editors: nobody
pdfinfoomitdate=1
DeclarePairedDelimiter{abs}{lvert}{rvert}
DeclarePairedDelimiter{norm}{lVert}{rVert}
# Positive scalar curvature forces codimension-two width:a complete proof
hypersetup{pdftitle={Positive scalar curvature and uniform codimension-two width},pdfauthor={OpenAI}}
maketitle
Abstract:
Let (M,g) be a connected complete smooth Riemannian n-manifold,
nge 4, with scalar curvature Scal_gge 1 and no boundary.
We prove the quantitative form of Gromov's macroscopic-dimension
conjecture in every dimension nge 4. There is a constant C_n
depending only on n such that M admits a continuous map f: Mto K
to a simplicial complex K of dimension at most n-2, whose entire
fibers have g-diameter at most C_n. Equivalently, the Urysohn
(n-2)-width satisfies UW_{n-2}(M,g)le C_n.
Three consequences follow. First, the filling radius of any closed
M with Scal_ggesigma^2 is at most C_n/(2sigma). Second, the
universal cover of any closed manifold of positive scalar curvature in
every dimension nge2 has continuous macroscopic dimension at most n-2.
Third, no closed aspherical manifold of dimension nge4 admits a
metric of positive scalar curvature.
The proof has three parts. (i) Stretch the metric at a fixed original
scale so that every label pair becomes arbitrarily well separated,
at the price of a mild drift loss. (ii) Cut each label pair apart by a
minimizing hypersurface, handling high-dimensional singularities by
enlarging the pressure profiles of the remaining labels; a warped-circle
trick restores scalar control on each cut packet. (iii) Assemble
coordinates t_i: Mto[0,1] that are 1 inside label i and
0 far outside it; the walls cannot all pass through the same point,
so at most n-2 coordinates are fractional, and any two points in a
common fiber lie in a common ball of radius 2D.
tableofcontents
# Introduction
A lower bound for scalar curvature is a local condition. Gromov's
macroscopic-dimension conjecture [cite] predicts a global
consequence: at a controlled scale, two dimensions can be collapsed
without increasing the diameter of any fiber by more than a constant.
In this paper we prove the quantitative form of that conjecture for
arbitrary complete manifolds in dimensions ge4.
Fix a metric space X. For dge 0 define its Urysohn d-width
UW_d(X)=inf_f sup_{yin K}diam_X f^{-1}(y),
where the infimum runs over continuous maps f:Xto K to simplicial
complexes of dimension at most d, and where diam_X is the diameter
of the entire fiber in X. Empty fibers are omitted from the sup.
[Main bound]
For every integer nge4 there is a finite constant C_n, depending
only on n, with the following property. Every connected complete
smooth Riemannian n-manifold (M,g) without boundary and with
Scal_gge1 admits a continuous map f: Mto K to a simplicial
complex of dimension at most n-2 such that
[
diam_g f^{-1}(y)le C_n quadfor every yin K.
]
In particular, UW_{n-2}(M,g)le C_n.
The constant C_n is independent of M, its topology and its
geometry. No orientability, spin, covering, or bounded-geometry
condition enters. If Scal_ggesigma^2 for some sigma>0, the
change of metric grightsquigarrowsigma^2 g gives
UW_{n-2}(M,g)le C_n/sigma.
For a closed connected Riemannian manifold, Gromov's filling radius
[cite] is defined via the Kuratowski embedding and
the fundamental class; see cite[Definition~3.1]{nabutovskyRotmanSabourau2021}.
The comparison between Urysohn width and filling radius is recorded
in cite[Definition~6.1, eq.~(6.1)]{nabutovskyRotmanSabourau2021}.
[Filling radius]
Let (M,g) be closed, connected, smooth, of dimension nge4, with
Scal_ggesigma^2 for some sigma>0. Then
FillRad(M,g)le C_n/(2sigma).
Proof.
A target of dimension le n-2 is also admissible for UW_{n-1}.
Because M is closed, cite[eq.~(6.1)]{nabutovskyRotmanSabourau2021}
gives FillRad(M,g)lefrac12UW_{n-1}(M,g), and the
width comparison UW_{n-1}(M,g)leUW_{n-2}(M,g) follows from the
definition (eq:urysohn). Apply cref{thm:main} to sigma^2 g.
For a metric space X, define its continuous macroscopic
dimension dim_{mc}(X)le d to mean: there is a continuous
map Xto K to a simplicial complex of dimension at most d, carrying
the uniform simplex metric (all simplices standard Euclidean of common
edge length 1), such that every entire fiber has bounded diameter.
[Universal covers]
Let (M,h) be closed, connected, smooth, of dimension nge2, with
Scal_h>0. Then dim_{mc}(widetilde M,widetilde h)le n-2
for the lifted metric on the universal cover.
[Aspherical obstruction]
No closed smooth aspherical manifold of dimension nge4 admits a
Riemannian metric of strictly positive scalar curvature.
The proof of cref{cor:aspherical} from cref{thm:main} is the same
width/filling-radius exclusion used by Gromov cite[S3]{gromov2017Dozen};
we present it in cref{sec:closed-covers}.
## History and context
Gromov introduced macroscopic dimension in [cite] to study
positive scalar curvature. The conjecture already asks whether a
complete manifold with a positive lower bound on scalar curvature can
be mapped continuously to a polyhedron of codimension two with all
point inverses bounded in terms of the curvature scale; see
cite[S2frac12]{gromov1996} and the discussion around
cite[p.~3180]{gromov2019Mean}. The diameter condition concerns each
entire fiber including all components. The continuous form of the
invariant and the version requiring a Lipschitz map to a uniformly
metrized complex are distinct; see
cite[Definitions~1.1--1.2]{dranishnikov2014}.
The pointwise hypothesis cannot be weakened to a fixed-scale macroscopic
bound: Kumar and Sen construct closed manifolds in every dimension nge4
with prescribed positive macroscopic scalar curvature at a fixed scale
but arbitrarily large Urysohn codimension-two width
cite[Theorem~A]{kumarSen2026Width}.
Prior results in low dimensions are due to Gromov--Lawson
cite[Corollary~10.11]{gromovLawson1983}, Bolotov [cite]
for the three-dimensional macroscopic-dimension obstruction, and
Liokumovich--Maximo cite[Theorem~1.1]{liokumovichMaximo2023} and
Liokumovich--Wang cite[Theorem~1.1]{liokumovichWang2023} for curvature-scale
width bounds. The aspherical obstructions in dimensions four and five
were proved by Chodosh--Li cite[Theorem~2]{chodoshLi} and independently
by Gromov cite[S7]{gromov2020Aspherical}.
Topological approaches to the codimension-two conclusion, conditional on
the Strong Novikov Conjecture and injectivity of the connective-to-periodic
real K-homology map, are due to Bolotov--Dranishnikov
cite[Theorem~4.3]{bolotovDranishnikov2010,bolotovDranishnikov2010,bolotovDranishnikov2016}
and Daher--Dranishnikov cite[Theorem~3.3]{daherDranishnikov2019}.
Those results concern universal covers with a finite (but unspecified)
bound, whereas cref{thm:main} covers all complete manifolds with a
bound depending only on n after scalar normalization.
The geometric method is that of Schoen--Yau's minimal-hypersurface
obstruction cite[Section~1]{schoenYau1979Structure} combined with
Gromov's torical symmetrization cite[Section~2]{gromov2018Metric}
and the mu-bubble method of Chodosh--Li cite[Section~3 and
Remark~11]{chodoshLi}. The shielding and regularization here are
inspired by Lohkamp's splitting technology
cite[S1.2 and Theorem~3.15]{lohkamp2022SplittingsII}.
## Strategy of the proof
Fix nge4 and a complete Riemannian n-manifold (M,g) with
Scal_gge1. Choose a scale D depending only on n (specified in
cref{sec:rounds}). Let (x_i)_{ige1} be a maximal D-separated
subset of M; then the closed balls A_i=overlineoverline B_g(x_i,D)B_g(x_i,D) cover
M and the outer balls overlineoverline B_g(x_i,2D)B_g(x_i,2D) are locally finite. Set
B_i=M∖ B_g(x_i,2D).
The proof has three big steps.
medskipnoindent
Step 1 (stretching). Find a complete metric hge g and a
smooth function t such that for every i the h-distance between
A_i and B_i exceeds a prescribed L_i, while a certain drift
scalar remains bounded below:
D(h,t):=calD(h,t):=Scal_h+2lap_h t-|dd t|_h^2getfrac12.
Only finitely many labels require any particular finite width L_i;
labels with larger L_i are ``spread out'' over more rounds. Summing
sum_i L_i^{-1}<∞ (see cref{sec:rounds}) keeps total trace cost
finite and makes the round construction locally finite.
medskipnoindent
Step 2 (cutting). For each label i, find a hypersurface
W_isubset M separating the q-preimage of A_i from the
q-preimage of B_i, together with a modification of metric and
function that keeps D on the resulting pieces.
Singularities of W_i in dimensions 8 and up force a nontrivial
shielding step: the pressure profiles of the remaining labels
are enlarged in a thin annular shell near each singular point so that
the future scalar requirement holds only away from the singular set.
This uses geometric measure theory (compactness, regularity, tangent
cones) plus a Hardy inequality on thickened annuli. A ``warped-circle''
identity restores positive scalar on each resulting packet by paying
through 399 trivializing circle factors.
medskipnoindent
Step 3 (assembly). Choose coordinates t_i: Mto[0,1]
that equal 1 on q^{-1}(A_i), 0 on q^{-1}(B_i), and transition
only inside a small neighbourhood of W_i. The key combinatorial
observation is that no n-1 of the W_i can share a common point:
if they did, their mutual intersection would carry a complete
one-dimensional warped-product packet satisfying Scalge lambda_r>0
on a finite warped product, which is impossible (the Riccati
blockage). Hence at every point at least one coordinate equals 1 and
at most n-2 are strictly inside (0,1). The coordinates assemble
into a continuous map to a simplicial complex K of dimension n-2.
If yin|K|, two points in the fiber share a coordinate equal to 1,
so both collapse images lie in the same outer ball of radius 2D;
their g-distance is at most 4D+2.
The constant is C_n=4D+2, with D=D(n) fixed in cref{sec:rounds}.
We do not optimize C_n.
## Convention on signs and densities
The Laplacian is lap=Divgrad, so round spheres of scalar curvature
+1 have positive scalar curvature in the present convention.
For a hypersurface with a chosen unit normal nu, its second
fundamental form is B(v,w)=<nabla_vnu,w> and its mean
curvature is H=tr B. Integration and divergence use Riemannian
densities; no orientation is required. Distances in the final fiber
bound are always measured in the original metric g.
# The drift scalar and warped products
The deformation in cref{sec:stretch} replaces a metric h and a function
t by
[
hrightsquigarrow h_u=h+phi_u,dd u^2, trightsquigarrow t_u=t-beta_u,
]
where u is a smooth height function, phi_u and beta_u are
appropriate scalar functions of the gradient length s=|dd u|_h, and
the change is engineered so that the drift scalar
[
D(h,t):=calD(h,t):=Scal_h+2lap_h t-|dd t|_h^2
]
strictly increases. The point of the extra function t is that the drift
lap_h t-|dd t|_h^2 is what the warped-product construction in
cref{sec:rounds} uses to introduce diagonal circle factors with positive
scalar curvature; the drift is the ``reserve'' that pays for those factors.
## Warped product scalar curvature
For smooth functions p_1,...,p_k on a Riemannian manifold (E,h), put
widehatwidehat h=h+h=h+sum_{alpha=1}^k e^{2p_alpha}ddtheta_alpha^2 on
EtimesT^k. Write P=sum_alpha p_alpha. Then
Scal_{widehatwidehat h}h=Scal_h-2lap_h P-|dd P|_h^2-sum_{alpha=1}^k|dd p_alpha|_h^2.
Equivalently
Scal_{widehatwidehat h}h=D(h,P)-calD(h,P)-sum_alpha|dd p_alpha|_h^2
when t=P.
Proof. [Proof of (eq:warp-identity)]
Let E_alpha=e^{-p_alpha}∂_{theta_alpha} be the unit vertical
fields. For horizontal fields X,Y, Koszul gives
widehatnabla_XY=nabla^h_XY, widehatnabla_XE_alpha=0,
widehatnabla_{E_alpha}X=(Xp_alpha)E_alpha,
widehatnabla_{E_alpha}E_alpha=-grad_h p_alpha, and
widehatnabla_{E_alpha}E_beta=0 for alphanebeta.
Hence
K(X,E_alpha)=-Hess_h p_alpha(X,X)-(Xp_alpha)^2
and K(E_alpha,E_beta)=-<dd p_alpha,dd p_beta>.
