
\cite{hs09}, the affine Moser--Trudinger and Morrey--Sobolev inequalities \cite{clyz09} and fractional affine inequalities 
\cite{DLTYY25,hl24,hl24-2}.

There is also a natural spectral and PDE side of the affine theory. Haddad et al.\ 
\cite{hjm21} developed affine Poincar\'e inequalities and affine spectral
inequalities. They introduced the affine $p$-Laplacian associated
with the affine $p$-energy. This operator arises as the Euler--Lagrange operator for
the affine $p$-Rayleigh quotient. It is affine invariant and coincides with the
usual $p$-Laplacian on radial functions. In the Hilbertian case $p=2$, this operator
coincides with the affine Laplacian introduced by Schindler and Tintarev
\cite{st18}. In this case, the affine Laplacian can be viewed as the usual Laplace
operator in a suitable coordinate system depending on the function. This gives a
natural PDE framework for extremals and eigenvalue problems related to affine
Sobolev-type inequalities.

\subsubsection*{Stability of the classical and affine Sobolev inequalities}

Lions \cite{l85} has proved qualitative stability for the classical Sobolev inequality in the sense that if the \emph{Sobolev
deficit}
\begin{align*}
\delta_{\rm Sob}(f):=\frac{\|\nabla f\|_p}{\|f\|_{p^*}}
  -S_{\rm Sob}
\end{align*}
is small, then the distance of $f$ to the set $\mathcal M_{\rm Sob}$ is small. A natural problem, raised by Brezis and Lieb \cite{bl85}, is whether there is a quantitative version of this stability result, in the sense that the Sobolev deficit controls a suitable distance from a function to the manifold $\mathcal M_{\rm Sob}$.
For $p=2$, this question was answered by Bianchi and Egnell \cite{be91}. They
proved that the deficit controls the squared $\dot W^{1,2}$-distance to
$\mathcal M_{\rm Sob}$. Their estimate is sharp both in the strength of the distance and in
the exponent. A stability estimate that reflects the optimal dependence of the constant on the dimension $n$ was recently obtained in \cite{deffl25}.

For $p\neq2$, the stability problem is more difficult, since the Hilbert space structure is no longer available. After preliminary works in \cite{cfmp09,fn19,n20} the sharp gradient stability estimate for the full range $1<p<n$ was proved by
Figalli and Zhang \cite{fz22}. More precisely, if $p\in(1,n)$, then 
there is a positive constant $c$, depending only on $n$ and $p$, 
such that, for all $f\in\dot{W}^{1,p}$,
\begin{align}\label{eq:figallizhang}
 \delta_{\rm Sob}(f)
  \ge
  c
  \inf_{v\in\mathcal M_{\rm Sob}}
  \left(
  \frac{\|\nabla f-\nabla v\|_p}{\|\nabla f\|_p}
  \right)^{\max\{2,p\}}.
\end{align}
Thus the classical Sobolev deficit controls the strongest natural distance to the family of extremal functions, namely the distance at the level of gradients. 

Moreover, the exponent $\max\{2,p\}$ is sharp. In particular, the fact that the exponent is strictly bigger than two for $p>2$ is surprising. This Figalli--Zhang phenomenon has since then also been observed in other stability inequalities, including \cite{glz25,fp24,wz25,fpr25}. As we have already mentioned after Theorem \ref{thm-stab}, it was noticed in \cite{fp24} and later independently in \cite{iz25} that the Figalli--Zhang bound can be strengthened for $p>2$ to
$$
\delta_{\rm Sob}(f)
  \ge
  c
  \inf_{v\in\mathcal M_{\rm Sob}}
  \left(
  \frac{\|\nabla f-\nabla v\|_p^p}{\|\nabla f\|_p^p} + \frac{\||\nabla v|^{(p-2)/2}(\nabla f-\nabla v)\|_2^2}{\|\nabla f \|_p^2}
  \right).
$$
The second term on the right side is important in some applications \cite{iz25}.

In passing we mention a completely different mechanism for stability of functional inequalities with an exponent strictly bigger than two, which is explored, for instance, in \cite{ens22,f24,fp24a}. See also the survey \cite{f22}.

So far, our discussion of stability has focused on the classical Sobolev inequality. There are also a few existing works in the affine case. Wang \cite{w13} studied
stability for the affine P\'olya--Szeg\H{o} principle and Nguyen \cite{n16} proved a stability result for
the affine Sobolev inequality on $BV(\mathbb R^n)$. Finding the optimal exponent in the latter case, however, is still an open problem. In addition to the work \cite{flz26b} by Fan et al., which we have already mentioned and where a slightly weaker form of Theorem \ref{thm-stab} was proved, we mention the sharp stability result for the affine fractional Sobolev inequality \cite{flz26} by the same authors.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsubsection*{Stability of critical points of the classical Sobolev inequalities}

Another natural direction is to study stability at the level of critical
points, rather than minimizers, of the Sobolev quotient. More precisely, one
may ask whether a function that almost solves the Euler--Lagrange equation
associated with the sharp Sobolev inequality must be quantitatively close to
an Aubin--Talenti bubble. In the classical notation, this means estimating the
distance from $u$ to the family of bubbles in terms of the $\dot W^{-1,p'}$-norm of the residual
$$
-\operatorname{div}(|\nabla u|^{p-2}\nabla u)
-
|u|^{p^*-2}u \,.
$$

The qualitative theory goes back to the compactness theorem of
Struwe \cite{s84}. In the case $p=2$, Struwe proved that a nonnegative
Palais--Smale sequence for the critical Sobolev functional decomposes, up to
a small error in $\dot W^{1,2}$, into a sum of weakly interacting Talenti
bubbles. This phenomenon is usually referred to as bubbling. The
corresponding compactness theory for the critical $p$-Laplacian was
later developed by Alves \cite{a02} and Mercuri and Willem \cite{mw10}. Thus,
under an a prior energy condition that excludes the appearance of more
than one bubble, the qualitative theory implies that an almost critical
point is close to a single bubble.

The first sharp quantitative result in this direction was obtained, for
$p=2$ and in the single-bubble regime, by Ciraolo et al.\
\cite{cfm18}. They proved a linear estimate of the $\dot W^{1,2}$-distance
from a bubble in terms of the Euler--Lagrange residual. The multiple-bubble
case was then studied by Figalli and Glaudo \cite{fg20}, who proved that the
same linear control holds in dimensions $3\le n\le5$, while it fails in
dimensions $n\ge6$. The sharp quantitative estimates in the higher
dimensional bubbling regime were later obtained by Deng et al.\ \cite{dsw25}. The bound is no longer linear due to the interaction of the bubbles.

For general $1<p<n$, the problem is substantially more delicate and still not completely solved. Recently, Liu and Zhang \cite{lz25}, following the
sharp stability framework of Figalli and Zhang \cite{fz22}, proved the sharp
single-bubble estimate
$$
\left\|\nabla u-\nabla v \right\|_{L^p}^{\max\{1,p-1\}}
\le
C
\left\|\operatorname{div}(|\nabla u|^{p-2}\nabla u)
+ |u|^{p^*-2}u \right\|_{W^{-1,p'}}
$$
whenever $u$ is sufficiently close to a single Aubin--Talenti bubble $v$. This
result gives the sharp quantitative stability of the Struwe-type
decomposition in the absence of bubbling. The exponent $\max\{1,p-1\}$ is optimal and is related to the exponent $\max\{2,p\}$ in \eqref{eq:figallizhang}.

Independently, Ciraolo and Gatti \cite{cg26} developed a different,
more PDE-oriented approach based on $P$-functions and integral identities.
Their method yields a quantitative closeness result for perturbations of the
critical $p$-Laplace equation, with a non-sharp exponent. Although the exponent
is not optimal, their approach is rather flexible; this has recently been
exploited by Antonini et al.\ \cite{acg26} to study classification,
Struwe-type decomposition, and stability for the anisotropic critical
$p$-Laplace equation. The analogue of the multi-bubble results by Figalli and Glaudo \cite{fg20} and Deng et al.\ \cite{dsw25} is still open.

Related to the `other' mechanism for non-quadratic stability for Sobolev-type inequalities mentioned before, there is a corresponding `other' mechanism for non-linear stability for critical points. In the setting of Caffarelli--Kohn--Nirenberg inequalities this is investigated in \cite{zz26,ww26}.


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Organization of this paper}

\subsubsection*{Outline of the proof of Theorem \ref{thm-stab}}

The overall strategy follows the sharp
gradient stability argument of Figalli--Zhang \cite{fz22}
(see also \cite{be91,fn19}). We first use a compactness argument to reduce the problem to functions that are close, up to
affine transformations, to the manifold $\mathcal M_{\rm aff}$ in Section \ref{sec-compact}. 
Then the main new difficulty is the expansion of the affine energy in the sense of 
\cite[Section 2]{fz22},
which is the objective of Section \ref{sec-expan}. Unlike the classical
Dirichlet energy, $\mathcal E_p$ is obtained from the directional quantities
$\{\|\nabla_\xi f\|_p^p\}_{\xi\in\mathbb S^{n-1}}$
through a nonlinear negative-power average on the sphere. Therefore, after expanding
the directional energies, one also has to expand the outer affine functional. This
produces an additional second-order term of variance type on the sphere 
[see \eqref{prop-expan-local-e2}]. Such a
term is absent in the classical Sobolev inequality. Its role is to detect the extra
trace-free affine directions in the tangent space of $\mathcal M_{\rm aff}$.

A key point is to control this new variance term. Following the affine Hessian
viewpoint used in the affine fractional setting of Fan et al.\ \cite{flz26}, we
establish an affine spectral gap inequality involving
this variance term in Section \ref{sec-spectral}
(precisely, Theorems \ref{thm-spectral} and \ref{thm-spectral-wei}). A similar analysis has appeared independently and concurrently in \cite{flz26b}. Combining these and the nonlinear
expansion, we obtain the desired lower bound for the affine deficit,
and finish the proof of Theorem \ref{thm-stab} in Section \ref{sec-proof}.
Some technical estimates are collected in two appendices.

%%%%%%%%%%%%%%%%%%%%%%%%

\subsubsection*{Outline of the proof of Theorem \ref{thm-bubble}}

Theorem \ref{thm-bubble} is a local formulation 
and the overall strategy of proof follows Liu and Zhang \cite[Theorem 1.2]{lz25}.
Since we have already assumed that the function under consideration is
close to the normalized affine Talenti bubble,
we expand the affine Laplacian around this bubble in 
Proposition \ref{prop-expan-lap}. The expansion is parallel to the one used
in the proof of Theorem \ref{thm-stab}: after expanding the directional terms 
and the outer affine functional,
one obtains the same variance operator on $\mathbb S^{n-1}$. Hence,
we need to establish the corresponding 
affine spectral gap inequality in Theorem \ref{thm-spectral-wei}.

There is, however, an additional point in the critical point problem. The normalized
affine bubbles solve the Euler--Lagrange equation with a \emph{fixed amplitude}, while the
spectral gap is naturally formulated after removing the full tangent space generated by the
amplitude, scaling, translation, and trace-free affine directions. To handle this
mismatch, we first perform the modulation on the enlarged cone of affine bubbles with
a free amplitude parameter. This gives the desired orthogonality with respect to this
full tangent space of directions. We then test the residual against the approximate bubble to
prove that the optimal amplitude is close to one [see \eqref{thm-bubble-e4}]. 
This is the affine counterpart of the
coefficient estimate in Figalli and Glaudo \cite[Proposition 3.11]{fg20}. Combining this amplitude estimate
with the above affine spectral gap inequality and the nonlinear expansion, we obtain the
desired quantitative stability estimate for the affine Euler--Lagrange equation.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Explicit parametrization}

When presenting our main results, Theorems \ref{thm-stab} and \ref{thm-bubble}, we have opted for an `abstract' formulation, in the sense that only the sets $\mathcal M_{\rm aff}$ and $\mathcal{M}_{\rm aff}^{(1)}$ appear and not their explicit form. However, when proving these results, we need the explicit form of all the elements of these sets. We collect all the relevant information here for future reference.

An important role is played by the function
\begin{equation}
    \label{eq:standardminimizer}
    U(x) := c_0 \, (1+|x|^{p'})^{\frac{p-n}{p}},
\end{equation}
where the constant $c_0>0$, depending on $n$ and $p$, is determined in such a way that
\begin{equation}
    \label{eq:standardminimizernorm}
    \|U\|_{p^*}=1.
\end{equation}
Then the theorem by Lutwak et al.\ \cite{lyz02} (see also \cite{hjm16}) says that $U$ is the unique optimizer of the affine Sobolev inequality \eqref{eq-affSob} up to symmetries. More precisely, for the set $\mathcal M_{\rm aff}$, defined in \eqref{eq:defmaff}, we have
\begin{align}\label{def-min}
\mathcal{M}_{\rm aff}=\left\{c\, |\det B|^{\frac{p-n}{np}}
U\left(B^{-1}(\cdot-x_0)\right):
c\in\mathbb{R}\setminus\{0\},\
B\in\operatorname{GL}(n),\ x_0\in\mathbb{R}^n\right\}.
\end{align}

This theorem should be compared with the Rodemich--Aubin--Talenti theorem, which characterizes the optimizers of the classical Sobolev inequality \eqref{eq-Sob}. Again $U_0$ is the unique minimizer up to symmetries of the inequality, but now the group of symmetries is smaller. We see that $\mathcal M_{\rm Sob}$, defined in \eqref{eq:defmsob}, is given by the same formula as \eqref{def-min}, but with matrices $B$ restricted to be of the form $B=\lambda Q$ with $\lambda>0$ and $Q\in O(n)$.

We prefer to work with a different parametrization of $\mathcal M_{\rm aff}$ than in \eqref{def-min}. For every $B\in \operatorname{GL}(n)$, 
by the polar decomposition, 
there are unique $\lambda\in(0,\infty)$, 
$S\in M_n(\mathbb{R})$ satisfying $S=S^T$ and $\operatorname{Tr} S=0$,
and $Q\in O(n)$ such that 
$B=\lambda Q e^{S}$.
Since $U$ is radial, 
it is invariant under $Q\in O(n)$ and hence, for every
$x_0\in\mathbb R^n$,
$$
U\left(B^{-1}(\cdot-x_0)\right)
=U\left(\lambda^{-1} e^{-S}(\cdot-x_0)\right).
$$
Therefore, by \eqref{def-min},
we obtain 
\begin{align}\label{def-min-equiv}
\mathcal{M}_{\rm aff}=\left\{cT_{\lambda,S,x}U:
c\in\mathbb{R}\setminus\{0\},\ \lambda\in(0,\infty),
\ S=S^T,\ \operatorname{Tr}S=0,\ x\in\mathbb{R}^n\right\},
\end{align}
where, for any $f\in L^1_{\rm loc}$,
\begin{align*}
T_{\lambda,S,x}f(\cdot):=\lambda^{-\frac{n-p}{p}}f\left(\lambda^{-1}e^{-S}
(\cdot-x)\right).
\end{align*}
Here and in what follows, 
we refer to $(c,\lambda,S,x)$ as the \emph{affine transformation parameters}
and 
$cT_{\lambda,S,x}$ as the \emph{affine transformation}, respectively.
The advantage
of this parametrization is that the parameters
$(c,\lambda,S,x)$ are uniquely determined by the corresponding extremal
function. Moreover, we define
$$
\mathscr{A}:=(\mathbb{R}\setminus\{0\})\times(0,\infty)
\times\left\{
A\in M_n(\mathbb{R}): A=A^T,\ \operatorname{tr}A=0
\right\}\times\mathbb{R}^n.
$$
and consider $\mathscr A$ as a finite-dimensional parameter space, endowed
with the topology induced from
$
\mathbb R\times(0,\infty)\times M_n(\mathbb R)\times\mathbb R^n.
$

After introducing these symbols, we can restate the result in Theorem \ref{thm-stab}
as follows: If $p\in[2,n)$, then there is a positive constant $c_{n,p}$, 
depending only on $n$ and $p$,
such that, for any $f\in \dot{W}^{1,p}$,
\begin{align}\label{eq-stab-e}
&{\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p\notag\\
&\quad\ge c_{n,p}
\inf_{(c,\lambda,S,x)\in\mathscr{A}}\left\{\|\nabla(T_{\lambda,S,x}f-cU)\|_{p}^p
+\int_{\mathbb{R}^n}
|c\nabla U|^{p-2}|\nabla(T_{\lambda,S,x}f-cU)|^2
\right\}.
\end{align}
Actually, it is stronger than \eqref{eq-stab}.

On the other hand, we can also restate Theorem \ref{thm-bubble}.
Since $T_{\lambda,S,x}$ preserves the $L^{p*}$-norm, \eqref{eq:standardminimizernorm} implies that 
the set $\mathcal M_{\rm aff}^{(1)}$, defined in \eqref{eq:defmaff1}, is given by
$$
\mathcal M_{\rm aff}^{(1)} = \left\{T_{\lambda,S,x}U:\lambda\in(0,\infty), \ S=S^T,\ \operatorname{Tr}S=0,\ x\in\mathbb{R}^n\right\}.
$$
Define 
\begin{align*}
P(f):=-\Delta_{p}^{\rm aff}(f)-S_{\rm aff}^p|f|^{p^*-2}f
\end{align*}
One can easily verify that, for all $\lambda\in(0,\infty)$,
symmetric matrix $S$ satisfying $\operatorname{Tr}S=0$,
and $x\in\mathbb{R}^n$,
\begin{align}\label{eq-P}
\int_{\mathbb{R}^n}P(T_{\lambda,S,x}f)g
=\int_{\mathbb{R}^n}P(f)T_{\lambda,S,x}^{-1}g,
\end{align}
here and in what follows, 
\begin{align*}
T_{\lambda,S,x}^{-1}f(\cdot):
=\lambda^{\frac{n-p}{p}}f\left(\lambda e^{S}
\cdot+x\right).
\end{align*}
Also define
\begin{align*}
d(f,\mathcal{M}_{\rm aff}^{(1)}):
=\inf\left\{\|\nabla (T_{\lambda,S,x}f-U)\|_p:
\lambda\in(0,\infty),
\ S=S^T,\ \operatorname{tr}S=0,\ x\in\mathbb{R}^n\right\}.
\end{align*}
Then, similar to \eqref{eq-stab-e}, 
we actually get the following improvement of 
Theorem \ref{thm-bubble}:
Let $p\in[2,n)$. There exist positive constants 
$c_{n,p}$ and $\varepsilon_{n,p}$, depending only on $n$ and $p$,
such that, for every positive $f\in\dot{W}^{1,p}$ such that 
$d(f,\mathcal{M}_{\rm aff}^{(1)})<\varepsilon_{n,p}$,
\begin{align}\label{thm-bubble-e6}
\delta_{\rm aff}(f)\ge c_{n,p}d(f,\mathcal{M}_{\rm aff}^{(1)})^{p-1}.
\end{align}

\emph{Notation.}
Let $\mathbb{N}:=\{1,2,\ldots\}$ and $\mathbb{Z}_+:=\mathbb{N}\cup\{0\}$.
In addition, we always denote by $C$ a
\emph{positive constant} which is independent
of the main parameters involved, but may vary from line to line.
We use $C_{\alpha,\dots}$ to denote a positive constant depending
on the indicated parameters $\alpha,\, \dots$.
The symbol $f\lesssim g$ means $f\le Cg$
and, if $f\lesssim g\lesssim f$, then we write $f\sim g$.

\section{Relative compactness}\label{sec-compact}

The main result of this section is the relative compactness
of affine Sobolev inequalities.
For $q\in[1,\infty)$, $\lambda\in(0,\infty)$,
$A\in \operatorname{SL}(n)$, and 
$x\in {\mathbb R}^n$, define the general affine transformation by setting,
for every $f\in L^1_{\rm loc}$,
\begin{align}\label{def-tran-g}
{\mathcal T}_{\lambda,A,x}^{(q)} f(\cdot)
:=
\lambda^{-\frac nq}f\left(\lambda^{-1}A^{-1}(\cdot-x)\right).
\end{align}
Note that ${\mathcal T}_{\lambda,A,x}^{(q)}$ is an isometry on $L^q$
and, for every $\lambda\in(0,\infty)$,
every symmetric matrix $S$ satisfying $\operatorname{Tr}{S}=0$, and
every $x\in\mathbb{R}^n$, 
\begin{align*}
{\mathcal T}_{\lambda, e^S,x}^{(p^*)}
=
T_{\lambda,S,x}.
\end{align*}
Unlike the affine transformations used in the equivalent 
definition of the extremal manifold in 
\eqref{def-min-equiv}, for a general function one cannot use the polar decomposition to remove 
the rotational part of a volume-preserving linear map. Indeed, 
such a reduction is possible for the bubble only because the bubble is radial. 
Therefore, in the compactness argument below, we shall use the full affine action.

\begin{theorem}\label{thm-comp}
Let $p\in[2,n)$. If a sequence
$\{u_k\}_{k\in\mathbb{N}}$ in $\dot{W}^{1,p}$ satisfies 
that, when $k\to\infty$, $\mathcal{E}_p(u_k)\to S_{\rm aff}$ 
and $\|u_k\|_{p^*}\to1$,
then there are $u\in\mathcal{M}_{\rm aff}$, $\{\lambda_k\}_{k\in\mathbb{N}}$
in $(0,\infty)$, 
$\{x_k\}_{k\in\mathbb{N}}$
in $\mathbb{R}^n$, and $\{A_k\}_{k\in\mathbb{N}}$ in 
$\operatorname{SL}(n)$ such that 
$\mathcal{T}^{(p^*)}_{\lambda_k,A_k,x_k}u_k$
converges to $u$ in $\dot{W}^{1,p}$ as $k\to\infty$.
\end{theorem}

To show this result, we need the profile decomposition 
in $\dot{W}^{1,p}$, which was established in \cite{j99}.
We first recall some necessary concepts.
Let $p\in(0,\infty)$.
For any $\mathfrak{g}:=(\lambda,x_0)\in(0,\infty)\times\mathbb{R}^n$,
define an operator $\mathfrak{g}$ on $L^1_{\operatorname{loc}}$ by setting,
for all $u\in L^1_{\operatorname{loc}}$,
\begin{align*}
\mathfrak{g}(u;p):=\lambda^{-\frac np}u(\lambda^{-1}[\cdot-x_0]);
\end{align*}
then the inverse mapping $\mathfrak{g}^{-1}$ is given by 
\begin{align*}
\mathfrak{g}^{-1}(u;p)=\lambda^{\frac np}u(\lambda\cdot+x_0).
\end{align*}
A sequence $\{\mathfrak{g}_k\}_{k\in\mathbb{N}}=\{(\lambda_k,x_k)\}_{k\in\mathbb{N}}$
is said to \emph{tend to infinity} if 
$|\log \lambda_n|+|x_k|\to\infty$ as $k\to\infty$.
Two sequences $\{\mathfrak{g}_{1,k}\}_{k\in\mathbb{N}}$
and $\{\mathfrak{g}_{2,k}\}_{k\in\mathbb{N}}$ taking the above form
is said to be \emph{orthogonal} if $\mathfrak{g}_{1,k}\mathfrak{g}^{-1}_{2,k}$
tends to infinity.
Then we have the following profile decomposition (see \cite{j99}
or \cite[Theorem B.1]{o22}).

\begin{lemma}\label{lem-pdecom}
Let $p\in(1,n)$. If $\{w_k\}_{k\in\mathbb{N}}$
is a bounded sequence in $\dot{W}^{1,p}$, then 
there are $\{V_j\}_{j\in\mathbb{N}}$
in $\dot{W}^{1,p}$, $\{g_{j,k}\}_{j,k\in\mathbb{N}}
=\{(\lambda_{j,k},x_{j,k})\}_{j,k\in\mathbb{N}}$
in $(0,\infty)\times\mathbb{R}^n$, and $\{r_{k}^J\}_{J,k\in\mathbb{N}}$
in $L^{p*}$ such that
\begin{enumerate}[{\rm(i)}] 
  \item\label{it1-pdecom} for all $J,k\in\mathbb{N}$,
  \begin{align*}
  w_k=\sum_{j=1}^{J}\mathfrak{g}_{j,k}(V_j;p^*)
  +r_k^{J};
  \end{align*}
  \item\label{it2-pdecom} For any $j_1,j_2\in\mathbb{N}$ with $j_1\ne j_2$,
  $\{\mathfrak{g}_{j_1,k}\}_{k\in\mathbb{N}}$
  and $\{\mathfrak{g}_{j_2,k}\}_{k\in\mathbb{N}}$ are orthogonal;
  \item\label{it3-pdecom} For any $J\in\mathbb{N}$ and $j\in\{1,\ldots,J\}$,
  $\mathfrak{g}_{j,k}^{-1}(r_k^J;p^*)\rightharpoonup0$ in $\dot{W}^{1,p}$
  as $k\to\infty$, and 
  \begin{align*}
  \lim_{J\to\infty}\limsup_{k\to\infty}
  \|r_k^J\|_{p^*}=0.
  \end{align*}
\end{enumerate}
\end{lemma}

We also need some technical lemmas
that are presented in Appendices \ref{App-B} and \ref{App-C}.

\begin{proof}[Proof of Theorem \ref{thm-comp}]
By the assumption on $\{u_k\}_{k\in\mathbb{N}}$, 
we know that $\{\mathcal{E}_p(u_k)\}_{k\in\mathbb{N}}$
is bounded. Consequently, from \cite[Theorem 1.2]{hl16}
[see also \eqref{eq-p}],
we deduce that, for each $k\in\mathbb{N}$, there is
$T_k\in\operatorname{SL}(n)$ such that $\|\nabla(u_k\circ T_k)\|_p\lesssim
\mathcal{E}_p(u_k)\lesssim1$.
This shows that $\{w_k\}_{k\in\mathbb{N}}:=\{u_k\circ T_k\}_{k\in\mathbb{N}}$
is a bounded sequence in $\dot{W}^{1,p}$.
Then, applying the profile decomposition, the conclusion of Lemma \ref{lem-pdecom}
holds for $\{w_k\}_{k\in\mathbb{N}}$ here. Assume all the symbols 
are the same as in Lemma \ref{lem-pdecom}.
By \cite[(3.1)]{o22} [see also \eqref{eq-pdecom}],
we obtain
\begin{align}\label{thm-comp-e1}
1=\lim_{k\to\infty}\|u_k\|_{p^*}^{p^*}=\lim_{k\to\infty}\|w_k\|_{p^*}^{p^*}
=\sum_{j=1}^{\infty}\|V_j\|_{p^*}^{p^*}.
\end{align}

Now, we turn to deal with the directional derivative terms.
For any $J,k\in\mathbb{N}$, define 
$G_k^J:=\sum_{j=1}^{J}\mathfrak{g}_{j,k}(V_j;p^*)$.
Then, from Lemma \ref{lem-pdecom}\eqref{it1-pdecom} and
the convexity of the function $t\mapsto t^p$, it follows that,
for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{thm-comp-e2}
\|\nabla_\xi w_k\|_p^p
\ge\|\nabla_\xi G_k^J\|_p^p+
p\int_{\mathbb{R}^n}|\nabla_\xi G_k^J|^{p-2}\nabla_\xi G_k^J
\nabla_\xi r_k^J.
\end{align}
Observe that
\begin{align}\label{thm-comp-e4}
\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)
&=\lambda_{j,k}^{-1-\frac{n}{p^*}}
\nabla_\xi V_j\left(\frac{\cdot-x_{j,k}}{\lambda_{j,k}}\right)\notag\\
&=\lambda_{j,k}^{-\frac np}
\nabla_\xi V_j\left(\frac{\cdot-x_{j,k}}{\lambda_{j,k}}\right)
\quad\text{by $n\left(\frac{1}{p}-\frac{1}{p^*}\right)=1$}\notag\\
&=\mathfrak{g}_{j,k}(\nabla_\xi V_j;p).
\end{align}
Thus, by Lemma \ref{lem-pdecom}\eqref{it2-pdecom} and 
\eqref{lem-decou-ee}, we find that, when $k\to\infty$,
$$
\left\|\sum_{j=1}^{J}|\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)|^{p-2}
\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)-|\nabla_\xi G_k^J|^{p-2}\nabla_\xi G_k^J\right\|_{p'}
\to0.
$$
This, combined with Lemma \ref{lem-pdecom}\eqref{it3-pdecom}, further implies that
$\|\nabla r^J_k\|_{p}$ is uniformly bounded in terms of $k$ and hence,
when $k\to\infty$,
\begin{align}\label{thm-comp-e3}
\int_{\mathbb{R}^n}|\nabla_\xi G_k^J|^{p-2}\nabla_\xi G_k^J
\nabla_\xi r_k^J
=\sum_{j=1}^{J}\int_{\mathbb{R}^n}|\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)|^{p-2}
\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)\nabla_\xi r_k^J+o(1).
\end{align}
From \eqref{thm-comp-e4} and a change of variables,
it follows that 
\begin{align}\label{thm-comp-e5}
\int_{\mathbb{R}^n}|\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)|^{p-2}
\nabla_\xi\mathfrak{g}_{j,k}(V_j;p^*)\nabla_\xi r_k^J
&=\int_{\mathbb{R}^n}|\mathfrak{g}_{j,k}(\nabla_\xi V_j;p)|^{p-2}
\mathfrak{g}_{j,k}(\nabla_\xi V_j;p)\nabla_\xi r_k^J\notag\\
&=\int_{\mathbb{R}^n}|\nabla_\xi V_j|^{p-2}
\nabla_\xi V_j\mathfrak{g}_{j,k}^{-1}(\nabla_\xi r_k^J;p)\notag\\
&=\int_{\mathbb{R}^n}|\nabla_\xi V_j|^{p-2}
\nabla_\xi V_j\nabla_\xi\mathfrak{g}_{j,k}^{-1}(r_k^J;p^*).
\end{align}
Note that 
$\Phi_{j,\xi}:u\mapsto
\int_{\mathbb{R}^n}|\nabla_\xi V_j|^{p-2}\nabla_\xi V_j\nabla_\xi u$
is a bounded linear functional on $\dot{W}^{1,p}$;
indeed, using H\"older's inequality,
we find that, for any 
$u\in\dot{W}^{1,p}$,
\begin{align}\label{thm-comp-e8}
|\Phi_{j,\xi}(u)|
\le \int_{\mathbb{R}^n}|\nabla V_j|^{p-1}|\nabla u|
\le\left(\int_{\mathbb{R}^n}|\nabla V_j|^p\right)^{1-\frac1p}
\left(\int_{\mathbb{R}^n}|\nabla u|^p\right)^{\frac1p}.
\end{align}
Consequently, Lemma \ref{lem-pdecom}\eqref{it3-pdecom},
combined with \eqref{thm-comp-e3} and \eqref{thm-comp-e5},
further implies that, for all $\xi\in\mathbb{S}^{n-1}$ and $J\in\mathbb{N}$,
\begin{align}\label{thm-comp-e6}
\lim_{k\to\infty}\int_{\mathbb{R}^n}|\nabla_\xi G_k^J|^{p-2}\nabla_\xi G_k^J
\nabla_\xi r_k^J=0.
\end{align}
On the other hand, by \eqref{lem-decou-e0} and \eqref{thm-comp-e4}, we conclude that,
when $k\to\infty$,
\begin{align*}
\|\nabla_\xi G_k^J\|_{p}^p
=\left\|\sum_{j=1}^{J}\mathfrak{g}_{j,k}(\nabla_\xi V_j;p)\right\|_p^p
=\sum_{j=1}^{J}\|\nabla_\xi V_j\|_p^p+o(1).
\end{align*}
From this, \eqref{thm-comp-e2}, and \eqref{thm-comp-e6}, we further infer that,
for all $J\in\mathbb{N}$,
when $k\to\infty$,
\begin{align}\label{thm-comp-e7}
\|\nabla_\xi w_k\|_p^p
\ge\sum_{j=1}^{J}\|\nabla_\xi V_j\|_p^p+o(1).
\end{align}
More precisely, the remainder term in \eqref{thm-comp-e7} is 
\begin{align*}
\rho_k^J(\xi):=\|\nabla_\xi G_k^J\|_{p}^p
-\sum_{j=1}^{J}\|\nabla_\xi V_j\|_p^p+p
\int_{\mathbb{R}^n}|\nabla_\xi G_k^J|^{p-2}\nabla_\xi G_k^J
\nabla_\xi r_k^J.
\end{align*}
It is easy to see that the sequence $\{\rho_k^J\}_{k\in\mathbb{N}}$
is in $C(\mathbb{S}^{n-1})$.
We have proved that $\rho_k^J\to0$ pointwise as $k\to\infty$.
But if we aim to further apply the 
nonlinear Brezis--Lieb lemma given in Lemma \ref{lem-BL}\eqref{it2-BL}, we still need the 
uniformity of this convergence.

Observe that $\mathfrak{g}_{j,k}(\cdot;p^*)$ keeps
$\|\cdot\|_{\dot{W}^{1,p}}$, we find that 
$\{G_{k}^J\}_{k\in\mathbb{N}}$ is a bounded sequence 
in $\dot{W}^{1,p}$.
Hence, applying H\"older's inequality similar to 
\eqref{thm-comp-e8} and the uniform boundedness 
of $\|\nabla r^J_k\|_{p}$ in terms of $k$,
we can easily conclude that $\{\rho_k^J\}_{k\in\mathbb{N}}$
is equicontinuous and is bounded in $C(\mathbb{S}^{n-1})$.
Then, from the Arzel\`a--Ascoli theorem,
it follows that, for any subsequence $\{\rho_{k_j}\}_{j\in\mathbb{N}}$,
there further is a subsequence $\{\rho^J_{k_{j_\ell}}\}_{\ell\in\mathbb{N}}$
and a function $\rho^J\in C(\mathbb{S}^{n-1})$ 
such that $\rho^J_{k_{j_\ell}}\to\rho^J$ in $C(\mathbb{S}^{n-1})$ as $\ell\to\infty$.
Noting that $\rho_k^J\to0$ pointwise as $k\to\infty$, we have 
$\rho^J\equiv0$.
Then, by the arbitrariness of the above subsequence, 
we know that the whole sequence $\{\rho_k^J\}_{k\in\mathbb{N}}$
converges to $0$ in $C(\mathbb{S}^{n-1})$
(that is, converges to $0$ uniformly).
On the other hand, from \eqref{thm-comp-e1}, 
it follows that there is a 
$J_0\in\mathbb{N}$ such that $V_{J_0}\ne 0$. 
Then $\xi\mapsto\|\nabla_\xi V_{J_0}\|_p$
is a norm on $\mathbb{R}^n$ and hence 
there is a $C_0\in(0,\infty)$ such that 
$\|\nabla_\xi V_{J_0}\|_p\ge C_0$ for all 
$\xi\in\mathbb{S}^{n-1}$
(see, for instance, \cite[(3.1)]{n16}).
Consequently, 
by the fact that $\{\rho_k^J\}_{k\in\mathbb{N}}$
converges to $0$ uniformly, we conclude that,
for any $\xi\in\mathbb{S}^{n-1}$,
any $J>J_0$, and any sufficiently large $k$,
\begin{align*}
\sum_{j=1}^{J}\|\nabla_\xi V_j\|_p^p+\rho_k^J(\xi)\ge
\frac{C_0^p}{2}.
\end{align*}
This, together with Lemma \ref{lem-BL}\eqref{it2-BL} and \eqref{thm-comp-e7},
further implies that, for all $J>J_0$, when $k\to\infty$,
\begin{align*}
\mathcal{E}_p(w_k)^p
&\ge\left(\int_{\mathbb{S}^{n-1}}\left[\sum_{j=1}^{J}
\|\nabla_\xi V_j\|_p^p+\rho_k^J(\xi)\right]^{-\frac np}\,d\xi\right)^{-\frac pn}\\
&=\left(\int_{\mathbb{S}^{n-1}}\left[\sum_{j=1}^{J}
\|\nabla_\xi V_j\|_p^p\right]^{-\frac np}\,d\xi\right)^{-\frac pn}
+o(1)\\
&\ge\sum_{j=1}^{J}\mathcal{E}_p(V_j)^p+o(1)
\quad\text{by the reverse Minkowski inequality}.
\end{align*}
Applying this, the assumption on $w_k$, and the arbitrariness of $J$, we then obtain 
\begin{align*}
S_{\rm aff}^p
&=\lim_{k\to\infty}\mathcal{E}_p(w_k)^p
\ge\sum_{j=1}^{\infty}\mathcal{E}_p(V_j)^p\\
&\ge S_{\rm aff}^p\sum_{j=1}^{\infty}\|V_j\|_{p^{*}}^{p}
\quad\text{by \eqref{eq-affSob}}\\
&\ge S_{\rm aff}^p
\left(\sum_{j=1}^{\infty}\|V_j\|_{p^*}^{p^*}\right)^{\frac{p}{p^*}}
\quad\text{by the concavity of $t\mapsto t^{\frac{p}{p^*}}$}\\
&=S_{\rm aff}^p\quad\text{by \eqref{thm-comp-e1}}.
\end{align*}
This shows all the above inequalities should be equalities.
Therefore, due to the strict concavity of $t\mapsto t^{\frac{p}{p^*}}$, 
there is a unique non-zero function in $\{V_j\}_{j\in\mathbb{N}}$, 
which we denote by $V$.
Then $\mathcal{E}_p(V)=S_{\rm aff}$ and $\|V\|_{p^*}=1$.
This implies $V\in\mathcal{M}_{\rm aff}$.
Moreover, by Lemma \ref{lem-pdecom}\eqref{it1-pdecom},
there are a sequence $\{\mathfrak{g}_k\}_{k\in\mathbb{N}}
=\{(\lambda_k,x_k)\}_{k\in\mathbb{N}}$
in $(0,\infty)\times\mathbb{R}^n$ and a sequence 
$\{r_k\}_{k\in\mathbb{N}}$ in $L^{p^*}$ such that,
for every $k\in\mathbb{N}$,
\begin{align*}
w_k=\mathfrak{g}_k(V;p^*)+r_k\quad
\text{and}\quad\mathfrak{g}_k^{-1}(r_k;p^*)\rightharpoonup0\ 
\text{in $\dot{W}^{1,p}$ as $k\to\infty$}.
\end{align*}
Hence, 
\begin{align*}
\mathfrak{g}_k^{-1}(w_k;p^*)=V+\mathfrak{g}_k^{-1}(r_k;p^*).
\end{align*}

Since $\mathfrak{g}_k^{-1}(w_k;p^*)$ takes the form 
$\mathcal{T}^{(p^*)}_{\lambda_k,A_k,x_k}u_k$, we infer that,
if we can show $$\mathfrak{g}_k^{-1}(r_k;p^*)\to0\ 
\text{in $\dot{W}^{1,p}$ as $k\to\infty$,}$$
then we complete the present proof.
Furthermore, by \eqref{def-min} and the fact that 
the mapping $\mathcal{T}^{(p^*)}_{\lambda_k,A_k,x_k}$ keeps both $\mathcal{E}_p(\cdot)$
and $\|\cdot\|_{p^*}$, and the weak convergence, 
we can assume that $V=U$. That is, 
we only need to prove that, if a sequence $\{\rho\}_{k\in\mathbb{N}}$
in $\dot{W}^{1,p}$ satisfying that,
when $k\to\infty$,
$\mathcal{E}_{p}(U+\rho_k)\to S_{\rm aff}$ and
$\rho_k\rightharpoonup0$
in $\dot{W}^{1,p}$, then
$\rho_k\to0$ in $\dot{W}^{1,p}$.
We now show this assertion. Indeed, 
applying $p\ge 2$ and Taylor's formula 
(see also \cite[Lemma 2.1(ii)]{fz22}), 
there is a positive constant $c_p$, depending only on 
$p$, such that, for any $x,y\in\mathbb{R}$,
$|x+y|^p\ge |x|^p+p|x|^{p-2}xy+c_p|y|^p$.
Thus, for all $\xi\in\mathbb{S}^{n-1}$ and $k\in\mathbb{N}$,
\begin{align*}
\|\nabla_\xi (U+\rho_k)\|_p^p
\ge \|\nabla_\xi U\|_p^p
+p\int_{\mathbb{R}^n}|\nabla_\xi U|^{p-2}
\nabla_\xi U\nabla_\xi \rho_k
+c_p\|\nabla_\xi \rho_k\|_p^p.
\end{align*}
Similarly to the analysis of $\{\rho_k^J\}_{k\in\mathbb{N}}$ 
above (and even more simply), we deduce that, when $k\to\infty$,
\begin{align*}
\mathcal{E}_p(U+\rho_k)^p
&\ge\left[\int_{\mathbb{S}^{n-1}}
\left( \|\nabla_\xi U\|_p^p
+c_p\|\nabla_\xi \rho_k\|_p^p\right)^{-\frac np}\,d\xi\right]^{-\frac pn}
+o(1)\\
&\ge S_{\rm aff}^p+o(1)\quad\text{by $\mathcal{E}_p(U)=S_{\rm aff}$}.
\end{align*}
This, combined with the assumption 
$\mathcal{E}_p(U+\rho_k)\to S_{\rm aff}$,
further implies that 
\begin{align}\label{thm-comp-e10}
\lim_{k\to\infty}\left[\int_{\mathbb{S}^{n-1}}
\left( \|\nabla_\xi U\|_p^p
+c_p\|\nabla_\xi \rho_k\|_p^p\right)^{-\frac np}\,d\xi\right]^{-\frac pn}
=S_{\rm aff}^p.
\end{align}
Noting that $U$ is radial, we conclude that $\|\nabla_\xi U\|_p^p=a$
with some absolute constant $a$ independent of $\xi$. 
Then, using the boundedness of $\{\rho_k\}_{k\in\mathbb{N}}$
in $\dot{W}^{1,p}$ and Taylor's formula, we find that,
for any $k\in\mathbb{N}$ and $\xi\in\mathbb{S}^{n-1}$, there is a $t\in(0,1)$
such that 
\begin{align}\label{thm-comp-e9}
a^{-\frac np}-
\left(a+c_p\|\nabla_\xi \rho_k\|_p^p\right)^{-\frac np}
=\frac npc_p\left(a+c_pt\|\nabla_\xi \rho_k\|_p^p\right)^{-1-\frac np}
\|\nabla_\xi\rho_k\|_p^p
\gtrsim\|\nabla_\xi\rho_k\|_p^p,
\end{align}
where the implicit positive constant depends only on $n$ and $p$. 
In addition, observe that  
$S_{\rm aff}=\mathcal{E}_p(U)=a^{\frac1p}|\mathbb{S}^{n-1}|^{-\frac1n}$.
From this and \eqref{thm-comp-e9}, it further follows that, 
when $k\to\infty$,
\begin{align*}
\|\nabla\rho_k\|_{p}^p
&\sim\int_{\mathbb{S}^{n-1}}\|\nabla_\xi\rho_k\|_{p}^p\,d\xi
\lesssim\int_{\mathbb{S}^{n-1}}
a^{-\frac np}-\left(a+c_p\|\nabla_\xi \rho_k\|_p^p\right)^{-\frac np}\,d\xi\\
&=S_{\rm aff}^{-n}-\int_{\mathbb{S}^{n-1}}
\left( \|\nabla_\xi U\|_p^p
+c_p\|\nabla_\xi \rho_k\|_p^p\right)^{-\frac np}\,d\xi\to0
\quad\text{by \eqref{thm-comp-e10}.}
\end{align*}
This shows $\lim_{k\to\infty}\rho_k=0$ in $\dot{W}^{1,p}$
and hence finishes the present proof.
\end{proof}

Moreover, this 
proof gives the following consequence, which may be of
independent interest.

\begin{corollary}
If $p\in[2,n)$, $U$ satisfies \eqref{eq:standardminimizer} and 
\eqref{eq:standardminimizernorm},
a sequence $\{\rho\}_{k\in\mathbb{N}}$
in $\dot{W}^{1,p}$ satisfies that,
when $k\to\infty$,
$$\mathcal{E}_{p}(U+\rho_k)\to S_{\rm aff}\quad\text{and}
\quad\rho_k\rightharpoonup0\
\text{in $\dot{W}^{1,p}$},
$$
then
$\rho_k\to0$ in $\dot{W}^{1,p}$. 
\end{corollary}

If, in this statement, the convergence of the affine energy is replaced by
$$
\|\nabla(U+\rho_k)\|_p\to S_{\rm Sob},
$$
then the conclusion follows immediately from the Radon--Riesz property
(or equivalently, the uniform convexity) of $L^p$. The corollary
above can therefore be viewed as an affine analogue of this fact.

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\section{Expansions for integral functionals}\label{sec-expan}

The goal of this section is to deal with the expansion 
of negative-exponent Lebesgue spaces and then 
obtain appropriate expansions for the affine Sobolev energy $\mathcal{E}_p(f)$
and the affine Euler--Lagrange residual $\delta_{\rm aff}$. 
This technique was perfected in \cite{fz22}, but some changes are needed because of the negative exponent that appears in the affine Sobolev energy.


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Expansion for Lebesgue spaces with 
negative exponents}

In this subsection, we assume $(\Omega,\mu)$
is a probability space and prove the following expansion.

\begin{proposition}\label{prop-neg-Lebe}
Let $\alpha\in(0,\infty)$.
Then there is a positive constant $C_{\alpha}$,
depending only on $\alpha$, such that
\begin{enumerate}[{\rm (i)}]
  \item\label{it0-neg-Lebe} for any $\delta\in(0,\frac12)$, $a\in(0,\infty)$,
and $U\in L^\infty(\Omega)$ such that 
$\|U\|_{L^\infty(\Omega)}\le a\delta$,
\begin{align}\label{prop-neg-Lebe-e1}
\left[\int_\Omega(a+U)^{-\alpha}\right]^{-\frac1\alpha}
\ge a+\int_{\Omega}U-
\frac{1+\alpha}{2}a^{-1}
\left[\int_{\Omega}U^2-\left(\int_{\Omega}U\right)^2\right]
-C_\alpha a^{-1}\delta\int_{\Omega}U^2;
\end{align}
  \item\label{it1-neg-Lebe} for any $\delta\in(0,\frac12)$, $a\in(0,\infty)$,
$U\in L^\infty(\Omega)$ satisfying
$\|U\|_{L^\infty(\Omega)}\le a\delta$,
and $V\in L^1(\Omega)$, 
\begin{align}\label{prop-neg-Lebe-polarized-e1}
&\left[\int_\Omega(a+U)^{-\alpha}\right]^{-1-\frac1\alpha}
\int_\Omega(a+U)^{-\alpha-1}V
\nonumber\\
&\qquad \ge
\int_\Omega V
-(1+\alpha)a^{-1}
\left[\int_{\Omega}
\left(U-\int_{\Omega}U\right)
\left(V-\int_{\Omega}V\right)
\right]
\nonumber\\
&\qquad\quad
-
C_\alpha a^{-1}\delta 
\left[
\int_\Omega |UV|
+
\left(\int_\Omega |U|\right)
\left(\int_\Omega |V|\right)
\right];
\end{align}
  \item\label{it2-neg-Lebe} for any $a\in(0,\infty)$
  and $U\in L^\infty(\Omega)$ satisfying
$\|U\|_{L^\infty(\Omega)}\le \frac a2$,
  \begin{align*}
  \left|\left[\int_\Omega(a+U)^{-\alpha}\right]^{-\frac1\alpha}
-a-\int_{\Omega}U\right|\le C_\alpha a^{-1}\|U\|_{L^\infty(\Omega)}^2.
  \end{align*}
\end{enumerate}
\end{proposition}

\begin{proof}
We first prove \eqref{it0-neg-Lebe}.
For any $t\in(0,\infty)$, define 
$\Psi(t):=\int_{\Omega}(a+tU)^{-\alpha}$
and $\Phi:=\Psi^{-\frac{1}{\alpha}}$.
Note that $$\left[\int_{\Omega}(a+U)^{-\alpha}\right]^{-\frac{1}{\alpha}}=\Phi(1).$$
From this and Taylor's formula, we infer that there is a 
$\xi\in[0,1]$ with
\begin{align}\label{prop-neg-Lebe-e2}
\left[\int_{\Omega}(a+tU)^{-\alpha}\right]^{-\frac{1}{\alpha}}
=\Phi(1)=\Phi(0)
+\Phi'(0)+\frac12\Phi''(0)
+\frac16\Phi'''(\xi).
\end{align}
Now, we calculate derivatives of both $\Phi$ and $\Psi$.
Indeed, for any $k\in\mathbb{N}$ and $t\in(0,\infty)$,
\begin{align*}
\Psi^{(k)}(t)=
(-1)^k\prod_{j=0}^{k-1}
(\alpha+j-1)\int_{\Omega}(a+tU)^{-\alpha-k}U^k,
\end{align*}
\begin{align*}
\Phi'=-\frac1\alpha\Psi^{-1-\frac1\alpha}\Psi',\quad 
\Phi''=\frac{1}{\alpha}\left(1+\frac1\alpha\right)
\Psi^{-2-\frac1\alpha}\Psi'^2
-\frac{1}{\alpha}\Psi^{-1-\frac1\alpha}\Psi'',
\end{align*}
and 
\begin{align}\label{prop-neg-Lebe-e4}
\Phi'''=-\frac{1}{\alpha}\left(1+\frac1\alpha\right)
\left(2+\frac1\alpha\right)
\Psi^{-3-\frac1\alpha}\Psi'^3+\frac{2}{\alpha}
\left(1+\frac1\alpha\right)\Psi^{-2-\frac1\alpha}
\Psi'\Psi''
-\frac1\alpha\Psi^{-1-\frac1\alpha}\Psi'''.
\end{align}
Hence, we find that 
the first three terms on the right-hand side
of \eqref{prop-neg-Lebe-e1} are exactly 
$\Phi(0)$, $\Phi'(0)$, and $\frac12\Phi''(0)$, respectively.
By this and \eqref{prop-neg-Lebe-e2}, to prove \eqref{prop-neg-Lebe-e1},
it suffices to show 
\begin{align}\label{prop-neg-Lebe-e3}
\left\|\Phi'''\right\|_{L^\infty[0,1]}\lesssim a^{-1}\delta\int_{\Omega}U^2.
\end{align}
For any $t\in[0,1]$, since $\delta\in(0,\frac12)$, from 
$(1-\delta)a\le|a+tU|\le (1+\delta)a$ ,
it follows that $|a+tU|\in(\frac12a,\frac32a)$.
Therefore, for any $t\in[0,1]$, $|\Psi(t)|\sim a^{-\alpha}$
and, for any $k\in\mathbb{N}$,
\begin{align}\label{prop-neg-Lebe-e5}
\left|\Psi^{(k)}(t)\right|
&\lesssim a^{-\alpha-k}\int_{\Omega}|U|^k
\le
\begin{cases}
\displaystyle{a^{-\alpha-1}}\int_{\Omega}|U|&\text{if $k=1$},\\
\displaystyle{a^{-\alpha-2}\delta^{k-2}}\int_{\Omega}U^2
&\text{if $k\ge2$}.
\end{cases}
\end{align}
Using these, we obtain, for any $t\in[0,1]$,
\begin{align*}
\left|\Psi(t)^{-3-\frac1\alpha}\Psi'(t)^3
\right|
&\lesssim a^{1+3\alpha}\left(a^{-\alpha-1}\int_{\Omega}|U|\right)^3
=a^{-2}\left(\int_{\Omega}|U|\right)^3\\
&\le a^{-2}\int_{\Omega}|U|\int_{\Omega}|U|^2
\quad\text{by H\"older's inequality}\\
&\le a^{-1}\delta\int_{\Omega}|U|^2\quad\text{by the assumption 
$\|U\|_{L^\infty}\le a\delta$}.
\end{align*}
Similarly (and even more simply), we deduce that
the remaining two terms 
on the right-hand side of \eqref{prop-neg-Lebe-e4} can also 
be controlled by $a^{-1}\delta\int_{\Omega}|U|^2$.
Consequently, \eqref{prop-neg-Lebe-e3} holds.
We then complete the proof of \eqref{prop-neg-Lebe-e1}.

We first prove \eqref{it1-neg-Lebe}.
For any $t\in(0,\infty)$, define 
$\Theta(t):=\int_\Omega(a+tU)^{-\alpha-1}V$
and $\Upsilon:=\Psi^{-1-\frac1\alpha}\Theta$.
Then
\begin{align*}
\left[\int_\Omega(a+U)^{-\alpha}\right]^{-1-\frac1\alpha}
\int_\Omega(a+U)^{-\alpha-1}V
=
\Upsilon(1).
\end{align*}
By Taylor's formula, there is a $\xi_1\in[0,1]$ with
\begin{align*}
\Upsilon(1)
=
\Upsilon(0)+\Upsilon'(0)+\frac12\Upsilon''(\xi_1).
\end{align*}
Now, we calculate derivatives of the above functions.
Indeed, for any $k\in\mathbb{N}$ and $t\in(0,\infty)$,
it holds that
\begin{align}\label{prop-neg-Lebe-polarized-e12}
\Theta^{(k)}(t)
=(-1)^{k}\prod_{j=1}^{k}(\alpha+j)
\int_{\Omega}(a+tU)^{-\alpha-k-1}U^kV,
\end{align}
\begin{align*}
\Upsilon'
=
-\left(1+\frac1\alpha\right)
\Psi^{-2-\frac1\alpha}\Psi'(t)\Theta
+
\Psi^{-1-\frac1\alpha}\Theta',
\end{align*}
and 
\begin{align*}
\Upsilon''
&=
\left(1+\frac1\alpha\right)
\left(2+\frac1\alpha\right)
\Psi^{-3-\frac1\alpha}\Psi'^2\Theta-
\left(1+\frac1\alpha\right)
\Psi^{-2-\frac1\alpha}\Psi''\Theta
\nonumber\\
&\quad
-
2\left(1+\frac1\alpha\right)
\Psi(t)^{-2-\frac1\alpha}\Psi'\Theta'
+
\Psi(t)^{-1-\frac1\alpha}\Theta''.
\end{align*}
Therefore,
the first two terms on the right-hand side
of \eqref{prop-neg-Lebe-polarized-e1} are exactly 
$\Upsilon(0)$ and $\Upsilon'(0)$, respectively.
Then it remains to show
\begin{align}\label{prop-neg-Lebe-polarized-e6}
\|\Upsilon''\|_{L^\infty[0,1]}
\lesssim a^{-1}\delta\left[
\int_\Omega |UV|
+\left(\int_\Omega |U|\right)
\left(\int_\Omega |V|\right)
\right].
\end{align}
Indeed, applying \eqref{prop-neg-Lebe-e5}, we find that 
\begin{align}\label{prop-expan-lap-e11}
&\left|\Psi^{-3-\frac1\alpha}\Psi'^2\Theta
\right|+\left|\Psi^{-2-\frac1\alpha}\Psi''\Theta\right|\notag\\
&\quad\lesssim\left(a^{-\alpha}\right)^{-3-\frac1\alpha}
\left(a^{-\alpha-1}\int_{\Omega}|U|\right)^2
\int_{\Omega}a^{-\alpha-1}|V|
+\left(a^{-\alpha}\right)^{-2-\frac1\alpha}
a^{-\alpha-1}\delta\int_{\Omega}|U|
\int_{\Omega}a^{-\alpha-1}|V|\notag\\
&\quad\lesssim a^{-1}\delta\int_{\Omega}|U|
\int_{\Omega}|V|.
\end{align}
On the other hand, from \eqref{prop-neg-Lebe-polarized-e12},
we infer that, for every $k\in\mathbb{N}$ and $t\in[0,1]$,
\begin{align*}
|\Theta^{(k)}(t)|
&\lesssim a^{-\alpha-k-1}\int_{\Omega}|U^kV|
\quad\text{by $a+tU\sim a$}\\
&\le \begin{cases}
\displaystyle{a^{-\alpha-1}}\delta\int_{\Omega}|V|&\text{if $k=1$},\\
\displaystyle{a^{-\alpha-2}\delta^{k-1}}\int_{\Omega}|UV|
&\text{if $k\ge2$}
\end{cases}
\quad\text{by $\|U\|_{L^\infty(\Omega)}\le a\delta$}.
\end{align*}
Applying this, \eqref{prop-neg-Lebe-e5},
and a similar argument used in the estimation of \eqref{prop-expan-lap-e11}, 
we obtain 
\begin{align*}
\left|\Psi^{-2-\frac1\alpha}\Psi'\Theta'\right|
+\left|\Psi^{-1-\frac1\alpha}\Theta''\right|
\lesssim\operatorname{RHS}\eqref{prop-neg-Lebe-polarized-e6}.
\end{align*}
This then finishes the proof of \eqref{prop-neg-Lebe-polarized-e6}
and hence \eqref{it1-neg-Lebe}.

Next, we show \eqref{it2-neg-Lebe}. The proof proceeds along very similar lines 
to that of \eqref{it0-neg-Lebe}.
Indeed, by an argument similar to that used 
in \eqref{prop-neg-Lebe-e3}, we conclude that
\begin{align*}
\left\|\Phi''\right\|_{L^\infty(\Omega)}
\lesssim \left(a^{-\alpha}\right)^{-2-\frac1\alpha}
\left(a^{-\alpha-1}\int_{\Omega}|U|\right)^2
+\left(a^{-\alpha}\right)^{-1-\frac1\alpha}
a^{-\alpha-2}\delta\int_{\Omega}|U|^2
\le a^{-1}\|U\|_{L^\infty(\Omega)}^2.
\end{align*}
This, together with 
Taylor's formula for $\Phi$ with second order remainder,
gives the desired estimate in \eqref{it2-neg-Lebe}.
Then the present proof is complete.
\end{proof}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Local expansion for affine Sobolev energy}\label{subsec-expan}

In this subsection, we establish the expansion of 
the affine Sobolev energy 
near a given function $f\in\dot{W}^{1,p}$,
which is important in the proof of Theorem \ref{thm-stab}.

It will suffice to consider the case where $f$ is \emph{radial}. Then $\|\nabla_\xi f\|_p$ is independent of $\xi$ and hence 
$\mathcal{E}_p(f)=|\mathbb{S}^{n-1}|^{-\frac1n}
\|\nabla_\xi f\|_{p}$.

For $g\in\dot{W}^{1,p}$ and $\xi\in\mathbb{S}^{n-1}$, define 
\begin{align}\label{def-L1}
L^{(1)}_\xi(f,g):=\int_{\mathbb{R}^n}|\nabla_\xi f|^{p-2}
\nabla_\xi f\nabla_\xi g,
\end{align}
\begin{align}\label{def-var}
\operatorname{Var}(f,g):=\int_{\mathbb{S}^{n-1}}\left[
L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}
L^{(1)}_u(f,g)\,du\right]^2 d\xi,
\end{align}
and 
\begin{align}\label{def-L2}
L^{(2)}_\xi(f,g):=&\int_{\mathbb{R}^n}
|\nabla_\xi f|^{p-2}(\nabla_\xi g)^2\notag\\
&\qquad+(p-2)\frac{\min\{|\nabla_\xi (f+g)|,|\nabla_\xi f|\}^{p-1}}{|\nabla_\xi f|}
\left(|\nabla_\xi(f+g)|-|\nabla_\xi f|\right)^2,
\end{align}
where, if $|\nabla_\xi f|=0$, then define
$\frac{\min\{|\nabla_\xi (f+g)|,|\nabla_\xi f|\}^{p-1}}{|\nabla_\xi f|}:=0$.
It is clear that, for every $i\in\{1,2\}$ and $\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{eq-L-dila}
 L^{(i)}_\xi(cf,g)=c^p L^{(1)}_\xi(f,c^{-1}g).
\end{align}
Moreover,
by the triangle inequality, for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\frac{\min\{|\nabla_\xi (f+g)|,|\nabla_\xi f|\}^{p-1}}{|\nabla_\xi f|}
\left(|\nabla_\xi(f+g)|-|\nabla_\xi f|\right)^2
\le|\nabla_\xi f|^{p-2}|\nabla_\xi g|^2;
\end{align*}
therefore, if $p\ge 2$, then using H\"older's inequality,
we obtain 
\begin{align}\label{eq-L2}
L^{(2)}_\xi(f,g)\le 
(p-1)\int_{\mathbb{R}^{n-1}}|\nabla_\xi f|^{p-2}
|\nabla_\xi g|^2
\le(p-1)\|\nabla_\xi f\|_p^{p-2}\|\nabla_\xi g\|_p^2.
\end{align}

Then we have the following local expansion of $\mathcal{E}_p(f+g)$.

\begin{proposition}\label{prop-expan-local}
Let $p\ge 2$. Then there are positive constants $c_{n,p}$,
$C_{n,p}$, and $\widetilde{C}_{n,p}$,
depending only on $n$ and $p$, such that, for any
$\varepsilon\in(0,c_{n,p})$ and $g\in\dot{W}^{1,p}$
satisfying
\begin{align}\label{prop-expan-local-e1}
\sup_{\xi\in\mathbb{S}^{n-1}}
\left[
\left(
\frac{L^{(2)}_\xi(f,g)}
{\|\nabla_\xi f\|_p^p}
\right)^{\frac12}
+
\frac{\|\nabla_\xi g\|_{p}^p}
{\|\nabla_\xi f\|_p^p}
\right]\le \varepsilon,
\end{align}
it holds that
\begin{align}\label{prop-expan-local-e2}
\mathcal{E}_p(f+g)^p
&\ge\mathcal{E}_p(f)^p+p|\mathbb{S}_{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)
+\frac{1-C_{n,p}\varepsilon}{2}L^{(2)}_\xi(f,g)\right]
\,d\xi\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal E_p(f)^{-p}
\operatorname{Var}(f,g)
+\widetilde C_{n,p}\varepsilon\|\nabla g\|_{p}^p.
\end{align}
\end{proposition}

\begin{remark}\label{rem-3.4}
Note that $f$ is radial and $\|\nabla_\xi g\|_p\le \|\nabla g\|_p|\xi|$ for all
$\xi\in\mathbb{S}^{n-1}$. 
By these and \eqref{eq-L2}, we conclude that, if $\varepsilon$ is sufficiently 
small, then, for all $g\in\dot{W}^{1,p}$ such that $\|\nabla g\|_p<\varepsilon\|\nabla f\|_p$,
it holds that 
$\sup_{\xi\in\mathbb{S}^{n-1}}\frac{\|\nabla_\xi g\|_p}{\|\nabla_\xi f\|_p}\lesssim\varepsilon$ and 
hence, for each $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\left(
\frac{L^{(2)}_\xi(f,g)}
{\|\nabla_\xi f\|_p^p}
\right)^{\frac12}
+
\frac{\|\nabla_\xi g\|_{p}^p}
{\|\nabla_\xi f\|_p^p}
&\lesssim\frac{\|\nabla_\xi g\|_p}{\|\nabla_\xi f\|_p}+
\left(\frac{\|\nabla_\xi g\|_p}{\|\nabla_\xi f\|_p}\right)^p\lesssim\varepsilon
\end{align*}
with the implicit positive constants depending only on $n$ and $p$.
This shows that, if $f+g$ is sufficiently close to $f$ in $\dot{W}^{1,p}$, 
then the expansion \eqref{prop-expan-local-e2} holds,
which is precisely what is required for the proof of Theorem \ref{thm-stab}.
Nevertheless, we choose to keep the present assumption \eqref{prop-expan-local-e1} 
to better highlight the parameter dependencies.
\end{remark}

\begin{proof}[Proof of Proposition \ref{prop-expan-local}]
We need to use several pointwise inequalities from Figalli--Zhang \cite{fz22}, which are exactly 
\cite[Lemma 2.1(ii) and (2.4)]{fz22}:
For any $\theta\in[0,\infty)$, there is a positive 
constant $\widetilde{c}_{p,\theta}$ (if $\theta=0$, then 
choose $\widetilde{c}_{p,\theta}=0$), depending only on $p$ and $\theta$, 
such that, for every $x,y\in\mathbb R$,
\begin{align}\label{eq:pointwise-FZ}
  |x+y|^p
  &\ge |x|^p+p|x|^{p-2}xy\notag\\
  &\quad+\frac{(1-\theta)p}{2}\left[|x|^{p-2}|y|^{2}
  +(p-2)\frac{\min\{|x+y|,|x|\}^{p-1}}{|x|}\left(|x+y|-|x|\right)^2\right]
  +\widetilde{c}_{p,\theta}|y|^p.
\end{align}
We first apply this inequality with $\theta=0$ and then obtain,
for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\|\nabla_\xi(f+g)\|_{p}^p
\ge\|\nabla_\xi f\|_p^p
+p\left[L^{(1)}_{\xi}(f,g)+\frac12L^{(2)}_\xi(f,g)\right]
=:a+U.
\end{align*}
Observe that, by H\"older's inequality and \eqref{prop-expan-local-e1},
we conclude that
\begin{align}\label{prop-expan-local-e3}
\left[L^{(1)}_\xi(f,g)\right]^2
\le\int_{\mathbb{R}^n}|\nabla_\xi f|^p\int_{\mathbb{R}^n}
|\nabla_\xi f|^{p-2}|\nabla_\xi g|^2\le aL^{(2)}_\xi(f,g)
\le a^2\varepsilon^2.
\end{align}
This implies 
$\|U\|_{L^\infty(\mathbb{S}^{n-1})}
\le ap\varepsilon(1+\frac12\varepsilon).$
Hence, by Proposition \ref{prop-neg-Lebe}, 
if $p\varepsilon(1+\frac12\varepsilon)<\frac12$, then there exists 
$C_1$, depending only on $n$ and $p$, such that 
\begin{align}\label{prop-expan-local-e4}
&\mathcal{E}_p(f+g)^p\notag\\
&\quad\ge\mathcal{E}_p(f)^p+p|\mathbb{S}_{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)
+\frac{1}{2}L^{(2)}_\xi(f,g)\right]\,d\xi\notag\\
&\qquad-\frac{p(n+p)}{2}|\mathbb{S}^{n-1}|^{-\frac{2p}{n}}
\mathcal{E}_p(f)^{-p}\notag\\
&\quad\qquad\times
\left[\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2
-\left(\fint_{\mathbb{S}^{n-1}}
L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2\right]\notag\\
&\qquad-C_1p\varepsilon\left(1+\frac12\varepsilon\right)
|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}\mathcal{E}_p(f)^{-p}
\int_{\mathbb{S}^n}\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2.
\end{align}
Using \eqref{prop-expan-local-e3} and 
the penultimate inequality in \eqref{prop-expan-local-e1}, we obtain 
\begin{align*}
\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2
\le \left(\sqrt{aL^{(2)}_\xi(f,g)}
+\frac12\varepsilon\sqrt{aL^{(2)}_\xi(f,g)}\right)^2
\le aL^{(2)}_{\xi}(f,g)\left(1+\frac12\varepsilon\right)^2.
\end{align*}
This further implies 
\begin{align*}
&p\varepsilon\left(1+\frac12\varepsilon\right)
|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}\mathcal{E}_p(f)^{-p}
\int_{\mathbb{S}^n}\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2
\lesssim p|\mathbb{S}^{n-1}|^{-1-\frac pn}
\varepsilon\int_{\mathbb{S}^{n-1}}L^{(2)}_\xi(f,g),
\end{align*}
where the implicit positive constant depends only on $n$ and $p$.