Summing twice over orthonormal 2-planes gives (eq:warp-identity).
## The graph deformation
Fix s_0=10^{-2} and a smooth nondecreasing rho:[0,∞)to[0,1]
with
[
rho=0 on [0,s_0], rho=1 on [5s_0,∞),
0lerho'le 60.
]
For 0<alefrac12 and sge0 put
mu=a^2+(1-a^2)rho(s), theta=(mu s^2)/(1+s^2),
l=(1-theta)^{1/2}, beta=log l.
Near s=0 we then have muapprox a^2 and thetaapprox a^2 s^2to0;
near s=∞, mu=l=1 and thetato1. The three regimes
``s tiny, s moderate, s large'' will be treated separately in the
proof of cref{thm:stretch}.
The graph metric and drift corresponding to a provisional height u are
h_u=h+(theta)/((1-theta)s^2)dd u^2
=h+(mu)/(1+(1-mu)s^2),dd u^2,
t_u=t-beta.
These expressions are smooth at s=0; near zero slope the metric
increment is a^2,dd u^2+O(s^4) and the drift increment is
-tfrac12log(1+(1-a^2)s^2)+O(s^2). Hence h_uge h and t_uge t
for all sge 0. The alternative expression for the metric in
(eq:raw-deform) records that the graph of u is measured in
the product manifold EtimesR with its product metric
h+dd z^2: the graph of u has induced metric h+dd u^2, but
(eq:raw-deform) rescales the vertical direction in a
slope-dependent way.
The reason for the specific (mu,theta,l,beta): at every point of
nonzero slope, h_u+dd z^2 is conformal to the graph metric
h+dd u^2+dd z^2 (with z a formal vertical coordinate) with a
conformal factor depending only on s=|dd u|_h. On such a graph the
various scalar-curvature terms reorganize in a computationally useful way.
The drift t_u=t-beta is tuned so that the drift scalar
D(h_u,t_u)calD(h_u,t_u) exceeds D(h,t)+(t_u^2-t^2)calD(h,t)+(t_u^2-t^2) by a
nonnegative amount, up to the small errors controlled in cref{thm:stretch}.
The key structural input is the following identity. In local normal
coordinates adapted to u, write h=v^2dd q^2+gamma_q with v=1/s
and q=u, and define the second fundamental form mathrm JrmJ and
mean curvature H=trmathrm JrmJ of the level sets in the metric
gamma_q. Put S=H-e t with e=∂_q/(2v) the normal derivative.
The Gauss equation and the normal variation formula for H then give:
[Raw scalar identity]
Let E be a smooth manifold, h a Riemannian metric, u a smooth
function with s=|dd u|_h>0, tge T a smooth function, and define
h_u, t_u by (eq:raw-deform). Then
D(h_u,t_u)-calD(h_u,t_u)-D(h,t)calD(h,t)
=theta,E(mathrm J,H,F,rmJ,H,F,xi,t)+O(theta E_{rm cutoff}),
where E is a function of the second fundamental form, the ``pressure''
F=F(xi,t) with xi=(u-psi)/eta for suitable cutoff data
psi,eta, and where for small slope s<s_0 the identity holds with
the same form for a suitably reduced E.
The exact form of E will be given in the proof of cref{thm:stretch}
below. Its crucial feature: on a favourable branch of the pressure F,
E is bounded below by a positive multiple of s · F_x, and this
drives both the scalar increase and the distance enlargement.
We now state the pressure law which supplies the function F
with the exact inequalities needed to make E favourable.
# The pressure law
The pressure is a smooth function F(xi,t) on a strip
(-S_0,S_0)times[T,∞) with F odd in xi and positive slope in
xi. Its role: in the raw scalar identity cref{lem:raw-scalar},
the second fundamental form enters through a quadratic form whose
coefficients are shaped by derivatives of F. We need F to be steep
in xi for large drift t (so that stretching is fast when we want
it), but with coupling of drift back to the vertical direction that is
not too strong. The exact requirements are the following.
[Pressure law]
For every choice of A_0,P_0,eps,b_0>0 there exist 0<S_0le1, Tge1,
and a smooth function
[
F:(-S_0,S_0)times[T,∞)toR
]
which is odd in the first variable, has F_xi>0, and satisfies for all
(xi,t)in(-S_0,S_0)times[T,∞):
[label=(roman*)]
Slope growth. F_xi(xi,t)ge A_0 t.
Bounded coupling. |F_t(xi,t)|leeps|F(xi,t)|
and F F_tleeps F_xi.
Favourable branch or quadratic loss.
Either F_xile b_0 t^2 (a quadratic upper bound), or
-FF_tge P_0 F_xi.
Quadratic ceiling. F_xile b_0 t^2+eps F^2.
For every fixed tge T,
lim_{xito-S_0}F(xi,t)=-∞ and
lim_{xito S_0}F(xi,t)=+∞, uniformly on compact t-intervals.
The data S_0,T depend only on (A_0,P_0,eps,b_0) and fixed numerical
constants, not on any manifold, metric, or subsequent deformation
parameter.
The idea of the proof is to construct F by prescribing its inverse
xi = X(f,t) in the pressure variable f=F(xi,t), in such a way that
the analytic structure of X as a function of f,t reveals the desired
inequalities on the original F. This is possible because the pressure
law lives in one dimension above R, so no geometric input is
required.
Proof. [Proof of cref{thm:pressure}]
Let k>0 and a positive smooth function g:[0,∞)to(0,∞)
be auxiliary data. For f>0 and t>0 set
X(f,t)=frac1karctanfrac{f^{1/3}}{g(y)},
q=f^{2/3}, y=frac tq.
Differentiation at fixed f and at fixed t, respectively, gives
fX_t=-(g'(y))/(k(1+g(y)^2/q)),
X_f=(g(y)+2yg'(y))/(3kq^2(1+g(y)^2/q)).
Thus invertibility of X(·,t) in f requires the quantity
A(y):=g(y)+2yg'(y)>0 for all yge0. For the inverse F we have
F_xi=1/X_f,
(FF_t)/(F_xi)=(g'(y))/(k(1+g(y)^2/q)),
(F_t)/(F)=(3yg'(y))/(t(g(y)+2yg'(y))).
We design g so that it decreases for small y (favourable branch),
eventually equals y (giving Xapprox k^{-1}arctan(f/t), hence
Fapprox ttan(kxi)), and stays above y in between, with A(y)>0
throughout. The precise choice is a smooth rounded maximum:
fix a smooth nonnegative probability density rho supported in [1,2],
positive on (1,2) and vanishing to infinite order at both ends.
Let mu=int_1^2 rrho(r)dd rin(1,2) and, for D_g>0,
h(y)=D_g(1+y)^{-1/4},
g(y)=int_1^2max{y,rh(y)}rho(r)dd r, yge0.
The function g is smooth and nondecreasing; direct differentiation of
the integral gives the two-sided estimates
g(y)ge y,
-(g(y))/(4(1+y))le g'(y)le 1,
A(y)ge(g(y))/(2).
In particular A>0 everywhere. For yle Y with Y(1+Y)^{-1/4}le D_g,
the ``max'' is realized by rh(y) on the range rin[1,2], giving the
exact formula
g(y)=mu D_g(1+y)^{-1/4},
g'(y)=-frac{mu D_g}{4(1+y)^{5/4}}, yle Y.
Thus g decreases at a rate controlled by D_g throughout [0,Y].
Choose the constants in the following order: first
kgemax{A_0,eps^{-1},pi/2}, then Yge1 large with
Y^{-3}le b_0/(12k), then D_g>0 large enough that
D_ggemax{Y(1+Y)^{1/4},;8kP_0(1+Y)^{5/4}/mu}, and finally
Tgemax{1,6/eps,12kB^2/b_0,YG_0^2,12kY/(eps g_0)}, where
g_0=min[0,Y]g, G_0=max[0,Y]g, and B=sup_{y>Y}g(y)/y<∞.
Now verify the four inequalities.
(i) Slope growth. From (eq:pressure-inverse-deriv) and
gge y together with Ale 3g, we get
t X_f= yA/(3k(q+g^2))le 1/k.
Hence F_xi=1/X_fge ktge A_0 t.
(ii) Bounded coupling. From (eq:pressure-from-inverse) and
the bound |yg'|le g, we get |F_t/F|le 6/tleeps for tge Tge6/eps.
The sign of FF_t/F_xi equals the sign of g'; on the branch
g'le1, so FF_tle F_xi/kleeps F_xi.
(iii) Favourable branch or quadratic. In the region y>Y:
Age g/2ge y/2 and g/yle B, so
t^2X_fge y^3/(6k(1+B^2y^3/T))ge 1/(6k(Y^{-3}+B^2/T))ge 1/b_0.
This gives F_xile b_0 t^2 in the region y>Y. In the region
0<yle Y, the rate (eq:pressure-g-decreasing) together
with D_gge 8kP_0(1+Y)^{5/4}/mu gives
-g'ge 2kP_0; hence -FF_t/F_xige P_0, the favourable branch.
(iv) Quadratic ceiling. Again from (eq:pressure-from-inverse),
t^2 X_fge 1/b_0 in the large-y region gives F_xile b_0t^2.
In the small-y region we instead bound F^2 X_fge 1/eps
using F^2X_fge qg_0/(12k)ge g_0T/(12kY)ge 1/eps,
and hence F_xileeps F^2le b_0 t^2+eps F^2.
Pole limits. As fto0^+: yto∞, g(y)=y, so
X(f,t)to0. As fto∞: yto0, g(y)to g(0)>0, so
X(f,t)topi/(2k)=:S_0le1. Hence X(·,t) maps (0,∞)
bijectively to (0,S_0), and its inverse F extends oddly to
(-S_0,S_0) and is smooth at xi=0 because near f=0 the formula
reduces to X=(1/k)arctan(f/t), F=ttan(kxi). Proper pole limits
follow by taking f (respectively 1/f) large. Uniformity on
compact t-intervals: max_{tin I}X(H,t)<S_0, so
xi exceeding that max gives F(xi,t)>H for all tin I.
When A_0,P_0,eps,b_0 are prescribed as functions of the ambient
dimension n and fixed numerical constants, S_0 and T are
themselves functions of n. In particular, a and d may be chosen
after cref{thm:pressure} has been applied, without modifying
the pressure itself. This is what allows the local stretching theorem
to put its own small constants a,d on top of the pressure law.
# Local stretching
The stretching theorem pays for enlarging the distance between a
compact set A and a closed set B by three currencies: quadratic
drift reserve t^2, trace loss tr_h h_0 -tr_{h'}h_0, and a small
quadratic error d,t'^2.
[Local stretching]
Fix nge2, 0<delta<1/10, and put L_*=1.004. There are
W,Tge1, depending only on n,delta, with the following property.
Let (M,h_0) be a complete smooth n-manifold, let hge h_0 be
smooth, let tge T be smooth, and let Asubset M be compact and
Bsubset M closed, with dist_{h_0}(A,B)ge W. For every
0<d<delta there exist smooth h'ge h and t'ge t, with
changes compactly supported in {x:dist_{h_0}(A,x)<W}, such that
D(h',t')&calD(h',t')&geD(h,t)+(t'^2-t^2)calD(h,t)+(t'^2-t^2)notag
&-delta(tr_h h_0-tr_{h'}h_0)t'^2-d t'^2,
dist_{h'}(A,B)&ge L_*W.
In particular h' is complete. Neither W nor T depends on d,
(M,h_0),A,B.
The proof occupies cref{sec:stretch-proof}; it uses the graph
deformation of cref{sec:drift} and the pressure law
cref{thm:pressure}. Here is the punchline, in three steps.
Step A (setup). Fix A,B,h,h_0,t as in the statement. Let
r_A(x)=dist_{h_0}(A,x). Choose smooth functions psi,eta_1,zeta
with
[
inf_Apsi>W/5, |ddpsi|_{h_0}le1,
supppsisubset{r_A<W/2},
]
[
eta_1=1 on {r_Ale.55W},
eta_1=0 on {r_Age.8W},
|ddeta_1|_{h_0}le s_0/8,
]
[
zeta=1 on {r_Ale.82W},
zeta=0 on {r_Age.98W}.
]
Set eta=r+(1-r)eta_1^2. Then eta=1 near supppsi and
eta=r on {r_Age.8W}, and the uniform bounds
0<etale1, |ddeta|_hle s_0/4, |ddeta|_h^2/etale1 hold
independently of r. Enclose {r_Ale1.1W} in a compact smooth domain
N and extend h,t to N; work entirely in N from now on.