On the other hand, 
\begin{align*}
&\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2
-\left(\fint_{\mathbb{S}^{n-1}}
L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2\\
&\quad=\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)+\frac12 L^{(2)}_\xi(f,g)\right)^2\\
&\quad=\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)^2
+\frac14\fint_{\mathbb{S}^{n-1}}
\left(L^{(2)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(2)}_\xi(f,g)\right)^2\\
&\qquad-\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)
\left(L^{(2)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(2)}_\xi(f,g)\right).
\end{align*}
From \eqref{prop-expan-local-e1}, we deduce that
\begin{align*}
\fint_{\mathbb{S}^{n-1}}
\left(L^{(2)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(2)}_\xi(f,g)\right)^2
\le\fint_{\mathbb{S}^{n-1}}\left(L_\xi^{(2)}(f,g)\right)^2
\le |\mathbb{S}^{n-1}|^{\frac pn-1}\mathcal{E}_p(f)^p\varepsilon^2
\int_{\mathbb{S}^{n-1}}L^{(2)}_\xi(f,g).
\end{align*}
Moreover, by \eqref{prop-expan-local-e3}, we find that,
for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{prop-expan-local-e9}
\left|L^{(1)}_{\xi}(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right|
\le 2\|L^{(1)}_\xi(f,g)\|_{L^\infty(\mathbb{S}^{n-1})}
\le 2a\varepsilon.
\end{align}
Thus, it holds that 
\begin{align*}
&\left|\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)\right|\\
&\quad\le 4a\varepsilon\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)
=4|\mathbb{S}^{n-1}|^{\frac pn-1}\mathcal{E}_p(f)^p\varepsilon
\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g).
\end{align*}
Incorporating the above estimates into \eqref{prop-expan-local-e4},
it follows that, for sufficiently small $\varepsilon$,
\begin{align}\label{prop-expan-local-e5}
&\mathcal{E}_p(f+g)^p\notag\\
&\quad\ge\mathcal{E}_p(f)^p+p|\mathbb{S}_{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)
+\frac{1-C_2\varepsilon}{2}L^{(1)}_\xi(f,g)\right]\,d\xi\notag\\
&\qquad-\frac{p(n+p)}{2}|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(f)^{-p}
\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)^2,
\end{align}
where $C_2$ depends only on $n$ and $p$.

To finish the proof, we still need to add a $C\|\nabla g\|_{p}^p$
on the right-hand side of \eqref{prop-expan-local-e5}. To achieve this,
we use \eqref{eq:pointwise-FZ} with $\theta=1$ and then conclude that
there exists a positive constant $C_3$, depending only on $p$, such that,
for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\|\nabla_\xi(f+g)\|_{p}^p
\ge\|\nabla_\xi f\|_p^p+pL^{(1)}_{\xi}(g)+C_3\|\nabla_\xi g\|_p^p
=:a+\widetilde{U}.
\end{align*}
From \eqref{prop-expan-local-e3} again, we infer that 
$\|\widetilde U\|_{L^\infty(\mathbb{S}^{n-1})}\le a\varepsilon(p+C_3)$.
Consequently, if $\varepsilon(p+C_3)<\frac12$, then, applying 
Proposition \ref{prop-neg-Lebe} again, 
we find that there exists $C_4$, depending only on $n$ and $p$, such that 
\begin{align}\label{prop-expan-local-e6}
\mathcal{E}_p(f+g)^p
&\ge\mathcal{E}_p(f)^p
+|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_p^p\right]\notag\\
&\quad-\frac{1+\frac np}{2}|\mathbb{S}^{n-1}|^{-\frac{2p}n}
\mathcal{E}_p(f)^{-p}\notag\\
&\qquad\times\left[\fint_{\mathbb{S}^{n-1}}
\left(pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2
-\left(\fint_{\mathbb{S}^{n-1}}
pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2\right]\notag\\
&\quad-C_4\varepsilon(p+C_3)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(f)^{-p}
\int_{\mathbb{S}^{n-1}}\left(pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2\notag\\
&\quad=:\mathcal{E}_p(f)^p
+|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_p^p\right]
-I_1(g)-I_2(g).
\end{align}
By Tonelli's theorem, we conclude that 
\begin{align}\label{eq-lp}
\int_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p\,d\xi=
\alpha_{n,p}^p\|\nabla g\|_p^p,
\end{align}
where $\alpha_{n,p}$ is as in \eqref{eq-alpha}.
Therefore, we only need to show that, for sufficiently small $\varepsilon$,
\begin{align}\label{prop-expan-local-e7}
I_1(g)+I_2(g)&\le \frac{p(n+p)}{2}|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(f)^{-p}
\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)^2\notag\\
&\quad+C_5\varepsilon\int_{\mathbb{S}^{n-1}}
\left[L^{(1)}_\xi(f,g)+\|\nabla_\xi g\|_{p}^p\right],
\end{align}
where the positive constant $C_5$ depends only on $n$ and $p$.
Indeed, if \eqref{prop-expan-local-e7} holds, then 
\eqref{prop-expan-local-e6} implies that there exist
two positive constants $C_6$ and $C_7$, depending only on $n$ and $p$,
such that 
\begin{align}\label{prop-expan-local-e8}
&\mathcal{E}_p(f+g)^p\notag\\
&\quad\ge\mathcal{E}_p(f)^p+p|\mathbb{S}_{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)
-\frac{C_6\varepsilon}{2}L^{(1)}_\xi(f,g)\right]\,d\xi\notag\\
&\qquad-\frac{p(n+p)}{2}|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(f)^{-p}
\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)^2
+C_7\|\nabla g\|_p^p.
\end{align}
Letting $(1-\varepsilon)\times\eqref{prop-expan-local-e5}+\varepsilon\times\eqref{prop-expan-local-e8}$,
we obtain the desired estimate \eqref{prop-expan-local-e2}.

Hence, we now prove \eqref{prop-expan-local-e7}.
We begin with the simpler term $I_2(g)$.
Indeed, using \eqref{prop-expan-local-e3} and 
the penultimate inequality in \eqref{prop-expan-local-e1}, we obtain 
\begin{align*}
\int_{\mathbb{S}^{n-1}}\left(pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2
&\le 2\int_{\mathbb{S}^{n-1}}\left(p^2\left[L^{(1)}_\xi(f,g)\right]^2
+C_3^2\|\nabla_\xi g\|_{p}^{2p}\right)\\
&\le 2\int_{\mathbb{S}^{n-1}}\left(p^2aL^{(1)}_\xi(f,g)
+C_3^2\varepsilon^2a\|\nabla_\xi g\|_p^p\right)\\
&=2(p^2+C_3^2\varepsilon^2)|\mathbb{S}^{n-1}|^{\frac pn}
\mathcal{E}_p(f)^p
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)+\|\nabla_\xi g\|_p^p\right].
\end{align*}
This further implies that, for sufficiently small $\varepsilon$,
\begin{align*}
I_2(g)\lesssim\varepsilon 
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)+\|\nabla_\xi g\|_p^p\right].
\end{align*}
Next, we deal with $I_1(g)$.
Note that 
\begin{align*}
&\fint_{\mathbb{S}^{n-1}}
\left(pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2
-\left(\fint_{\mathbb{S}^{n-1}}
pL^{(1)}_\xi(f,g)+C_3\|\nabla_\xi g\|_{p}^p\right)^2\\
&\quad=p^2\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)^2
+C_3^2\fint_{\mathbb{S}^{n-1}}
\left(\|\nabla_\xi g\|_p^p-\fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p\right)^2\\
&\qquad+2pC_3\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)
\left(\|\nabla_\xi g\|_p^p-\fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p\right)
\end{align*}
By \eqref{prop-expan-local-e1} and \eqref{prop-expan-local-e9}, we conclude that 
\begin{align*}
\fint_{\mathbb{S}^{n-1}}
\left(\|\nabla_\xi g\|_p^p-\fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p\right)^2
\le \fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_{p}^{2p}\le 
|\mathbb{S}^{n-1}|^{\frac pn-1}\mathcal{E}_p(f)^p
\varepsilon^2\int_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_{p}^p
\end{align*}
and 
\begin{align*}
&\left|\fint_{\mathbb{S}^{n-1}}
\left(L^{(1)}_\xi(f,g)-\fint_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\right)
\left(\|\nabla_\xi g\|_p^p-\fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p\right)\right|\\
&\quad\le 4a\varepsilon\fint_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_{p}^p
=4|\mathbb{S}^{n-1}|^{-\frac pn-1}\mathcal{E}_p(f)^{p}\varepsilon
\int_{\mathbb{S}^{n-1}}\|\nabla_\xi g\|_p^p.
\end{align*}
Combining these estimates, we obtain \eqref{prop-expan-local-e7}
for sufficiently small $\varepsilon$.
Hence, we complete the proof of the present proposition.
\end{proof}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Local expansion for affine $p$-Laplacian}\label{subsec-expan2}

Next, we consider the expansion of the affine $p$ Laplacian, which is important in the proof of Theorem \ref{thm-bubble}.

Recall that, to study the stability of critical points of the classical Sobolev inequality, 
Liu and Zhang \cite[Lemma 2.1(2)]{lz25} derived the following 
expansion: If $p\ge 2$, then there exists 
$c_p\in(0,1)$, depending only on $p$, such that, 
for any $\theta\in (0,\infty)$, there exists a positive 
constant $\widetilde{c}_{p,\theta}$, depending only on $p$ and $\theta$, 
such that, for every $x,y\in\mathbb R$,
\begin{align}\label{eq:pointwise-LZ}
  |x+y|^{p-2}(x+y)y
  &\ge |x|^{p-2}xy\notag\\
  &\quad+(1-\theta)
  \frac{\min\{|x|,\max\{c_{p}^{\frac1{p-1}}|x|,|x+y|\}\}^{p-1}}{|x|}|y|^2\notag\\
  &\quad+(1-\theta)(p-2)
  \frac{\min\{|x+y|,|x|\}^{p-1}}{|x|}\left(|x+y|-|x|\right)^2
  +\widetilde{c}_{p,\theta}|y|^p.
\end{align}

As in the previous subsection we assume that $f\in \dot W^{1,p}$ is \emph{radial}. For any $g\in\dot{W}^{1,p}$ and $\xi\in\mathbb{S}^{n-1}$, define 
\begin{align}\label{eq-L2w}
\widetilde{L}^{(2)}_\xi(f,g)
:=&\int_{\mathbb{R}^n}
\frac{\min\{|\nabla_\xi f|,\max\{c_{p}^{\frac1{p-1}}|\nabla_\xi f|,|\nabla_\xi(f+g)|\}\}^{p-1}}
{|\nabla_\xi f|}|\nabla_\xi g|^2\notag\\
&\quad+(p-2)\frac{\min\{|\nabla_\xi (f+g)|,|\nabla_\xi f|\}^{p-1}}{|\nabla_\xi f|}
\left(|\nabla_\xi(f+g)|-|\nabla_\xi f|\right)^2,
\end{align}
where $c_{p}$ is as in \eqref{eq:pointwise-LZ}.
It is clear that $\widetilde{L}^{(2)}$ has the same 
dilation property as in \eqref{eq-L-dila} 
and $\widetilde{L}^{(2)}_\xi(f,g)\le L^{(2)}_\xi(f,g)$.

\begin{proposition}\label{prop-expan-lap}
Let $p\ge 2$. Then there exist positive constants $c_{n,p}$,
$C_{n,p}$, and $\widetilde C_{n,p}$,
depending only on $n$ and $p$, such that, for any
$\varepsilon\in(0,c_{n,p})$ and $g\in\dot W^{1,p}$ satisfying
\begin{align}\label{prop-expan-lap-e1}
\sup_{\xi\in\mathbb{S}^{n-1}}
\left[
\left(
\frac{\widetilde{L}^{(2)}_\xi(f,g)}
{\|\nabla_\xi f\|_p^p}
\right)^{\frac12}
+
\frac{\|\nabla_\xi g\|_{p}^p}
{\|\nabla_\xi f\|_p^p}
\right]\le \varepsilon,
\end{align}
it holds that
\begin{align}\label{prop-expan-lap-e2}
-\int_{\mathbb{R}^n}
\Delta_p^{\rm aff}(f+g)g
&\ge 
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g
+(1-C_{n,p}\varepsilon)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(f,g)\,d\xi\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal E_p(f)^{-p}
\operatorname{Var}(f,g)
+\widetilde C_{n,p}\varepsilon\|\nabla g\|_{p}^p.
\end{align}
\end{proposition}

\begin{remark}\label{rem-3.3}
Similar to Remark \ref{rem-3.4}, 
if $\varepsilon$ is sufficiently 
small, then, for all $g\in\dot{W}^{1,p}$ such that $\|\nabla g\|_p<\varepsilon\|\nabla f\|_p$,
it holds that 
$\operatorname{LHS}\eqref{prop-expan-lap-e1}\lesssim\varepsilon$
with the implicit positive constants depending only on $n$ and $p$.
This shows that, if $f+g$ is sufficiently close to $f$ in $\dot{W}^{1,p}$, 
then the expansion \eqref{prop-expan-lap-e2} holds.
\end{remark}


\begin{proof}[Proof of Proposition \ref{prop-expan-lap}]
We will choose $\theta=0$ and $\theta=1$ in \eqref{eq:pointwise-LZ} respectively,
derive two estimates, and then take a convex combination of them.

First, letting $\theta=0$ in \eqref{eq:pointwise-LZ}, we find that, 
for any $\xi\in\mathbb{S}^{n-1}$, 
\begin{align*}
\int_{\mathbb{R}^n}|\nabla_\xi (f+g)|^{p-2}
\nabla_\xi(f+g)\nabla_\xi g
&\ge L^{(1)}_\xi(f,g)+\widetilde{L}^{(2)}_\xi(f,g)=:V_0(\xi).
\end{align*}
By this and the definition of the affine Laplacian, we obtain 
\begin{align*}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
\ge \mathcal{E}_p(f+g)^{n+p}
\int_{\mathbb{S}^{n-1}}\|\nabla_\xi (f+g)\|_p^{-n-p}
V_0(\xi)\,d\xi.
\end{align*}
Define $a:=\|\nabla_\xi f\|_p^p$
and $U(\xi):=\|\nabla_\xi (f+g)\|_p^p-\|\nabla_\xi f\|_p^p$ for 
all $\xi\in\mathbb{S}^{n-1}$;
since $f$ is radial, $a$ is independent of the choice of $\xi$.
Then
\begin{align}\label{prop-expan-lap-e3}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
\ge \left[\int_{\mathbb{S}^{n-1}}(a+U)^{-\frac np}
\right]^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}(a+U)^{-1-\frac np}V_0.
\end{align}
To apply Proposition \ref{prop-neg-Lebe}\eqref{it1-neg-Lebe}, 
we now estimate $\|U\|_{L^\infty(\mathbb{S}^{n-1})}.$
From the definition of $\widetilde{L}^{(2)}$ and $p\ge 2$, 
we deduce that, for all $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\widetilde{L}^{(2)}_\xi(f,g)\ge c_{p}^{-1}\int_{\mathbb{R}^n}
|\nabla_\xi f|^{p-2}(\nabla_\xi g)^2,
\end{align*}
where $c_p$ is as in \eqref{eq:pointwise-LZ}.
This, together with H\"older's inequality and 
\eqref{prop-expan-lap-e1}, 
further implies 
\begin{align}\label{prop-expan-lap-e4}
\left[L^{(1)}_\xi (f,g)\right]^2
\le \int_{\mathbb{R}^n}|\nabla_\xi f|^p\int_{\mathbb{R}^n}
|\nabla_\xi f|^{p-2}|\nabla_\xi g|^2\le 
c_{p}^{-1}a \widetilde{L}^{(2)}_\xi(f,g)\le c_p^{-1} a^2\varepsilon^2.
\end{align}
Moreover, by Taylor's formula and $p\ge 2$ again, we obtain, for any $x,y\in\mathbb{R}$,
there exists $t\in[0,1]$ such that 
\begin{align*}
\left||x+y|^p-|x|^p-p|x|^{p-2}xy\right|
=p(p-1)|x+ty|^{p-2}y^2
\lesssim
|x|^{p-2}y^2+|y|^p,
\end{align*}
where the implicit positive constant depends only on $p$.
Thus, if let $R(\xi):=U(\xi)-pL^{(1)}_\xi(f,g)$
for every $\xi\in\mathbb{S}^{n-1}$, then, from 
\eqref{prop-expan-lap-e1} and \eqref{prop-expan-lap-e4}, 
it follows that, for $\varepsilon\in(0,1)$, 
\begin{align}\label{prop-expan-lap-e5}
|R(\xi)|\lesssim
\left[
\widetilde{L}^{(2)}_\xi(f,g)+\|\nabla_\xi g\|_p^p
\right]\lesssim a\varepsilon
\quad\text{and}\quad
|U(\xi)|\lesssim a\varepsilon,
\end{align}
where the implicit positive constants depend only on $p$.
Therefore, for sufficiently small $\varepsilon$,
using \eqref{prop-expan-lap-e3} and Proposition \ref{prop-neg-Lebe} with $\alpha=\frac np$,
we have 
\begin{align}\label{prop-expan-lap-e6}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
&\ge
|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}V_0(\xi)\,d\xi\notag\\
&\quad
-\left(1+\frac np\right)
|\mathbb{S}^{n-1}|^{-1-\frac pn}a^{-1}
\int_{\mathbb{S}^{n-1}}
\left(U-\fint_{\mathbb{S}^{n-1}}U\right)
\left(V_0-\fint_{\mathbb{S}^{n-1}}V_0\right)\notag\\
&\quad
-C_{1}\varepsilon|\mathbb{S}^{n-1}|^{-1-\frac pn}a^{-1}
\left[
\int_{\mathbb{S}^{n-1}}|UV_0|
+
\left(\int_{\mathbb{S}^{n-1}}|U|\right)
\left(\fint_{\mathbb{S}^{n-1}}|V_0|\right)
\right]\notag\\
&=:|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(f,g)+\widetilde{L}^{(2)}_\xi(f,g)\right]\,d\xi
-I_1(g)-I_{2}(g),
\end{align}
where $C_{1}$ is a positive constant depending only on $n$ and $p$.

We first deal with $I_1$. Note that
\begin{align*}
&\int_{\mathbb{S}^{n-1}}
\left(U-\fint_{\mathbb{S}^{n-1}}U\right)
\left(V_0-\fint_{\mathbb{S}^{n-1}}V_0\right)d\xi \\
&\quad =
p\int_{\mathbb{S}^{n-1}}
\left[
L^{(1)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}L^{(1)}_u(f,g)\,du
\right]^2\,d\xi \\
&\qquad
+
\int_{\mathbb{S}^{n-1}}
\left[
\widetilde{L}^{(2)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_u(f,g)\,du
\right]\\
&\qquad\quad\times\left\{
pL^{(1)}_\xi(f,g)+R(\xi)
-\fint_{\mathbb{S}^{n-1}}\left[pL^{(1)}_u(f,g)+R(u)\right]\,du\right\}\,d\xi \\
&\qquad
+
\int_{\mathbb{S}^{n-1}}
\left[
R(\xi)-\fint_{\mathbb{S}^{n-1}}R(u)\,du
\right]
\left[
L^{(1)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}L^{(1)}_u(f,g)\,du
\right]d\xi .
\end{align*}
Observe that the first term on the right-hand side
is exactly $p\operatorname{Var}(f,g)$.
We now estimate the last two terms. By \eqref{prop-expan-lap-e4}
and \eqref{prop-expan-lap-e5}, we find that
$|pL^{(1)}_\xi(f,g)+R(\xi)|\lesssim a\varepsilon$,
which, combined with $L^(2)(f,g)\ge 0$, further implies
\begin{align*}
&\left|\int_{\mathbb{S}^{n-1}}
\left[
\widetilde{L}^{(2)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_u(f,g)\,du
\right]
\left\{
pL^{(1)}_\xi(f,g)+R(\xi)
-\fint_{\mathbb{S}^{n-1}}\left[pL^{(1)}_u(f,g)+R(u)\right]\,du\right\}\,d\xi\right| \\
&\quad\lesssim a\varepsilon
\int_{\mathbb{S}^{n-1}}
\left|
\widetilde{L}^{(2)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_u(f,g)\,du
\right|\,d\xi\lesssim a\varepsilon
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(f,g)\,d\xi ,
\end{align*}
where the implicit positive constants depend only on $n$ and $p$.
Similarly, one has 
\begin{align*}
&\left|
\int_{\mathbb{S}^{n-1}}
\biggl[
R(\xi)-\fint_{\mathbb{S}^{n-1}}R(u)\,du
\biggr]
\biggl[
L^{(1)}_\xi(f,g)
-\fint_{\mathbb{S}^{n-1}}L^{(1)}_u(f,g)\,du
\biggr]d\xi
\right|\\
&\quad\lesssim a\varepsilon
\int_{\mathbb{S}^{n-1}}
\left|
R(\xi)-\fint_{\mathbb{S}^{n-1}}R(u)\,du
\right|d\xi\lesssim a\varepsilon
\int_{\mathbb{S}^{n-1}}
\left[
\widetilde{L}^{(2)}_\xi(f,g)+\|\nabla_\xi g\|_p^p
\right]d\xi .
\end{align*}
Combining the above estimates and the fact 
$\mathcal{E}_p(f)^p=|\mathbb{S}^{n-1}|^{-\frac pn}a$
(since $f$ is radial), we obtain,
for sufficiently small $\varepsilon$, 
\begin{align}\label{prop-expan-lap-e7}
|I_1(g)|&\le
(n+p)|\mathbb{S}^{n-1}|^{-1-\frac {2p}n}\operatorname{Var}(f,g)
+C_{2}\varepsilon
\int_{\mathbb{S}^{n-1}}
\left[
\widetilde{L}^{(2)}_\xi(f,g)+\|\nabla_\xi g\|_p^p
\right]d\xi,
\end{align}
where $C_{2}$ is a positive constant depending only on $n$ and $p$.
This is the desired estimate of $I_1$. 
On the other hand, using the similar argument and the 
good pointwise estimates of $L^{(1)}_\xi$, 
$\widetilde{L}^{(2)}_\xi$, and $R$ obtained above, 
we can also find that, for sufficiently small $\varepsilon$,
\begin{align*}
|I_2(g)|\lesssim 
\varepsilon
\int_{\mathbb{S}^{n-1}}
\left[\widetilde{L}^{(2)}_\xi(f,g)+\|\nabla_\xi g\|_p^p\right]\,d\xi
\end{align*}
with the implicit positive constant depending only on $n$ and $p$.
This, together with \eqref{prop-expan-lap-e6}, \eqref{prop-expan-lap-e7},
and \eqref{eq-lp},
further implies that, for sufficiently small $\varepsilon$,
\begin{align*}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
&\ge
|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\,d\xi
+(1-C_{3}\varepsilon)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(f,g)\,d\xi\notag\\
&\quad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac {2p}n}
\mathcal{E}_p(f)^{-p}\operatorname{Var}(f,g)
-C_{4}\varepsilon\|\nabla g\|_p^p,
\end{align*}
where $C_{3}$ and $C_4$ are positive constants depending only on $n$ and $p$.
Furthermore, since 
$f$ is radial, from the fact $\mathcal{E}_p(f)^p=|\mathbb{S}^{n-1}|^{-\frac pn}a$
and the definition of the affine Laplacian, we infer that
\begin{align}\label{prop-expan-lap-e12}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g
&=\mathcal{E}_p(f)^{n+p}\int_{\mathbb{S}^{n-1}}
\|\nabla_\xi f\|_p^{n+p}\int_{\mathbb{R}^n}
|\nabla_\xi f|^{p-2}\nabla_\xi f\nabla_\xi g\notag\\
&=|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(f,g)\,d\xi.
\end{align}
Thus, it holds that, for sufficiently small $\varepsilon$,
\begin{align}\label{prop-expan-lap-e9}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
&\ge
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g
+(1-C_{3}\varepsilon)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(f,g)\,d\xi\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac {2p}n}
\mathcal{E}_p(f)^{-p}\operatorname{Var}(f,g)
-C_{4}\varepsilon\|\nabla g\|_p^p.
\end{align}

We next apply the same argument with $\theta=1$. In this case,
\eqref{eq:pointwise-LZ} shows that there exists a
positive constant $\widetilde{c}_p$, depending only
on $p$, such that
\begin{align*}
\int_{\mathbb{R}^n}|\nabla_\xi (f+g)|^{p-2}
\nabla_\xi(f+g)\nabla_\xi g
\ge
L^{(1)}_\xi(f,g)+\widetilde{c}_p\|\nabla_\xi g\|_p^p.
\end{align*}
Repeating the argument used above with 
$V_0$ replaced by $L^{(1)}_\xi(f,g)+\widetilde{c}_p\|\nabla_\xi g\|_p^p$,
one has, for sufficiently small $\varepsilon$, 
\begin{align}\label{prop-expan-lap-e10}
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f+g)g
&\ge
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g
-C_5\varepsilon|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(f,g)\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac {2p}n}\int_{\mathbb{S}^{n-1}}
\mathcal{E}_p(f)^{-p}\operatorname{Var}(f,g)+C_6\|\nabla g\|_p^p,
\end{align}
where $C_{5}$ and $C_6$ are positive constants depending only on $n$ and $p$.
Now, letting $(1-\varepsilon)\times\eqref{prop-expan-lap-e9}+
\varepsilon\times\eqref{prop-expan-lap-e10}$,
we obtain the desired estimate \eqref{prop-expan-lap-e2}.
Then the present proof is complete.
\end{proof}

Moreover, to show the stability of critical points,
we also need the following upper estimate.

\begin{proposition}\label{prop-expan-E}
Let $p\ge 2$. Then there exist positive constants $c_{n,p}$ and 
$C_{n,p}$,
depending only on $n$ and $p$, such that, for any
$\varepsilon\in(0,c_{n,p})$ and $g\in\dot W^{1,p}$ satisfying
\begin{align}\label{prop-expan-E-e1}
\sup_{\xi\in\mathbb{S}^{n-1}}
\frac{\|\nabla_\xi g\|_{p}}
{\|\nabla_\xi f\|_p}\le \varepsilon,
\end{align}
it holds that
\begin{align}\label{prop-expan-E-e2}
\left|\mathcal{E}_p(f+g)^p
-\mathcal{E}_p(f)
+p\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g\right|
\le C_{n,p}\|\nabla f\|_p^{p-2}\|\nabla g\|_{p}^2.
\end{align}
\end{proposition}

\begin{proof}
We aim to apply Proposition \ref{prop-neg-Lebe}\eqref{it2-neg-Lebe}.
As in the proof of Proposition \ref{prop-expan-lap}, 
define $a:=\|\nabla_\xi f\|_p^p$ (which is independent of $\xi$)
and $U(\xi):=\|\nabla_\xi (f+g)\|_p^p-\|\nabla_\xi f\|_p^p$ for 
all $\xi\in\mathbb{S}^{n-1}$.
From Taylor's formula, we deduce that, for every $x,y\in\mathbb{R}$,
there exists $t\in[0,1]$ such that
\begin{align*}
\left||x+y|^p-|x|^p\right|
=p|x+ty|^{p-1}|y|\lesssim
(|x|^{p-1}+|y|^{p-1})|y|,
\end{align*}
where the implicit positive constant depends only on $n$ and $p$.
Therefore, by H\"older's inequality and \eqref{prop-expan-E-e1},
it holds that, for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{prop-expan-E-e3}
|U(\xi)|&\lesssim\int_{\mathbb{R}^n}
\left[|\nabla_\xi f|^{p-1}|\nabla_\xi g|
+|\nabla_\xi g|^p\right]\notag\\
&\le \|\nabla_\xi f\|_p^{p-1}\|\nabla_\xi g\|_p
+\|\nabla_\xi f\|_p^{p}
\lesssim(\varepsilon+\varepsilon^p)a
\end{align}
with the implicit positive constant depending only on $n$ and $p$.
Then, using this and Proposition \ref{prop-neg-Lebe}\eqref{it2-neg-Lebe}, 
we conclude that, if $\varepsilon$ is sufficiently small, then 
\begin{align}\label{prop-expan-E-e4}
\left|\mathcal{E}_p(f+g)^p
-\mathcal{E}_p(f)-
|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}U(\xi)\,d\xi\right|
&\lesssim a^{-1}\|U\|_{L^\infty(\mathbb{S}^{n-1})}^2.
\end{align}
Since $f$ is radial, we infer that $a=\|\nabla_\xi f\|_p^p\sim\|\nabla f\|_p$.
Combining this and \eqref{prop-expan-E-e3}, we obtain, 
for sufficiently small $\varepsilon$,
\begin{align}\label{prop-expan-E-e5}
\operatorname{LHS}\eqref{prop-expan-E-e4}
\lesssim \|\nabla f\|_p^{-p}
\left(\|\nabla f\|_p^{p-1}\|\nabla g\|_p\right)^2
=\|\nabla f\|_p^{p-2}\|\nabla g\|_p^2.
\end{align}
On the other hand, using \eqref{prop-expan-lap-e12}, we find that,
for sufficiently small $\varepsilon$,
\begin{align*}
&\left|p\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(f)g
+|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}U(\xi)\,d\xi\right|\\
&\quad=\left||\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[U(\xi)-p
L^{(1)}_\xi(f,g)\right]\,d\xi\right|\\
&\quad\lesssim\int_{\mathbb{S}^{n-1}}
\left[
\widetilde{L}^{(2)}_\xi(f,g)+\|\nabla_\xi g\|_p^p
\right]\,d\xi\quad\text{by \eqref{prop-expan-lap-e5}}\\
&\quad\lesssim\|\nabla f\|_p^{p-2}\|\nabla g\|_p^2
\quad\text{by \eqref{eq-L2}}.
\end{align*}
This, together with \eqref{prop-expan-E-e5},
gives \eqref{prop-expan-E-e2} and finishes the present proof.
\end{proof}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\section{Affine spectral gap inequality}\label{sec-spectral}

The main objective of this section is to establish an affine spectral 
gap inequality. We first introduce some 
symbols. 
Recall that $U(x):=c_0(1+|x|^{p'})^{\frac{p-n}{p}}$
satisfying $\|U\|_{p^*}=1$.
Then $U$ satisfies 
\begin{align}\label{eq-pU}
-\Delta_pU:=-\operatorname{div}\left(|\nabla U|^{p-2}\nabla U\right)
={S}_{\rm Sob}^{p}U^{p^*-1};
\end{align}
see \cite{a76,t76,cnv04}.

For any positive $V=c_0T_{\lambda_0,S,x_0}U\in\mathcal{M}_{\rm aff}$ with 
affine transformation parameters $(c_0,\lambda_0,S,x_0)\in\mathscr{A}$,
define the weighted Lebesgue space $L^{2}(V^{p^*-2})$ by setting 
\begin{align*}
L^2(V^{p^*-2}):=
\left\{f:\|f\|_{L^{2}(V^{p^*-2})}:=
\left(\int_{\mathbb{R}^n}|f|^2V^{p^*-2}\right)^{\frac12}<\infty\right\}.
\end{align*}
For any
$\varphi,\psi\in L^2(V^{p^*-2})$, we say that 
$\varphi\perp \psi$ in $L^2(V^{p^*-2})$ if 
\begin{align*}
\langle\varphi,\psi\rangle:=
\int_{\mathbb{R}^n}V^{p^*-2}\varphi\psi=0;
\end{align*}
moreover, define the tangent space of $\mathcal{M}_{\rm aff}$ at $V$ 
by setting 
\begin{align}\label{eq-tangent}
T_V\mathcal{M}_{\rm aff}:=
\left\{V,\ \partial_i V,\ \frac{n-p}{p}V+x\cdot\nabla V,\
\langle Bx,\nabla V\rangle:i=1,\ldots,n,\ B=B^{T},\ \operatorname{Tr}B=0\right\}.
\end{align}
Let $C_{{\rm c},V}^1$ denote
the set of compactly supported functions of class $C^1$ 
that are constant in a neighborhood of $x_0$.
Define the weighted Sobolev space $\dot{W}^{1,2}(|\nabla V|^{p-2})$
to be the closure of $C_{{\rm c},{V}}^1$ with respect to 
\begin{align*}
\|\varphi\|_{\dot{W}^{1,2}(|\nabla V|^{p-2})}
:=\left(\int_{\mathbb{R}^n}|\nabla \varphi|^2|\nabla V|^{p-2}\right)^{\frac12}.
\end{align*}

We now recall the classical spectral gap inequality,
which was established in \cite{fn19,fz22}.
For simplicity, in the remainder of this section, we assume $V=U$, 
with the understanding that the orthogonality condition 
$\perp$ holds in $L^2(U^{p^*-2})$ and will not be explicitly mentioned hereafter.
Define the linearized $p$-Laplacian $\mathcal{L}_U$ by setting 
\begin{align*}
\mathcal{L}_U(\varphi):=-\operatorname{div}
\left(|\nabla U|^{p-2}\nabla\varphi+(p-2)
|\nabla U|^{p-4}(\nabla U\cdot\nabla\varphi)\nabla U\right).
\end{align*}
Then one has the following spectral gap inequality for $\mathcal{L}_U$
(see \cite[Proposition 3.6]{fz22}).

\begin{lemma}\label{fz-spectral}
There exists $\delta\in(0,\infty)$,
depending only on $n$ and $p$, such that, for any function $\varphi\in \dot{W}^{1,2}(|\nabla U|^{p-2})$
satisfying $\varphi\perp\{U,\ \partial_i U,\ \frac{n-p}{p}U+x\cdot\nabla U:i=1,\ldots,n\}$,
\begin{align*}
\langle \mathcal{L}_U\varphi,\varphi\rangle
\ge\left[(p^*-1){S}_{\rm Sob}^p+\delta\right]
\int_{\mathbb{R}^n}U^{p^*-2}\varphi^2\,dx.
\end{align*}
\end{lemma}

In this section, we prove the following affine spectral gap inequality.

\begin{theorem}\label{thm-spectral}
There exists $\delta\in(0,\infty)$,
depending only on $n$ and $p$, such that, for any $\varphi\in \dot{W}^{1,2}(|\nabla U|^{p-2})$
satisfying $\varphi\perp T_U\mathcal{M}_{\rm aff}$,
\begin{align}\label{thm-spectral-e0}
&(p-1)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^n}
|\nabla_\xi U|^{p-2}(\nabla_\xi\varphi)^2\,dx\,d\xi
-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi)\notag\\
&\qquad\ge\left[(p^*-1)S_{\rm aff}^p+\delta\right]
\int_{\mathbb{R}^n}U^{p^*-2}\varphi^2\,dx,
\end{align}
where $L^{(1)}_\xi$ is as in \eqref{def-L1} with 
$f$ replaced by $U$.
\end{theorem}

To prove this, we need the spherical harmonic functions, which 
are homogeneous harmonic polynomials restricted to the
sphere. To be exact, for any $\ell\in\mathbb{Z}_+$,
$Y_{\ell}$ is called an \emph{$\ell$ order spherical harmonic function}
if it is $\ell$ homogeneous [that is, for any $\lambda\in(0,\infty)$,
$Y_\ell(\lambda\cdot)=\lambda^\ell Y_\ell(\cdot)$]
polynomial such that 
\begin{align}\label{eq-sphere}
-\Delta_{\mathbb S^{n-1}}Y_{\ell,m}
  =\mu_\ell Y_{\ell,m},\quad \mu_\ell:=\ell(\ell+n-2),
\end{align}
where $\Delta_{\mathbb{S}^{n-1}}$ denotes the Laplace--Beltrami operator
on $\mathbb{S}^{n-1}$.
This family of functions forms an orthonormal basis for $L^2(\mathbb{S}^{n-1})$.
Moreover, we denote by $\mathcal{H}_\ell(\mathbb{S}^{n-1})
:=\{Y_{\ell,m}\}_m$ the set of all $\ell$ order spherical harmonic functions.
For more properties about spherical harmonic functions,
we refer to \cite{dx13}.
We first establish the following identities.

\begin{lemma}\label{lem-L2}
Let $\alpha_{n,p}$ be as in \eqref{eq-alpha}.
\begin{enumerate}[{\rm(i)}]
  \item\label{lem-L2-it1} For any $\varphi\in\dot{W}^{1,2}(|\nabla U|^{p-2})$,
\begin{align}\label{lem-L2-e0}
\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^n}
|\nabla_\xi U|^{p-2}(\nabla_\xi\varphi)^2\,dx\,d\xi
=\frac{\alpha_{n,p}^p}{p-1}\langle\mathcal{L}_U\varphi,\varphi\rangle.
\end{align}
  \item\label{lem-L2-it2} For any $Y_{\ell,m}\in\mathcal{H}_\ell(\mathbb{S}^{n-1})$
and any differentiable function $f$ in $(0,\infty)$,
\begin{align*}
\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^n}
|\nabla_\xi U|^{p-2}(\nabla_\xi [fY_{\ell,m}])^2\,dx\,d\xi
=\frac{\alpha_{n,p}^p}{p-1}
\int_{0}^{\infty}|\partial_r U|^{p-2}
r^{n-1}\left[(p-1)f'^2+\frac{\mu_\ell}{r^2}f^2\right]\,dr.
\end{align*}
\end{enumerate}
\end{lemma}

To show this lemma, we first calculate 
a spherical integral explicitly.

\begin{lemma}\label{lem-wu}
If $p\in(1,\infty)$, then,
for any $u\in\mathbb{R}^n$ and $\omega\in\mathbb{S}^{n-1}$,
\begin{align}\label{lem-wu-e0}
\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^{p-2}
|u\cdot\xi|^2\,d\xi=\alpha_{n,p}^p
\left[|u\cdot\omega|^2+\frac{|u|^2-|u\cdot\omega|^2}{p-1}\right].
\end{align}
\end{lemma}

\begin{proof}
For $u:=(u_1,\ldots,u_n)\in\mathbb{R}^n$ and $\omega\in\mathbb{S}^{n-1}$, define
\begin{align*}
F(\omega,u):=\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^{p-2}
|u\cdot\xi|^2\,d\xi.
\end{align*}
Choose $A\in O(n)$ such that $A\omega=e_n:=(0,\ldots,0,1)$.
Then, by a change of variables, we conclude that 
\begin{align}\label{lem-wu-e1}
F(\omega,u)
&=\int_{\mathbb{S}^{n-1}}
\left|e_n\cdot(A\xi)\right|^{p-2}
\left|(Au)\cdot(A\xi)\right|^2\,d\xi\notag\\
&=\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}
|(Au)\cdot\xi|^2\,d\xi=F(e_n,Au).
\end{align}
Therefore, we first consider the case $\omega=e_n$. In this case,
\begin{align}\label{lem-wu-e2}
F(e_n,u)
&=\int_{\mathbb{S}^{n-1}}
|\xi_n|^{p-2}\left(\sum_{i=1}^n u_i\xi_i\right)^2\,d\xi
=\sum_{i=1}^{n-1}u_i^2\int_{\mathbb{S}^{n-1}}
|\xi_n|^{p-2}\xi_i^2\,d\xi
+u_n^2\int_{\mathbb{S}^{n-1}}|\xi_n|^p\,d\xi\notag\\
&=\sum_{i=1}^{n-1}u_i^2\int_{\mathbb{S}^{n-1}}
|\xi_n|^{p-2}\xi_i^2\,d\xi+u_n^2\alpha_{n,p}^p.
\end{align}
Observe that, for any $i\in\{1,\ldots,n-1\}$,
$\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}\xi_i^2\,d\xi=
\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}\xi_1^2\,d\xi$.
Indeed, choose $B\in O(n)$ such that 
$Be_i=e_1$ and $Be_n=e_n$; then using a change of variables, 
we immediately find that this claim holds.
Hence, 
\begin{align*}
(n-1)\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}\xi_1^2\,d\xi
&=\sum_{i=1}^{n-1}\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}\xi_i^2\,d\xi\\
&=\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}(1-\xi_n^2)\,d\xi
=\alpha_{n,p-2}^{p-2}-\alpha_{n,p}^p.
\end{align*}
Since we assume that $p\in(1,\infty)$,
from \eqref{eq-alpha} and the basic property $\Gamma(\alpha+1)=\alpha\Gamma(\alpha)$
for the gamma function, we infer that 
$\alpha_{n,p-2}^{p-2}=\frac{n+p-2}{p-1}\alpha_{n,p}^{p}$.
This then implies, for any $i\in\{1,\ldots,n-1\}$,
\begin{align*}
\int_{\mathbb{S}^{n-1}}|\xi_n|^{p-2}\xi_i^2\,d\xi
=\frac{1}{n-1}\left(\alpha_{n,p-2}^{p-2}-\alpha_{n,p}^p\right)
=\frac{1}{p-1}\alpha_{n,p}^p.
\end{align*}
Combining this and \eqref{lem-wu-e2}, we further obtain 
\begin{align*}
F(e_n,u)=\alpha_{n,p}^p
\left[u_n^2+\frac{1}{p-1}(|u|^2-u_n^2)\right],
\end{align*}
which shows \eqref{lem-wu-e0} for $u=e_n$.
Furthermore, using this, the fact 
$(Au)_n=(Au)\cdot e_n=u\cdot(A^Te_n)=u\cdot\omega$, and 
\eqref{lem-wu-e1}, we conclude \eqref{lem-wu-e0} for general $u$.
This finishes the proof.
\end{proof}


\begin{proof}[Proof of Lemma \ref{lem-L2}]
We first prove \eqref{lem-L2-it1}. Using polar coordinates, we have 
\begin{align}\label{eq-nabla}
\nabla=\partial_r \omega+\frac1r\nabla_{\mathbb{S}^{n-1}},
\quad\omega=\frac{x}{|x|}\in\mathbb{S}^{n-1}.
\end{align}
Since $U$ is radial, it follows that 
$\nabla U=\partial_r U\omega$, $\nabla\varphi=\partial_r\varphi\omega
+\frac1r\nabla_{\mathbb{S}^{n-1}}\varphi$, and 
$\omega\perp\nabla_{\mathbb{S}^{n-1}}\varphi$.
This further implies 
\begin{align}\label{lem-L2-e1}
\langle\mathcal{L}_U\varphi,\varphi\rangle
&=\int_{\mathbb{R}^n}|\nabla U|^{p-2}
(\nabla\varphi)^2+(p-2)|\nabla U|^{p-4}
(\nabla U\cdot\nabla\varphi)^2\notag\\
&=\int_{0}^{\infty}\int_{\mathbb{S}^{n-1}}
\left\{|\partial_rU|^{p-2}
\left[|\partial_r\varphi|^2+\frac{1}{r^2}
|\nabla_{\mathbb{S}^{n-1}}\varphi|^2\right]
+(p-2)|\partial_r U|^{p-4}(\partial_r U\partial_r\varphi)^2\right\}
r^{n-1}\,d\omega\,dr\notag\\
&=\int_{0}^{\infty}|\partial_r U|^{p-2}r^{n-1}
\int_{\mathbb{S}^{n-1}}\left[(p-1)|\partial_r\varphi|^2+\frac{1}{r^2}
|\nabla_{\mathbb{S}^{n-1}}\varphi|^2\right]\,d\omega\,dr.
\end{align}
Moreover, note that $\nabla_\xi U=\partial_r U(\omega\cdot\xi)$.
Applying polar coordinates for the inner $\int_{\mathbb{R}^n}$ and 
Lemma \ref{lem-wu}, we obtain 
\begin{align*}
\operatorname{LHS}\eqref{lem-L2-e0}
&=\int_{0}^{\infty}|\partial_r U|^{p-2}r^{n-1}
\int_{\mathbb{S}^{n-1}}\int_{\mathbb{S}^{n-1}}
|\omega\cdot\xi|^{p-2}|\nabla\varphi\cdot\xi|^2\,d\omega\,d\xi\,dr\\
&=\alpha_{n,p}^p\int_{0}^{\infty}|\partial_rU|^{p-2}r^{n-1}
\int_{\mathbb{S}^{n-1}}\left(|\partial_r\varphi|^2+\frac{1}{p-1}
\left[|\nabla\varphi|^2-
\left|\nabla\varphi\cdot(\partial_r\varphi\omega)\right|^2\right]\right)\,d\omega\,dr\\
&=\alpha_{n,p}^p\int_{0}^{\infty}|\partial_rU|^{p-2}r^{n-1}
\int_{\mathbb{S}^{n-1}}\left(|\partial_r\varphi|^2+\frac{1}{(p-1)r^2}
|\nabla_{\mathbb{S}^{n-1}}\varphi|^2\right)\,d\omega\,dr\\
&\qquad\text{by \eqref{eq-nabla} and $\omega\perp\nabla_{\mathbb{S}^{n-1}}\varphi$}\\
&=\frac{\alpha_{n,p}^p}{p-1}\langle\mathcal{L}_U\varphi,\varphi\rangle
\quad\text{by \eqref{lem-L2-e1}}\\
&=\operatorname{RHS}\eqref{lem-L2-e0}.
\end{align*}
This finishes the proof \eqref{lem-L2-it1}. 

Once we show \eqref{lem-L2-it1}, \eqref{lem-L2-it2} is a 
direct consequence of the representation of $\mathcal{L}_U$ in polar coordinates 
(see \cite[(3.10)]{fn19}).
For the convenience of the reader, we also give some details.
Indeed, if $\varphi=fY_{\ell,m}$, then
$\partial_r\varphi=f'Y_{\ell,m}\omega$ and 
$\nabla_{\mathbb{S}_{n-1}}\varphi=f\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}$.
Combining this, the facts that $\int_{\mathbb{S}^{n-1}}Y_{\ell,m}^2=1$ and 
\begin{align*}
\int_{\mathbb{S}^{n-1}}|\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}|^2
&=-\int_{\mathbb{S}^{n-1}}(\Delta_{\mathbb{S}^{n-1}}Y_{\ell,m})Y_{\ell,m}\\
&=\mu_\ell\int_{\mathbb{S}^{n-1}}Y_{\ell,m}^2\quad\text{by \eqref{eq-sphere}}\\
&=\mu_\ell,
\end{align*}
and \eqref{lem-L2-e1}, we conclude that 
\begin{align*}
\langle\mathcal{L}_U(fY_{\ell,m}),fY_{\ell,m}\rangle
=\int_{0}^{\infty}|\partial_r U|^{p-2}
r^{n-1}\left[(p-1)f'^2+\frac{\mu_\ell}{r^2}f^2\right]\,dr.
\end{align*}
This, together with \eqref{lem-L2-it1},
proves \eqref{lem-L2-it2}.
\end{proof}

Recall that $L^{(1)}_\xi$ is defined as in \eqref{def-L1}
with $f$ replaced by $U$.
We have the following conclusion.

\begin{lemma}\label{lem-L1}
Let $\ell\in\mathbb{Z}_+$.
Then there exists $\beta_{\ell,p}\in\mathbb{R}$
such that
\begin{align}\label{lem-L1-e0}
|\beta_{\ell,p}|=
\begin{cases}
\alpha_{n,p}^p&\text{if $\ell=0$},\\
{\displaystyle\alpha_{n,p}^p
\prod_{j=0}^{k-1}\left|\frac{2j-p}{2j+n+p}\right|}& 
\text{if $\ell=2k$ for $k\in\mathbb{N}$},\\
0&\text{otherwise}
\end{cases}
\end{align}
and, for any $Y_{\ell,m}\in\mathcal{H}_\ell(\mathbb{S}^{n-1})$,
any $\xi\in\mathbb{S}^{n-1}$, and any differentiable function $f$ in $(0,\infty)$,
\begin{align}\label{lem-L1-ee}
L^{(1)}_{\xi}(fY_{\ell,m})
=\beta_{\ell,p}\int_{0}^{\infty}|\partial_rU|^{p-2}
\partial_rUr^{n-1}
\left[f'(r)+\frac{\mu_\ell}{pr}f(r)\right]\,dr
Y_{\ell,m}(\xi).
\end{align}
\end{lemma}

\begin{proof}
Let $\varphi:=fY_{\ell,m}$.
Then, by \eqref{eq-nabla}, we find that 
$\nabla\varphi=f'Y_{\ell,m}\omega+\frac{f}r\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}$.
This further implies that, for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{lem-L1-e1}
L^{(1)}_\xi(\varphi)
&=\int_{0}^{\infty}|\partial_rU|^{p-2}
\partial_rUr^{n-1}f'\int_{\mathbb{S}^{n-1}}
|\omega\cdot\xi|^pY_{\ell,m}\,d\omega\,dr\notag\\
&\quad+\int_{0}^{\infty}|\partial_rU|^{p-2}
\partial_rUr^{n-1}\frac{f}{r}\int_{\mathbb{S}^{n-1}}
|\omega\cdot\xi|^{p-2}(\omega\cdot\xi)\nabla_{\mathbb{S}^{n-1}}
Y_{\ell,m}\cdot\xi\,d\omega\,dr.
\end{align}
Note that, if we consider $|\omega\cdot\xi|^p$ as a function of 
$\omega\in\mathbb{R}^n$, then $\nabla|\omega\cdot\xi|^p
=p|\omega\cdot\xi|^{p-2}(\omega\cdot\xi)\xi$.
This, combined with \eqref{eq-nabla} and the fact 
$\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}\perp\omega$, further implies, for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
(\nabla_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p)
\cdot\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}
=(\nabla|\omega\cdot\xi|^p)\cdot\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}
=p|\omega\cdot\xi|^{p-2}(\omega\cdot\xi)\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}\cdot\xi.
\end{align*}
Hence, from this and integration by parts, we deduce that, 
 for any $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
&\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^{p-2}
(\omega\cdot\xi)\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}\cdot\xi\,d\omega\\
&\quad=\frac{1}{p}\int_{\mathbb{S}^{n-1}}
(\nabla_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p)
\cdot\nabla_{\mathbb{S}^{n-1}}Y_{\ell,m}\,d\omega
=-\frac1p\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
\Delta_{\mathbb{S}^{n-1}}Y_{\ell,m}\\
&\quad=\frac{\mu_\ell}{p}\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{\ell,m}\,d\omega\quad\text{by \eqref{eq-sphere}}.
\end{align*}
Using this and \eqref{lem-L1-e1}, we obtain 
\begin{align*}
L^{(1)}_\xi(\varphi)=\int_{0}^{\infty}|\partial_rU|^{p-2}
\partial_rUr^{n-1}
\left[f'(r)+\frac{\mu_\ell}{pr}f(r)\right]\,dr
\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{\ell,m}\,d\omega.
\end{align*}

Therefore, to complete the proof, we only need to 
calculate $\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{\ell,m}\,d\omega$.
Indeed, for $\ell=0$, since $Y_{0,m}$ is a constant, then 
$\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{0,m}\,d\omega=\alpha_{n,p}^pY_{0,m}$.
Moreover, if $\ell=2k$ for some $k\in\mathbb{N}$,
then, from \cite[p.\,3054]{r26}, we infer that
there exists $\beta_{\ell,p}$ satisfying \eqref{lem-L1-e0}
such that $\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{\ell,m}\,d\omega=\beta_{\ell,p}Y_{\ell,m}$.
Finally, for a odd number $\ell\in\mathbb{N}$,
applying the Funk--Hecke formula (see \cite[Theorem 1.2.9]{dx13}),
we have $\int_{\mathbb{S}^{n-1}}|\omega\cdot\xi|^p
Y_{\ell,m}\,d\omega=\beta_{\ell,p}Y_{\ell,m}$ and 
\begin{align*}
\beta_{\ell,p}=|\mathbb{S}^{n-2}|\int_{-1}^{1}
|t|^p\frac{C_\ell^{\frac{n-2}{2}}(t)}{C_\ell^{\frac{n-2}{2}}(1)}
(1-t^2)^{\frac{n-3}{2}}\,dt,
\end{align*}
where $C^{\alpha}_\ell$ is the Gegenbauer polynomial.
Noting that Gegenbauer polynomials have the property:
$C_\ell^\alpha(-t)=(-1)^\ell C_\alpha^\lambda(t)$
(see, for instance, Szeg\"o \cite[p.\,80, (4.7.4)]{s75}),
it then follows that 
$\beta_{\ell,p}=0$ when $\ell\ge1$ is odd. 
Hence, we complete the proof.
\end{proof}

Therefore, if we expand
$\varphi\in \dot{W}^{1,2}(|\nabla U|^{p-2})$ in terms of spherical harmonics as
\begin{align*}
\varphi=\sum_{\ell=0}^{\infty}\sum_{m}
\left(\int_{\mathbb{S}^{n-1}}\varphi Y_{\ell,m}\right)
Y_{\ell,m}=:\sum_{\ell=0}^{\infty}\sum_{m}\varphi_{\ell,m},
\end{align*}
then, from the orthogonality of $\{Y_{\ell,m}\}$, it follows that
$\langle \mathcal{L}_U\varphi,\varphi\rangle=\sum_{\ell=0}^{\infty}\sum_{m}
\langle \mathcal{L}_U\varphi_{\ell,m},\varphi_{\ell,m}\rangle$.
On the other hand, by \eqref{lem-L1-ee} and the 
fact that $Y_{\ell,m}$ has zero mean for all $\ell\ge1$, we find that
\begin{align*}
\int_{\mathbb{S}^{n-1}}L^{(1)}_u(\varphi)\,du
=\int_{\mathbb{S}^{n-1}}L^{(1)}_u(\varphi_{0,m})\,du.
\end{align*}
Moreover, $Y_{0,m}$ is a constant. Consequently,
by \eqref{lem-L1-ee} and the orthogonality of $\{Y_{\ell,m}\}$ again,
we have 
\begin{align}\label{eq-L1}
\operatorname{Var}(U,\varphi)
&=\sum_{\ell=1}^\infty
\sum_{m}\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(\varphi_{\ell,m})\right]^2\,d\xi
=\sum_{\ell=0}^{\infty}\sum_{m}\operatorname{Var}(U,
\varphi_{\ell,m}).
\end{align}
This, combined with \eqref{lem-L2-e0}, further implies 
\begin{align*}
\operatorname{LHS}\eqref{thm-spectral-e0}
=\sum_{\ell=0}^{\infty}\sum_{m}R_{U,\ell,m}(\varphi),
\end{align*}
where 
\begin{align*}
R_{U,\ell,m}(\varphi):=&\,
(p-1)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^n}
|\nabla_\xi U|^{p-2}(\nabla_\xi\varphi_{\ell,m})^2\,dx\,d\xi\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi_{\ell,m}).
\end{align*}
Thus, to prove Theorem \ref{thm-spectral},
we only need to estimate $R_{U,\ell,m}$ for all $\ell$ and $m$.

\begin{proposition}\label{prop-spectral}
Let $\ell\in\mathbb{Z}_+$ and $\varphi\in\dot{W}^{1,2}(|\nabla U|^{p-2})$. There exists a positive constant $\delta\in(0,\infty)$,
depending only on $n$ and $p$, such that the following statements hold.
\begin{enumerate}[{\rm(i)}]
  \item\label{it0-spectral} If $\ell=0$ and 
  $\varphi\perp\{U,\ \frac{n-p}{p}U+x\cdot\nabla U\}$,
 then 
 \begin{align}\label{eq-spectral}
 R_{U,\ell,m}(\varphi)\ge \left[(p^*-1)S_{\rm aff}^p+\delta\right]
 \int_{\mathbb{R}^{n}}U^{p^*-2}\varphi_{\ell,m}^2\,dx.
 \end{align}
 \item\label{it1-spectral} If $\ell=1$ and
 $\varphi\perp\{\partial_i U:i=1,\ldots,n\}$, then \eqref{eq-spectral} holds.
 \item\label{it2-spectral} If $\ell=2$ and 
 $\varphi\perp\{\langle Bx,\nabla U\rangle:B=B^{T},\ \operatorname{Tr}B=0\}$,
 then \eqref{eq-spectral} holds.
 \item\label{it3-spectral} If $\ell\ge3$, then 
 \eqref{eq-spectral} holds for all $\varphi$.
\end{enumerate}
\end{proposition}

\begin{proof}[Proof of \eqref{it0-spectral} and \eqref{it1-spectral} of 
Proposition \ref{prop-spectral}]
These assertions are 
the direct consequence of the classical spectral gap 
of $\mathcal{L}_U$ (that is, Lemma \ref{fz-spectral}).
Indeed, from \eqref{eq-L1} and Lemma \ref{lem-L1}, we infer that,
for $\ell=0$ and $\ell=1$, 
$\operatorname{Var}(U,\varphi_{\ell,m})=0$.
Therefore, by \eqref{lem-L2-e0} and \eqref{eq-relaSob}, we have 
\begin{align}\label{it0-spectral-e1}
R_{U,\ell,m}(\varphi)
=|\mathbb{S}_{n-1}|^{-1-\frac pn}\alpha_{n,p}
\langle \mathcal{L}_U(\varphi_{\ell,m}),\varphi_{\ell,m}\rangle
=\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
\langle \mathcal{L}_U(\varphi_{\ell,m}),\varphi_{\ell,m}\rangle.
\end{align}

We first consider $\ell=0$. 
Note that both $U$ and $\frac{n-p}{p}U+x\cdot\nabla U$ are radial and hence 
are orthogonal to $Y_{\ell,m}$ for all $\ell\ge1$. From this and 
the orthogonality assumption on $\varphi$ in \eqref{it0-spectral},
it follows that $\varphi_{0,m}\perp\{U,\ \frac{n-p}{p}U+x\cdot\nabla U\}$;
moreover, since $\varphi_{0,m}$ is radial, 
we deduce that $\varphi_{0,m}\perp\{\partial_i U:i=1,\ldots,n\}$ automatically.
Combining these, \eqref{it0-spectral-e1}, and Lemma \ref{fz-spectral},
we then obtain \eqref{eq-spectral} for $\ell=0$.

Next, we deal with $\ell=1$. Observe that,
for any $i\in\{1,\ldots,n\}$, $\partial_i U=\nabla U\cdot e_i=\partial_rU(\omega\cdot e_i)$,
which is exactly a constant multiple of a first order spherical harmonic function
on $\mathbb{S}^{n-1}$. Thus, $\partial_i U$ is orthogonal to 
$\varphi_{\ell,m}$ for all $\ell\ne 1$. 
Combining this and the orthogonality assumption on $\varphi$ in \eqref{it1-spectral},
we find that 
$\varphi_{1,m}\perp\{\partial_i U:i=1,\ldots,n\}$;
on the other hand, as mentioned above, 
$\varphi_{1,m}\perp\{U,\frac{n-p}{p}U+x\cdot\nabla U\}$ automatically.
Hence, applying \eqref{it0-spectral-e1} and Lemma \ref{fz-spectral} again,
we conclude \eqref{eq-spectral} for $\ell=1$.
This then finishes the proof of \eqref{it0-spectral} 
and \eqref{it1-spectral} of Proposition \ref{prop-spectral}.
\end{proof}

To prove the remaining part of Proposition \ref{prop-spectral}, 
we still need some lemmas. Indeed, from polar coordinates
(see Lemmas \ref{lem-L2} and \ref{lem-L1}), we can 
deduce the following radial representation: for all $\ell\in\mathbb{N}$
and $\varphi=fY_{\ell,m}$ with a differentiable $f$ in $(0,\infty)$,
\begin{align}\label{eq-radialR}
R_{U,\ell,m}(\varphi)
&=\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
\int_{0}^{\infty}|\partial_r U|^{p-2}r^{n-1}
\left[(p-1)f'^2+\frac{\mu_\ell}{r^2}f^2\right]\,dr\notag\\
&\quad-b_{\ell,p}\left(\int_{0}^{\infty}
|\partial_r U|^{p-2}\partial_r U
r^{n-1}\left[f'+\frac{\mu_\ell}{pr}f\right]\,dr\right)^2,
\end{align}
where $b_{\ell,p}:=(p+n)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}\mathcal{E}_p(U)^{-p}
\beta_{\ell,p}^2$ and $\beta_{\ell,p}$ is as in Lemma \ref{lem-L1}.
Moreover, as $U$ is radial, we can regard $U$ as 
a function in $(0,\infty)$ and write $\partial_r U$ simply as $U'$ in what follows for convenience.
Define the operator $L^{(\ell)}_U$ on $L^2(U^{p^*-2}r^{n-1},(0,\infty))$ 
by setting, for any differentiable $f$ in $(0,\infty)$,
\begin{align*}
L_{U}^{(\ell)}(f):=
\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
U^{2-p^*}\left\{-(p-1)r^{1-n}\left(f'|U'|^{p-2}r^{n-1}\right)'
+\frac{\mu_\ell}{r^2}|U'|^{p-2}f\right\}
-(p^*-1)S_{\rm aff}^p I.
\end{align*}
This is exactly a multiple of the Sturm--Liouville type operator
appearing in the classical spectral analysis of $\mathcal{L}_U$
(see \cite{fn19,fz22}).
Indeed, we need the following properties of $L^{(0)}_U$;
see \cite[Proposition 3.1]{fn19} for $p\ge 2$ and see
\cite[Appendix B]{fz22} for other cases.

\begin{lemma}\label{lem-fzspe}
For any $p\in(1,n)$,
it holds that 
\begin{align*}
L^{(0)}_U(U)=-(p^*-p)S_{\rm aff}^pU\quad\text{and}\quad
L^{(0)}_U\left(\frac{n-p}{p}U+rU'\right)=0.
\end{align*} 
\end{lemma}

In addition, if we define 
\begin{align*}
\psi_\ell:=U^{2-p^*}
\left[-r^{1-n}\left(|U'|^{p-2}U'r^{n-1}\right)'
+\frac{\mu_\ell}{pr}|U'|^{p-2}U'\right],
\end{align*}
then, by integration by parts, we find that
\begin{align*}
\left(\int_{0}^{\infty}
|\partial_r U|^{p-2}\partial_r U
r^{n-1}\left[f'+\frac{\mu_\ell}{pr}f\right]\,dr\right)^2
=\langle f,\psi_\ell\rangle_0^2,
\end{align*}
here and thereafter, for any $f,g\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
define 
\begin{align*}
\langle f,g\rangle_0:=\int_{0}^{\infty}fgU^{p^*-2}r^{n-1}.
\end{align*}
Applying Lemma \ref{lem-fzspe}, we 
immediately obtain the following property of $L^{(\ell)}_U$
for $\ell\ge 1$, which is important to prove
the remaining part of Proposition \ref{prop-spectral}.

\begin{lemma}\label{lem-psiz}
For any $\ell\in\mathbb{N}$, 
\begin{align*}
L^{(\ell)}_U(rU')
=p\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
\psi_\ell.
\end{align*}
\end{lemma}

\begin{proof}
Using Lemma \ref{lem-fzspe} and a direct calculation, we find that,
for any $\ell\in\mathbb{N}$,
\begin{align}\label{lem-psiz-e1}
L^{(\ell)}_{U}(rU')&=
L^{(0)}_U(rU')+\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
\left(U^{2-p^*}\frac{\mu_\ell}{r}|U'|^{p-2}U'\right)\notag\\
&=p\left(\frac{S_{\rm aff}}{S_{\rm Sob}}\right)^p
U^{2-p^*}\left({S}_{\rm Sob}^pU^{p^*-1}+\frac{\mu_\ell}{pr}|U'|^{p-2}
U'\right).
\end{align}
On the other hand, \eqref{eq-pU} and \eqref{eq-nabla} show that 
\begin{align*}
{S}_{\rm Sob}^pU^{p^*-1}&=-\operatorname{div}
\left(|\nabla U|^{p-2}\nabla U\right)\\
&=-\left(\partial_r\omega+\frac{1}{r}\nabla_{\mathbb{S}^{n-1}}\right)
\cdot\left(|U'|^{p-2}U'\right)\\
&=\left(|U'|^{p-2}U'\right)'+\frac{n-1}{r}|U'|^{p-2}U'
\quad\text{by $-\operatorname{div}_{\mathbb{S}^{n-1}}\omega
:=\nabla_{\mathbb{S}^{n-1}}\cdot\omega=n-1$}\\
&=r^{1-n}\left(|U'|^{p-2}U'r^{n-1}\right)'.
\end{align*}
Combining this and \eqref{lem-psiz-e1},
we complete the present proof.
\end{proof}

We also need the following identity.

\begin{lemma}\label{lem-psizin}
For any $\ell\in\mathbb{N}$,
\begin{align*}
\langle rU',\psi_\ell\rangle_0
=\left(1+\frac{\mu_\ell-n}{p}\right)
|\mathbb{S}^{n-1}|^{-1} S^{p}_{\rm Sob}.
\end{align*}
\end{lemma}

\begin{proof}
Noting that $(rU')'=U'+rU''$, we have 
\begin{align*}
\langle rU',\psi_\ell\rangle_0
&=\int_{0}^{\infty}
|U'|^{p-2}U'r^{n-1}
\left[\left(1+\frac{\mu_\ell}{p}U'+rU''\right)\right]\,dr\\
&=\left(1+\frac{\mu_\ell}{p}\right)
\int_{0}^{\infty}|U'|^{p}r^{n-1}\,dr+\int_{0}^{\infty}
|U'|^{p-2}U'U''r^n\,dr.
\end{align*}
Since $|U'|^{p-2}U'U''=\frac1p(|U'|^p)'$, from 
$|U'|^p(0)=|U'|^p(\infty)=0$ and integration by parts,
we infer that 
\begin{align*}
\int_{0}^{\infty}|U'|^{p-2}U'U''r^{n}\,dr
=-\frac np\int_{0}^{\infty}
|U'|^{p}r^{n-1}\,dr.
\end{align*}
Hence, by the facts that $U$ is radial and $U$ is an extremal function of 
the classical Sobolev inequality (that is, $\|\nabla U\|_p={S}_{\rm Sob}\|U\|_{p^*}$),
we further obtain 
\begin{align*}
\langle rU',\psi_\ell\rangle_0
=\left(1+\frac{\mu_\ell-n}{p}\right)
\int_{0}^{\infty}|U'|^{p}r^{n-1}\,dr
=\left(1+\frac{\mu_\ell-n}{p}\right)
|\mathbb{S}^{n-1}|^{-1}{S}^{p}_{\rm Sob}.
\end{align*}
This finishes the present proof.
\end{proof}

Also, we need the following lemma to deal with 
the rank-one operator in \eqref{eq-radialR}.