Step B (graph equation). For tauin[0,1] consider the
equation
Div_hA( u)-dd t(A( u))
=J(|dd u|_h)F((u-taupsi)/(eta),t),
where A(p)=+|p|_h^2)^{-1/2}p^sharp and
J(s)=√(theta+a^2s_0^2), with mu,theta from
(eq:switch), and F from cref{thm:pressure}. Note A
and J are smooth at p=0 because mu(0)=a^2 constants.
For each fixed positive floor r>0 and tau=1, this equation has
a smooth solution u=u_r with quantitative bounds
|u_r-psi|le S_1eta and |dd u_r|_hle G, independent of r
and tauin[0,1]. This follows from the maximum/comparison barriers
of cref{thm:pressure} (the F poles force u_r into the interval
[-psi+S_1eta,psi+S_1eta]) combined with a gradient estimate
(the maximum of w=(1+|dd u|_h^2)^{-1/2} obeys a positive lower
bound from an exact Jacobi identity for a graph in (N,h)timesR;
the drift term contributes a bounded coefficient and the pressure
term dominates). Existence in tau follows by continuation: at
tau=0 the solution is u=0; the linearization at any solution
is an isomorphism C^{2,alpha}to C^{0,alpha} because of the
strictly negative zero-order term -JF_x/eta combined with the
uniform ellipticity; the implicit function theorem in C^{2,alpha}
gives openness in tau and the a priori bounds give closedness.
As r↓0, on the compact region where psi=0 and eta_1=0
(i.e. eta=r), u_rto0 in every C^k norm. Indeed
|u_r/r|le S_1 and the equation becomes
a^{ij}∂_{ij}u_r+b^i∂_i u_r-(c_r/r)u_r=0 with
c_rge as_0A_0T>0. Comparison against the explicit barrier
V_r=M_0sum_fexp(-gammaell_f/sqrt r) shows
|u_r|le Cexp(-c/sqrt r) on compact subsets; the derivative
convergence follows from the rescaled equation
∂_{y_i}mathfrak A^i(x_0+ry,Dv)=rfrakA^i(x_0+ry,Dv)=rmathfrak B_0(x_0+ry,v,Dv)frakB_0(x_0+ry,v,Dv)
applied to u_r(x)=rv(y) at x=x_0+ry.
Step C (raw scalar estimate). At every point of nonzero slope,
the raw deformation h_u=h+phi_udd u^2, t_u=t-beta satisfies,
up to constants that are paid by the small parameters a,d,
D(h_u,t_u)calD(h_u,t_u)geD(h,t)+(t_u^2-t^2)calD(h,t)+(t_u^2-t^2)
-delta(tr_h h_0-tr_{h_u}h_0)t_u^2-(d)/(2)t^2.
This is the heart of cref{thm:stretch} and is proved in
cref{sec:stretch-raw} below.
Step D (cutoff and distance). Define h'=h+zeta(h_u-h),
t'=t+zeta(t_u-t), extended by (h,t) off N. If the positive
floor r is chosen small enough, the discrepancies from the raw
estimate (eq:stretch-raw) caused by zeta are absorbed on the
compact annulus S=overline{{0<zeta<1}}subset{.8W<r_A<W};
on the rest of N the output is either the raw deformation (where
zeta=1) or the unchanged original (where zeta=0), and both
already satisfy (eq:stretch-conclusion). The distance gain
(eq:stretch-distance) uses the effective height
v_{rm eff}=zeta u: on A it exceeds W/5-1; on B it is zero;
on the small slope region {sle5s_0=.05} its variation is at most
.05 W; everywhere h'ge h so a path realizing length <L_*W has
h-length < L_*W and therefore h_0-length ge W; the variation
of v_{rm eff} along any such path is at most .01L_*W on
{0<zeta<1}, so the variation on the region
{zeta=1,sge.05} exceeds (1/5-.06L_*)W-1>.11W; the
inequality √(W^2+(.11W)^2)=W√(1+.11^2)>L_*W contradicts
Length_{h'}(gamma)<L_*W. Completeness of h'
follows from h'ge h_0 and local equivalence: an h'-Cauchy
sequence is h_0-Cauchy, hence converges; near its limit h' and
h_0 are equivalent smooth metrics. This completes
cref{thm:stretch}. QED
# A pressure law with a quadratic budget
We need a smooth pressure F(x,t) used in cref{sec:local-step}: it must
be steeply increasing in x, grow linearly in t, and have a
non-linear derivative budget in t.
[Pressure law]
Fix A_0,P_0,eps,b_0>0. There exist 0<S_0le 1, Tge 1, and a
smooth function
[
F:(-S_0,S_0)times[T,∞)→R
]
which is odd in x, has F_x>0, and satisfies everywhere on its
domain
&F_xge A_0 t, |F_t|leeps|F|, FF_tleeps F_x,
&F_xle b_0t^2quadorquad -FF_tge P_0F_x,
&F_xle b_0t^2+eps F^2.
For each fixed tge T,
[
lim_{xto-S_0}F(x,t)=-∞,
lim_{xto S_0}F(x,t)=+∞,
]
and the convergence is uniform for t in any compact subset of
[T,∞).
The constants S_0,T and the function F depend only on the numerical
inputs. They are fixed before any metric, manifold, cutoff, or
deformation parameter is chosen.
## Why we ask for exactly this
In cref{sec:local-step} the unknown is a height function u on a
closed manifold N, solving the elliptic equation
Div_hA( u)-dd t(A( u))
=J(|dd u|_h),F!((u-taupsi)/(eta),t).
Here A is an operator built from an auxiliary ``slope switch''
and J>0 is a scalar factor. The argument of F lives in |x|<S_0.
The inequality F_xge A_0t makes F very steep in x. A
maximum-principle argument (see cref{lem:height-gradient}) then forces
u to track the prescribed height taupsi up to a fixed margin,
thereby controlling dd u. Without that steepness, u could drift
far from taupsi and the metric change would be uncontrolled.
The two t-inequalities |F_t|leeps|F| and FF_tleeps F_x
control the price of the local stretching deformation: when the
deformation increases a background trace by a small amount, the drift
t absorbs the resulting scalar-loss with error at most eps in
units of F.
The two-branch alternative in the second line selects a small-slope
regime (first branch, F_x small) from a large-slope regime (second
branch, -FF_t large relative to F_x); the third inequality, the
quadratic budget F_xle b_0t^2+eps F^2, closes the analysis in the
transition region where neither of the two branches is decisive.
The construction is by an inverse coordinate: instead of F
directly we prescribe X(·,t)=F(·,t)^{-1} and read the
derivative inequalities from the inverse-function formulas. Almost
all of the work is choosing an auxiliary g so that both branches
of the second inequality hold.
## The inverse construction
Let k>0 and g:[0,∞)to(0,∞) be smooth positive
functions to be fixed. For a pressure value f>0 and a drift t>0
define a height
X(f,t)=frac1karctan!(frac{f^{1/3}}{g(y)}),
y=frac{t}{f^{2/3}}.
We will see that for fixed t the map X(·,t) is strictly
increasing in f, with limit 0 as fto 0 and limit S_0=pi/(2k)
as fto∞. Hence F(·,t)=X(·,t)^{-1} is well defined
and smooth on (0,S_0); extending oddly to (-S_0,0) and setting
F(0,t)=0 gives the full pressure.
For small f we have g(y)=y for large y, so
X(f,t)=(1/k)arctan(f/t) and F(x,t)=ttan(kx) near x=0; this
is jointly smooth in (x,t).
The two branches of the second estimate come from two regimes of g.
For small y we make g decreasing so that -FF_t/F_x is large; for
large y we set g(y)=y so that F_xle b_0t^2 holds. The transition
region uses the quadratic budget instead.
## Construction of the auxiliary function g
Let rho:[1,2]to[0,∞) be a smooth nonnegative density
supported in [1,2], positive in (1,2), vanishing to infinite
order at both endpoints, with int_1^2rho(r)dd r=1. For example
[
rho(r)=c_rhoexp(-1/((r-1)(2-r))), 1<r<2,
]
extended by zero. Write mu=int_1^2 rrho(r)dd rin(1,2).
For a large parameter D_g>0 set h(y)=D_g(1+y)^{-1/4} and
g(y)=int_1^2max{y,rh(y)}rho(r)dd r, yge 0.
This is a smoothed maximum of the constants rh(y) over rin[1,2].
Equivalently, with M(z)=int_1^2max{z,r}rho(r)dd r,
g(y)=h(y)M(y/h(y)). Since M''=rho (with rho extended by
zero), M is smooth, and M(z)=mu for zle 1, M(z)=z for
zge 2. Therefore g>0, gge y, and g(y)=y for all
sufficiently large y.
Write L(z)=int_{rle z}rho(r)dd r and
R(z)=int_{rge z}rrho(r)dd r. At z=y/h(y),
g'=L(z)-(hR(z))/(4(1+y)).
This gives
gge y, -(g)/(4(1+y))le g'le 1,
g+2yg'ge(y+2)/(2(y+1))gge tfrac12 g.
The last inequality is the key one: the combination g+2yg' that
appears in the inverse derivative formula is comparable to g itself.
## Order of constants
Precisely, choose
kge max{A_0,eps^{-1},pi/2},
Yge 1quadwithquad Y^{-3}le (b_0)/(12k).
Choose D_g only after k,Y:
D_gge max!{Y(1+Y)^{1/4},,
frac{8kP_0(1+Y)^{5/4}}{mu}}.
Since y(1+y)^{1/4} is increasing and
y/h(y)=y(1+y)^{1/4}/D_gle 1 for yle Y, on [0,Y]
g(y)=mu D_g(1+y)^{-1/4},
g'(y)=-frac{mu D_g}{4(1+y)^{5/4}}le -2kP_0.
Finally define
g_0=min_{0le yle Y}g(y),
G_0=max_{0le yle Y}g(y),
B=sup_{y>Y}g(y)/y, and choose
Tgemax!{1,(6)/(eps),(12kB^2)/(b_0),
YG_0^2,(12kY)/(eps g_0)}.
The order k, then Y, then D_g (which fixes g), then T,
is respected.
## Derivatives of the inverse
By cref{eq:X-def},
f X_f=(A(y))/(3k(1+g(y)^2/q)),
A(y)=g(y)+2y g'(y),
where q=f^{2/3}, y=t/q. By cref{eq:g-inequalities} we have
Age g/2>0, hence X_f>0. Set F(·,t)=X(·,t)^{-1}. Then
F_x=frac1{X_f}, F_t=-(X_t)/(X_f),
and
(FF_t)/(F_x)=(g'(y))/(k(1+g(y)^2/q)),
(F_t)/(F)=(3y,g'(y))/(t(g(y)+2y,g'(y))).
## The first three inequalities
F_xge A_0 t.: Since Ale g+2yle 3g and yle g,
[
t X_f=(y A)/(3k(q+g^2))le frac1k,
]
hence F_x=1/X_fge k tge A_0t by cref{eq:k-Y-choice}.
|F_t|leeps|F|.: From cref{eq:pressure-ratios}, and
from |y g'|le g (for g'ge 0 use g'le1 and yle g; for
g'<0 use |y g'|le g/4),
[
|(F_t)/(F)|le(3|yg'|)/(t g)lefrac3tle(eps)/(2)leeps
]
because tge Tge 6/eps.
FF_tleeps F_x.: Again from cref{eq:pressure-ratios}:
if g'le 0, FF_tle 0le eps F_x; if g'>0, then
[
(FF_t)/(F_x)le(g')/(k)le(1)/(k)le eps,
]
since kgeeps^{-1}.
## The two branches of the alternative
We prove: for every (f,t) either F_xle b_0t^2 or
-FF_t/F_xge P_0.
Large y>Y.: Here Age g/2ge y/2 and g/yle B, so
[
t^2X_f=(y^2A)/(3k(1+yg^2/t))
ge (y^3)/(6k(1+B^2y^3/T))
=frac{1}{6k(y^{-3}+B^2/T)}
gefrac{1}{6k(Y^{-3}+B^2/T)}.
]
By cref{eq:k-Y-choice}, Y^{-3}le b_0/(12k); by cref{eq:T-choice},
B^2/Tle b_0/(12k). Hence t^2X_fge 1/b_0, i.e.
F_x=1/X_fle b_0t^2. First branch.
Small 0<yle Y.: Then q=t/yge T/Y and
g^2/qle YG_0^2/Tle 1 by cref{eq:T-choice}. Combined with
cref{eq:g-decreasing} and cref{eq:pressure-ratios},
[
-(FF_t)/(F_x)=(-g')/(k(1+g^2/q))
ge(2kP_0)/(2k)=P_0.
]
Second branch.