\begin{lemma}\label{lem-rankone}
Let $\mathcal{H}$ be a real Hilbert space, $Q$ be a symmetric
bilinear form on $\mathcal{H}$, and $L\in\mathcal{H}^*$. 
If, for any $f\in\mathcal{H}$,
\begin{align*}
Q(f,f)\ge L(f)^2,
\end{align*} 
where the equality holds if $f=f_0$,
and there exists $\delta\in(0,\infty)$ such that, for any $f\in \mathcal{H}$,
$Q(f,f)\ge\delta\|f\|_{\mathcal{H}}^2$,
then, for any $f$ satisfying $f\perp f_0$,
\begin{align*}
Q(f,f)\ge L(f)^2+\delta\|f\|_{\mathcal{H}}^2.
\end{align*}
\end{lemma}

\begin{proof}
We first claim that, for any $f\in\mathcal{H}$,
$Q(f,f_0)=L(f)L(f_0)$. 
Indeed, using the assumption in the present lemma, we obtain,
for any $t\in\mathbb{R}$ and $f\in\mathcal{H}$,
\begin{align*}
0\le Q(tf+f_0,tf+f_0)-|L(tf+f_0)|^2
=t^{2}\left[Q(f,f)-|L(f)|^2\right]
+2t\left[Q(f,f_0)-L(f)L(f_0)\right].
\end{align*}
Hence, from the arbitrariness of $t\in\mathbb{R}$,
we infer the above claim.
Now, for any $f,g\in\mathcal{H}$, define 
$G(f,g):=Q(f,g)-\delta\langle f,g\rangle_{\mathcal{H}}$.
Then using the assumption in the present lemma and 
Cauchy--Schwarz inequality for bilinear forms, we conclude that,
for any $f\in\mathcal{H}$,
\begin{align}\label{lem-rankone-e1}
G(f,f)G(f_0,f_0)\ge G(f,f_0)^2.
\end{align}
Note that $G(f_0,f_0)=Q(f_0,f_0)-\delta\|f_0\|_{\mathcal{H}}^2
\le Q(f_0,f_0)=L(f_0)^2$
and, by the above claim, it holds that, for any $f\perp f_0$,
\begin{align*}
G(f,f_0)^2=Q(f,f_0)^2=L(f)^2L(f_0)^2.
\end{align*}
Combining these and \eqref{lem-rankone-e1},
we then complete the proof.
\end{proof}

Now, we are able to prove \eqref{it2-spectral} and \eqref{it3-spectral}
of Proposition \ref{prop-spectral}.

\begin{proof}[Proof of Proposition \ref{prop-spectral}\eqref{it2-spectral}]
We first analyse the operator $L^{(2)}_U$.
Since many of the arguments can be generalized to arbitrary $\ell \ge 2$ 
and will be needed later, we choose to present them in a general form here.
Indeed, from the spectral analysis of $L^{(\ell)}_U$
in \cite{fn19,fz22}, we infer that the first eigenvalue 
of $L^{(1)}_U$ is $0$ and hence the first eigenvalue of $L^{(\ell)}_U$,
$\ell\ge 2$, is positive. 
Therefore, using the Rayleigh quotient characterization,
we conclude that there exists
$\delta\in(0,\infty)$, depending only on $n$ and $p$, such that,
for all $\ell\ge2$ and $f\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align}\label{it2-spectral-e1}
\langle L^{(\ell)}_Uf,f\rangle_0\ge \delta\langle f,f\rangle_0.
\end{align}
Hence, $L^{(\ell)}_U$ is invertible and,
if we apply the Cauchy--Schwarz inequality on
the inner product $\langle\cdot,\cdot\rangle^\sim
:=\langle L^{(\ell)}_U\cdot,\cdot\rangle_0$, then we find that,
for all $f,g\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align}\label{it2-spectral-e2}
\langle L^{(\ell)}_Uf,g\rangle_0^2\le
\langle L^{(\ell)}_Uf,f\rangle_0\langle L^{(\ell)}_Ug,g\rangle_0.
\end{align}
Letting $g:=(L^{(\ell)}_U)^{-1}\psi_\ell$ and 
using Lemmas \ref{lem-psiz} and \ref{lem-psizin}, we obtain, for all 
$f\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align}\label{it2-spectral-e3}
\langle f,\psi_\ell\rangle_0^2&\le
\langle \psi_\ell,(L^{(\ell)}_U)^{-1}\psi_\ell\rangle_0
\langle L^{(\ell)}_Uf,f\rangle_0\notag\\
&=\frac{\mu_\ell+p-n}{p^2}\left(\frac{{S}_{\rm Sob}}{S_{\rm aff}}\right)^p
|\mathbb{S}^{n-1}|^{-1}{S}_{\rm Sob}^p
\langle L^{(\ell)}_Uf,f\rangle_0.
\end{align}
In particular, if $\ell=2$, then, from Lemma \ref{lem-L1} and 
\eqref{eq-relaSob}, we infer that 
$$\frac{\mu_\ell+p-n}{p^2}\left(\frac{{S}_{\rm Sob}}{S_{\rm aff}}\right)^p
|\mathbb{S}^{n-1}|^{-1}{S}_{\rm Sob}^p=b_{2,p}^{-1}.$$
This shows that, for all $f\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align}\label{it2-spectral-e4}
\langle L^{(2)}_Uf,f\rangle_0\ge b_{2,p}
\langle f,\psi_2\rangle_0^2.
\end{align}
Moreover, by Lemma \ref{lem-psiz}, we know that 
$g:=(L^{(\ell)}_U)^{-1}\psi_\ell$ and $rU'$ are linearly dependent.
Thus, from the equality condition for the Cauchy--Schwarz inequality
\eqref{it2-spectral-e2}, it follows that 
the `$\ge$' in \eqref{it2-spectral-e4} becomes an `$=$'
when $f=rU'$.
Using this, \eqref{it2-spectral-e1}, \eqref{it2-spectral-e4},
and Lemma \ref{lem-rankone}, we further conclude that,
if $f\perp rU'$ in  $L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
then 
\begin{align}\label{it2-spectral-e5}
\langle L^{(2)}_Uf,f\rangle_0\ge \delta\langle f,f\rangle_0
+b_{2,p}\langle f,\psi_2\rangle_0^2.
\end{align}

Now, we turn to $R_{U,2,m}$. Observe that,
for any symmetric matrix $B$ satisfying $\operatorname{Tr}B=0$,
$\langle Bx,\nabla U\rangle=rU'\langle B\omega,\omega\rangle$,
while $\langle B\omega,\omega\rangle$ is precisely the 
form of the second order spherical harmonic function.
Consequently, for any $\ell\ne 2$, $\varphi_{\ell,m}$
is orthogonal to $\{\langle Bx,\nabla U\rangle:B=B^{T},\ \operatorname{Tr}B=0\}$
automatically. This, combined with the 
orthogonality assumption on $\varphi$ in \eqref{it2-spectral},
further implies that $\int_{\mathbb{S}^{n-1}}\varphi Y_{\ell,m}$ 
is orthogonal to $rU'$.
Applying this and \eqref{it2-spectral-e5},
we find that \eqref{eq-spectral} holds with $\ell=2$.
This finishes the present proof.
\end{proof}

\begin{proof}[Proof of Proposition \ref{prop-spectral}\eqref{it3-spectral}]
By Lemma \ref{lem-L1}, when $\ell$ is odd, $b_{\ell,p}=0$.
Hence, if $\ell\ge 3$ is odd, then \eqref{it2-spectral-e1}
is equivalent to \eqref{eq-spectral},
which completes the proof when $\ell$ is odd.
Next, we consider the case where $\ell\ge 4$ is even.
In this case, from \eqref{it2-spectral-e3}, it follows that,
for all $f\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align*}
\langle L^{(\ell)}_Uf,f\rangle_0
-b_{\ell,p}
\langle f,\psi_\ell\rangle_0^2
\ge
\left[1-\frac{\mu_\ell+p-n}{p^2}
\left(\frac{{S}_{\rm Sob}}{S_{\rm aff}}\right)^p
|\mathbb{S}^{n-1}|^{-1}{S}_{\rm Sob}^p
b_{\ell,p}\right]\langle L^{(\ell)}_Uf,f\rangle_0.
\end{align*}
Therefore, if we can prove that, for all even $\ell\ge 4$,
\begin{align*}
\lambda_{\ell,p}:=
\frac{\mu_\ell+p-n}{p^2}
\left(\frac{{S}_{\rm Sob}}{S_{\rm aff}}\right)^p
|\mathbb{S}^{n-1}|^{-1}{S}_{\rm Sob}^p
b_{\ell,p}
\end{align*}
has a upper bound strictly smaller than $1$,
then, applying \eqref{it2-spectral-e1}, we find that there exists 
$\widetilde{\delta}\in(0,\infty)$, depending only on $n$ and $p$,
such that, for all $f\in L^2(U^{p^*-2}r^{n-1},(0,\infty))$,
\begin{align*}
\langle L^{(\ell)}_Uf,f\rangle_0-b_{\ell,p}\langle f,\psi_\ell\rangle_0^2
\ge \widetilde{\delta}\langle f,f\rangle_0,
\end{align*}
which is equivalent to \eqref{eq-spectral} and hence 
completes the proof.

Now, we estimate $\lambda_{\ell,p}$. Indeed, 
by \eqref{eq-relaSob} and Lemma \ref{lem-L1}, we have,
for any $k\in\mathbb{N}$,
\begin{align*}
\lambda_{2k,p}=\frac{n+p}{p^2}(\mu_{2k}+p-n)
\prod_{j=0}^{k-1}\left(\frac{2j-p}{2j+n+p}\right)^2.
\end{align*}
Then 
\begin{align*}
\frac{\lambda_{2k+2,p}}{\lambda_{2k,p}}-1
&=\frac{\mu_{2k+2}+p-n}{\mu_{2k}+p-n}
\left(\frac{2k-p}{2k+n+p}\right)^2-1\\
&=\frac{(4k+n)[n(n+p) - 2k(2k+n)(n+2p-2)]}
{(\mu_{2k}+p-n)(2k+n+p)^2}<0.
\end{align*}
Therefore, $\{\lambda_{2k,p}\}_{k\ge1}$ is strictly decreasing and hence,
for any $k\ge 2$,
$\lambda_{2k,p}< \lambda_{2,p}=1.$
This is the desired upper bound of $\lambda_{\ell,p}$ and 
finishes the present proof.
\end{proof}

Finally, applying Theorem \ref{thm-spectral}, we obtain 
the following weighted spectral gap inequality, which 
matches the expansions given in Subsection \ref{subsec-expan} well 
and plays a key role in the proofs of Theorems \ref{thm-stab} 
and \ref{thm-bubble} later.

\begin{theorem}\label{thm-spectral-wei}
Let $p\in[2,n)$ and $\delta$ be as in Theorem \ref{thm-spectral}.
If $C_0\in(0,\infty)$ and $\gamma\in(0,\delta)$, then there exists
$\rho_0\in(0,\infty)$ such that, for any $\rho\in(0,\rho_0)$ and 
$\varphi\in\dot{W}^{1,p}$
satisfying $\|\varphi\|_{\dot{W}^{1,p}}<\rho$ and 
$\varphi\perp T_U\mathcal{M}_{\rm aff}$,
\begin{align*}
&(1-C_0\rho)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(U,\varphi)
\,d\xi-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi)\notag\\
&\qquad\ge\left[(p^*-1)S_{\rm aff}^p+\gamma\right]
\int_{\mathbb{R}^n}U^{p^*-2}\varphi^2\,dx
+\rho\int_{\mathbb{R}^n}
|\nabla U|^{p-2}(\nabla\varphi)^2,
\end{align*}
where $\operatorname{Var}$ and $\widetilde{L}^{(2)}$ are as, 
respectively, in \eqref{def-var} and \eqref{eq-L2w}.
\end{theorem}

\begin{remark}\label{rem-spectral}
If $L^{(2)}$ is as in \eqref{def-L2}, then, from the obvious pointwise inequality 
$L^{(2)}_\xi\ge \widetilde{L}^{(2)}_\xi$ for all
$\xi\in\mathbb{S}^{n-1}$, we infer that the conclusion of 
Theorem \ref{thm-spectral-wei} also holds with 
$\widetilde{L}^{(2)}$ replaced by $L^{(2)}$.
\end{remark}

\begin{proof}
We prove this theorem by contradiction. 
If this theorem does not hold, then there exist
$\{\rho_j\}_{j\in\mathbb{N}}$ in $(0,\frac1{2C_0})$
such that $\lim_{j\to\infty}\rho_j=0$ and 
$\{\varphi_j\}_{j\in\mathbb{N}}$ in 
$\dot{W}^{1,p}$ such that, for any $j\in\mathbb{N}$,
$\|\varphi_j\|_{\dot{W}^{1,p}}<\rho_j$,
$\varphi_j\perp T_U\mathcal{M}_{\rm aff}$,
but 
\begin{align}\label{eq-spectral-1}
&(1-C_0\rho_j)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(U,\varphi_j)
\,d\xi-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi_j)\notag\\
&\qquad<\left[(p^*-1)S_{\rm aff}^p+\gamma\right]
\int_{\mathbb{R}^n}U^{p^*-2}\varphi_j^2\,dx
+\rho_j\int_{\mathbb{R}^n}
|\nabla U|^{p-2}(\nabla\varphi_j)^2,
\end{align}
For any $j\in\mathbb{N}$, let 
\begin{align*}
\varepsilon_j:=
\sqrt{(1-C_0\rho_j)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(U,\varphi_j)}.
\end{align*}
Then $\varepsilon_j\in(0,\infty)$. Indeed, $\varepsilon_j<\infty$
is easily concluded by \eqref{eq-L2}.
On the other hand, since $p\ge 2$,  
we deduce that, for any 
$\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{eq-spectral-2}
\widetilde{L}^{(2)}_{\xi}(U,\varphi_j)
\gtrsim\int_{\mathbb{R}^n}|\nabla_\xi U|^{p-2}
|\nabla_\xi\varphi_j|^2;
\end{align}
combining this and Lemmas \ref{lem-L2} and 
\ref{fz-spectral}, we further obtain 
\begin{align*}
\varepsilon_j^2
\gtrsim\sqrt{1-C_0\rho_j}
\langle\mathcal{L}_U\varphi_j,\varphi_j\rangle
\gtrsim\sqrt{1-C_0\rho_j}\int_{\mathbb{R}^n}U^{p^*-2}\varphi_j^2.
\end{align*}
Hence, if $\varepsilon_j=0$, then $\varphi_j=0$, which contradicts 
\eqref{eq-spectral-1}.

Since $\varepsilon_j\in(0,\infty)$, for any $j\in\mathbb{N}$, define 
$\widehat{\varphi}_j:=\frac{\varphi_j}{\varepsilon_j}$.
Then, for every $j\in\mathbb{N}$, 
by Lemma \ref{lem-L2}\eqref{lem-L2-it1}, the assumptions $p\ge 2$
and $\rho_j<\frac{1}{2C_0}$, and \eqref{eq-spectral-2}, we know that 
\begin{align*}
\|\widehat{\varphi}_j\|_{\dot{W}^{1,2}(|\nabla U|^{p-2})}
&\lesssim\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^{n-1}}
|\nabla_\xi U|^{p-2}(\nabla_\xi \widehat{\varphi}_j)^2
\lesssim\frac{1}{1-C_0\rho_j}\lesssim1 .
\end{align*}
This, together with the compact embedding between
$\dot{W}^{1,2}(|\nabla U|^{p-2})$ and $L^2(U^{p^*-2})$
(see \cite[Proposition 3.2]{fz22}), implies that 
there exists a 
subsequence, still denoted by $\{\widehat{\varphi}_{j}\}_{j\in\mathbb{N}}$,
and a function $\varphi_0\in\dot{W}^{1,2}(|\nabla U|^{p-2})$
such that, when $j\to\infty,$
$\widehat{\varphi}_{j}\rightharpoonup\varphi_0$
in $\dot{W}^{1,2}(|\nabla U|^{p-2})$ and 
$\widehat{\varphi}_{j}\to\varphi_0$ in $L^2(U^{p^*-2})$.
Moreover, we have the following two observations
on $\{\varphi_j\}_{j\in\mathbb{N}}$.

Firstly, since $\|\nabla\varphi_j\|_p\to0$ as $j\to\infty$,
we can find a subsequence $\{\varphi_{j_k}\}_{k\in\mathbb{N}}$
such that $|\nabla\varphi_{j_k}|\to0$ almost everywhere 
as $k\to\infty$.
Secondly, for a fixed $R\in(1,\infty)$, define 
$K_R:=B({\bf0},R)\setminus B({\bf0},\frac1R)$.
Then, from the structure of $U$, it follows that, for any 
$j\in\mathbb{N}$, 
\begin{align*}
\int_{K_R}|\nabla\widehat{\varphi}_j|^2
\sim\int_{K_R}|\nabla U|^{p-2}|\nabla\widehat{\varphi}_j|^2
\le \|\widehat{\varphi}_j\|_{\dot{W}^{1,2}(|\nabla U|^{p-2})}
\lesssim 1,
\end{align*}
where the implicit positive constants 
may depend on $R$ but are independent of $j$.
Consequently, by the reflexive of $L^2(K_R;\mathbb{R}^n)$,
there exists a sequence $\{\widehat{\varphi}_{j_k}\}_{k\in\mathbb{N}}$
and a function $u\in L^2(K_R;\mathbb{R}^n)$ such that
$\nabla\widehat{\varphi}_{j_k}\rightharpoonup
u$ in $L^2(K_R;\mathbb{R}^n)$ as $k\to\infty$.
This, combined with the fact
$\widehat{\varphi}_{j}\to\varphi_0$ in $L^2(U^{p^*-2})$,
further implies that, for any $\Phi\in C_{\rm c}^1(K_R;\mathbb{R}^n)$, 
\begin{align*}
\int_{K_R}u\cdot\Phi=\lim_{k\to\infty}\int_{K_R}\nabla\widehat{\varphi}_{j_k}\cdot\Phi
=-\lim_{k\to\infty}\int_{K_R}\widehat{\varphi}_{j_k}\operatorname{div}
\Phi=-\int_{K_R}\varphi_0\operatorname{div}\Phi.
\end{align*}
This shows $u=\nabla\varphi_0$ and 
hence $\nabla\widehat{\varphi}_{j_k}\rightharpoonup
\nabla\varphi_0$ in $L^2(K_R;\mathbb{R}^n)$ as $k\to\infty$.

In summary, for a fixed $R\in(1,\infty)$, we can choose a subsequence, still denoted by 
$\{\varphi_j\}_{j\in\mathbb{N}}$, such that, when 
$j\to\infty$, 
\begin{enumerate}[{\rm(i)}]
  \item\label{it1-varphi} $\widehat{\varphi}_{j}\rightharpoonup\varphi_0$
in $\dot{W}^{1,2}(|\nabla U|^{p-2})$ and 
$\widehat{\varphi}_{j}\to\varphi_0$ in $L^2(U^{p^*-2})$;
  \item\label{it2-varphi} $|\nabla\varphi_j|\to0$ almost everywhere in $\mathbb{R}^n$;
  \item\label{it3-varphi} $\nabla\widehat{\varphi}_{j}\rightharpoonup
\nabla\varphi_0$ in $L^2(K_R;\mathbb{R}^n)$.
\end{enumerate}
Note that, for any $j\in\mathbb{N}$,
\begin{align}\label{eq-spectral-3}
\frac{(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_{\xi}(U,\varphi_j)\,d\xi}{\varepsilon_j^2}
&\ge(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}\int_{\mathbb{R}^n}
\omega_{j,\xi}^{(1)}|\nabla_\xi\widehat{\varphi}_j|^2\notag\\
&\quad+(p-2)(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}
\int_{\mathbb{R}^n}\omega_{j,\xi}^{(2)}\left(
\frac{|\nabla_\xi(U+\varphi_j)|-|\nabla_\xi U|}{\varepsilon_j}\right)^2\notag\\
&=:{\rm I}_j+(p-2){\rm II}_j,
\end{align}
where $$\omega_{j,\xi}^{(1)}:=\frac{\min\{|\nabla_\xi U|,\max\{c_p^{\frac{1}{p-1}}
|\nabla_\xi (U+\varphi_j)|\}\}^{p-1}}
{|\nabla_\xi U|}\quad\text{and}\quad
\omega_{j,\xi}^{(2)}:=\frac{\min\{|\nabla_\xi (U+\varphi_j)|,|\nabla_\xi U|\}^{p-1}}
{|\nabla_\xi U|}$$
with $c_p$ as in \eqref{eq:pointwise-LZ}. We now deal with the terms 
${\rm I}_j$ and ${\rm II}_j$ respectively. Since their underlying arguments are largely analogous but ${\rm II}_j$ is structurally more complicated, we shall present the detailed estimates 
for ${\rm II}_j$ first, and then briefly point out the necessary modifications for ${\rm I}_j$.
To this end, define $\Psi_j:=\widehat{\varphi}_j-\varphi_0$.
Then, from the Newton--Leibniz formula, we infer that,
for any $j\in\mathbb{N}$, 
\begin{align*}
{\rm II}_{j}
&=(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}
\int_{K_R}\omega_{j,\xi}
 \left(\int_{0}^{1}\frac{\nabla_\xi(U+t\varphi_j)}
{|\nabla_\xi(U+t\varphi_j)|}\nabla_\xi\widehat{\varphi}_j\,dt\right)^2\\
&\ge (1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}
\int_{K_R}\omega_{j,\xi} \left(\int_{0}^{1}\frac{\nabla_\xi(U+t\varphi_j)}
{|\nabla_\xi(U+t\varphi_j)|}\,dt\right)^2
\left[(\nabla_\xi\varphi_0)^2
+2\nabla_\xi\varphi_0\nabla_\xi\Psi_j\right]\\
&=:{\rm II}_{j,1}+2{\rm II}_{j,2}.
\end{align*}
Applying \eqref{it2-varphi} and 
the Lebesgue domination convergence theorem, we conclude that,
when $j\to\infty$, $\omega_{j,\xi}\to|\nabla_\xi U|^{p-2}$ 
and  $\int_{0}^{1}\frac{\nabla_\xi(U+t\varphi_j)}{|\nabla_\xi(U+t\varphi_j)|}\,dt\to
\frac{\nabla_\xi U}{|\nabla_\xi U|}$
almost everywhere in $\mathbb{R}^n$.
Hence, by this, $|\omega_{j,\xi}|\le|\nabla_\xi U|^{p-2}$,
and the Lebesgue domination convergence theorem and 
$\lim_{j\to\infty}\rho_j=0$ again, we find that 
\begin{align}\label{eq-spectral-5}
\lim_{j\to\infty}{\rm II}_{j,1}=
\int_{\mathbb{S}^{n-1}}\int_{K_R}
|\nabla_\xi U|^{p-2}(\nabla_\xi \varphi_0)^2.
\end{align}
On the other hand, for any $j\in\mathbb{N}$ and $\xi\in\mathbb{S}^{n-1}$,
it holds that 
\begin{align*}
|A_{j,\xi}|:=
\left|\omega_{j,\xi}\int_{0}^{1}
\frac{\nabla_\xi(U+t\varphi_j)}{|\nabla_\xi(U+t\varphi_j)|}
(\nabla_\xi\varphi_0)\xi\right|
\le|\nabla_\xi U|^{p-2}|\nabla_\xi\varphi_0|
\lesssim|\nabla\varphi_0|,
\end{align*}
where the implicit positive constant is independent of $j$ and 
$\xi$. From this and the structure of $U$, 
we deduce that $A_{j,\xi}\in L^2(K_R;\mathbb{R}^n)$
and $\|A_{j,\xi}\|_{L^2(K_R;\mathbb{R}^n)}\lesssim1$
with the implicit positive constant independent of $j$ and $\xi$.
Therefore, applying \eqref{it3-varphi}, we find that, for any $\xi\in\mathbb{S}^{n-1}$,
$\int_{K_R}\nabla\Psi_j\cdot A_{j,\xi}\to0$ as $j\to\infty$.
Furthermore, since 
\begin{align*}
\left|\int_{K_R}\nabla\Psi_j\cdot A_{j,\xi}\right|
\le\left(\int_{K_R}|\nabla\Psi_j|^2\right)^{\frac12}
\left(\int_{K_R}|A_{j,\xi}|^2\right)^{\frac12}\lesssim1
\end{align*}
uniformly in terms of both $j$ and $\xi$, from Lebesgue 
domination convergence theorem and the assumption 
$\lim_{j\to\infty}\rho_j=0$, we infer that 
\begin{align*}
\lim_{j\to\infty}
{\rm II}_{j,2}=\lim_{j\to\infty}(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}\int_{K_R}
\nabla\Psi_j\cdot A_{j,\xi}=0.
\end{align*}
This, together with 
\eqref{eq-spectral-5}, further implies that 
\begin{align}\label{eq-spectral-9}
\liminf_{j\to\infty}{\rm II}_{j}\ge
\int_{\mathbb{S}^{n-1}}\int_{K_R}
|\nabla_\xi U|^{p-2}(\nabla_\xi \varphi_0)^2,
\end{align}
which is the desired estimate of ${\rm II}_j$.
For ${\rm I}_j$, repeating the same argument:
expanding $\widehat{\varphi}_j=\varphi_0+\Psi_j$
and noting that $\omega_{j,\xi}^{(1)}\to |\nabla_\xi U|^{p-2}$
as $j\to\infty$, we can also obtain 
\begin{align*}
\liminf_{j\to\infty}{\rm I}_j
\ge\int_{\mathbb{S}^{n-1}}\int_{K_R}
|\nabla_\xi U|^{p-2}|\nabla_\xi \varphi_0|^2.
\end{align*}
This, combined with \eqref{eq-spectral-3} and 
\eqref{eq-spectral-9}, further implies that 
\begin{align}\label{eq-spectral-6}
\liminf_{j\to\infty}
\frac{(1-C_0\rho_j)\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_{\xi}(U,\varphi_j)\,d\xi}{\varepsilon_j^2}
\ge(p-1)\int_{\mathbb{S}^{n-1}}\int_{K_R}
|\nabla_\xi U|^{p-2}(\nabla_\xi\varphi_0)^2.
\end{align} 

Next, we prove that, when $j\to\infty$,
\begin{align}\label{eq-spectral-7}
\operatorname{Var}(U,\widehat{\varphi}_j)
\to\operatorname{Var}(U,\varphi_0).
\end{align}
Indeed, by Minkowski's inequality, we obtain
\begin{align*}
\left|\operatorname{Var}(U,\widehat{\varphi}_j)^{\frac12}
-\operatorname{Var}(U,{\varphi}_0)^{\frac12}\right|
&\le\operatorname{Var}(U,\widehat{\varphi}_j-\varphi_0)^{\frac12}\\
&\le\left[\int_{\mathbb{S}^{n-1}} L^{(1)}_{\xi}(U,\widehat{\varphi}_j
-\varphi_0)^2\,d\xi\right]^{\frac12}
=\left[\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(U,\Psi_j)^2\,d\xi\right]^{\frac12}.
\end{align*}
Note that $L^{(1)}_\xi(U,\cdot)$ is a bounded linear functional
on $\dot{W}^{1,2}(|\nabla U|^{p-2})$. Indeed, from 
H\"older's inequality, it follows that, for any 
$f\in\dot{W}^{1,2}(|\nabla U|^{p-2})$,
\begin{align*}
|L^{(1)}_\xi(U,f)|
\le\int_{\mathbb{R}^n}|\nabla U|^{p-1}|\nabla f|
\le\left(\int_{\mathbb{R}^n}|\nabla U|^p\right)^{\frac12}
\left(\int_{\mathbb{R}^n}|\nabla U|^{p-2}|\nabla f|^2\right)^{\frac12}
\sim\|f\|_{\dot{W}^{1,2}(|\nabla U|^{p-2})},
\end{align*}
where the positive equivalence constants are independent of $\xi$.
Hence, using this, \eqref{it1-varphi}, and the Lebesgue 
domination theorem, we obtain
$\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(U,\Psi_j)^2\,d\xi\to0$ as $j\to\infty$,
which shows \eqref{eq-spectral-7}. Then, 
combining \eqref{eq-spectral-1}, \eqref{eq-spectral-6},
\eqref{eq-spectral-7}, \eqref{it1-varphi}, 
and the boundedness of $\{\widehat{\varphi}_j\}_{j\in\mathbb{N}}$
in $\dot{W}^{1,2}(|\nabla U|^{p-2})$,
we conclude that 
\begin{align}\label{eq-spectral-8}
&(p-1)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\int_{K_R}
|\nabla_\xi U|^{p-2}(\nabla_\xi \varphi_0)^2
-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi_0)\notag\\
&\qquad\le\left[(p^*-1)S_{\rm aff}^p+\gamma\right]
\int_{\mathbb{R}^n} U^{p^*-2}\varphi_0^2.
\end{align}
On the other hand, by \eqref{it1-varphi} and 
the orthogonality assumptions on $\{\varphi_j\}_{j\in\mathbb{N}}$,
we know $\varphi_0\perp T_U\mathcal{M}_{\rm aff}$; hence,
applying Theorem \ref{thm-spectral} and \eqref{eq-spectral-8}
together with the arbitrariness of $R$, we further obtain 
\begin{align*}
\left[(p^*-1)S_{\rm aff}^p+\gamma\right]
\int_{\mathbb{R}^n} U^{p^*-2}\varphi_0^2
\ge \left[(p^*-1)S_{\rm aff}^p+\delta\right]
\int_{\mathbb{R}^n} U^{p^*-2}\varphi_0^2.
\end{align*}
Since $\gamma<\delta$, it holds that $\varphi_0=0$ almost everywhere.
However, \eqref{eq-spectral-1} also shows that,
for any $j\in\mathbb{N}$,
\begin{align*}
&1-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\widehat{\varphi}_j)\\
&\quad<\left[(p^*-1)S_{\rm aff}^p+\gamma\right]
\int_{\mathbb{R}^n} U^{p^*-2}\widehat\varphi_j^2
+\rho_j\int_{\mathbb{R}^n}
|\nabla U|^{p-2}(\nabla\widehat\varphi_j)^2,
\end{align*}
which, combined with \eqref{eq-spectral-7}, \eqref{it1-varphi},
and the boundedness of $\{\widehat{\varphi}_j\}_{j\in\mathbb{N}}$
in $\dot{W}^{1,2}(|\nabla U|^{p-2})$ again,
further implies that, if $\varphi_0=0$, then
$1\le 0$. It is impossible!
This then proves the present theorem.
\end{proof}


%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\section{Proofs of Theorems \ref{thm-stab} and \ref{thm-bubble}}\label{sec-proof}

In this section, we prove Theorems \ref{thm-stab} and \ref{thm-bubble} following the ideas of \cite{fz22} and \cite{lz25}, respectively. 

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Proof of Theorem \ref{thm-stab}}

We need the following two lemmas
about the compactness and the orthogonality, respectively.

\begin{lemma}\label{lem-min}
If $p\in[2,n)$, then, for any $\varepsilon\in(0,\infty)$,
there is a positive constant $\delta$ such that,
for any $f\in\dot{W}^{1,p}$ satisfying 
\begin{align*}
{\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p<\delta\mathcal{E}_p(f)^p,
\end{align*}
it holds that
\begin{align*}
\inf_{(c,\lambda,S,x)\in\mathscr{A}}\|\nabla(T_{\lambda,S,x}f-cU)\|_{p}^p
<\varepsilon\mathcal{E}_p(f)^p.
\end{align*}
\end{lemma}

\begin{proof}
Assume the desired conclusion does not true.
Then, by the homogeneity, there is a positive constant $\varepsilon_0$
and a sequence $\{u_k\}_{k\in\mathbb{N}}$ in $\dot{W}^{1,p}$
such that $\mathcal{E}_p(u_k)=1$ for every $k\in\mathbb{N}$
and 
\begin{align}\label{lem-min-e3}
\lim_{k\to\infty}\left[{\mathcal{E}_p(u_k)}^p
-S_{\rm aff}^p\|u_k\|_{p^*}^p\right]=0,
\end{align}
but, for every $k\in\mathbb{N}$,
\begin{align}\label{lem-min-e4}
\inf_{(c,\lambda,S,x)\in\mathscr{A}}\|\nabla(T_{\lambda,S,x}u_k-cU)\|_{p}
\ge\varepsilon_0.
\end{align}
By \eqref{lem-min-e3} and 
Theorem \ref{thm-comp}, we find that,
for each $k\in\mathbb{N}$, there exist $\lambda_k\in(0,\infty)$,
$A_k\in \operatorname{SL}(n)$,
and $x_k\in\mathbb{R}^n$ such that, when $k\to\infty$,
$v_k:=\mathcal{T}^{(p^*)}_{\lambda_k,A_k,x_k}u_k\to u$ in $\dot{W}^{1,p}$ for some 
$u\in\mathcal{M}_{\rm aff}$.
Hence, from \eqref{def-min-equiv}, it follows that $u=c_0 T^{-1}_{\lambda_0,S_0,x_0}U$
for some $(c_0,\lambda_0,S_0,x_0)\in\mathscr{A}$.
Then, applying the polar decomposition and the fact that 
$\|\cdot\|_{\dot{W}^{1,p}}$ is invariant under
$O(n)$ transformations, we conclude that 
\eqref{lem-min-e4} also holds with $u_k$ replaced by 
$v_k$. 
Therefore, using a change of variables and the 
operator norm inequality, we further obtain, for all $k\in\mathbb{N}$,
\begin{align}\label{lem-min-e2}
\varepsilon_0
\le \inf_{(c,\lambda,S,x)\in\mathscr{A}}\|\nabla(T_{\lambda,S,x}v_k-cU)\|_{p}
\le\|\nabla(T_{\lambda_0,S_0,x_0}v_k-u)\|_p
\le 
e^{-\|S_0\|_{\mathcal{L}(\mathbb{R}^n)}}\|\nabla(v_k-u)\|_{p}.
\end{align}
Letting $k\to\infty$, we conclude $\varepsilon_0\le0$. It is impossible.
This then implies that the desired conclusion holds and finishes the proof.
\end{proof}


\begin{lemma}\label{lem-orth}
Let $p\in(1,n)$. Assume that $v_0=c_0T_{\lambda_0,{S_0},x_0}U$ is a positive function
with parameters $(c_0,\lambda_0,S_0,x_0)\in\mathscr{A}$,
and $\mathcal{U}$ is a compact neighborhood of $(c_0,\lambda_0,S_0,x_0)$ in
$\mathscr{A}$. Then there exist a positive constant 
  $\varepsilon_0$ and a positive function $\omega$ in 
  $(0,\infty)$ satisfying $\lim_{t\to0}\omega(t)=0$ such that,
  if $\varepsilon\in(0,\varepsilon_0)$ and 
  $f\in\dot{W}^{1,p}$ satisfies $\|\nabla f-\nabla v_0\|_p<\varepsilon$,
  then there exist parameters $(c_1,\lambda_1,S_1,x_1)$
  in the interior of 
  $\mathcal{U}$ such that $v:=c_1T_{\lambda_1,{S_1},x_1}U$ is positive,
  $\|\nabla f-\nabla v\|_p<\omega(\varepsilon)$
  and, for any $\eta\in T_v\mathcal{M}_{\rm aff}$,
  $f-v\perp\eta$ in $L^2(v^{p^*-2})$.
\end{lemma}

\begin{proof}
This proof is essentially the same as the one given by 
\cite[Lemma 4.1]{fz22}; nevertheless, we outline the necessary details below.
For $f\in\dot{W}^{1,p}$ and any
$w\in\mathcal{M}_{\rm aff}$, define the functional $F_f(w)$ by setting 
\begin{align*}
F_f(w):=\frac{1}{p^*}\int_{\mathbb{R}^n}|w|^{p^*}
-\frac{1}{p^*-1}\int_{\mathbb{R}^n}|w|^{p^*-2}wf.
\end{align*}
If $f=v_0$, then, by H\"older's inequality and the property of 
the function 
\begin{align*}
(0,\infty)\ni s\mapsto\frac1{p^*}s^{p^*}-
\frac{1}{p^*-1}As^{p^*-1}
\end{align*}
for $A\in(0,\infty)$, we find that 
$F_{v_0}(w)$ is uniquely minimized at $w=v_0$.
For general $f$, since the mapping
$\mathscr{A}
\ni (c,\lambda,S,x)\mapsto cT_{\lambda,S,x}(U)$
is continuous from $\mathscr{A}$ to $\dot{W}^{1,p}$, we infer that 
\begin{align*}
\mathcal{F}_f:\ \mathscr{A}\ni (c,\lambda,S,x)\mapsto F_f(cT_{\lambda,S,x}U)
\end{align*}
is also continuous.
By this continuity and the compactness of $\mathcal{U}$, 
we conclude that
there exists $(c_1,\lambda_1,S_1,x_1)\in\mathcal{U}$ such that 
\begin{align}\label{lem-min-e1}
F_f(c_1T_{\lambda_1,{S_1},x_1}U)=
\mathcal{F}_f(c_1,\lambda_1,S_1,x_1)=
\inf_{(c,\lambda,S,x)\in\mathcal{U}}\mathcal{F}_f(c,\lambda,S,x)=
\inf_{(c,\lambda,S,x)\in\mathcal{U}}F_f(cT_{\lambda,S,x}U)
\end{align}
is attained. Moreover, if $f$ is close to $v_0$,
then, from a standard compactness argument and the observations that 
$F_{v_0}(w)$ is uniquely minimized at $w=v_0$,
$f\mapsto F_{f}(w)$ is continuous in $\dot{W}^{1,p}$,
and $(c,\lambda,S,x)\in\mathscr{A}$ are uniquely determined by the corresponding extremal
function, 
we deduce that the parameters $(c_1,\lambda_1,S_1,x_1)$
such that the infimum in \eqref{lem-min-e1} is close 
to $(c_0,\lambda_0,S_0,x_0)$ in $\mathscr{A}$.
Hence, if $\varepsilon$ is sufficiently small, then,
for every $f\in\dot{W}^{1,p}$ such that $\|\nabla f-\nabla v_0\|_p<\varepsilon$,
$v:=c_1T_{\lambda_1,S_1,x_1}U$ is also positive and is close to $f$ in $\dot{W}^{1,p}$.
This shows the existence of the desired $\varepsilon_0$ and 
$\omega$. Furthermore, if $f$ is close to $v_0$,
then $(c_1,\lambda_1,S_1,x_1)$ is in the interior of $\mathcal{U}$;
thus, for every direction vector $\mathbf{e}$ in $\mathscr{A}$,
\begin{align*}
\left.\frac{d}{dt}\right|_{t=0}
\mathcal{F}_f\big((c_1,\lambda_1,S_1,x_1)+t\mathbf{e}\big)=0.
\end{align*}
This, together with the definition of 
the tangent space given in \eqref{eq-tangent},
further implies that, 
for any $\eta\in T_v\mathcal{M}_{\rm aff}$,
$\int_{\mathbb{R}^n}v^{p^*-2}(v-f)\eta=0$,
which gives the desired orthogonality.
We then complete the present proof.
\end{proof}

\begin{proof}[Proof of Theorem \ref{thm-stab}]
We actually prove the improved version \eqref{eq-stab-e}.
In the present proof, 
unless otherwise specified, the implicit positive constants 
depend only on $n$ and $p$.
By \cite[Theorem 1.2]{hl16} [see also \eqref{eq-p}] and 
the affine invariance of \eqref{eq-stab}, we 
can assume $\mathcal{E}_p(f)=1$ and 
$\|\nabla f\|_p\lesssim1$.
Then, applying H\"older's inequality, we obtain 
\begin{align*}
&\inf_{(c,\lambda,S,x)\in\mathscr{A}}\left\{\|\nabla(T_{\lambda,S,x}f-cU)\|_{p}^p
+\int_{\mathbb{R}^n}
|c\nabla U|^{p-2}|\nabla(T_{\lambda,S,x}f-cU)|^2\right\}\\
&\quad\le\|\nabla(f-cU)\|_{p}^p
+\int_{\mathbb{R}^n}
|\nabla U|^{p-2}|\nabla(f-cU)|^2\\
&\quad\le \|\nabla(f-cU)\|_{p}^p+\|\nabla U\|_p^{p-2}
\|\nabla(f-cU)\|_{p}^2\lesssim1.
\end{align*}
Therefore, to show \eqref{eq-stab-e}, we only need to  
consider the case where ${\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p$ is sufficiently small.
Then, from Lemma \ref{lem-min}, 
we deduce that there exist sufficiently small $\varepsilon$
and affine transformation parameters $(c_0,\lambda_0,S_0,x_0)\in\mathscr{A}$ such that 
$\|\nabla(T_{\lambda_0,S_0,x_0}f-c_0U)\|_p<\varepsilon$.
By replacing $f$ with $-f$ if necessary, we may assume that $c_0 > 0$.
Moreover, from the classical Sobolev inequality 
\eqref{eq-Sob} and $\|T_{\lambda_{0},S_0,x_0}f\|_{p^*}=\|U\|_{p^*}=1$,
we deduce $c_0\sim 1$ and hence 
$$\left\|\nabla\left(c^{-1}_0T_{\lambda_0,S_0,x_0}f-U\right)\right\|_p\lesssim\varepsilon.$$
By this and Lemma \ref{lem-orth} for the 
parameters $(1,1,{\bf 0},{\bf 0})\in\mathscr{A}$ and their fixed 
small compact neighborhood $\mathcal{U}\subset\mathscr{A}$, 
there exist a positive function $\omega$ in $(0,\infty)$ satisfying 
$\lim_{t\to0}\omega(t)=0$ and a positive function $v=c_1T_{\lambda_1,S_1,x_1}U$ with 
$(c_1,\lambda_1,S_1,x_1)\in\mathcal{U}$
such that
\begin{align*}
\left\|\nabla\left(c_0^{-1}T_{\lambda_{0},S_0,x_0}f-v\right)
\right\|_p\lesssim\omega(\varepsilon)\quad\text{and}
\quad c_{0}^{-1}T_{\lambda_{0},S_0,x_0}f-v\perp T_v\mathcal{M}_{\rm aff}
\quad \text{in}\quad L^2(v^{p^*-2}).
\end{align*}
Since $(c_1,\lambda_1,S_1,x_1)$ is close to 
$(1,1,{\bf 0},{\bf 0})$ in $\mathscr{A}$, 
similarly to the last inequality 
of \eqref{lem-min-e2}, we have
\begin{align*}
\left\|\nabla\left(c_0^{-1}c_{1}^{-1}T_{\lambda_1, S_1,x_1}^{-1}
T_{\lambda_{0},S_0,x_0}f-U\right)\right\|_p
&\lesssim\left\|\nabla\left(c_{0}^{-1}
T_{\lambda_{0},S_0,x_0}f-c_1T_{\lambda_1,S_1,x_1}U\right)\right\|_p
\lesssim\omega(\varepsilon),
\end{align*}
Moreover, applying a change of variables, one can easily see that 
the orthogonality is invariant under the 
affine transformation; that is, 
\begin{align*}
c_0^{-1}c_{1}^{-1}T_{\lambda_1, S_1,x_1}^{-1}
T_{\lambda_{0},S_0,x_0}f-U\perp T_U\mathcal{M}_{\rm aff}
\quad \text{in}\quad L^2(U^{p^*-2}).
\end{align*}

Consequently, by the affine invariance of 
\eqref{eq-stab-e} and the behavior of $\omega$ near $0$, we only need to consider the 
case where $f=U+\varphi$, 
where $\varphi\perp T_{U}\mathcal{M}$ and 
$\|\nabla\varphi\|_p<\varepsilon<\varepsilon_0$ for fixed and sufficiently 
small $\varepsilon_0,\varepsilon$. Then, by Proposition \ref{prop-expan-local} and 
Remark \ref{rem-3.4}, there exist two positive constants 
$C_{n,p}$ and $\widetilde{C}_{n,p}$, depending only on $n$ and $p$,
such that 
\begin{align}\label{thm-stab-e1}
\mathcal{E}_p(f)^p
&\ge \mathcal{E}_p(U)^p
+p|\mathbb{S}_{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\left[L^{(1)}_\xi(U,\varphi)
+\frac{1-C_{n,p}\varepsilon_0}{2}L^{(2)}_\xi(U,\varphi)\right]
\,d\xi\notag\\
&\qquad-\frac{p(n+p)}{2}|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}\operatorname{Var}(U,\varphi)
+\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_{p}^p,
\end{align}
where $L^{(1)}$ and $L^{(2)}$ are as in
\eqref{def-L1} and \eqref{def-L2}, respectively,
with $f$ replaced by $U$.
On the other hand, from the expansion for 
$|a+b|^{p^*}$ obtained in \cite[Lemma 3.2]{fn19} (see also 
\cite[Lemma 2.4]{fz22}):
for any $\varepsilon_0\in(0,\infty)$, there exists a positive constant 
$C_{p,\varepsilon_0}$, depending only on $p$ and $\varepsilon_0$,
such that, for all $a,b\in\mathbb{R}$ with $a\ne 0$,
\begin{align*}
    |a + b|^{p^*} \leq & |a|^{p^*} + p^*|a|^{p^*-2}ab
     + \left( \frac{p^*(p^* - 1)}{2} + \varepsilon_0 \right)
     |a|^{p^*-2}|b|^2 + C_{p,\varepsilon_0}|b|^{p^*};
\end{align*}
therefore, we have 
\begin{align*}
\|f\|_{p^*}^{p^*}
&\le\int_{\mathbb{R}^n}\left\{U^{p^*}
+p^{*}U^{p^*-1}\varphi
+\left[\frac{p^*(p^*-1)}{2}+\varepsilon_0\right]U^{p^*-2}\varphi^2
+C_{p,\varepsilon_0}|\varphi|^{p^*}\right\}.
\end{align*}
Combining this and the concavity of $t\mapsto t^{\frac{p}{p^*}}$
and $\|U\|_{p^*}=1$, we conclude that 
\begin{align*}
\|f\|_{p^*}^p
\le 1+p
\int_{\mathbb{R}^n}\left\{U^{p^*-1}\varphi
+\left[\frac{(p^*-1)}{2}+\frac1{p^*}\varepsilon_0\right]U^{p^*-2}\varphi^2
+\widetilde{C}_{p,\varepsilon_0}|\varphi|^{p^*}\right\},
\end{align*}
where the positive constant $\widetilde{C}_{p,\varepsilon_0}$ depends only on $p$ and $\varepsilon_0$.
This, together with \eqref{thm-stab-e1} and $\mathcal{E}_p(U)=S_{\rm aff}$, 
further implies
\begin{align}\label{thm-stab-e3}
&{\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p\notag\\
&\quad\gtrsim p\left[|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}L^{(1)}_\xi(U,\varphi)-
\int_{\mathbb{R}^n}U^{p^*-1}\varphi\right]\notag\\
&\qquad+\frac p2
\left\{|\mathbb{S}^{n-1}|^{-1-\frac pn}(1-C_{n,p}\varepsilon_0)
\int_{\mathbb{S}^{n-1}}L^{(2)}_\xi(U,\varphi)-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(U)^{-p}
\operatorname{Var}(U,\varphi)\right.\notag\\
&\left.\qquad-\left[(p^*-1)S_{\rm aff}^p+\frac{S_{\rm aff}^p}{p^*}
{\varepsilon_0}\right]\int_{\mathbb{R}^n}
U^{p^*-2}\varphi^2
\right\}\notag\\
&\qquad+\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^p\widetilde{C}_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}\notag\\
&\quad=:p{\rm I}_1+\frac p2{\rm I}_2(\varepsilon_0)+
\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^p\widetilde{C}_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}.
\end{align}
By the fact that $U$ is an extremal function, it holds that 
\begin{align*}
\left.\frac{d}{dt}\right|_{t=0}
\left[\mathcal{E}_p(U+t\varphi)^p
-S_{\rm aff}^p\|U+t\varphi\|_{p^*}^p\right]=0.
\end{align*}
Applying this, the facts that 
$\mathcal{E}_p(U)=S_{\rm aff}$, $\|U\|_{p^*}=1$,
and $\mathcal{E}_p(U)=|\mathbb{S}^{n-1}|^{-\frac1n}\|\nabla_\xi U\|_p$
for all $\xi\in\mathbb{S}^{n-1}$ (which is a direct consequence of 
$U$ being radial), we can find that 
the term ${\rm I}_1$ defined in \eqref{thm-stab-e3} vanishes.
Moreover, from the affine spectral gap inequality 
Theorem \ref{thm-spectral-wei} and Remark \ref{rem-spectral}, it follows that there exists 
$\gamma\in(0,\infty)$, depending only on $n$ and $p$, such that, for 
sufficiently small $\varepsilon_0$,
\begin{align*}
{\rm I}_2(\varepsilon_0)
\ge \left(\gamma-\frac{S_{\rm aff}^p}{p^*}\varepsilon_0\right)
\int_{\mathbb{R}^n}U^{p^*-2}\varphi^2
+\varepsilon_0\int_{\mathbb{R}^n}|\nabla U|^{p-2}
|\varphi|^2.
\end{align*}
Hence, when $\varepsilon_0$ is sufficiently small, it also holds 
$
{\rm I}_2(\varepsilon_0)\ge \varepsilon_0\int_{\mathbb{R}^n}|\nabla U|^{p-2}
|\varphi|^2$.
On the other hand, 
when $\varepsilon$ is sufficiently small, applying 
\eqref{eq-Sob} and the fact the exponent $p^*>p$, we find that 
\begin{align}\label{thm-stab-e4}
\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^pC_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}
\ge \widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^{p}S_{\rm Sob}^{-p^*}C_{p,\varepsilon_0}
\|\nabla\varphi\|_{p^*}^{p^*}\gtrsim\|\nabla\varphi\|_p^p.
\end{align}
These further implies 
\begin{align*}
{\mathcal{E}_p(f)}^p
-S_{\rm aff}^p\|f\|_{p^*}^p&\gtrsim\|\nabla\varphi\|_p^p
+\int_{\mathbb{R}^n}|\nabla U|^{p-2}
|\varphi|^2\\
&\ge \inf_{(c,\lambda,S,x)\in\mathscr{A}}\left\{\|\nabla(T_{\lambda,S,x}f-cU)\|_{p}^p
+\int_{\mathbb{R}^n}
|c\nabla U|^{p-2}|\nabla(T_{\lambda,S,x}f-cU)|^2\right\}.
\end{align*}
This shows \eqref{eq-stab-e} and finishes the proof of Theorem 
\ref{thm-stab}.
\end{proof}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\subsection{Proof of Theorem \ref{thm-bubble}}

\begin{proof}[Proof of Theorem \ref{thm-bubble}]
We actually prove the improved version \eqref{thm-bubble-e6}.
In the present proof, 
unless otherwise specified, the implicit positive constants 
depend only on $n$ and $p$.
Since we only study the case where
$d(f,\mathcal{M}_{\rm aff}^{(1)})$ is 
sufficiently small, to prove \eqref{thm-bubble-e6},
we can also assume that $\delta_{\rm aff}$
is small. Assume that $d(f,\mathcal{M}_{\rm aff}^{(1)})<\varepsilon$
for sufficiently small $\varepsilon$.
Then there exist 
$\lambda_0\in(0,\infty)$,
symmetric matrix $S_0$ 
satisfying $\operatorname{Tr}S=0$,
and $x_0\in\mathbb{R}^n$ such that 
$\|\nabla(T_{\lambda_0,S_0,x_0}f-U)\|_p<\varepsilon$.
By this and Lemma \ref{lem-orth} for the 
parameters $(1,1,{\bf 0},{\bf 0})\in\mathscr{A}$ and their fixed 
small compact neighborhood $\mathcal{U}\subset\mathscr{A}$, 
there exist a positive function $\omega$ in $(0,\infty)$ satisfying 
$\lim_{t\to0}\omega(t)=0$ and a positive function $v=c_1T_{\lambda_1,{S_1},x_1}U$ with 
$(c_1,\lambda_1,S_1,x_1)\in\mathcal{U}$
such that
\begin{align*}
\left\|\nabla\left(T_{\lambda_{0},S_0,x_0}f-v\right)
\right\|_p\le\omega(\varepsilon)\quad\text{and}
\quad T_{\lambda_{0},S_0,x_0}f-v\perp T_v\mathcal{M}_{\rm aff}
\quad \text{in}\quad L^2(v^{p^*-2}).
\end{align*}
Let $g:=T_{\lambda_1,{S_1},x_1}^{-1}T_{\lambda_{0},S_0,x_0}f$
and $\varphi:=g-c_1U$.
Since $(c_1,\lambda_1,S_1,x_1)$ is close to 
$(1,1,{\bf 0},{\bf 0})$ in $\mathscr{A}$, it holds that 
\begin{align*}
\|\nabla\varphi\|_p
=\left\|\nabla\left(T_{\lambda_1,{S_1},x_1}^{-1}
T_{\lambda_{0},S_0,x_0}f-c_1U\right)\right\|_p
\le |\lambda_1|\|e^{-S_1}\|_{\mathcal{L}(\mathbb{R}^n)}
\left\|\nabla\left(T_{\lambda_{0},S_0,x_0}f-v\right)
\right\|_p\le C_{n}\omega(\varepsilon),
\end{align*}
where $C_{n}$ is a positive constant 
depending only on $n$, and in the second inequality we used the operator norm inequality for $e^{-S_1}$.
Moreover, by the behavior of $\omega$ near $0$,
we know that, for any sufficiently small $\varepsilon_0\in(0,\infty)$,
there exists $\varepsilon_1\in(0,1)$ such that, for every 
$\varepsilon\in(0,\varepsilon_1)$, 
$C_n\omega(\varepsilon)<\varepsilon_0$. In what follows,
we fix sufficiently small $\varepsilon,\varepsilon_0,\varepsilon_1$ that satisfies this property
at the moment.

Then, by Proposition \ref{prop-expan-lap} and 
Remark \ref{rem-3.3}, there exist two positive constants 
$C_{n,p}$ and $\widetilde{C}_{n,p}$, depending only on $n$ and $p$,
such that 
\begin{align*}
-\int_{\mathbb{R}^n}
\Delta_p^{\rm aff}(g)\varphi
&\ge 
-\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(c_1U)\varphi
+(1-C_{n,p}\varepsilon_0)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(c_1U,\varphi)\,d\xi\notag\\
&\qquad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal E_p(c_1U)^{-p}
\operatorname{Var}(c_1U,\varphi)
+\widetilde C_{n,p}\varepsilon_0\|\nabla \varphi\|_{p}^p.
\end{align*}
where $\operatorname{Var}$ and $\widetilde{L}^{(2)}$ are as, 
respectively, in \eqref{def-var} and \eqref{eq-L2w}.
On the other hand, from the expansion for 
$|a+b|^{p^*-2}(a+b)b$ obtained in \cite[Lemma 2.2]{lz25}:
for any $\varepsilon_0\in(0,\infty)$, there exists a positive constant 
$C_{p,\varepsilon_0}$, depending only on $p$ and $\varepsilon_0$,
such that, for all $a,b\in\mathbb{R}$ with $a\ne 0$,
\begin{align*}
    |a + b|^{p^*-2}(a+b)b \leq & |a|^{p^*-2}ab
    +(p^*-1+\varepsilon_0)|a|^{p^*-2}b^2
    + C_{p,\varepsilon_0}|b|^{p^*}.
\end{align*}
Abbreviating
\begin{align*}
P(g):=-\Delta_{p}^{\rm aff}(g)-S_{\rm aff}^p|g|^{p^*-2}g,
\end{align*}
we infer that
\begin{align*}
\int_{\mathbb{R}^n}
P(g)\varphi
&=-\int_{\mathbb{R}^n}
\Delta_p^{\rm aff}(g)\varphi
-S_{\rm aff}^p\int_{\mathbb{R}^n}|g|^{p^*-2}g\varphi\\
&\ge \int_{\mathbb{R}^n}
P(c_1U)\varphi
+\left\{(1-C_{n,p}\varepsilon_0)|\mathbb{S}^{n-1}|^{-1-\frac pn}
\int_{\mathbb{S}^{n-1}}\widetilde{L}^{(2)}_\xi(c_1U,\varphi)\,d\xi\right.\notag\\
&\quad-(n+p)|\mathbb{S}^{n-1}|^{-1-\frac{2p}{n}}
\mathcal{E}_p(c_1U)^{-p}
\operatorname{Var}(c_1U,\varphi)\notag\\
&\left.\quad-\left[(p^*-1)S_{\rm aff}^p-{S_{\rm aff}^p}{\varepsilon}\right]\int_{\mathbb{R}^n}
|c_1U|^{p^*-2}\varphi^2
\right\}\notag\\
&\quad+\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^pC_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}\notag\\
&=:{\rm I}_1+{\rm I}_2(\varepsilon_0)+
\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^pC_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}.
\end{align*}
By the fact $-\Delta_p^{\rm aff}U=S_{\rm aff}^pU^{p^*-1}$,
we obtain 
\begin{align*}
P(c_1U)=-c_1^{p-1}\Delta_p^{\rm aff}(U)-S_{\rm aff}c_1^{p^*-1}U^{p^*-1}
=S_{\rm aff}^p\left(c_1^{p-1}-c_1^{p^*-1}\right)U^{p^*-1}.
\end{align*}
Applying a change of variables, one can easily see that 
the orthogonality is invariant under the 
affine transformation; that is, 
$\varphi\perp T_U\mathcal{M}_{\rm aff}$
in $L^2(U^{p^*-2})$.
This further implies that $${\rm I}_1=\int_{\mathbb{R}^n}P(c_1U)\varphi=0.$$
Moreover, from the affine spectral gap inequality 
Theorem \ref{thm-spectral-wei} and the dilation properties 
of $\widetilde{L}^{(2)}$ [see \eqref{eq-L-dila}], it follows that there exists 
$\gamma\in(0,\infty)$, depending only on $n$ and $p$, such that
\begin{align*}
{\rm I}_2(\varepsilon_0)
\ge \left[c_1\gamma+(c_1^{p-2}-1)(p^*-1)
S_{\rm aff}^p-S_{\rm aff}^p
\varepsilon_0\right]
\int_{\mathbb{R}^n}U^{p^*-2}\varphi^2.
\end{align*}
Hence, when $\varepsilon_0>0$ and 
the compact neighborhood $\mathcal{U}\subset\mathscr{A}$ are both sufficiently small, 
it also holds 
$
{\rm I}_2(\varepsilon_0)\ge 0$.
Furthermore, 
when $\varepsilon_1>0$ is sufficiently small, applying 
\eqref{eq-Sob} and \eqref{thm-stab-e4}, we find that 
$\widetilde{C}_{n,p}\varepsilon_0\|\nabla \varphi\|_p^p
-S_{\rm aff}^pC_{p,\varepsilon_0}\|\varphi\|_{p^*}^{p^*}
\gtrsim\|\nabla\varphi\|_p^p.$
This, combined with \eqref{eq-P}, further implies 
\begin{align}\label{thm-bubble-e4a}
\|\nabla\varphi\|_p^p&\lesssim 
\int_{\mathbb{R}^n}
P(g)\varphi
=\int_{\mathbb{R}^n}P\left(f\circ e^{S_1-S_0}\right)
T_{\lambda_0,{\bf 0},x_1}T^{-1}_{\lambda_1,{\bf 0},x_1}\varphi\notag\\
&\le \left\|P\left(f\circ e^{S_1-S_0}\right)\right\|_{\dot{W}^{-1,p'}}
\|\nabla\varphi\|_{p}\le 
\delta_{\rm aff}(f)\|\nabla\varphi\|_{p},
\end{align}
where we used the fact that if the trace-zero symmetric matrix
in the affine transformation is zero, then 
this transformation preserves the $\dot{W}^{1,p}$ norm.
Then we have 
\begin{align}\label{thm-bubble-e5}
d(f,\mathcal{M}_{\rm aff}^{(1)})^{p-1}
\le \|\nabla (g-U)\|_p^{p-1}
\lesssim\|\nabla \varphi\|_p^{p-1}
+|c_1-1|^{p-1}\lesssim
\delta_{\rm aff}(f)+|c_1-1|^{p-1}.
\end{align}

Therefore, we still need to estimate $|c_1-1|$.
Note that 
\begin{align*}
\int_{\mathbb{R}^{n}}
P(c_1U)(c_1U)
&=\int_{\mathbb{R}^{n}}
P(g)g
-\left[\int_{\mathbb{R}^{n}}
P(g)g
-\int_{\mathbb{R}^{n}}
P(c_1U)(c_1U)\right]\\
&=:{\rm I}_3-{\rm I}_4.
\end{align*}
By $\|\nabla(T_{\lambda_0,S_0,x_0}f-U)\|_p<\varepsilon$,
the fact that $(c_1,\lambda_1,S_1,x_1)$ is close to 
$(1,1,{\bf 0},{\bf 0})$, and a similar change of variables
argument used in \eqref{thm-bubble-e4a}, we conclude that 
\begin{align*}
|{\rm I}_3|\le \left\|P\left(f\circ e^{S_1-S_0}\right)\right\|_{\dot{W}^{-1,p'}}
\|\nabla g\|_{p}\lesssim\delta_{\rm aff}(f).
\end{align*}
On the other hand, 
\begin{align*}
{\rm I}_4
=\mathcal{E}_{p}(g)^p-\mathcal{E}_p(c_1U)^p
-S_{\rm aff}^p
\left(\|g\|_{p^*}^{p^*}-\|c_1U\|_{p^*}^{p^*}\right).
\end{align*}
From the fact that $\|\nabla\varphi\|_p$ is sufficiently small and 
Proposition \ref{prop-expan-E}, we deduce that
\begin{align*}
\left|\mathcal{E}_p(g)^p
-\mathcal{E}_p(c_1U)^p
+p\int_{\mathbb{R}^n}\Delta_p^{\rm aff}(c_1U)\varphi\right|
\lesssim\|\nabla c_1 U\|_p^{p-2}\|\nabla \varphi\|_{p}^2
\lesssim\|\nabla \varphi\|_{p}^2.
\end{align*}
Moreover, by  $-\Delta_p^{\rm aff}U=S_{\rm aff}^pU^{p^*-1}$ and 
$\varphi\perp T_U\mathcal{M}_{\rm aff}$
in $L^2(U^{p^*-2})$,
we have 
\begin{align*}
\int_{\mathbb{R}^n}\Delta_p^{\rm aff}U\varphi
=\int_{\mathbb{R}^n}U^{p^*-1}\varphi=0.
\end{align*}
Then 
\begin{align}\label{thm-bubble-e2}
\left|\mathcal{E}_p(g)^p
-\mathcal{E}_p(c_1U)^p\right|
\lesssim\|\nabla c_1 U\|_p^{p-2}\|\nabla \varphi\|_{p}^2
\lesssim\|\nabla \varphi\|_{p}^2.
\end{align}
Similarly, we can get 
\begin{align*}
\left|\|g\|_{p^*}^{p^*}-\|c_1U\|_{p^*}^{p^*}\right|
=\left|\int_{\mathbb{R}^n}
|g|^{p^*}-|c_1U|^{p^*}-|c_1U|^{p^*-2}(c_1U)\varphi\right|
\lesssim\|U\|_{p^*}^{p^*-2}\|\varphi\|_{p^*}^2
+\|\varphi\|_{p^*}^p
\lesssim\|\nabla \varphi\|_p^2.
\end{align*}
Combining this and \eqref{thm-bubble-e2}, we 
obtain 
$|{\rm I}_4|\lesssim\|\nabla\varphi\|_p^2$.
This, together with the estimate of ${\rm I}_3$, 
further implies 
\begin{align}\label{thm-bubble-e3}
\left|\int_{\mathbb{R}^{n}}
P(c_1U)(c_1U)\right|
\lesssim\delta_{\rm aff}(f)
+\|\nabla\varphi\|_p^2.
\end{align}
On the other hand, 
note that $\mathcal{E}_p(U)=S_{\rm aff}$
and $\|U\|_{p^*}=1$.
Consequently, 
\begin{align*}
\int_{\mathbb{R}^{n}}
P(c_1U)(c_1U)
=\mathcal{E}_{p}(c_1U)^p-S_{\rm aff}^p\|c_1U\|_{p^*}^{p^*}
=(c_1^{p}-c_1^{p^*})
S_{\rm aff}^p.
\end{align*}
From $\lim_{c\to 1}\frac{c^p-c^{p^*}}{1-c}
=p^*-p>0$
and $c_1$ is close $1$, 
it follows that $|c_1^{p}-c_1^{p^*}|\gtrsim|c_1-1|$.
Hence, by \eqref{thm-bubble-e3}, 
we find that 
\begin{align}\label{thm-bubble-e4}
|c_1-1|\lesssim\delta_{\rm aff}(f)
+\|\nabla\varphi\|_p^2,
\end{align}
which, combined with \eqref{thm-bubble-e4a},
implies that, if $\delta_{\rm aff}(f)$ 
is sufficiently small, then 
$$
|c_1-1|^{p-1}\lesssim\delta_{\rm aff}(f)^{p-1}
+\|\nabla\varphi\|_p^{2p-2}
\lesssim\delta_{\rm aff}(f)
+\|\nabla\varphi\|_p^{p-1}
\lesssim \delta_{\rm aff}(f).
$$
Using this and \eqref{thm-bubble-e5},
we obtain \eqref{thm-bubble-e6}.
This finishes the proof of Theorem 
\ref{thm-bubble}.
\end{proof}

\section{Sharpness of the exponents in Theorems \ref{thm-stab} and \ref{thm-bubble}}\label{Sec-sharp}

In this section, we are devoted to proving 
the sharpness of the decreasing exponents in Theorems 
\ref{thm-stab} and \ref{thm-bubble}.
The argument follows from the same
constructions as in \cite[Remark 1.2]{fz22} and 
\cite[Section 5.1]{fp24}. 
Recall that $U:=c_0(1+|x|^{p'})^{\frac{p-n}{p}}$
satisfy $\|U\|_{p^*}=1$ and 
fix $\varphi\in C_{\rm c}^\infty(B({\bf0},1))$ be nonzero.