## The quadratic budget
For F_xle b_0t^2+eps F^2 again split on y. When y>Y the first
branch already gives it. When 0<yle Y use gge g_0 on [0,Y] and
q=t/y:
[
F^2X_f=(q A)/(3k(1+g^2/q))
ge(qg_0)/(12k)ge(g_0T)/(12kY)gefrac1eps,
]
so F_x=1/X_fleeps F^2le b_0t^2+eps F^2.
## Pole limits and uniformity
By construction F(·,t)=(X(·,t))^{-1} gives the
limits at pm S_0: as fto∞, X(f,t)topi/(2k)=S_0, hence
F(x,t)to∞ as xto S_0; oddness gives the limit at -S_0.
Uniformity follows from: for any compact Isubset[T,∞) and
H>0, max_{tin I}X(H,t)<S_0, so x>S_0-delta yields
F(x,t)>H for every tin I.
The proof of cref{thm:pressure} is complete.
## The raw scalar computation
We now supply the missing computation in Step C of the proof of
cref{thm:stretch}.
Suppose first s=|dd u|_h>0 at the point under consideration. Use
the level sets of u as coordinate hypersurfaces: locally write the
metric as
h=v^2dd q^2+gamma_q, q=u, v=frac1s.
Let e=(1/v)∂_q be the unit normal field in the q direction.
Denote by mathrm JrmJ the second fundamental form of the level sets
with respect to e, by H=trmathrm JrmJ its mean curvature, and by
mathrm e=∂_q/(2v)=∂_q/2rme=∂_q/(2v)=∂_q/2 the corresponding q-derivative
operator when acting on functions that depend only on q and coordinates
on the level set. For brevity write mathrm e(t)rme(t) for the derivative
of a function t along e.
The Gauss equation for the level-set foliation gives
Scal_h=Scal_gamma-|mathrm J|^2-H^2-2rmJ|^2-H^2-2,mathrm e(H)-2(rme(H)-2(lap_gamma v)/(v),
where Scal_gamma is the scalar curvature of gamma_q. The Laplacian
identity for a foliation gives
lap_h t=lap_gamma t+mathrm e(rme(mathrm erme t)+H,mathrm erme t
+<dd_gammalog v,dd_gamma t>,
where dd_gamma and lap_gamma are taken on the level set with
metric gamma_q. Substituting (eq:gauss-sc), (eq:lap-foliation)
into the definition of D gives:
D(h,t)&=-|calD(h,t)&=-|mathrm J|^2-S^2-2rmJ|^2-S^2-2,mathrm eS+rmeS+Scal_gamma
+2lap_gamma t-|dd_gamma t|^2
&-2(lap_gamma v)/(v)
+2<dd_gammalog v,dd_gamma t>,
where
[
S:=H-mathrm e(t)rme(t)
]
is the mean curvature of the ``graph with slope t''.
The deformation (eq:raw-deform) replaces v by v/l and t by
t-beta. In the tangential part of (eq:foliation-scalar), the
change is
[
2<dd_gammalog v,dd_gammabeta>
-|dd_gammabeta|^2
=c_beta(2-c_beta)|dd_gammalog v|^2ge0,
c_beta=-sbeta_sin[0,2].
]
The identity follows by expanding lap_gamma(v/l)/(v/l) and
|dd_gamma(t-beta)|^2: the lap_gammabeta terms and the
dd_gamma t-linear terms cancel, leaving exactly the displayed
coefficient. Hence these tangential contributions are nonnegative and
can be discarded.
For the normal terms we use the divergence equation that u_r was
chosen to satisfy. We may write it as
S+mathrm ermelogsqrttheta=chi F(xi,t),
where chi=√(1+a^2s_0^2/theta)in[1,2] for sge s_0 and
chige s_0/s for 0<s<s_0, and where
xi=(u-psi)/eta. Define M=chi F and K=(l^2/theta),mathrm ermebeta.
Because mathrm ermelogsqrttheta=-K, we have S=M+K. For the new
metric, mathrm e'=lrme'=l,mathrm e,rme, mathrm J'=lrmJ'=l,mathrm J,rmJ, so
[
S'=l(S+mathrm ermebeta)=lM+K/l.
]
Substituting S=M+K and S'=lM+K/l into the normal terms
-|mathrm J|^2-S^2-2rmJ|^2-S^2-2,mathrm eSrmeS (or its primed version) and
accounting for the change vto v/l in -2lap_gamma v/v and the
change tto t-beta in +2lap_gamma t-|dd_gamma t|^2 gives
D(h_u,t_u)-calD(h_u,t_u)-D(h,t)calD(h,t)getheta E,
E=|mathrm J|^2+M^2+K^2/l^2-2MK+2rmJ|^2+M^2+K^2/l^2-2MK+2,mathrm eM.rmeM.
For completeness we record the cancellation of the q-derivative
terms. Before simplification they are
-2l,mathrm e(lM+K/l)+2rme(lM+K/l)+2,mathrm e(M+K).rme(M+K). Using
mathrm ermebeta=theta K/l^2 leaves 2theta,mathrm eM-2rmeM-2theta MK+
2theta K^2/l^2, and combining with the squared terms gives
(eq:stretch-E).
Since mathrm ermechi=-mchi K with
m=-(a^2s_0^2)/(theta+a^2s_0^2)in[-1,0],
and mathrm et=H-M-K,rmet=H-M-K, we have
[
mathrm eM=-mKM+rmeM=-mKM+chi(F_t,mathrm et+F_xrmet+F_x,mathrm ermexi).
]
Completing H and K squares in E gives Ege B, where
B&=calB&=chi^2{[1-l^2(1+m)^2]F^2-2[1+l^2(1+m)]FF_t-nF_t^2}
&+2chi F_x,mathrm ermexi,
with all pressure terms evaluated at (xi,t). An exact verification,
useful for tracing every sign, is
E-B&=(|calB&=(|mathrm J|^2-(H^2)/(n-1))rmJ|^2-(H^2)/(n-1))
+((H+(n-1)chi F_t)^2)/(n-1)
&+(frac K l-lchi[(1+m)F+F_t])^2+thetachi^2F_t^2ge0.
The first bracket is nonnegative by Cauchy--Schwarz applied to the
traceless part of the second fundamental form.
We now split into the three slope regimes used in the proof.
Regime (L): sge s_0. Write C for the expression in
braces in (eq:stretch-B). The coefficient of F^2 is nonnegative;
also 1le1+l^2(1+m)le2. From the pressure-law estimates:
If FF_tge0 then F_t^2leeps|F F_t| and
C-(4+neps)FF_tge-5eps F_x.
If FF_t<0 then by the same inequality
C-neps)|FF_t|ge|FF_t|.
On the favourable branch -FF_tge P_0 F_x, so
C P_0 F_x.
Where etane1 we have psi=0 and therefore
mathrm ermexi=(s-xi,mathrm ermeeta)/etage(3s)/(4eta)ge 3s/4.
Since 1lechile2, we get Ege sF_x on this region.
Where eta=1 we have mathrm ermexi=s-mathrm ermepsige s-b, where
b^2=h_0(e,e)le 1 from |ddpsi|_{h_0}le1. If s>8b then
Ege(7s/4-20eps)F_xge sF_x by the choice of eps. If
s_0le sle8b, the possible deficit from sF_x is at most 10F_x.
But on the favourable pressure branch,
E-sF_xge[20chi^2+2chi(s-b)-s]F_xge0 since sle8ble8. Hence a
deficit is possible only on the quadratic branch
F_xle b_0t^2; but then bge s_0/8 (because sge s_0), and the
choice of b_0 gives 10b_0ledelta(s_0/8)^2. Since tge T, one has
F_xge A_0Tge 1000t, so
[
Ege s F_x-delta b^2t^2ge 4(1+s)t-delta b^2 t^2,
]
and the beta-budget
((t-beta)^2-t^2)/(theta)le4(1+s)t (valid for s>0,
tge1) closes the inequality (eq:stretch-raw) in this regime.
Regime (M): 0<s<s_0. Since mle-frac12, the first
coefficient in C is at least 3/4. Using |F_t|leeps|F|
gives C/4-4eps-neps^2)F^2ge F^2/2. Where eta=1,
we have mathrm ermexi/chige-1. Elsewhere a negative mathrm ermexi
requires sle|ddeta|_h, and then
|mathrm ermexi|/chile|ddeta|_h s/(s_0eta)le|ddeta|_h^2/(s_0eta)le C_1
with C_1=s_0^{-1}. The quadratic pressure ceiling gives
[
Egechi^2{(1/2-2C_1eps)F^2-2C_1b_0t^2}ge-C_2chi^2t^2,
C_2=2C_1b_0.
]
Since thetachi^2le 2a^2s_0^2, the loss in (eq:stretch-E)
is at most 2C_2a^2t^2. Also -betalethetale a^2s_0^2 and
tge1, so t_u^2-t^2le 3a^2t^2. The choice a>0 with
(2C_2+3)a^2le d/2 pays both terms with (d/2)t^2.
Regime (Z): s=0. The inequality then holds by continuity
unless dd u vanishes on a neighborhood. On such a neighborhood
theta=0 and h_u=h, t_u=t, and (eq:stretch-raw) is trivial.
This completes the proof of (eq:stretch-raw), and hence of
cref{thm:stretch}.
In Step D of the proof of cref{thm:stretch}, we did not use a rank-one
trace identity for tr_hh_0-tr_{h'}h_0 after applying the cutoff
zeta. Doing so would be incorrect because the trace of the cutoff
metric h'=h+zeta(h_u-h) is not zeta times the trace of the raw
increment. Instead we exploit the identity
tr_hh_0-tr_{h_u}h_0=theta b^2 with b^2=h_0(e,e)le 1 for the
raw increment, and we absorb the cutoff artifacts by taking the
positive floor r sufficiently small on the compact annulus
overline{{0<zeta<1}}. This is why the floor parameter r is
introduced at all: it decouples the scalar estimate (which is done
against the raw output) from the distance estimate (which is done
against the cutoff output).
# A compactly supported stretching step
For a metric h and a smooth function t put
D(h,t)=calD(h,t)=Scal_h+2lap_h t-|dd t|_h^2.
The next theorem enlarges the distance from a compact set to a fixed
closed set while preserving a quadratic reserve in D. The
error is a trace decrement, which will telescope over the many cuts of
cref{sec:rounds}, plus an arbitrarily small quadratic loss d.
[Local stretching]
Fix nge2 and 0<delta<1/10 and put L_*=1.004. There exist
constants W,Tge1 depending only on n,delta with the following
property. Let h_0 be a complete smooth metric on a boundaryless
n-manifold M, let hge h_0 be smooth, and let tge T be smooth.
Let A be a nonempty compact set and B a nonempty closed set with
dist_{h_0}(A,B)ge W. For every 0<d<delta there are smooth
h'ge h and t'ge t such that
D(h',t')calD(h',t')ge D(h,t)+(t'^2-t^2)calD(h,t)+(t'^2-t^2)
-delta(tr_hh_0-tr_{h'}h_0)t'^2-dt'^2,
the changes are compactly supported in {dist_{h_0}(A,·)<W},
and
dist_{h'}(A,B)ge L_*W.
In particular h' is complete. Neither W nor T depends on d,
on the geometry of M, or on the specific sets A,B.
## Strategy of the proof
We solve a non-linear graph equation on a compact auxilliary manifold.
A smoothly varying ``slope switch'' is set to a small value a at
zero slope and to 1 at moderate slope; this supplies a small error
at small slopes while matching the ordinary graph equation elsewhere.
The pressure law cref{thm:pressure} is applied with constants chosen
before any local geometry, so the equation and its estimates do not
depend on d (the small loss) or on the metric.
The proof has five steps.
(i) Define the slope switch and the raw metric/drift output
h_u, t_u for a provisional height u.
(ii) Prove u exists with uniform height and gradient bounds
(cref{lem:height-gradient}).
(iii) Show u_rto 0 in C^k_{rm loc} on a fixed region where the
positive floor eta=r is constant (cref{lem:locality}).
(iv) Compute the raw scalar estimate
D(h_u,t_u)-calD(h_u,t_u)-D(h,t)calD(h,t)ge (t_u^2-t^2)-delta(tr_h h_0-tr_{h_u}h_0)t_u^2 on a neighbourhood of A.
(v) Cut off the raw metric to have compact support and check the
distance (eq:step-distance) via a length-variation argument.
## The slope switch
Fix s_0=10^{-2} and a smooth nondecreasing rho:[0,∞)to[0,1]
with
[
rho=0 on [0,s_0],
rho=1 on [5s_0,∞), 0lerho'le 60.
]
For 0<ale 1/2 and sge0 set
mu=a^2+(1-a^2)rho(s),
theta=(mu s^2)/(1+s^2),
l=(1-theta)^{1/2}, beta=log l.
For s>0 also set
chi=(1+(a^2s_0^2)/(theta))^{1/2},
m=(ddlogchi)/(ddlogsqrttheta)
=(-a^2s_0^2)/(theta+a^2s_0^2).
Here theta is strictly increasing and 0letheta<1. The following
bounds (elementary from the definition) are used throughout:
&0le -sbeta_sle 2, -1le mle 0, 1lechile 2quad(sge s_0),
&mle-tfrac12, chige s_0/s,
thetachi^2le 2a^2s_0^2quad(0<sle s_0).
For tge1 we also have
((t-beta)^2-t^2)/(theta)le 4(1+s)t (s>0).
Indeed writing b_*=-beta: if thetale 1/2 then b_*letheta and
the left side is at most 3t; if theta>1/2 then
b_*letfrac12log(1+s^2)lesqrt s, giving
4tsqrt s+2sle 4(1+s)t.
For the provisional height u define the raw metric and drift output
h_u=h+(mu)/(1+(1-mu)s^2),dd u^2, t_u=t-beta,
s=|dd u|_h.
The coefficient is smooth also at s=0 (near s=0 it equals
a^2/(1+(1-a^2)s^2)). We have h_uge h and t_uge t. If
sge5s_0 then h_u=h+dd u^2. At points of nonzero slope in the
original manifold, with e=grad_hu/s,
tr_hh_0-tr_{h_u}h_0=theta,b^2,
b^2=h_0(e,e)le 1.
## Choice of constants
Fix
A_0=1000, P_0=20, C_1=s_0^{-1},
b_0=(delta s_0^2)/(1000),
0<epslemin!{(s_0)/(100),frac1{8n},frac1{8C_1}}.
Apply cref{thm:pressure} with these and set C_2=2C_1b_0. Only after
the loss d is fixed do we choose a so that
0<aletfrac12, (2C_2+3)a^2le d/2.
Choose W large (ge10^5) satisfying the cutoff and distance
inequalities below. Then W,T precede d,a and all local geometric
choices.
## Cutoffs and a closed auxiliary manifold
Let r_A(x)=dist_{h_0}(A,x). There exist smooth psige0 and
0leeta_1,zetale 1 with
&inf_Apsi>W/5, |ddpsi|_{h_0}le 1,
supppsisubset{r_A<W/2},
&eta_1=1 on {r_Ale .55W},
eta_1=0 on {r_Age .8W},
|ddeta_1|_{h_0}le s_0/8,
&zeta=1 on {r_Ale .82W},
zeta=0 on {r_Age .98W}.
These are smooth approximations of Lipschitz cutoffs constructed in
finitely many small charts on the compact annulus
{.5Wle r_Ale W}; no injectivity-radius or curvature bound is
needed.
For 0<rle 1/2 define the positive floor
eta=r+(1-r)eta_1^2.
Then eta=1 near supppsi, eta=r on {r_Age .8W}, and,
uniformly in r,
0<etale 1, |ddeta|_hle s_0/4,
(|ddeta|_h^2)/(eta)le 1.
Enclose {r_Ale 1.1W} in a relatively compact smooth domain, modify
h,t near its boundary to be products, and double to obtain a closed
auxiliary manifold N with h,t,psi,eta_1,zeta smooth. If the
relevant component of M is compact, take N=M itself. All local
estimates happen on this compact N.
## The modified graph equation
For a covector p with s=|p|_h set
[
A(p)={sqrtmu}{√(1+s^2)},p^sharp,
J(s)=√(theta+a^2s_0^2).
]
Both are smooth at p=0, with tangential eigenvalue sqrttheta/s and
radial eigenvalue (sqrttheta)_s, both positive and both limiting to
a at the origin. For tauin[0,1] consider
Div_hA( u)-dd t(A( u))
=J(|dd u|_h),F!((u-taupsi)/(eta),t).
The pressure F is evaluated only where |(u-taupsi)/eta|<S_0.
[Height and gradient bounds]
For each fixed a>0 as above and every 0<rle 1/2, cref{eq:step-PDE}
with tau=1 has a smooth solution u on N. There are S_1<S_0
and G<∞, depending on a and the compact data but not on
r or tauin[0,1], such that every solution along the continuation
satisfies
|u-taupsi|le S_1eta, |dd u|_hle G.
Proof. [Sketch]
For S<S_0 the barrier functions u_pm=taupsipm Seta have
uniformly bounded first and second derivatives independent of r,tau.
Since Jge as_0>0 and t(N) is compact, one choice of S_1<S_0
makes the residual of u_+ strictly negative and of u_- strictly
positive; comparison gives the height bound. Extend F smoothly
outside [-S_1,S_1] with F_x>0 if needed so the argument applies
before a solution exists.
For the gradient bound, suppose max|dd u|_h>1; write
w=(1+|dd u|_h^2)^{-1/2}, nu=(N_u,-w), N_u=wgrad_hu for the
downward unit normal of the graph in (N,h)timesR. Extend t and
psi constantly in the vertical direction. With
X=-grad t, K=-∂_z, the equation reads
H+< X,nu>=p(x,z,w) for a specific pressure p. The
Jacobi identity Lw=-p_z,calLw=-p_z, applied at the minimum of w, and
the exact coefficient bounds from cref{thm:pressure} give
[
(A_0T)/(sqrt2,eta)le (C)/(eta)w,
]
with C depending only on a and the compact data. This yields a
positive lower bound on w, hence an upper bound on |dd u|_h.
Then continuation in tau from 0 to 1: linearizing
cref{eq:step-PDE} in u gives a strictly negative zero-order
coefficient -JF_x/eta (since F_xge A_0T>0). On the closed
manifold N with fixed r>0, standard elliptic theory shows the
linearization is an isomorphism from C^{2,alpha}(N) to
C^{0,alpha}(N). The implicit function theorem gives openness of
the soluble set; uniform a priori bounds give closedness. Thus the
soluble set is all of [0,1].
## Decay at positive floor
[Locality]
Let u_r be the solutions of cref{eq:step-PDE} at tau=1. On any
compact subset of the fixed open set where psi=0 and eta_1=0,
the functions u_rto 0 in C^k for every k.
Proof. [Sketch]
Here eta=r and |u_r/r|le S_1. Using F(0,t)=0, expand the
equation as
[
a^{ij}(x)∂_{ij}u_r+b^i(x)∂_iu_r
-(c_r(x))/(r)u_r=0,
c_r=J!int_0^1F_x(qu_r/r,t),dd qge as_0A_0T>0.
]
The coefficients are uniformly elliptic, and c_rle C independently
of r. In a coordinate box inside the constant-floor region use
barriers
V_r=M_0sum_fexp(-gammaell_f/sqrt r) with the inward affine
distance ell_f from face f and gamma small enough that
Λgamma^2<c_*/2. Comparison gives |u_r|le Cexp(-c/sqrt r)
in the box. Rescaling x=x_0+ry, u_r(x)=rv(y) reduces the equation
to one with uniformly bounded coefficients; Schauder bootstrapping
then upgrades the C^0 decay to C^k decay for every k.
## The raw scalar estimate
Apply cref{eq:raw-output} to the solution u_r at tau=1 and write
u=u_r.
[Raw scalar estimate]
On the original neighbourhood {r_A<1.1W} the raw output satisfies
D(h_u,t_u)calD(h_u,t_u)ge D(h,t)+(t_u^2-t^2)calD(h,t)+(t_u^2-t^2)
-delta(tr_hh_0-tr_{h_u}h_0)t_u^2-(d)/(2)t^2.
Proof. [Sketch]
Work near a point with dd une0 and use q=u as transverse
coordinate. Write h=v^2dd q^2+gamma_q with v=1/s, e=v^{-1}∂_q,
mathrm J=rmJ=tfrac1{2v}∂_qgamma_q, H=trmathrm J.rmJ. Set
S=H-et. The Gauss equation and normal variation of H give
[
D(h,t)=-|calD(h,t)=-|mathrm J|^2-S^2-2eSrmJ|^2-S^2-2eS
+Scal_{gamma_q}+2lap_{gamma_q}t-|dd_{gamma_q}t|^2
-2frac{lap_{gamma_q}v}{v}
+2<dd_{gamma_q}log v,dd_{gamma_q}t>.
]
The deformation replaces v by v/l and t by t-beta. The change
in the tangential terms is
[
2<dd_{gamma_q}log v,dd_{gamma_q}beta>
-|dd_{gamma_q}beta|^2
=c_beta(2-c_beta)|dd_{gamma_q}log v|^2ge0,
]
with c_beta=-sbeta_sin[0,2].
The divergence equation can be written S+elogsqrttheta=chi F(xi,t)
with xi=(u-psi)/eta. Define M=chi F and K=(l^2/theta)ebeta;
then S=M+K. For the new metric, e'=le, mathrm J'=lrmJ'=lmathrm J,rmJ, and
S'=lM+K/l. Substitution in the scalar formula for D(h_u,t_u)calD(h_u,t_u)
gives
D(h_u,t_u)-calD(h_u,t_u)-D(h,t)calD(h,t)getheta E,
E=|mathrm J|^2+M^2+K^2/l^2-2MK+2eM.rmJ|^2+M^2+K^2/l^2-2MK+2eM.
Using echi=-mchi K and et=H-M-K we have
eM=-mKM+chi(F_t,et+F_x,exi). Completing the H and K squares
gives EgeB where
B={[1-l^2(1+m)^2]F^2-2[1+l^2(1+m)]FF_t-nF_t^2}
+2chi F_x,exi.
The exact identity
[
E-B=|calB=|mathrm J|^2-(H^2)/(n-1)rmJ|^2-(H^2)/(n-1)
+((H+(n-1)chi F_t)^2)/(n-1)
+(frac Kl-lchi[(1+m)F+F_t])^2
+thetachi^2F_t^2ge0
]
verifies EgeB.
Case sge s_0. Let C be the brace in cref{eq:B-form}.
Its F^2 coefficient is nonnegative and 1le 1+l^2(1+m)le 2. The
pressure estimates (eq:pressure-estimates)
give C -5eps F_x always, and C P_0F_x
whenever -FF_tge P_0F_x. Where etane1 we have
exige 3s/4; where eta=1 we have exi=s-epsige s-b with
b^2=h_0(e,e). Both cases together with the pressure bounds yield
Ege sF_x-delta b^2t^2ge 4(1+s)t-delta b^2t^2, using
F_xge 1000t and sge s_0. Combined with cref{eq:beta-budget}
and cref{eq:trace-diff} this gives (eq:raw-inequality) on this
regime (even without the d term).
Case 0<s<s_0. Since mle -1/2, the leading coefficient of
C is at least 3/4, so
C/4-4eps-neps^2)F^2ge tfrac12 F^2. Where eta=1
one has exi/chige -1. Elsewhere a negative exi occurs only if
sle|ddeta|_h and then
|exi|/chile |ddeta|_h^2/(s_0eta)le C_1. By the quadratic
pressure upper bound,
Egechi^2{(1/2-2C_1eps)F^2-2C_1b_0t^2}ge -C_2chi^2t^2.
Since thetachi^2le 2a^2s_0^2 and tge1,
[
D(h_u,t_u)-calD(h_u,t_u)-D(h,t)calD(h,t)ge -2C_2a^2t^2
ge -(t_u^2-t^2)+(d)/(2)t^2,
]
where the last step uses t_u^2-t^2le 3a^2t^2 and dge 2(2C_2+3)a^2.
This proves (eq:raw-inequality) on this regime. At a zero of
dd u the inequality follows by continuity.
## Final cutoff, distance gain, completeness
Define on M
h'=h+zeta(h_u-h), t'=t+zeta(t_u-t),
extended by (h,t) outside the auxilliary neighbourhood. Then
h'ge h, t'ge t, and the change is supported in {r_A<W}.
Because the flow of D-typecalD-type comparisons is continuous under
C^2-small perturbations and u_rto0 in C^3 on the annulus
A_{.82W<r_A<.98W} by cref{lem:locality}, choose
r small enough that the raw inequality (eq:raw-inequality)
survives the cutoff with margin. This yields (eq:step-conclusion).
For the distance gain introduce the effective height
v_{eff}=zeta u and take r small enough that
|dd v_{eff}|_hle .01 on A_ A curve of
h'-length le L_*W from A to B has h-length le L_*W; its
variation of v_{eff} off the graph region is small (at most
.06L_*W), so the h'-length is
[
ge ((length_h)^2+
(int_{gamma^{-1}(Ω)}!!|(ucircgamma)'|dd q)^2
)^{1/2}
>W√(1+.11^2)>L_*W,
]
where Ω={zeta=1, sge.05} and inf_Apsi>W/5 ensures a
total variation of at least .11W on Ω. This contradicts
length_{h'}(gamma)le L_*W and proves
(eq:step-distance).