\subsection{Sharpness of the exponents on $\dot{W}^{1,p}$ norm}

We first consider the exponents on the $\dot{W}^{1,p}$ norm. 
Let $\mathbf{e}_1:=(1,0,\ldots,0)\in\mathbb{R}^n$.
Fix $\{R_\varepsilon\}_{\varepsilon\in(1,\infty)}$
in $(0,\infty)$.
For every $\varepsilon\in(0,1)$, 
define $\varphi_\varepsilon(\cdot):=\varphi(\cdot-R_\varepsilon \mathbf{e}_1)$
and $f_\varepsilon:=U+\varepsilon\varphi_\varepsilon$.
We first provide the following upper estimate for the deficits.

\begin{lemma}
For every $\varepsilon\in(0,1)$, 
choose $R_\varepsilon\in(1,\infty)$ such that
\begin{align}\label{eq:small-tail}
\sup_{|x-R_\varepsilon \mathbf{e}_1|<1}
\left(
|U(x)|+|\nabla U(x)|+|\nabla^2U(x)|
\right)
\le \varepsilon.
\end{align}
Then there exist positive constants $c$ and $C$, depending only on $n$, $p$, $U$, and
$\varphi$, such that, for every $\varepsilon\in(0,c)$,
$${\mathcal{E}_p(f_\varepsilon)}^p
-S_{\rm aff}^p\|f_\varepsilon\|_{p^*}^p\le C\varepsilon^p\quad
\text{and}\quad
\delta_{\rm aff}(f_\varepsilon)
\le C\varepsilon^{p-1}.$$
\end{lemma}

\begin{proof}
In the present proof, 
unless otherwise specified, the implicit positive constants 
depend only on $n$, $p$, $U$, and $\varphi$. 
Due to \eqref{eq-holder},
the first assertion is a direct consequence of 
\cite[Remark 1.2]{fz22}. Hence, we only prove the estimate on 
$\delta_{\rm aff}$. 
Let $q:=(p^*)'$.
Then, by the Sobolev inequality \eqref{eq-Sob}
and duality, we obtain, for any $R\in L^q$,
$\|R\|_{\dot W^{-1,p'}}
\lesssim\|R\|_q.$
For every symmetric matrix $S$ satisfying 
$\operatorname{Tr}S=0$,
by this and a change of variables combined with 
$\det e^S=1$, we conclude that, for every $\varepsilon\in(0,1)$,
\begin{align*}
\left\|P(f_\varepsilon\circ e^S)\right\|_{\dot W^{-1,p'}}
&\lesssim\left\|P(f_\varepsilon\circ e^S)\right\|_q
=\left\|P(f_\varepsilon)\circ e^S\right\|_q
=\left\|P(f_\varepsilon)\right\|_q.
\end{align*}
Taking the supremum over all such $S$, we obtain
\begin{align}
\delta_{\rm aff}(f_\varepsilon)
\lesssim\left\|P(f_\varepsilon)\right\|_q.
\label{eq:delta-Lq}
\end{align}
Therefore, to complete the present proof, we only need to prove
$\|P(f_\varepsilon)\|_q\lesssim\varepsilon^{p-1}.$

Now, fix $\varepsilon\in(0,1)$.
Since $U$ is radial, $\|\nabla_\xi U\|_p$ is independent of $\xi\in\mathbb{S}^{n-1}$.
Define $a:=\|\nabla_\xi U\|_p^p$ for any 
$\xi\in\mathbb{S}^{n-1}$.  
Then, from the argument similar to that used in \eqref{prop-expan-E-e3} and 
the fact that $\operatorname{supp}(\varphi_\varepsilon)\subset
B(R_\varepsilon\mathbf{e}_1,1)$, we deduce that, for every 
$\xi\in\mathbb{S}^{n-1}$,
\begin{align}\label{eq:directional-p-bound}
\left|\|\nabla_\xi f_\varepsilon\|_p^p-a\right|
&\lesssim
\int_{B(R_\varepsilon \mathbf{e}_1,1)}
\left[\varepsilon
|\nabla_\xi U|^{p-1}|\nabla_\xi\varphi_\varepsilon|
+
\varepsilon^p|\nabla_\xi\varphi_\varepsilon|^p
\right]
\notag\\
&\lesssim\varepsilon^p\quad\text{by \eqref{eq:small-tail}}.
\end{align}
 Hence, if $\varepsilon$ is sufficiently small, then,
 for all $\xi\in\mathbb{S}^{n-1}$,
$\frac{a}{2}
\le \|\nabla_\xi f_\varepsilon\|_p^p
\le 2a.$
Applying this, the mean value theorems, respectively,
for $$t\mapsto (a+t\rho)^{-1-\frac np}\quad\text{and}\quad
t\mapsto\left(\int_{\mathbb{S}^{n-1}}\left[a+tV(\xi)\,d\xi\right]^{-\frac np}\right)^{1-\frac pn},$$
and \eqref{eq:directional-p-bound},
we find that, for all $\xi\in\mathbb{S}^{n-1}$,
\begin{align*}
\left|\|\nabla_\xi f_\varepsilon\|_p^{-n-p}-
\|\nabla_\xi U\|_p^{-n-p}\right|
\lesssim a^{-2-\frac np}
\sup_{\xi\in\mathbb{S}^{n-1}}
\left|\|\nabla_\xi f_\varepsilon\|_p^p-a\right|
\lesssim\varepsilon^p
\end{align*}
and
\begin{align*}
\left|{\mathcal E}_p(f_\varepsilon)-{\mathcal E}_p(U)\right|
\lesssim a^{-\frac np}
\sup_{\xi\in\mathbb{S}^{n-1}}
\left|\|\nabla_\xi f_\varepsilon\|_p^p-a\right|
\lesssim\varepsilon^p.
\end{align*}
These further imply 
\begin{align}\label{eq:coefficient-bound}
\left|
{\mathcal E}_p(f_\varepsilon)^{n+p}
\|\nabla_\xi f_\varepsilon\|_p^{-n-p}
-
{\mathcal E}_p(U)^{n+p}\|\nabla_\xi U\|_p^{-n-p}
\right|
\lesssim\varepsilon^p
\end{align}
for all $\xi\in{\mathbb S}^{n-1}$.

Now, we turn to estimate $P(f_\varepsilon)$.
Since $P(U)=0$, we infer that 
\begin{align}\label{eq:split-P}
P(f_\varepsilon)
=-
\left(\Delta_p^{\rm aff}f_\varepsilon-\Delta_p^{\rm aff}U\right)
-S_{\rm aff}^p
\left[
|f_\varepsilon|^{p^*-2}f_\varepsilon-U^{p^*-1}
\right].
\end{align}
We first estimate the affine Laplacian term. By the definition of $\Delta_p^{\rm aff}$,
we obtain
\begin{align*}
&\Delta_p^{\rm aff}f_\varepsilon-\Delta_p^{\rm aff}U
\notag\\
&\quad=\operatorname{div}
\left(
\int_{{\mathbb S}^{n-1}}
{\mathcal E}_p(U)^{n+p}\|\nabla_\xi U\|_p^{-n-p}
\left[
|\nabla_\xi f_\varepsilon|^{p-2}\nabla_\xi f_\varepsilon
-
|\nabla_\xi U|^{p-2}\nabla_\xi U
\right]\xi\,d\xi
\right)
\notag\\
&\qquad+\operatorname{div}
\left(
\int_{{\mathbb S}^{n-1}}
\left[
{\mathcal E}_p(f_\varepsilon)^{n+p}
\|\nabla_\xi f_\varepsilon\|_p^{-n-p}
-
{\mathcal E}_p(U)^{n+p}\|\nabla_\xi U\|_p^{-n-p}
\right]
|\nabla_\xi f_\varepsilon|^{p-2}\nabla_\xi f_\varepsilon\,\xi\,d\xi
\right)\notag\\
&\quad=:I_\varepsilon+II_\varepsilon.
\end{align*}
We first deal with $I_\varepsilon$.
By the support assumption, 
$I_\varepsilon$ vanishes outside $B(R_\varepsilon \mathbf{e}_1,1)$. 
On the other hand, using 
\eqref{eq:small-tail} and the smoothness of $\varphi$,
we obtain, for any $x\in B(R_\varepsilon \mathbf{e}_1,1)$,
\begin{align*}
&\left|
\nabla
\left[
|\nabla_\xi f_\varepsilon(x)|^{p-2}\nabla_\xi f_\varepsilon(x)
-
|\nabla_\xi U(x)|^{p-2}\nabla_\xi U(x)
\right]
\right|
\\
&\quad=(p-1)
\left|
|\nabla_\xi f_\varepsilon(x)|^{p-2}\nabla\nabla_\xi f_\varepsilon(x)
-
|\nabla_\xi U(x)|^{p-2}\nabla\nabla_\xi U(x)
\right|
\\
&\quad
\lesssim|\nabla_\xi f_\varepsilon(x)|^{p-2}
|\varepsilon\nabla\nabla_\xi\varphi_\varepsilon|
+
\left||\nabla_\xi f_\varepsilon(x)|^{p-2}-|\nabla_\xi U(x)|^{p-2}\right|
|\nabla\nabla_\xi U(x)|
\lesssim\varepsilon^{p-1}.
\end{align*}
Thus, it holds that 
\begin{align*}
|I_\varepsilon|
&\le 
\int_{{\mathbb S}^{n-1}}
{\mathcal E}_p(U)^{n+p}\|\nabla_\xi U\|_p^{-n-p}
\left|\nabla\left[
|\nabla_\xi f_\varepsilon|^{p-2}\nabla_\xi f_\varepsilon
-
|\nabla_\xi U|^{p-2}\nabla_\xi U
\right]\right|\,d\xi
\lesssim\varepsilon^{p-1}{\bf 1}_{B(R_\varepsilon \mathbf{e}_1,1)}
\end{align*}
and hence $\|I_\varepsilon\|_q\lesssim\varepsilon^{p-1}$.
In addition, for $II_\varepsilon$, from the choice of $\varphi$
and \eqref{eq:small-tail},
it follows that 
\begin{align}\label{eq-secon}
\left|\nabla
\left(|\nabla_\xi f_\varepsilon(x)|^{p-2}\nabla_\xi f_\varepsilon\right)\right|
&=(p-1)\left||\nabla_\xi f_\varepsilon|^{p-2}\nabla\nabla_\xi f_\varepsilon\right|
\lesssim\left(|\nabla_\xi U|+|\nabla_\xi\varphi_\varepsilon|\right)^{p-2}
\left(|\nabla\nabla_\xi U|+|\nabla\nabla_\xi \varphi_\varepsilon|\right)\notag\\
&\lesssim|\nabla_\xi U|^{p-2}|\nabla\nabla_\xi U|
+{\bf 1}_{B(R_\varepsilon\mathbb{e}_1,1)}.
\end{align}
By the construction of $U$, we find that, when $|x|\to\infty$,
\begin{align*}
|\nabla_\xi U(x)|^{p-2}|\nabla\nabla_\xi U(x)|
=O\left(|x|^{\frac{p-n}{p-1}-1}\right)^{p-2}
O\left(|x|^{\frac{p-n}{p-1}-2}\right)
=O\left(|x|^{-n}\right).
\end{align*}
Hence, combining this and \eqref{eq-secon},
we conclude that $\|\nabla
(|\nabla_\xi f_\varepsilon(x)|^{p-2}\nabla_\xi f_\varepsilon)\|_q\lesssim1.$
Then, by \eqref{eq:coefficient-bound}, we obtain 
$\|II_\varepsilon\|_q\lesssim\varepsilon^p$.
Consequently, we get
\begin{align}\label{eq:laplacian-bound}
\left\|
\Delta_p^{\rm aff}f_\varepsilon-\Delta_p^{\rm aff}U
\right\|_q
\lesssim\varepsilon^{p-1}.
\end{align}

Next, we consider the Lebesgue term in \eqref{eq:split-P}. 
Following the strategy used in the estimation of $I_\varepsilon$, 
we treat the domains $B(R_\varepsilon \mathbf{e}_1,1)$ and its complement separately; 
utilizing the inequality
\begin{align*}
\left||a+b|^{p^*-2}(a+b)-|a|^{p^*-2}a\right| \lesssim |a|^{p^*-2}|b| +|b|^{p^*-1}
\end{align*}
(which is a direct consequence of Taylor's formula) in $B(R_\varepsilon \mathbf{e}_1,1)$, 
we deduce that $$\left||f_\varepsilon|^{p^*-2}f_\varepsilon-U^{p^*-1} \right| \lesssim 
\varepsilon^{p^*-1}{\bf 1}_{B(R_\varepsilon\mathbf{e}_1,1)}$$ and hence
\begin{align}\label{eq:zero-order-bound}
\left\| |f_\varepsilon|^{p^*-2}f_\varepsilon-U^{p^*-1} \right\|_q \lesssim\varepsilon^{p^*-1}.
\end{align}
Since $p^*>p$, from \eqref{eq:split-P},
\eqref{eq:laplacian-bound}, and \eqref{eq:zero-order-bound}, 
we infer that, for sufficiently small $\varepsilon$, 
$\|P(f_\varepsilon)\|_q
\lesssim\varepsilon^{p-1}.$
This, together with \eqref{eq:delta-Lq}, further implies
$\delta_{\rm aff}(f_\varepsilon)
\lesssim\varepsilon^{p-1}$
and finishes the present proof.
\end{proof}

On the other hand, we now estimate the 
distance from below.

\begin{lemma}\label{lem-pterm}
For every $0<\varepsilon<1$, choose
$R_\varepsilon$ such that
\begin{align}\label{lem-d-e1}
\sup_{|y|\ge \frac{R_\varepsilon}{4}} |U(y)|\le \varepsilon^2.
\end{align}
Then there exist positive constants $c$ and $C$, depending only on $n$, $p$, $U$, and
$\varphi$, such that, for every $\varepsilon\in(0,c)$,
$$d(f_\varepsilon,\mathcal{M}_{\rm aff}^{(1)})\ge
\inf_{(c,\lambda,S,x)\in\mathscr{A}}\|\nabla(T_{\lambda,S,x}f-cU)\|_{p}
\ge C\varepsilon.$$
\end{lemma}

\begin{proof}
We only need to show the second estimate. 
We prove this by contradiction.  Suppose that it is false. 
Then there exist
$\{\varepsilon_j\}_{j\in\mathbb{N}}$ in $(0,1)$, 
$\{R_j\}_{j\in\mathbb{N}}:=\{R_{\varepsilon_j}\}_{j\in\mathbb{N}}$
in $(1,\infty)$, and affine transformation parameters 
$\{(c_j,\lambda_j,A_j,x_j)\}_{j\in\mathbb{N}}\in\mathscr{A}$
such that, when $j\to\infty$, 
$\varepsilon_j\to0$ and $f_j:=f_{\varepsilon_j}$ satisfies
\begin{align}\label{lem-d-e3}
\frac1{\varepsilon_j}
\left\|\nabla\left(T_{\lambda_j,S_j,x_j}f_j-c_j U\right)\right\|_p
\to0.
\end{align}
Applying this and the Sobolev inequality [\eqref{eq-Sob}],
we have, when $j\to\infty$,
$\|T_{\lambda_j,S_j,x_j}f_j-c_j U\|_{p^*}
\to0.$
Since $T_{\lambda_j,S_j,x_j}$ is an isometry on $L^{p^*}$, we deduce that,
when $j\to\infty$,
\begin{align*}
\left\|T_{\lambda_j,S_j,x_j}U-c_j U\right\|_{p^*}
&\le
\left\|T_{\lambda_j,S_j,x_j}f_j-c_j U\right\|_{p^*}
+
\varepsilon_j
\left\|T_{\lambda_j,S_j,x_j}\varphi(\cdot-R_j \mathbf{e}_1)\right\|_{p^*}
\\
&=
\left\|T_{\lambda_j,S_j,x_j}f_j-c_jU\right\|_{p^*}
+
\varepsilon_j\|\varphi\|_{p^*}
\to0.
\end{align*}
By this and Corollary \ref{cor-escape}, 
we find that, when $j\to\infty$,
\begin{align}\label{lem-d-e2}
c_j\to 1,\quad\lambda_j\to1,\quad
S_j\to{\bf 0}, \quad\text{and}\quad x_j\to{\bf 0}.
\end{align}

For each $j\in\mathbb{N}$, define 
$
E_j:=x_j+\lambda_je^{S_j}B(R_j{\bf e}_1,1).$
Applying \eqref{lem-d-e2}, we conclude that, 
for sufficiently large $j\in\mathbb{N}$,
$E_j\subset\{y: |y|\ge \frac{R_j}4\}$.
Therefore, by \eqref{lem-d-e1} and a change of variables, we obtain, for 
sufficiently large $j\in\mathbb{N}$, $\|U\|_{L^{p^*}(E_j)}
\lesssim\varepsilon_j^2$ and 
\begin{align*}
\left\|T_{\lambda_j,S_j,x_j}U\right\|_{L^{p^*}(E_j)}
=\|U\|_{L^{p^*}(B(R_j\mathbf{e}_j,1))}
\lesssim\varepsilon_j^2;
\end{align*}
from these and the Sobolev inequality, it follows that
\begin{align*}
\left\|\nabla\left(T_{\lambda_j,S_j,x_j}f_j-c_jU\right)\right\|_p
&\gtrsim
\left\|T_{\lambda_j,S_j,x_j}f_j-c_jU\right\|_{L^{p^{*}}(E_j)}
\\
&\ge
\varepsilon_j
\left\|T_{\lambda_j,S_j,x_j}\varphi_{\varepsilon_j}
\right\|_{L^{p^*}(E_j)}
-
\left\|T_{\lambda_j,S_j,x_j}U-c_jU\right\|_{L^{p^*}(E_j)}
\\
&\ge\varepsilon_j\|\varphi\|_{p^*}
-\|U\|_{L^{p^*}(E_j)}-|c_j|\|U\|_{L^{p^*}(B(R_j\mathbf{e}_j,1))}
\gtrsim\varepsilon_j.
\end{align*}
This contradicts \eqref{lem-d-e3} and hence 
finishes the present proof.
\end{proof}

Based on these two lemmas, we can immediately conclude 
the sharpness of the exponents $p$ and $p-1$
on the $\dot{W}^{1,p}$ norm, respectively, 
in Theorems \ref{thm-stab} and \ref{thm-bubble}.

\subsection{Sharpness of the exponent on weighted quadratic term in 
Theorem \ref{thm-stab}}

In this subsection, we focus on the weighted quadratic term in 
Theorem \ref{thm-stab}. To this end, further assume that, for every 
$\psi\in T_U\mathcal{M}_{\rm aff}$,
\begin{align}\label{eq-second}
\int_{B({\bf0},1)}
U^{p^*-2}\varphi\psi=0.
\end{align}
Indeed, we can choose $\psi$ in the orthocomplement
of $T_U\mathcal{M}_{\rm aff}$ in $L^2(B({\bf0},1),U^{p^*-2})$.
For any $\varepsilon\in(0,1)$, 
define 
$f_\varepsilon:=U+\varepsilon \varphi$
and 
\begin{align*}
d_\varepsilon:=
\inf_{(c,\lambda,S,x)\in\mathscr{A}}\left\{\|\nabla(T_{\lambda,S,x}f_\varepsilon-cU)\|_{p}^p
+\int_{\mathbb{R}^n}
|c\nabla U|^{p-2}|\nabla(T_{\lambda,S,x}f_\varepsilon-cU)|^2
\right\}.
\end{align*}

\begin{lemma}\label{lem-second}
Let $p\in(2,n)$. 
Then there exist positive constants $c$, depending only on $n$, $p$, $U$, and
$\varphi$, such that, for every $\varepsilon\in(0,c)$,
$d_\varepsilon\sim\varepsilon^2$
with the positive equivalence constants
depend only on $n$, $p$, $U$, and $\varphi$, 
and the main part in $d_\varepsilon$ is 
the quadratic term.
\end{lemma}

\begin{proof}
In the present proof, 
unless otherwise specified, the implicit positive constants 
depend only on $n$, $p$, $U$, and $\varphi$. 
By the definition and H\"older's inequality, we find that, for every $\varepsilon\in(0,1)$,
\begin{align}\label{eq-second-e1}
d_\varepsilon
&\le\left\|\nabla(f_\varepsilon-U)\right\|_p^p
+\int_{\mathbb{R}^n}
|\nabla U|^{p-2}|\nabla (f_\varepsilon-U)|^2\notag\\
&\le \left\|\nabla(f_\varepsilon-U)\right\|_p^p
+\|\nabla U\|_p^{p-2}
\left\|\nabla(f_\varepsilon-U)\right\|_p^2
\lesssim\varepsilon^p+\varepsilon^2\lesssim\varepsilon^2.
\end{align}

Now, we consider the lower estimate by contradiction. 
Assume that the lower bound is not true.
Then there exist $\{\varphi_j\}_{j\in\mathbb{N}}$
in $(0,1)$ and $\{(c_j,\lambda_j,S_j,x_j)\}_{j\in\mathbb{N}}$
in $\mathscr{A}$
such that, when $j\to\infty$,
$\varepsilon_j\to0$ and, for every $j\in\mathbb{N}$,
$f_j:=f_{\varepsilon_j}$ satisfies 
\begin{align}\label{eq-second-e2}
\|\nabla(T_{\lambda_j,S_j,x_j}f_j-c_jU)\|_{p}^p
+\int_{\mathbb{R}^n}
|c_j\nabla U|^{p-2}|\nabla(T_{\lambda_j,S_j,x_j}f_j-c_jU)|^2
<\frac{1}{j}\varepsilon_j.
\end{align}
Then $\|\nabla(T_{\lambda_j,S_j,x_j}f_j-c_jU)\|_{p}\to0$ as $j\to\infty$.
Applying this and an argument similar to that used in Lemma \ref{lem-pterm},
we find that, when $j\to\infty$,
$(c_j,\lambda_j,S_j,x_j)\to(1,1,{\bf0},{\bf 0})$ in $\mathscr{A}$.
This implies that $U_j:=c_jT_{\lambda_j,S_j,x_j}^{-1}U$
converge to $U$ uniformly in $B({\bf0},1)$. Thus,
for sufficiently large $j$, 
$U_j\ge\frac12U$ holds in $B({\bf0},1)$. 
Moreover, using Taylor's formula for the parameter space, we obtain,
when $j\to\infty$, 
\begin{align*}
U_j=U+\sum_\ell {\bf a}_{j,\ell}\psi_\ell
+o(|\mathbf{a}_j|),
\end{align*}
where ${\bf a}_{j,\ell}$ is one component of 
${\bf a}_j:=(c_j,\lambda_j,S_j,x_j)$ and $\psi_\ell$ is the corresponding 
element in the tangent space $T_U\mathcal{M}_{\rm aff}$. 
From this, the embedding $\dot{W}^{1,2}(|\nabla U|^{p-2})\hookrightarrow
L^2(U^{p^*-2})$, and a change of variables, 
it follows that, for sufficiently large $j$, 
\begin{align*}
\int_{\mathbb{R}^n}
|c_j\nabla U|^{p-2}|\nabla(T_{\lambda_j,S_j,x_j}f_j-c_jU)|^2
&\gtrsim\int_{\mathbb{R}^n}
(c_j U)^{p^*-2}|T_{\lambda_j,S_j,x_j}f_j-c_jU|^2\\
&\ge \int_{B({\bf0},1)}U_j^{p^*-2}|f_j-U_j|^2\\
&\gtrsim\int_{B({\bf0},1)}
U^{p^*-2}\left(\varepsilon_j\varphi-\sum_\ell
{\bf a}_{j,\ell}\psi_\ell\right)^2+o(|{\bf a}_j|^2).
\end{align*}
This, combined with the orthogonality condition \eqref{eq-second},
further implies, for sufficiently large $j\in\mathbb{N}$,
\begin{align*}
\int_{\mathbb{R}^n}
|c_j\nabla U|^{p-2}|\nabla(T_{\lambda_j,S_j,x_j}f_j-c_jU)|^2
\gtrsim\varepsilon_j^2
\int_{B({\bf0},1)}U^{p^*-2}\varphi^2
+\int_{B({\bf0},1)}U^{p^*-2}\left(\sum_\ell
{\bf a}_{j,\ell}\psi_\ell\right)^2
\gtrsim\varepsilon_j^2,
\end{align*}
which contradicts \eqref{eq-second-e2}.
Therefore, we obtain the lower estimate $d_\varepsilon\gtrsim
\varepsilon^2$.
Furthermore, 
by the estimation of \eqref{eq-second-e1}, we know that 
the first term is controlled by $\varepsilon^p$.
Hence, since $p>2$,  
the second term (namely the quadratic term) is the main part.
This finishes the present proof.
\end{proof}

Now, we have constructed a family of functions such that the distance 
$d_\varepsilon$ is dominated by the quadratic term and is of order $\varepsilon^2$. Therefore,
to prove the sharpness of the quadratic term, we only need to 
verify ${\mathcal{E}_p(f_\varepsilon)}^p
-S_{\rm aff}^p\|f_\varepsilon\|_{p^*}^p\lesssim\varepsilon^2$.
By \eqref{eq-holder}, it suffices to show 
\begin{align}\label{eq-second-e3}
\delta(\varepsilon):=
\|\nabla f_\varepsilon\|_p^p
-S_{\rm Sob}^p\|f_\varepsilon\|_{p^*}^p\lesssim\varepsilon^2.
\end{align}
Indeed, since $U$ is the extremal function of \eqref{eq-Sob}
and $f_\varepsilon=U+\varepsilon\varphi$, we infer that 
$\delta(0)=0=\delta'(0)$. This shows \eqref{eq-second-e3}
and hence the sharpness.

\section{Appendix}

\setcounter{subsection}{0}                     
\renewcommand{\thesubsection}{\Alph{subsection}}   

\subsection{Vanishing under escaping affine transformations}

\setcounter{theorem}{0}                         
\renewcommand{\thetheorem}{\thesubsection.\arabic{theorem}}

We record the following elementary compactness observation. 
While the corresponding statement for translations and dilations is 
standard in concentration-compactness arguments, the appearance 
of anisotropic affine deformations in our setting requires us to state a version 
for general linear maps. Recall that
the general affine transformation $\mathcal{T}$ is deinfed 
in \eqref{def-tran-g}.

\begin{lemma}\label{lem-escape}
Let $q\in[1,\infty)$ and $f\in L^q({\mathbb R}^n)$.  Let
$\{\lambda_j\}_{j\in\mathbb{N}}\in(0,\infty)$,
$\{A_j\}_{j\in\mathbb{N}}$ in  
$\operatorname{SL}(n)$, and $\{x_j\}_{j\in\mathbb{N}}$
in ${\mathbb R}^n$ be such that, when $j\to\infty$,
\begin{align*}
|\log\lambda_j|+
\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}+|x_j|\to\infty .
\end{align*}
Then ${\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f\to0$ 
locally in measure as $j\to\infty$.
\end{lemma}

\begin{proof}
By the density argument, 
we can assume $f\in C_{\rm c}$.  
To show the present lemma for such $f$, we only need to prove
that, for every compact set $K\subset {\mathbb R}^n$ and
every $t\in(0,\infty)$,
\begin{align}\label{lem-escape-e1}
\lim_{j\to\infty}
\left|\left\{x\in K:
\left|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)\right|>t\right\}\right|=0.
\end{align}
Assume that both 
$\operatorname{supp}f$ and $K$ are contained in 
$B({\bf0},R)$ for some $R\in(0,\infty)$. 
Moreover, it is enough to prove \eqref{lem-escape-e1}
along subsequences. Indeed,
if \eqref{lem-escape-e1} does not hold, 
then there exists a subsequence, still denoted by $\{(\lambda_j,A_j,x_j)\}_{j\in\mathbb{N}}$, 
and $\varepsilon_0\in(0,\infty)$ such that, for all $j$,
\begin{align}\label{lem-escape-e2}
\left|\left\{x\in K:
\left|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)\right|>t\right\}\right|\ge\varepsilon_0.
\end{align}
The argument below for subsequences shows that the left-hand side of 
\eqref{lem-escape-e1} converges to zero along a further subsequence,
which contradicts \eqref{lem-escape-e2}. 
Thus, after passing to a
subsequence and still denoting it by $\{(\lambda_j,A_j,x_j)\}_{j\in\mathbb{N}}$, 
it suffices to consider the following four cases.

\emph{Case 1.} $\lim_{j\to\infty}
\lambda_j=\infty.$
In this case, for any $x\in\mathbb{R}^n$,
\begin{align*}
\left|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)\right|
\le
\lambda_j^{-\frac nq}\|f\|_\infty
\to0
\end{align*}
as $j\to\infty$.
This further implies \eqref{lem-escape-e1}.

\emph{Case 2.} $\lim_{j\to\infty}
\lambda_j=0.$ In this case, by the support assumption,
we find that, for all $j\in\mathbb{N}$,
$\operatorname{supp}{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f
\subset x_j+\lambda_jA_jB({\bf 0},R)$. 
Applying this, we further have, when $j\to\infty$,
\begin{align*}
\left|\left\{x\in K:
\left|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)\right|>t\right\}\right|
\le 
\left|x_j+\lambda_jA_jB({\bf 0},R)\right|
=
\lambda_j^n |B({\bf 0},R)|
\to0,
\end{align*}
which proves \eqref{lem-escape-e1} in this case.

\emph{Case 3.} The sequence 
$\{|\log\lambda_j|\}_{j\in\mathbb{N}}$ is bounded, but 
\begin{align}\label{lem-escape-e3}
\lim_{j\to\infty}\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}=\infty.
\end{align}  Let
$s_j^{(1)}\ge \cdots \ge s_j^{(n)}>0$ denotes the singular value of $A_j$.
Then 
$$s_j^{(1)}=\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}
\quad\text{and}\quad
s_{j}^{(n)}=\frac{1}{\|A^{-1}_j\|_{\mathcal{L}(\mathbb{R}^n)}}
=\frac{1}{\|A^{-T}_j\|_{\mathcal{L}(\mathbb{R}^n)}}
=\min_{|v|=1}|A_j^{T}v|.$$
Applying this,
we conclude that, for every $j\in\mathbb{N}$,
\begin{align*}
1=|\det A_j|
=\prod_{k=1}^{n}s_j^{(k)}
\ge s_{j}^{(1)}\left[s_j^{(n)}\right]^{n-1}
=\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}
\left[s_j^{(n)}\right]^{n-1}.
\end{align*}
From this and \eqref{lem-escape-e3}, it follows that
$\lim_{j\to\infty}s_j^{(n)}=0$.
Moreover, there exists a unit vector $\nu_j$ such that
$|A_j^T\nu_j|=s_j^{(n)}$,
which implies that, for every $j\in\mathbb{N}$ and 
$y\in x_j+\lambda_jA_jB({\bf 0},R)$,
$|\nu_j\cdot(y-x_j)|
\le\lambda_j
Rs_j^{(n)}.$
 Hence, by the support assumption and the boundedness of $\{|\log\lambda_j|\}_{j\in\mathbb{N}}$, 
we obtain, when $j\to\infty$, 
\begin{align*}
\left|\left\{x\in K:
\left|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)\right|>t\right\}\right|
&\le \left|B({\bf 0},R)\cap \left[x_j+\lambda_jA_jB({\bf 0},R)\right]\right|\\
&\le\left|B({\bf 0},R)\cap\left\{y: |\nu_j\cdot(y-x_j)|\le \lambda_jRs_j^{(n)}\right\}\right|
\lesssim s_j^{(n)}\to0.
\end{align*}
Then the desired convergence holds.

\emph{Case 4.} 
The sequences $\{|\log\lambda_j|\}_{j\in\mathbb{N}}$ 
and $\{\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}\}_
{j\in\mathbb{N}}$ are both bounded.
In this case, it holds that 
$\lim_{j\to\infty}|x_j|=\infty$.
In addition, there exists $M\in(0,\infty)$ such that,
for all $j\in\mathbb{N}$,
$x_j+\lambda_j A_jB({\bf 0},R)\subset x_j+B({\bf0},MR)$.
Thus, if $j$ is sufficiently large, 
then $K\cap [x_j+\lambda_jA_jB({\bf 0},R)]=\emptyset$ and hence 
$\{x\in K:|{\mathcal T}_{\lambda_j,A_j,x_j}^{(q)}f(x)|>t\}|=\emptyset$.
This proves the desired convergence in \eqref{lem-escape-e1}.