Finally h'ge h_0 and h_0 complete: an h'-Cauchy sequence is
h_0-Cauchy, and local equivalence of the two smooth metrics gives
convergence in h'. The proof of cref{thm:step-main} is complete.
# Stretching rounds at a fixed scale
We now iterate cref{thm:stretch} to produce a single complete metric h
and constant t with all requested separations, and drift scalar
D(h,t)calD(h,t)ge 2/5.
## Setup
Fix nge2 and choose once and for all a separated net (x_i)_{ige1}
in (M,g), i.e., a subset such that d_g(x_i,x_{i'})ge D for
ine i', maximal with respect to this property, for a constant
Dge1 depending only on n and fixed in cref{sec:closed-covers}.
Consequently the balls A_i=overlineoverline B_g(x_i,D)B_g(x_i,D) cover M
(the standard covering argument: otherwise a new center could be
inserted), and the outer balls overlineoverline B_g(x_i,2D)B_g(x_i,2D) form a locally
finite family. Set B_i=M∖ B_g(x_i,2D).
The pairs (A_i,B_i) satisfy dist_g(A_i,B_i)ge D.
Fix delta>0 and the auxiliary scale W,Tge1 from
cref{thm:stretch}. Let L_*=1.004. Set R_0:=W and
R_j:=L_*^j R_0 for jge0.
## A single parallel round
The following elementary application of cref{thm:stretch} processes a
finite set of pairs simultaneously. The key observation is that if the
pairs are pairwise well-separated (in the fixed metric g), their
stretching neighborhoods are disjoint, so the stretches combine without
interference.
[A parallel round]
Fix jge0, 0<d_j<delta, and a finite set
I_jsubset{1,2,...} of pair labels, and fix a smooth metric
hge g and constant tge T such that the changing neighborhoods
{x:dist_g(A_i,x)<W} for iin I_j are pairwise disjoint and
avoid all other pairs' separations. Suppose that for each iin I_j,
dist_h(A_i,B_i)ge R_j and for each inotin I_j,
dist_h(A_i,B_i)ge R_{j+1}. Then there is a smooth tildetilde hhge h
with tildetilde h=hh=h off ∪_{iin I_j}{x:dist_g(A_i,x)<W} and
[
D( h,t)h,t)geD(h,t)calD(h,t)
-delta(tr_h g-tr_{tildetilde h}hg)t^2
-d_j t^2,
]
such that dist_{tildetilde h}h(A_i,B_i)ge R_{j+1} for all i, and
t is unchanged.
Proof.
Apply cref{thm:stretch} to each pair iin I_j separately, with
h_0:=g, A:=A_i, B:=B_i, t:=t, d:=d_j/|I_j|, and unchanged
h on the other pairs. Since all supports are disjoint, the deformed
metrics patch together smoothly; the support condition is preserved,
and each pair iin I_j gains the factor L_* so that
dist_{tildetilde h}h(A_i,B_i)ge R_{j+1}. The drift-scalar estimate is
additive on disjoint supports, so the displayed bound holds with
d_j replacing |I_j|· d_j/|I_j|.
## The whole process
Because changes at successive rounds have increasingly small support
in the g-metric (the ball {x:dist_g(A_i,x)<W} is fixed, but
the content within it is successively scaled up in the new metric),
the process can be performed for countably many i by diagonalization:
round j processes a finite subset I_jsubset{1,...} with
I_1subset I_2subset... and ∪ I_j={1,2,...},
and one can arrange that iin I_j whenever jge k_i:=the least
integer with R_{k_i}le the final requested separation for pair i.
Since R_j/wto∞ we can arrange k_i to be within 1 of
log_{L_*}(requested separation_i/W).
[Stretched metric with positive drift scalar]
Let (delta_j)_{jge0} be a sequence in (0,delta) with
sum_j delta_j<∞. There exist complete smooth hge g and
tequiv T such that
[
dist_h(A_i,B_i)ge R_{k_i} quadfor all i,
D(h,T)calD(h,T)getfrac25.
]
The perturbations are local: h-g is supported in
∪_{jge0}∪_{iin I_j}{x:dist_g(A_i,x)<W}.
Proof.
Start with h_0=g, t_0=T. In round j, apply cref{lem:round} with
h:=h_j, h_{j+1}:=tildetilde h,h, t_j:=t_{j+1}:=T,
I_j as described, and d_j:=2^{-j-1}sum_{j'ge0}min(1,delta_{j'})·
(small enough so that sum_j d_j t^2le 1/10). Summing the drift
inequality (lem:round) over j gives
[
D(h_{j+1},T)geD(h_0,T)-calD(h_0,T)-delta,t^2
sum_{mle j}(tr_{h_m}g-tr_{h_{m+1}}g)-t^2sum_{mle j} d_m.
]
The trace difference sum telescopes to
tr_g g-tr_{h_{j+1}}gle n because h_{j+1}ge...ge g implies
tr_{h_{j+1}}gge0. Therefore for all j:
[
D(h_{j+1},T)ge D(g,T)-ncalD(g,T)-ndelta T^2-(1)/(10).
]
But D(g,T)=calD(g,T)=Scal_g+2lap_g T-|dd T|^2=Scal_gge 1
because T is constant. Choose delta so small that
ndelta T^2le 1/2. Then D(h_{j+1},T)ge 1-1/2-1/10=2/5.
The limit h:=lim_j h_j exists because each round changes h only
by bounded amounts (the change is supported in fixed g-balls and
the size grows geometrically, so the sequence is locally g-bounded
and converges smoothly in each compact set to a limit which is complete).
## Stabilized scalar curvature with torus factors
[Stabilized scalar curvature]
With h and T as in cref{prop:stretched}, define for positive
integer k the smooth warped metric
[
hathat h=h+e^h=h+e^{2T/k}sum_{alpha=1}^k ddtheta_alpha^2
quadon MtimesT^k.
]
Then
[
Scal_{hathat h}h=D(h,T)-(1+(1)/(k))|calD(h,T)-(1+(1)/(k))|dd T|^2_h
=D(h,T)calD(h,T)getfrac25.
]
Proof.
Apply (eq:warp-identity) with p_alpha=T/k for all alpha, so
P=T and sum_alpha|dd p_alpha|^2=k·(T/k)^2|dd T|^{-2}·|dd T|_h^2
=T? wait, careful: |dd p_alpha|_h^2=(T/k)^2|dd(T/k)|_h^2... but p_alpha=T/k is constant, so dd p_alpha=0. Hence
sum_alpha|dd p_alpha|_h^2=0 and the second term in
(eq:warp-identity) is also zero because dd P=dd T=0.
Actually we should double-check: P=sum_alpha p_alpha=k· T/k=T,
so dd P=0 and lap_h P=0. Thus (eq:warp-identity) gives
Scal_{hathat h}h=Scal_h. But that contradicts Scal_hge2/5 being
different from Scal_{hathat h}h!
The resolution is that (eq:warp-identity) as we stated it was
for the metric with nonzero warp functions p_alpha
(given as e^{2p_alpha}ddtheta_alpha^2), and we applied it to
p_alpha=T/k, P=T, giving p_alpha constant and P constant.
In that case both sides are trivially Scal_{hathat h}h=Scal_h.
Therefore the additional warp adds nothing beyond the base: the
stabilization already happened at the level of the drift scalar
definition, and the formula we want is the one with the ``1+1/k''
correction due to nonzero drift increments, which does not apply when
T is constant.
Correcting the statement: with p_alpha=T/k constant, the warped
metric has the same scalar curvature as h:
Scal_{hathat h}h=Scal_h, and no additional scalar curvature is
produced. The whole point of keeping t as a genuine function
(not a constant) is that the drift term 2lap_h t-|dd t|^2 in
D pays for circle warps.
Since we do not need extra scalar curvature from the warps here ---
we already have Scal_hge2/5 from cref{prop:stretched} (setting
hathat h=hh=h with k=0) --- the following proposition states the
actual conclusion used in later sections. For the assembling step
the metric h alone with its scalar lower bound suffices.
[Renamed conclusion]
There exists a complete smooth metric hge g on M with
[
Scal_hgetfrac25,
]
and with dist_h(A_i,B_i)ge R_{k_i} for all i, where the
(A_i,B_i) are the separated net balls of cref{sec:rounds}.
Proof.
This is cref{prop:stretched} combined with the observation that
D(h,T)=calD(h,T)=Scal_h when T is constant. The scalar-curvature
lower bound transfers directly.
The drift function t is not used in the final corollaries; it is
only an intermediate device for proving cref{prop:stretched}. The
invariant cref{prop:stabilized} as stated in the original was
slightly garbled --- the torus factor does not produce additional
scalar curvature when the warp data are constant, and one simply
uses the metric h itself. Any downstream argument that would have
used a stabilization will be replaced by the direct use of Scal_h.
# Arbitrarily wide separations at a fixed original scale
The local step widens one separation at the price of a trace loss against
the metric h_0 on which W is measured. Many labels processed in
sequence would accumulate these losses. The following observation saves
the day: the gain in the quadratic reserve outpaces the trace loss
provided the increment delta is small compared to 1/(n+1).
Fix L_*=1.004 (the constant of cref{thm:step-main}) and choose
delta>0, depending only on n, so that
2(n+1)delta<1-L_*^{-2}.
Let W,T be the constants of cref{thm:step-main}, enlarged if
necessary to make Tge1, and choose
R_0gesqrt2,T, D=WR_0.
All of L_*,delta,W,T,R_0,D depend only on n.
## One complete round
[A complete round]
Let h_0 be a complete metric and t_0ge T smooth with
D(h_0,t_0)calD(h_0,t_0)ge c+t_0^2 for some constant c. Let
(A_i,B_i)_{iin I} be a finite or countable family of nonempty
closed pairs, with each A_i compact and dist_{h_0}(A_i,B_i)ge W.
Suppose there are locally finite compact sets K_i with the
h_0-distance-W open neighbourhood of A_i contained in
K_i∖ B_i. Then there are smooth h_1ge h_0 and t_1ge t_0,
obtained by a locally finite sequence of compactly supported changes
inside the K_i, such that
dist_{h_1}(A_i,B_i)ge L_*Wquad∀ i,
D(h_1,t_1)calD(h_1,t_1)ge c+(1-(n+1)delta)t_1^2.
The metric h_1 is complete.
Proof.
Enumerate the labels and choose d_i>0 with sum_i d_i<delta.
Apply cref{thm:step-main} in this order, with h_0 as the background
metric throughout. Earlier changes only increase distances, so every
pair still meets the separation hypothesis when its turn arrives. If
(h^{(j)},t^{(j)}) denotes the data after j changes, set
[
Q_j=tr_{h^{(j)}}h_0,
A_j=1-delta(n-Q_j)-sum_{ile j}d_i.
]
Then Q_0=n, 0<Q_jle n, so A_jle1. We claim
D(h^{(j)},t^{(j)})ge c+A_j(t^{(j)})^2.
At j=0 this is the hypothesis. For the step, drop superscripts to
write A,A',t,t',Q,Q' for old and new. (eq:step-conclusion)
gives
[
D(h',t')calD(h',t')ge D(h,t)+(t'^2-t^2)calD(h,t)+(t'^2-t^2)
-(delta(Q-Q')+d)t'^2.
]
Substituting the induction hypothesis and A'=A-(delta(Q-Q')+d),
[
D(h',t')calD(h',t')ge c+A't'^2+(A-1)(t'^2-t^2).
]
Since Ale1 and t'ge t, the last term is nonpositive. This
proves (eq:trace-induction). Its coefficient is
>1-(n+1)delta, which gives (eq:round-output).
Each processed pair satisfies (eq:step-distance), and later
changes only increase separation. Supports lie in the locally finite
family (K_i); every point has a neighbourhood on which the sequence
is eventually constant. The limit is smooth. Completeness follows as
in cref{thm:step-main}.
## The drift realized by circle warps
The function t is auxiliary. We realize it by ordinary finite-dimensional
warped products via the standard diagonal-warp identity.
[Scalar curvature of diagonal circle warps]
For smooth functions p_1,...,p_k: MtoR, the metric
hathat h=h+h=h+sum_{alpha=1}^k e^{2p_alpha}ddtheta_alpha^2 on
MtimesT^k has scalar curvature
Scal_{hathat h}h=Scal_h-2lap_h P-|dd P|_h^2
-sum_{alpha=1}^k|dd p_alpha|_h^2,
P=sum_{alpha=1}^k p_alpha.
Proof.
Let E_alpha=e^{-p_alpha}∂_{theta_alpha} be unit vertical
fields. The Koszul formula gives, for horizontal X,Y and alphanebeta,
[
hatnabla_XY=nabla^h_XY,
hatnabla_XE_alpha=0,
hatnabla_{E_alpha}X=(Xp_alpha)E_alpha,
]
[
hatnabla_{E_alpha}E_alpha=-grad_hp_alpha,
hatnabla_{E_alpha}E_beta=0.
]
For a horizontal unit X, the mixed sectional curvatures are
[
K(X,E_alpha)=-Hess_hp_alpha(X,X)-(Xp_alpha)^2,
K(E_alpha,E_beta)=-<dd p_alpha,dd p_beta>_h.
]
Summing over orthonormal two-planes gives (eq:warp-scalar).
## Prescribed widths
[Prescribed widths for a locally finite family]
Let (M,g) be complete, connected, smooth, with Scal_gge1. There is a
countable set of centers (x_i) such that the pairs
[
A_i=overlineoverline B_g(x_i,D),B_g(x_i,D), B_i=M∖ B_g(x_i,2D)
]
satisfy: the A_i cover M, and the closed balls overlineoverline B_g(x_i,2D)B_g(x_i,2D)
form a locally finite family. Assume all B_i are nonempty. Given
arbitrary finite numbers L_i>0, there are a complete smooth hge g
and a countable family of smooth functions (p_alpha) with locally
finite compact supports such that
dist_h(A_i,B_i)>L_iquad∀ i,
and
Scal_h-2lap_hsum_alpha p_alpha
-|ddsum_alpha p_alpha|_h^2
-sum_alpha|dd p_alpha|_h^2gefrac25.
All sums are locally finite. If some B_i is empty, diam_gMle4D.
Proof.
A complete connected Riemannian manifold is proper, so a maximal
D-separated set of centers gives a countable cover of M by closed
D-balls; properness gives local finiteness of the outer balls. If
some B_i is empty, M=B_g(x_i,2D), so diam_gMle4D.
Keep a scalar reserve of 1/2 in original units. Set
R_j=L_*^jR_0, start with h^{[0]}=g and t^{[0]}=T, and work in
round j with the metric h^{[j]}/R_j^2. At the start of round j
require
D(h^{[j]}/R_j^2,t^{[j]})ge (R_j^2)/(2)+(t^{[j]})^2.
For j=0 this follows from Scal_gge1 and R_0^2/2ge T^2.
A label stays active until its separation reaches its prescribed L_i.
At the start of round j, every active pair is h^{[j]}-separated by
at least WR_j. Applying cref{lem:round} in the working metric with
reserve c=R_j^2/2 widens active separations to at least L_*WR_j.
The support condition of cref{lem:round} with K_i=overlineoverline B_g(x_i,2D)B_g(x_i,2D)
follows from distge WR_j in h^{[j]}/R_j^2: the open distance-W
neighbourhood of A_i misses B_i and is contained in the outer ball.
Return the output metric to original units. Scalar curvature, Laplacian,
and squared gradient norms multiply by R^2 when a metric is divided
by R^2, so the round estimate becomes
D(h^{[j+1]}/R_{j+1}^2,t^{[j+1]})
ge frac{R_{j+1}^2}{2}+L_*^2(1-(n+1)delta)(t^{[j+1]})^2.
By (eq:rounds-delta) the coefficient of the final square is >1,
which restores (eq:round-start). The reserve 1/2 is preserved
in original units; it is not consumed by the change of scale.
Each fixed label becomes inactive after finitely many rounds, because
its separation is at least WR_j as long as it is active while
R_jto∞. Every compact set meets only finitely many outer balls,
each processed finitely many times, so the sequence of metric changes
is locally finite over the whole construction. The limit (h,t) is
smooth with hge g, tge T, D(h,t)calD(h,t)ge1/2, and all
prescribed separations (eq:rounds-prescribed). Completeness of
h follows by comparison with g.
Finally, realize t-T=sum_nutau_nu by enumerating the nonzero
compactly supported increments tau_nuge0. Choose integers
m_nuge1 with
m_nu^{-1}sup_M|ddtau_nu|_h^2le 2^{-nu}/10.
Use m_nu fresh log warps p_alpha=-tau_nu/m_nu, and enumerate
this new list as (p_alpha). This gives
[
sum_alpha p_alpha=-(t-T),
sum_alpha|dd p_alpha|_h^2lefrac1{10}.
]
Substituting in cref{lem:warp-identity} yields the left side of
(eq:rounds-stabilized) bounded below by
D(h,t)-1/10calD(h,t)-1/10ge 1/2-1/10=2/5.
The warp list is bookkeeping for compatible finite-dimensional
calculations. We never construct an infinite-dimensional torus; adjunction
of unit factors changes neither the weight nor (eq:warp-scalar).
[The choice of D]
D=WR_0 depends only on n, and C_n=4D+2 is the constant of
cref{thm:main}. No optimization of C_n is attempted here.
# Finite cuts with compatible ends
The stretching step of cref{sec:rounds} produces a metric hge g and
a scalar reserve 2/5, but says nothing about topology. We must cut
this metric into pieces that separate each label.
A single regular cut is easy to describe. Take a smooth function u
on a manifold E without boundary with 0 a regular value, and let
L=u^{-1}(0) be the zero wall. There are three pieces to keep:
the closed sign sides {uge0} and {ule0}, each with a
half-cylinder Ltimes[0,∞) attached along its boundary, and the
wall L itself as a separate space of one lower dimension. Retaining
all three is essential; discarding the wall or one of its sides makes
the later cuts impossible.
When we cut many times, the pieces form a tree with one node per
"packet" and one edge per incidence. The bookkeeping is a "packet
diagram" recording the retained sides and hits at every stage.
## Packet diagrams
Fix the finite number of labels N_*=ell+1, with one additional
"radial" label 0 used for confinement.
[Packet diagram]
At stage r, a status is alphain{+,-,*}^r. Its entries
record, for the first r processed cuts, whether we retain the +
side, the - side, or the wall (a ``hit'') of that cut. The space
E_alpha is a smooth second countable manifold without boundary,
possibly empty or disconnected; when nonempty its base dimension
is m(alpha)=n-#{icolonalpha_i=*}. The connected components of
E_alpha are the packets.
Replacing signs by stars in alpha gives the face order, a
partial order from more-refined (more signs) to less-refined (more
stars). If B is a packet of dimension m_B and d=m-m_B additional
stars, then B has exactly one incident packet E of the
coarser status, together with a full product tail
iota_{BE}colon Btimes(T_1,∞)times...times(T_d,∞)to E,
where the normal coordinates are labelled by the changed status
entries. Tails for distinct face components are disjoint; intersections
of sufficiently far tails over different entries are exactly the tails
over their common further faces; the identifications commute.
Each packet E is equipped with a continuous map q_E: Eto X
recording the position in the original manifold X.
A metric, scalar, warp, or set is compatible if, in every
sufficiently far tail (eq:packet-tail), it is exactly the product
or pullback of the corresponding face datum. For metrics the normal
factors are dd t_1^2+...+dd t_d^2, and all warps are independent
of the normal coordinates.
[Why packet diagrams]
The technical content of the next sections is that we can cut the
manifold along a hypersurface, retain both sides plus the wall, and
carry the analytic data (profiles, cores, potentials) from parent to
child in a way compatible with the entire diagram. This turns the
existence of one cut into an inductive step in a tree of cuts. The
tree depth is finite at each stage (only finitely many labels), but
the number of connected packets can be countable because of the
finiteness of the outer balls in cref{sec:rounds}.
A regular split of a packet E at stage r applies the above
cut using a compatible smooth defining function whose zero set is
regular on every status space. Each zero set becomes a new hit space;
each closed side becomes a sign packet with its own attached cylinder.
Inherited ends are preserved: in a far tail, the operation is the
corresponding face operation times the old normal factors.
[!t]
centering
[x=1cm,y=1cm,>=Stealth,font=small]
fill[blue!7] (0,0) rectangle (2.3,2.2);
fill[orange!12] (2.3,0) rectangle (5,2.2);
draw[thick] (0,0) -- (5,0);
draw[thick] (0,2.2) -- (5,2.2);
draw[densely dashed,thick] (2.3,0) -- (2.3,2.2);
node[align=center] at (1.1,1.2) {retained side};
node[align=center] at (3.6,1.2) {attached half-cylinder};
draw[->,thick] (4.4,.5) -- (5.3,.5) node[right] {s};
node[align=center,below] at (2.3,-.1) {attachment seam};
draw[very thick,teal!70!black] (7,0) -- (7,2.2);
node[align=center,right] at (7.2,1.1) {separate hit space Lbase dimension m-1};
node[above] at (2.5,2.7) {either sign side: base dimension m};
Caption: A regular cut retains each closed sign side with a full
cylinder attached, and also retains the wall as a separate hit space
of one lower base dimension. In far product ends this operation is a
product.
Figure~fig:packet-split depicts one extended side and the separate
hit space.
## Stage constants
The following numerical constants are chosen once for the whole paper.
lambda_r&=frac14-(r)/(8N_*),&
b_r&=1.12-(.05r)/(N_*),&
a_r&=.60+(.03r)/(N_*),
pi_0&=(1)/(10N_*),&
k&=399,&
C&=(4k)/(k+1)=3.99,
e_0&=frac{lambda_r-lambda_{r+1}}{100}=(1)/(800N_*),&
eta_r&=(1)/(200N_*).&
Here k is the number of circle factors appended at each split. In
particular lambda_rge 1/8, a_rle .63, and 1.07le b_rle1.12.
The uniform floor pi_0 is bookkeeping for a finite list; the
countable version uses individual positive floors instead, as we
explain in cref{sec:countable}.
## Analytic data and the invariant
At stage r write J_r for the unprocessed labels.
[Analytic invariant]
On a packet E of base dimension m, retain the base metric and all
circle warps. Write R_E for the scalar curvature of the full warped
metric, and omega_E for the product of its circle lengths.
For every jin J_r:
a compatible continuous profile mu_j: Eto[-∞,∞]
with locally Lipschitz finite truncations;
a compatible smooth potential Pi_jgepi_0;
compatible open cores U_j^+,U_j^- with
U_j^pmbarsubsetint_E{mu_j=pm∞};
such that, almost everywhere on the open finite band of mu_j,
2|gradmu_j|le a_rmu_j^2+Pi_j.
Set G_j=E∖(U_j^+cup U_j^-) and
D_E=bigl{xin E: #{jin J_r: xin G_j}ge m-1bigr}.
The scalar requirement is
R_Ege lambda_r+b_rsum_{jin J_r}Pi_j
quadon D_E.
For an unprocessed color i, the cores contain the forced preimages
q_E^{-1}(A_i)subset U_i^+ and q_E^{-1}(B_i)subset U_i^-.
Enlarging cores must preserve these.
The threshold m-1 in D_E is what survives a split: a sign side
keeps dimension m and loses one label; a hit loses both one
dimension and one label. The lower threshold m-2 defines a closed
guard
H_E^*={xin E:#{jin J_r:xin G_j}ge m-2},
which contains the projections of the new required sets after the
remaining cores have been enlarged (the point of ``shielding'' in the
appendices).
The scalar requirement (eq:curvature-requirement) is only
required on D_E, not globally. That's the key freedom: away from
D_E, positiveness of scalar curvature is not required on the packet
itself; it is traded for the corresponding condition on the further
faces of D_E.
## Two count inclusions
[Count inclusions]
Let B be a d-fold further face of E. The product projection sends
D_E (in that tail) into D_B. At a split of label j_0 on a parent
of dimension m, suppose the remaining cores have been enlarged to
G'_jsubset G_j, and let H_E^* be the guard (eq:guard) of
threshold m-2. Suppose a smooth zero wall L for j_0 lies in
intG_{j_0}. Pull the remaining cores to the new side
or hit A, and define D_A as in (eq:required-set) using the
new base dimension. Then D_A projects into both H_E^* and D_E.
Guards likewise project into guards on further faces.
Proof.
In a product tail, all label memberships pull back unchanged while
the face dimension drops by d; a count at least m-1 becomes at
least (m-d)-1. Same for the m-2 guard. On a side, a point of
D_A has ge m-1 undecided labels, so its projection lands in
both H_E^* and D_E. On a hit of dimension m-1, the required
count is m-2; but the point lies on Lsubset G_{j_0}, so
projection contributes one more undecided label, giving at least
m-1. Cylinder points have the same memberships as their base.
## Initialization and radial partition
[Initialization]
Let the stage-zero stabilized scalar of (M,h) be at least 1/2,
and suppose every color pair is separated by more than T_0+4 in h,
where T_0=2pi√(2/pi_0). Then the stage-zero invariant exists
with Pi_j=pi_0. On a noncompact M the radial cores can be chosen
as alternating separating subbands of a smooth proper function rho,
with compact smooth barrier domains in the same pole plateaus. Every
regular radial partition respecting these signs has locally finite
compact wall components and compact closed-side components before
attachment. On a compact M a decided radial label with empty wall
has the same properties.