Combining the above cases, we complete the present proof.
\end{proof}

As a consequence, we have the following 
convergence of bubbles in terms of parameters.

\begin{corollary}\label{cor-escape}
Let $p\in(1,n)$.  Assume that the 
sequence of affine transformation parameters 
$\{(c_j,\lambda_j,S_j,x_j)\}_{j\in\mathbb{N}}\in\mathscr{A}$
satisfies 
\begin{align*}
\lim_{j\to\infty}\left\|T_{\lambda_j,S_j,x_j}U-c_j U\right\|_{p^*}=0.
\end{align*}
Then, when $j\to\infty$,
$(c_j,\lambda_j,S_j,x_j)\to(1,1,{\bf0},{\bf0})$
in $\mathscr{A}$.
\end{corollary}

\begin{proof}
Since the $L^{p^*}$ norm is invariant under $T_{\lambda_j,S_j,x_j}$,
we infer that $c_j\to 1$ as $j\to\infty$.
Hence, in what follows, we can assume $c_j=1$ for every 
$j\in\mathbb{N}$.
For every $j\in\mathbb{N}$, define $A_j:=e^{S_j}$.  Then
$T_{\lambda_j,S_j,x_j}U={\mathcal T}_{\lambda_j,A_j,x_j}^{(p^*)}U.$
If
\begin{align*}
|\log\lambda_j|+
\|A_j\|_{\mathcal{L}(\mathbb{R}^n)}+|x_j|\to\infty
\end{align*}
as $j\to\infty$, 
then, using Lemma \ref{lem-escape}, we 
find that $\{T_{\lambda_j,S_j,x_j}U\}_{j\in\mathbb{N}}$ converges to zero
locally in measure.  On the other hand, the convergence of 
$\{T_{\lambda_j,S_j,x_j}U\}_{j\in\mathbb{N}}$
in $L^{p^*}$ implies the 
local convergence in measure to $U$.  It is impossible since $U$ is not identically
zero. Therefore, there exist a positive constant $M$ and 
a subsequence, still denoted by 
$\{(\lambda_j,S_j,x_j)\}_{j\in\mathbb{N}}$, such that,
for every $j\in\mathbb{N}$,
\begin{align*}
|\log\lambda_j|+
\|S_j\|_{\mathcal{L}(\mathbb{R}^n)}+|x_j|\le M.
\end{align*}
From this, we further deduce that,  
there exist a subsequence, still denoted by 
$\{(\lambda_j,S_j,x_j)\}_{j\in\mathbb{N}}$,
and $(\lambda,S,x)$ with $\lambda>0$, such that,
when $j\to\infty$,
$\lambda_j\to\lambda$, $S_j\to S$, and 
$x_j\to x$.
By the Lebesgue dominated convergence theorem
and the construction of $U$,
we then conclude that 
\begin{align*}
T_{\lambda_j,S_j,x_j}U\to T_{\lambda,S,x}U
\end{align*}
strongly in $L^{p^*}$. Therefore $T_{\lambda,S,x}U=U$ 
and hence $\lambda=1$, 
$S={\bf 0}$, and $x={\bf 0}$.
This establishes the conclusion of Corollary \ref{cor-escape} for a subsequence. 
Then, arguing by contradiction similarly as in the proof of Lemma \ref{lem-escape}, 
we deduce that the conclusion actually holds for the full sequence, which completes the proof.
\end{proof}

\subsection{Nonlinear Brezis--Lieb lemmas}\label{App-B}

\setcounter{theorem}{0}                         
\renewcommand{\thetheorem}{\thesubsection.\arabic{theorem}}

Recall that the well-known Brezis--Lieb lemma (see \cite[Theorem 1.9]{ll01}):
if $(\Omega,\mu)$ is a measure space, 
$p\in(0,\infty)$, and $\{f_k\}_{k\in\mathbb{N}}$ is a bounded sequence in 
$L^p(\Omega)$ such that $f_k\to f$ almost everywhere as $k\to\infty$, then
\begin{align}\label{eq-BL}
\lim_{k\to\infty}\left\||f_k|^p-|f|^p-|f-f_k|^p\right\|_{L^1(\Omega)}=0.
\end{align}

In this section, we establish the following two nonlinear Brezis--Lieb lemmas,
which are important in our proof of relative compactness.

\begin{lemma}\label{lem-BL}
Let $(\Omega,\mu)$ be a measure space.
\begin{enumerate}[{\rm(i)}]
  \item\label{it1-BL} If $p\in(1,\infty)$,
  $\{f_k\}_{k\in\mathbb{N}}$ is a bounded sequence in 
  $L^p(\Omega)$, and $f_k\to f$ almost everywhere as $k\to\infty$,
  then 
  \begin{align*}
  \lim_{k\to\infty}\left\||f_k|^{p-2}-|f|^{p-2}f-|f_k-f|^{p-2}(f_k-f)
  \right\|_{L^{p'}{(\Omega)}}=0.
  \end{align*}
  \item\label{it2-BL} Further assume $(\Omega,\mu)$ is a finite measure space and 
  the nonnegative measurable sequences 
  $\{a_k\}_{k\in\mathbb{N}}$ and $\{b_{k}\}_{k\in\mathbb{N}}$
  and the measurable function $a$ satisfy that there exist positive constants 
  $c_0,c_1,c_2$ such that $a\ge c_0$ and $a_k\in[c_1,c_2]$ for any 
  $k\in\mathbb{N}$. If $\|a_k-a-b_k\|_{L^1(\Omega)}\to0$ as $k\to\infty$,
  then, for any $\alpha\in(0,\infty)$,
  \begin{align*}
  \lim_{k\to\infty}\left|\left(\int_{\Omega}a_n^{-\alpha}\right)^{-\frac1\alpha}
  -\left(\int_{\Omega}(a+b_n)^{-\alpha}\right)^{-\frac1\alpha}\right|=0.
  \end{align*}
\end{enumerate}
\end{lemma}

To show Lemma \ref{lem-BL}\eqref{it1-BL}, we need the following 
inequalities.

\begin{lemma}\label{lem-pine}
Let $p\in(1,\infty)$.
For any $\varepsilon\in(0,\infty)$, there exists a positive constant 
$C_{\varepsilon}$ such that, for any $a,b\in\mathbb{C}$,
\begin{align}\label{lem-pine-e0}
\left||a+b|^p-|b|^{p}\right|
\le\varepsilon|b|^p+C_{\varepsilon}|a|^p
\end{align}
and
\begin{align}\label{lem-pine-ee}
\left||a+b|^{p-2}(a+b)-|b|^{p-2}b\right|^{p'}
\le\varepsilon|b|^p+C_{\varepsilon}|a|^p.
\end{align}
\end{lemma}

\begin{proof}
The inequality \eqref{lem-pine-e0} is well known;
see, for instance, \cite[p.\,699]{Bre02}.
Next, we show \eqref{lem-pine-ee} by contradiction.
Indeed, assume \eqref{lem-pine-ee} does not hold.
Then there exist a positive constant $\varepsilon_0$ 
and two complex sequences $\{a_k\}_{k\in\mathbb{N}}$ and $\{b_{k}\}_{k\in\mathbb{N}}$
such that, for all $k\in\mathbb{N}$,
\begin{align}\label{lem-pine-e1}
\left||a_k+b_k|^{p-2}(a_k+b_k)-|b_k|^{p-2}b_k\right|^{p'}
>\varepsilon_0|b_k|^p+k|a_k|^p.
\end{align}
Then $b_k$ should not be zero and hence, by the homogeneity,
we can assume $|b_k|=1$ for all $k\in\mathbb{N}$.
This, together with the triangle inequality, implies that, for any $k\in\mathbb{N}$,
\begin{align*}
\left||a_k+b_k|^{p-2}(a_k+b_k)-|b_k|^{p-2}b_k\right|^{p'}
\le\left[(|a_k|+1)^{p-1}+1\right]^{p'}.
\end{align*}
From this and \eqref{lem-pine-e1}, we infer that,
for all $k\in\mathbb{N}$,
\begin{align*}
k|a_k|^p\le\left[(|a_k|+1)^{p-1}+1\right]^{p'}.
\end{align*}
This shows that $\{a_k\}_{k\in\mathbb{N}}$ can not have a positive lower bound.
Hence, we can choose the subsequences,
still denoted by $\{a_k\}_{k\in\mathbb{N}}$ and $\{b_k\}_{k\in\mathbb{N}}$,
such that, when $k\to\infty$, $a_k\to0$ and $b_k\to b_0$ for some $b_0\in\mathbb{C}$.
Then, letting $k\to\infty$ in \eqref{lem-pine-e1},
we obtain $0\ge \varepsilon_0$. It is impossible. 
This then proves the present lemma.
\end{proof}

Now, we give the proof of Lemma \ref{lem-BL}.

\begin{proof}[Proof of Lemma \ref{lem-BL}]
We first consider \eqref{it1-BL}. To this end, fix $\varepsilon\in(0,1)$.
Applying Lemma \ref{lem-pine},
we find that there exists two positive constants $C_{\varepsilon,1}$
and $C_{\varepsilon,2}$ such that, for all $k\in\mathbb{N}$,
\begin{align*}
F_k:=&\,\left||f_k|^{p-2}-|f|^{p-2}f-|f_k-f|^{p-2}(f_k-f)\right|^{p'}\\
\le&\,(1+\varepsilon)\left||f_k|^{p-2}-|f_k-f|^{p-2}(f_k-f)\right|^{p'}
+C_{\varepsilon,1}|f|^p\quad\text{by \eqref{lem-pine-e0}}\\
\le&\,\varepsilon(1+\varepsilon)|f_k-f|^p
+\left[C_{\varepsilon,1}+(1+\varepsilon)C_{\varepsilon,2}\right]
|f|^p\quad\text{by \eqref{lem-pine-ee}}.
\end{align*}
Thus, by Fatou's lemma, we obtain 
\begin{align}\label{lem-BL-e1}
\limsup_{k\to\infty}\int_{\Omega}
\left[F_k-\varepsilon(1+\varepsilon)|f_k-f|^p\right]
\le \int_{\Omega}\limsup_{k\to\infty}
\left[F_k-\varepsilon(1+\varepsilon)|f_k-f|^p\right]=0.
\end{align}
On the other hand,
from the boundedness of $\{f_k\}_{k\in\mathbb{N}}$
in $L^p(\Omega)$, it follows that there exists 
$M\in(0,\infty)$ such that, for all $k\in\mathbb{N}$,
\begin{align*}
\int_{\Omega}F_k&=
\int_{\Omega}
\left[F_k-\varepsilon(1+\varepsilon)|f_k-f|^p\right]+
\varepsilon(1+\varepsilon)\int_{\Omega}|f_k-f|^p\\
&\le \int_{\Omega}
\left[F_k-\varepsilon(1+\varepsilon)|f_k-f|^p\right]+M\varepsilon.
\end{align*}
Hence, applying this and \eqref{lem-BL-e1}, we conclude 
$\limsup_{k\to\infty}\int_{\Omega}F_k\le M\varepsilon$.
Then the arbitrariness of $\varepsilon$, together 
with the non-negativity of $F_k$,
further implies 
$\int_{\Omega}F_k\to0$ as $k\to\infty$.
This shows \eqref{it1-BL}.

Next, we prove \eqref{it2-BL}. It is much simpler.
Indeed, from the assumptions, we infer that 
$a_k\ge c_1$ and $a+b_k\ge c_0$ for all $k\in\mathbb{N}$.
Using this and Taylor's formula, we find that, 
for any $k\in\mathbb{N}$, there exists 
$t\in(0,1)$ such that
\begin{align*}
\left|a_k^{-\alpha}-(a+b_k)^{-\alpha}\right|
=\alpha|ta_k+(1-t)(a+b_k)|^{-\alpha-1}
|a_k-a-b|
\lesssim|a_k-a-b|,
\end{align*}
where the implicit positive constant 
is independent of $k$ and
is uniform on $\Omega$.
Therefore, using $\lim_{k\to\infty}\|a_k-a-b\|_{L^1(\Omega)}=0$,
we get
\begin{align}\label{lem-BL-e2}
\lim_{k\to\infty}\left\|a_k^{-\alpha}-(a+b_k)^{-\alpha}\right\|_{L^1(\Omega)}=0.
\end{align}
Since $a_k\le c_2$ for all $k\in\mathbb{N}$, 
it follows that $\int_{\Omega}a_k^{-\alpha}\ge c_2^{-\alpha}|\Omega|$.
Then, by \eqref{lem-BL-e2}, we know that
$\int_{\Omega}(a+b_k)^{-\alpha}\gtrsim 1$ uniformly 
for sufficiently large $k$.
Thus, using Taylor's formula again, we obtain, for sufficiently large $k$,
there exist $t\in(0,1)$ such that 
\begin{align*}
&\left|\left(\int_{\Omega}a_k^{-\alpha}\right)^{-\frac1\alpha}
-\left(\int_{\Omega}(a+b_k)^{-\alpha}\right)^{-\frac1\alpha}\right|\\
&\quad\le\frac{1}{\alpha}
\left[t\left(\int_{\Omega}a_k^{-\alpha}\right)
+(1-t)\left(\int_{\Omega}(a+b_k)^{-\alpha}\right)\right]^{-1-\frac1\alpha}
\left\|a_k^{-\alpha}-(a+b_k)^{-\alpha}\right\|_{L^1(\Omega)}\\
&\quad\lesssim\left\|a_k^{-\alpha}-(a+b_k)^{-\alpha}\right\|_{L^1(\Omega)}
\end{align*}
with the implicit positive constant independent of $k$.
Then \eqref{lem-BL-e2} further implies 
\eqref{it2-BL}. 
This finishes the proof of the present lemma.
\end{proof}

\subsection{Asymptotic decoupling formulae}\label{App-C}

\setcounter{theorem}{0}                         
\renewcommand{\thetheorem}{\thesubsection.\arabic{theorem}}

Let $p\in(1,n)$ and $\{w_k\}_{k\in\mathbb{N}}$
be a bounded sequence in $\dot{W}^{1,p}$.
Then $\{w_k\}_{k\in\mathbb{N}}$ has the profile decomposition 
(see Lemma \ref{lem-pdecom}). Moreover, \cite[(3.1)]{o22} tells us 
the sequence $\{V_j\}_{j\in\mathbb{N}}$ given in Lemma \ref{lem-pdecom}
satisfies the following asymptotic decoupling:
\begin{align}\label{eq-pdecom}
\lim_{k\to\infty}\|w_k\|_{p^*}^{p^*}
=\sum_{j=1}^{\infty}\|V_j\|_{p^*}^{p^*}.
\end{align}
But in this paper, we also need to deal with the asymptotic decoupling 
of directional derivatives.
Hence, we establish the following two general 
asymptotic decoupling formulae.

\begin{lemma}
Let $p\in(1,\infty)$, $J\in\mathbb{N}$, and $\{f_j\}_{j=1}^J$ 
be a sequence in $L^p$.
If the sequence $$\{\mathfrak{g}_{j,k}\}_{j=1,\ldots J,\,k\in\mathbb{N}}
=\{(\lambda_{j,k},x_{j,k})\}_{j=1,\ldots J,\,k\in\mathbb{N}}$$ in 
$(0,\infty)\times\mathbb{R}^n$ satisfies, for any $j_1\ne j_2$,
$\{\mathfrak{g}_{j_1,k}\}_{k\in\mathbb{N}}$
and $\{\mathfrak{g}_{j_2,k}\}_{k\in\mathbb{N}}$ are orthogonal, then 
\begin{align}\label{lem-decou-e0}
\lim_{k\to\infty}\left|\left\|\sum_{j=1}^{J}
\mathfrak{g}_{j,k}(f_j;p)\right\|_{p}^p
-\sum_{j=1}^{J}\|f_j\|_p^p\right|=0
\end{align}
and 
\begin{align}\label{lem-decou-ee}
\lim_{k\to\infty}
\left\|\sum_{j=1}^{J}
|\mathfrak{g}_{j,k}(f_j;p)|^{p-2}\mathfrak{g}_{j,k}(f_j;p)
-\left|\sum_{j=1}^{J}\mathfrak{g}_{j,k}(f_j;p)\right|^{p-2}
\sum_{j=1}^{J}\mathfrak{g}_{j,k}(f_j;p)
\right\|_{p'}=0.
\end{align}
\end{lemma}

\begin{proof}
We first prove \eqref{lem-decou-e0} by induction.
This is obvious when $J=1$. Now, assume \eqref{lem-decou-e0} holds 
for $J\in\mathbb{N}$ and we aim to show this for $J+1$.
For any $k\in\mathbb{N}$, define 
$F_k:=\sum_{j=1}^{J}\mathfrak{g}_{j,k}(f_j;p)$;
then, by a change of variables, we have 
\begin{align*}
\left\|\sum_{j=1}^{J+1}\mathfrak{g}_{j,k}(f_j;p)\right\|_p^p
=\left\|F_k+\mathfrak{g}_{J+1,k}(f_{J+1};p)\right\|_p^p
=\left\|\mathfrak{g}_{J+1,k}^{-1}(F_k;p)+f_{J+1}\right\|_p^p.
\end{align*}
By the orthogonality assumption on $\mathfrak{g}_{j,k}$, we 
can easily show that, for all $j\in\{1,\ldots, J\}$,
the function $\mathfrak{g}_{J+1,k}^{-1}\mathfrak{g}_{j,k}(f_j;p)$
converges locally in measure as $k\to\infty$.
Hence, there exists a subsequence of $\{\mathfrak{g}_{J+1,k}^{-1}(F_k;p)\}_{k\in\mathbb{N}}
=\{\sum_{j=1}^{J}\mathfrak{g}_{J+1,k}^{-1}(f_j;p)\}_{k\in\mathbb{N}}$,
denoted by $\{\mathfrak{g}_{J+1,k_\ell}^{-1}(F_{k_\ell};p)\}_{\ell\in\mathbb{N}}$,
such that $\mathfrak{g}_{J+1,{k_\ell}}^{-1}(F_{k_\ell};p)\to$ almost everywhere as $\ell\to\infty$.
Since $\mathfrak{g}_{j,k}(\cdot;p)$ keeps the $L^p$ norm, from 
the Brezis--Lieb lemma [see \eqref{eq-BL}], we deduce that, when $\ell\to\infty$,
\begin{align}\label{lem-decou-e1}
\left\|\mathfrak{g}_{J+1,k_\ell}^{-1}(F_{k_\ell};p)+f_{J+1}\right\|_p^p
=\|F_{k_\ell}\|_p^p+\|f_{J+1}\|_p^p+o(1).
\end{align}
Indeed, this also holds for the whole sequence. Otherwise, 
there exist positive constant $\varepsilon_0$ and 
a subsequence $\{\mathfrak{g}_{J+1,k_m}^{-1}(F_{k_m};p)\}_{m\in\mathbb{N}}$
such that, for all $m\in\mathbb{N}$,
\begin{align*}
\left|\left\|\mathfrak{g}_{J+1,k_m}^{-1}(F_{k_m};p)+f_{J+1}\right\|_p^p
-\|F_{k_m}\|_p^p-\|f_{J+1}\|_p^p\right|\ge\varepsilon_0.
\end{align*}
Repeating the argument used above, we can further choose 
a subsequence satisfying \eqref{lem-decou-e1}. It is impossible.
Therefore, by the assumption that \eqref{lem-decou-e0} holds 
for $J$, we find that, when $k\to\infty$,
\begin{align*}
\left\|\sum_{j=1}^{J+1}\mathfrak{g}_{j,k}(f_j;p)\right\|_p^p
&=\left\|\mathfrak{g}_{J+1,k}^{-1}(F_{k};p)+f_{J+1}\right\|_p^p
=\|F_{k}\|_p^p+\|f_{J+1}\|_p^p+o(1)\\
&=\sum_{j=1}^{J+1}\|f_j\|_p^p+o(1).
\end{align*}
This shows that \eqref{lem-decou-e0} also holds for 
$J+1$. Then we complete the proof of \eqref{lem-decou-e0}.
Furthermore, a similar induction argument, combined with 
the nonlinear Brezis--Lieb lemma established in Lemma 
\ref{lem-BL}\eqref{it1-BL}, also yields \eqref{lem-decou-ee}.
This then finishes the present proof.
\end{proof}


\smallskip
\noindent\textbf{Acknowledgements}\quad
The authors acknowledge the use of AI tools during the exploratory stage of this project.
All mathematical arguments and proofs in the final manuscript were checked
and written by the authors.

\begin{thebibliography}{99} 

\bibitem{a02} 
C. O. Alves, 
Existence of positive solutions for a problem with lack of compactness involving the $p$-Laplacian, 
Nonlinear Anal. 51 (2002), 1187--1206.

\vspace{-.3cm}

\bibitem{acg26}
C. A. Antonini, G. Ciraolo and M. Gatti,
On the anisotropic critical $p$-Laplace equation: classification, decomposition,
and stability results, arXiv:2604.13758.

\vspace{-.3cm}

\bibitem{a76}
T. Aubin, 
Probl\`emes isoperimetriques et espaces de Sobolev,
J. Differ. Geom. 11 (1976), 573--598.

\vspace{-.3cm}

\bibitem{be91}
G. Bianchi and H. Egnell,
A note on the Sobolev inequality,
J. Funct. Anal. 100 (1991), 18--24.

\vspace{-.3cm}

\bibitem{bfr26}
K. J. B\"or\"oczky, A. Figalli and J. P. G. Ramos,
Isoperimetric Inequalities, Brunn--Minkowski Theory and Minkowski-Type 
Monge-Amp\`ere Equations on the Sphere, Zur. Lect. Adv. Math.,
European Mathematical Society (EMS), Berlin, 2026.

\vspace{-.3cm}

\bibitem{Bre02}
H. Brezis,
How to recognize constant functions. A connection with Sobolev spaces,
Russian Math. Surveys 57 (2002), 693--708.

\vspace{-0.3cm}

\bibitem{bl85}
H. Brezis and E. H. Lieb,
Sobolev inequalities with remainder terms,
J. Funct. Anal. 62 (1985), 73--86.

\vspace{-.3cm}

\bibitem{cfmp09}
A. Cianchi, N. Fusco, F. Maggi and A. Pratelli,
The sharp Sobolev inequality in quantitative form,
J. Eur. Math. Soc. (JEMS) 11 (2009), 1105--1139.

\vspace{-.3cm}

\bibitem{clyz09}
A. Cianchi, E. Lutwak, D. Yang and G. Zhang,
Affine Moser--Trudinger and Morrey--Sobolev inequalities,
Calc. Var. Partial Differential Equations 36 (2009), 419--436.

\vspace{-.3cm}

\bibitem{cfm18} 
G. Ciraolo, A. Figalli and F. Maggi, 
A quantitative analysis of metrics on $\mathbb R^n$ with almost constant positive scalar curvature, 
with applications to fast diffusion flows, 
Int. Math. Res. Not. IMRN 2018 (2018), 6780--6797. 

\vspace{-.3cm}

\bibitem{cg26} G. Ciraolo and M. Gatti, 
On the stability of the critical $p$-Laplace equation, 
J. Funct. Anal. 291 (2026), Paper No. 111575.

\vspace{-.3cm}

\bibitem{cnv04}
D. Cordero-Erausquin, B. Nazaret and C. Villani,
A mass-transportation approach to sharp Sobolev and 
Gagliardo--Nirenberg inequalities,
Adv. Math. 182 (2004), 307--332. 

%\vspace{-.3cm}
%
%\bibitem{dgpyyz24}
%F. Dai, L. Grafakos, Z. Pan, D. Yang, W. Yuan and Y. Zhang,
%The Bourgain--Brezis--Mironescu formula on ball Banach
%function spaces,
%Math. Ann. 388 (2024), 1691--1768.

\vspace{-.3cm}

\bibitem{dx13}
F. Dai and Y. Xu,
Approximation Theory and Harmonic Analysis on Spheres and Balls,
Springer Monogr. Math, Springer, New York, 2013.

\vspace{-.3cm}

\bibitem{dsw25} 
B. Deng, L. Sun and J.-C. Wei, 
Sharp quantitative estimates of Struwe's decomposition, 
Duke Math. J. 174 (2025), 159--228. 

\vspace{-.3cm}

\bibitem{deffl25}
J. Dolbeault, M. J. Esteban, A. Figalli, R. L. Frank and M. Loss, Sharp stability for Sobolev and log-Sobolev inequalities, with optimal dimensional dependence, 
Camb. J. Math., 13 (2025), 359--430,

\vspace{-.3cm}

\bibitem{DLTYY25}
O. Dom\'inguez, Y. Li, S. Tikhonov, D. Yang and W. Yuan,
New approach to
affine Moser--Trudinger inequalities via Besov polar projection bodies,
Math. Ann. 392 (2025), 3319--3366.

\vspace{-.3cm}

\bibitem{ens22}
M. Engelstein, R. Neumayer and L. Spolaor, 
Quantitative stability for minimizing Yamabe metrics,
Trans. Amer. Math. Soc. Ser. B 9 (2022), 395--414.

\vspace{-.3cm}

\bibitem{flz26}
S. Fan, G.-D. Li and J. J. Zhang,
Sharp stability for the affine fractional Sobolev inequality,
arXiv:2605.03732.

\vspace{-.3cm}

\bibitem{flz26b}
S. Fan, G.-D. Li and J. J. Zhang,
Sharp Quantitative Stability for the Affine \(p\)-Sobolev Inequality, Part I: The Case \(2\le p<n\),
arXiv:2606.09555

\vspace{-.3cm}

\bibitem{fg20}
A. Figalli and F. Glaudo, 
On the sharp stability of critical points of the Sobolev inequality,
Arch. Ration. Mech. Anal. 237 (2020), 201--258.

\vspace{-.3cm}

\bibitem{fn19}
A. Figalli and R. Neumayer,
Gradient stability for the Sobolev inequality: the case $p\ge2$,
J. Eur. Math. Soc. (JEMS) 21  (2019), 319--354.

\vspace{-.3cm}

\bibitem{fz22}
A. Figalli and Y. R.-Y. Zhang,
Sharp gradient stability for the Sobolev inequality,
Duke Math. J. 171 (2022), 2407--2459.

\vspace{-.3cm}

\bibitem{f22}
R. L. Frank, 
Degenerate stability of some Sobolev inequalities, 
Ann. Inst. H. Poincar'e Anal. Non Lin'eaire 39 (2022), no. 6, 1459--1484.

\vspace{-.3cm}

\bibitem{f24}
R. L. Frank, 
The sharp Sobolev inequality and its stability: An introduction. In: Geometric and Analytic Aspects of Functional Variational Principles. Cetraro, Italy 2022 (A. Cianchi, V. Maz’ya, and T. Weth, eds.), pp. 1--64, Springer, Cham, 2024.

\vspace{-.3cm}

\bibitem{fp24a}
R. L. Frank and J. W. Peteranderl,
Degenerate stability of the Caffarelli--Kohn--Nirenberg inequality
along the Felli–Schneider curve, 
Calc. Var. Partial Differential Equations 63 (2024), no. 44.

\vspace{-.3cm}

\bibitem{fp24}
R. L. Frank and J. W. Peteranderl,
The sharp $\sigma_2$-curvature inequality on the sphere in quantitative form, arXiv:2412.12819.

\vspace{-.3cm}

\bibitem{fpr25}
R. L. Frank, J. W. Peteranderl and Larry Read,
Sharp quantitative integral inequalities for harmonic extensions,
https://arxiv.org/abs/2508.09940.

\vspace{-.3cm}

\bibitem{g02}
R. J. Gardner,
The Brunn--Minkowski inequality,
Bull. Amer. Math. Soc. (N.S.) 39 (2002), 355--405.

\vspace{-.3cm}

\bibitem{hs09}
C. Haberl and F. E. Schuster,
Asymmetric affine $L_p$ Sobolev inequalities,
J. Funct. Anal. 257 (2009), 641--658.

\vspace{-.3cm}

\bibitem{glz25}
A. Guerra, X. Lamy and K. Zemas, 
Sharp quantitative stability of the M\"obius group among sphere-
valued maps in arbitrary dimension, 
Trans. Amer. Math. Soc. 378 (2025), 1235--1259.

\vspace{-.3cm}

\bibitem{hjm16}
J. Haddad, C. H. Jim\'enez and M. Montenegro,
Sharp affine Sobolev type inequalities via the $L_p$
Busemann--Petty centroid inequality,
J. Funct. Anal. 271 (2016), 454--473.

\vspace{-.3cm}

\bibitem{hjm21}
J. Haddad, C. H. Jim\'enez and M. Montenegro,
From affine Poincar\'e inequalities to affine spectral inequalities,
Adv. Math. 386 (2021), Paper No. 107808, 35 pp.

\vspace{-.3cm}

\bibitem{hl24}
J. Haddad and M. Ludwig,
Affine fractional $L^p$ Sobolev inequalities,
Math. Ann. 388 (2024), 1091--1115.

\vspace{-.3cm}

\bibitem{hl24-2}
J. Haddad and M. Ludwig,
Affine fractional Sobolev and isoperimetric inequalities,
J. Differential Geom. 129 (2025), 695--724.

\vspace{-.3cm}

\bibitem{hl16}
Q. Huang and A.-J. Li,
Optimal Sobolev norms in the affine class,
J. Math. Anal. Appl. 436 (2016), 568--585.

\vspace{-.3cm}

\bibitem{iz25}
L.~I.~Ignat, E.~Zuazua,
Optimal convergence rates for the finite element approximation of the Sobolev constant,
arXiv:2504.09637.

\vspace{-.3cm}

\bibitem{j99}
S. Jaffard, 
Analysis of the lack of compactness in the critical Sobolev embeddings,
J. Funct. Anal. 161 (1999), 384--396.

\vspace{-.3cm}

\bibitem{ll01}
E. H. Lieb and M. Loss,
Analysis, 2nd ed, Grad. Stud. Math. 14, 
American Mathematical Society (AMS), Providence, RI, 2001.

\vspace{-.3cm}

\bibitem{l85}
P.-L. Lions,
The concentration-compactness principle in the calculus of variations. The limit case. I,
Rev. Mat. Iberoamericana 1 (1985), no. 1, 145--201.

\vspace{-.3cm}

\bibitem{lz25}
G. Liu and Y. R.-Y. Zhang,
Sharp stability for critical points of the Sobolev inequality 
in the absence of bubbling,
arXiv:2503.02340.

\vspace{-.3cm}

\bibitem{lyz02}
E. Lutwak, D. Yang and G. Zhang,
Sharp affine $L_p$ Sobolev inequalities,
J. Differ. Geom. 62 (2002), 17--38.

\vspace{-.3cm}

\bibitem{mw10} 
C. Mercuri and M. Willem, 
A global compactness result for the $p$-Laplacian involving critical nonlinearities, 
Discrete Contin. Dyn. Syst. 28 (2010), 469--493. 

\vspace{-.3cm}

\bibitem{n20}
R. Neumayer, 
A note on strong-form stability for the Sobolev inequality,
Calc. Var. Partial Differ. Equ. 59 (2020), Paper No. 25, 8 pp.

\vspace{-.3cm}

\bibitem{n16}
V. H. Nguyen, 
New approach to the affine P\'olya--Szeg\"o principle and 
the stability version of the affine Sobolev inequality,
Adv. Math. 302 (2016), 1080--1110.

\vspace{-.3cm}

\bibitem{o22}
M. Okumura,
Profile decomposition in Sobolev spaces and decomposition of integral functionals II: 
homogeneous case,
arXiv:2109.08177.

\vspace{-.3cm}

\bibitem{r26}
C. Reuter,
Local fixed point results for centroid body operators,
Trans. Am. Math. Soc. 379 (2026), 3043--3061.

\vspace{-.3cm}

\bibitem{r66}
E. Rodemich, 
The Sobolev inequality with best possible constant. 
Analysis Seminar Caltech, Spring 1966.

\vspace{-.3cm}

\bibitem{st18}
I. Schindler and C. Tintarev,
Compactness properties and ground states for the affine Laplacian,
Calc. Var. Partial Differential Equations 57 (2018), Paper No. 48,
14 pp.

\vspace{-.3cm}

\bibitem{s14}
R. Schneider,
Convex Bodies: The Brunn--Minkowski Theory,
2nd expanded ed.,
Encycl. Math. Appl. 151,
Cambridge University Press, Cambridge, 2014.

\vspace{-.3cm}

\bibitem{s84} 
M. Struwe, 
A global compactness result for elliptic boundary value problems involving limiting nonlinearities, 
Math. Z. 187 (1984), 511--517. 

\vspace{-.3cm}

\bibitem{s75}
G. Szeg\"o,
Orthogonal Polynomials, 4th ed,
Colloq. Publ., Am. Math. Soc. 23, 
American Mathematical Society (AMS), Providence, RI, 1975.

\vspace{-.3cm}

\bibitem{t76}
G. Talenti,
Best constant in Sobolev inequality,
Ann. Mat. Pura Appl. (4) 110 (1976), 353--372.

\vspace{-.3cm}

\bibitem{wz25}
G. Wang and M. Zhang, 
Stability of spinorial Sobolev inequalities on $\mathbb S^n$,
arXiv:2508.09047.

\vspace{-.3cm}

\bibitem{w12}
T. Wang,
The affine Sobolev--Zhang inequality on $BV(\mathbb R^n)$,
Adv. Math. 230 (2012), 2457--2473.

\vspace{-.3cm}

\bibitem{w13}
T. Wang,
The affine P\'olya--Szeg\"o principle: equality cases and stability,
J. Funct. Anal. 265 (2013), 1728--1748.

\vspace{-.3cm}

\bibitem{ww26}
J. Wei, Y. Wu,
Stability of the Caffarelli--Kohn--Nirenberg inequality along the Felli--Schneider curve: critical points at infinity.
Proc. Lond. Math. Soc. (3) 132 (2026), no. 4, Paper No. e70146, 102 pp.

\vspace{-.3cm}

\bibitem{z99}
G. Zhang,
The affine Sobolev inequality,
J. Differ. Geom. 53 (1999), 183--202.

\vspace{-.3cm}

\bibitem{zz26}
Y. Zhou, W. Zou,
Degenerate stability of critical points of the Caffarelli--Kohn--Nirenberg inequality along the Felli--Schneider curve.
J. Lond. Math. Soc. (2) 113 (2026), no. 2, Paper No. e70454, 37 pp.

\end{thebibliography}

\bigskip

\noindent Rupert L. Frank

\smallskip

\noindent Mathematisches Institut, Ludwig-Maximilians Universit\"at M\"unchen,
Theresienstr.~39, 80333 M\"unchen, Germany;
Munich Center for Quantum Science and Technology,
Schellingstr.~4, 80799 M\"unchen, Germany;
Mathematics 253-37, Caltech,
Pasadena, CA 91125, USA

\smallskip

\noindent {\it E-mail}: \texttt{r.frank@lmu.de} 

\bigskip

\noindent Yinqin Li and Dachun Yang

\medskip

\noindent Laboratory of Mathematics and Complex Systems
(Ministry of Education of China),
School of Mathematical Sciences, Beijing Normal University,
Beijing 100875, The People's Republic of China

\smallskip

\noindent{\it E-mails:} \texttt{yinqli@mail.bnu.edu.cn} (Y. Li)

\noindent\phantom{\it E-mails:} \texttt{dcyang@bnu.edu.cn} (D. Yang)


\end{document}