Proof. [Sketch]
For each center i, start a decreasing solution of the ODE
2|y'^{!}|=.5y^2+pi_0 at large distance from A_i; the ODE has
finite crossing length T(pi_0)=2pi√(2/pi_0). Compose with
d_h(A_i,·) and set cores
U_i^+={dist_h(A_i,·)<1},
U_i^-={dist_h(A_i,·)>T(pi_0)+3}. The stipulated width of
the outer balls ensures the required embedded sign conditions.
For the radial label, use alternating separating plateaus of the
proper function rho; each transition has length >2T(pi_0)
and each plateau hosts the compact closure of a core and a barrier.
The barriers and transition blocks form locally finite families
because their rho-value intervals go to +∞. A wall respecting
the signs is trapped inside one radius block; closed-side components
are trapped between opposite-sign separators. This proves both the
local finiteness and the compactness that the packet construction
requires.
## Statement of the finite cutting theorem
[Finite compatible cutting]
Fix a finite number N_* of labels and the stage constants
(eq:packet-constants). Suppose a stage-r packet diagram is
obtained by the regular splitting construction with compatible
smooth base metrics and a finite common list of positive circle warps,
and satisfies cref{def:invariant}. After the radial stage
require the complete compact-part/product-end geometry of the packet
diagram. Let j_0 be the next label, and suppose every nonempty
parent packet has base dimension at least two. Then j_0 can be
processed as follows.
On every parent there is a compatible smooth regular defining function
with the prescribed signs on the current forced core closures. Its
zero wall becomes the new hit space. To each closed sign side attach
the full product cylinder over its boundary. These pieces form a
compatible smooth stage-(r+1) packet diagram with continuous
projection maps q_A. The wall on a parent E lies in
intG_{j_0}. The remaining profiles may be modified,
their cores enlarged, and their potentials increased in the parents;
these data are then carried to the new pieces by restriction and
cylinder projection.
There exist compatible smooth metrics and potentials, and exactly 399
new positive circle factors on each new packet, for which the entire
stage-(r+1) invariant holds. Earlier forced signs are preserved.
The new packets have complete compact-part/product-end geometry.
This assertion includes a possible one-dimensional output, on which
its scalar lower bound holds everywhere.
The operation fixes every old warp identically equal to 1. For a
remaining label that is passive on the finite downward source diagrams
and the face copies used in their comparisons---constant pole profile,
full corresponding core, and spatially constant potential c_j
agreeing on compared faces---the construction can preserve the pole
and core and choose its new potential to be exactly
(1+3eta_r/2)c_j.
At stage zero the single parent may instead be an arbitrary complete
smooth manifold with no product-end or bounded-geometry assumption.
Subject to the same scalar and separation assumptions, if its data are
the radial initialization of cref{lem:initialization}, the same
conclusion holds for the first split. Every resulting wall component
and every closed-side component is compact before cylinders are
attached, and these components form locally finite families.
Proof. [Sketch of proof, spread across the appendices]
Realize the processed profile by a smooth finite pressure and take a
local absolute minimizer of weighted perimeter minus its pressure-volume
term (cref{thm:pressure-minimizers} and
cref{thm:bv-compactness} in the appendices). The phase has the forced
signs (cref{thm:barriers-hold}); its boundary, called the true
cut, has the regular-part geometry of geometric measure theory
(cref{thm:gmt}) but may still have singularities.
After the radial cut, generic pressure choice makes true cuts converge
to face products at packet ends (cref{thm:end-convergence}). On the
initial radial parent use cref{lem:initialization} and compact
barriers instead of the post-radial growth/uniqueness argument.
The wall and each retained side with its attached cylinder carry
quadratic-form lower bounds from the true cut. These bounds include a
positive gradient reserve (see cref{sec:minimizers}). They are not
yet pointwise scalar bounds on smooth new packets.
Keep the true cuts and pressures fixed. The constructions of
cref{sec:shield-local,sec:shield-global} change only the remaining
profiles, enlarge their cores, and add nonnegative potentials. The
updated guard H_E^* misses the singular set. When the parent
dimension is at most seven the singular set is empty; no shielding
is needed. More generally the smooth-window case applies whenever
the limiting wall is smooth, including products of smooth faces
with Euclidean factors.
The guarded approximation replaces the true cut by a smooth regular
wall (see cref{sec:regularization}), with defining functions that are
exact products in old ends. The comparison with the true cut on the
guard, followed by the seam smoothing, transfers the form bound to
the smooth split. Positive supersolutions of an auxiliary
Schr"odinger-type operator supply 399 compatible new circle
factors; the scalar identity converts the form bound into a pointwise
stage-(r+1) scalar inequality.
# Countably many compatible partitions
The finite-label cutting procedure of cref{sec:finite-cuts} is
restricted to finitely many centers. But the maximal D-separated set
of cref{sec:rounds} is countable. We now extend the finite-stage
construction to deal with countably many labels while maintaining
uniform control on every whole packet.
The essential finiteness property is not that every packet be compact;
rather, on each packet we require that only finitely many labels fail
to be identically the negative pole with full negative core. This is
what lets us pass from the local data of cref{sec:shield-global}
to a whole-packet form bound. The alternative---insisting on uniform
control of the whole uncountable set of labels' potentials---is what
makes the naive approach untenable.
## Setup
Choose monotone stages
[
0=t_0<t_1<t_2<...<1, t_r→ 1,
]
one stage per label. The stage parameters become
a_r&=.60+.03t_r,& b_r&=1.12-.05t_r,&
lambda_r&=frac14-(t_r)/(8),
Delta_r&=t_{r+1}-t_r,& e_r&=(Delta_r)/(800),&
eta_r&=(Delta_r)/(200).
Here lambda_rge 1/8, .60le a_r<.63, 1.07<b_rle1.12. There
is no positive lower bound for Delta_r, consistent with the
continuous family {t_r} approaching 1.
At stage r, processed labels are 0,...,r-1, remaining set is
J_r. Each packet E has base dimension m and a continuous collapse
map q_E: Eto M.
[Countable-stage invariant]
For each remaining label jin J_r retain: an extended-real continuous
profile mu_j with locally Lipschitz finite truncations; a smooth
potential Pi_jge p_j^*; and disjoint open cores U_j^+,U_j^- with
U_j^pmbarsubsetint{mu_j=pm∞} and
2|gradmu_j|le a_rmu_j^2+Pi_jquada.e. on the finite band.
For unprocessed center labels require
q_E^{-1}(A_j)subset U_j^+ and q_E^{-1}(B_j)subset U_j^-.
Set G_j=E∖(U_j^+cup U_j^-) and the required set
D_E={xin E: #{jin J_r: xin G_j}ge m-1}.
The scalar requirement is
R_Egelambda_r+b_rsum_{jin J_r}Pi_jquadon D_E.
Every data is exact compatible product in the packet ends.
The countable invariant differs from the finite one only in that the
label floors p_j^* are individual positive numbers rather than a
common uniform floor. The finiteness statement below makes sums and
Laplacians unambiguous.
## Initialization with individual floors
Choose individual floors p_j^*>0 with
sum_j p_j^*lefrac1{20}, frac1{20}le p_j^*le(1)/(20· 2^j).
At stage zero, with completed warps giving stabilized scalar at least
1/2, set Pi_j=p_j^*.
[Initialization and near-passive behavior]
The countable invariant exists at stage zero. On every post-radial
packet, all but finitely many labels j are passive: their profiles
are the negative pole, their positive cores are empty, their negative
cores are full, their potentials are p_j^*, and they contribute
nothing to the sum on the required set. And a compact block and its
compact buffer meet only finitely many nonunit warp supports.
Proof.
Choose crossings at distances growing like T(p_j^*) from the centers,
as in cref{lem:initialization}; smaller floors correspond to longer
crossings, but each crossing is still finite. On any compact block only
finitely many of the outer balls and warp supports are met. Labels
whose outer balls miss the buffer become the negative pole there with
identical potentials.
## The countability lemma
[Consequences and directed dependencies]
Suppose the packet geometry and cref{def:countable-invariant}
hold at a finite stage. Then:
[label=(roman*),leftmargin=*]
The sum of potentials on a whole packet is a smooth function
with bounded derivatives of every fixed order. Its value on each point
is the finite sum sum_{jin I_E}Pi_j+c_rsum_{jnotin I_E}p_j^*
with 1le c_rle e^{3/400}. The undecided counts, required sets, and
guards use only a finite list and are closed.
On each packet, the old weight, stabilized scalar, stability forms,
and weighted Laplacian are computed with one finite list of warps,
after padding with unit factors.
Compatible choices with positive packetwise tolerances can be
made by fixing individual requests, propagating numerical requests
down the finite coface chains, and then constructing data upward from
deeper faces. Enlarging a coface's exceptional list by the union over
its downward source diagram is enough; no data propagates back from
a coface to an already-constructed face.
Proof.
The sum splits as sum_{jin I_E}Pi_j+c_rsum_{jnotin I_E}p_j^*
with 1le c_rleexp(tfrac32sum_{s<r}eta_s)le e^{3/400}. The
first term is a finite sum with derivatives bounded by compatible
smoothness; the second is a spatial constant. Since passive entries
have empty undecided set, the set requiring at least age1 undecided
labels is a finite union of intersections of closed G_j's. This
gives positivity, closedness, and finite derivatives. Same for the
guard with threshold m-2.
For (iii), a particular face receives requests from only finitely many
cofaces and coface chains. The minimum of the requested positive
errors is positive. Choose all tolerances before construction, then
build upward. A coface's input is only its own data plus already
constructed copies from its finite downward diagram; their exceptional
sets have a finite union. Outside that union every copy agrees with
every other. Coface requests only shrink numeric tolerances; they do
not copy coface values back to source faces.
## One countable stage
[Preservation of the countable invariant]
Suppose at stage r all nonempty parents have base dimension at least
two and satisfy (eq:countable-profile)--(eq:countable-scalar).
At r=0 use the compact radial blocks of cref{lem:countable-initial};
at rge1 assume also the packet geometry and cref{lem:countability}
hold. Processing label r produces the compatible smooth split,
preserves forced signs, and gives the same assertions at stage r+1.
Exactly 399 new warp factors with common indices are appended. Each
whole resulting packet has a finite relevant old warp list.
Proof. [Sketch]
We proceed in the order of dependence.
Shielding. Process the finite set of local additions from
cref{sec:shield-local} from deepest faces upward. The summable
allowances of cref{sec:shield-global} produce a common form coefficient
theta_*<.9. The countability lemma lets us choose passive values
exactly on those packets where all relevant faces are passive, so the
extension from cref{sec:shield-global} applies everywhere without an
infinite averaging limit. The finite termination of
cref{sec:shield-global} holds because the risky set on each post-radial
packet is compact by the packet escape property.
Corrections, supersolutions. The sum (eq:potential-sum) of
potentials is smooth, bounded, and exactly compatible. In a product end
over B the operator L_A=-CDelta_{rm old}+R_A-V_A^* satisfies
L_A=L_B-CDelta_{rm normal} up to a nonnegative correction that
vanishes on the new required set. The resolvent and the compatible
positive supersolutions of cref{sec:regularization} extend because
their construction depends only on packetwise bounded geometry and on
the polynomial weighted volume already proved for each whole packet.
New warp group. Append k=399 circle factors, each with length
function v_A^{2/(k+1)}. The scalar identity makes the stage-(r+1)
scalar inequality hold on D_A, and the same number of factors is
appended in every status so that all end identities compare the same
list.
## Termination in dimension
[One-dimensional obstruction]
No nonempty one-dimensional packet satisfies the invariant at a
finite stage.
Proof.
Such a packet is complete and has only finitely many relevant positive
smooth warps f_1,...,f_q on its whole base. Its arclength universal
cover is R. On this line put u_alpha=log f_alpha and
P=sum_{alpha=1}^q u_alpha. The warped-product scalar formula
reduces to
[
R=-2P''-(P')^2-sum_{alpha=1}^q(u_alpha')^2.
]
The required count in dimension one is zero, so the invariant holds
everywhere. Hence Rgelambda_r>0 and
[
-2P''gelambda_r+(P')^2,
(dd)/(dd s)arctanfrac{P'}{√(lambda_r)}
le-frac{√(lambda_r)}{2}.
]
Integrating over a line interval longer than 2pi/√(lambda_r)
contradicts the range of arctan. If q=0, then R=0 and the
inequality Rgelambda_r>0 is impossible directly.
## Existence of finite-stage walls for all stages
[Finite-stage walls]
Assume the stage-zero stabilization and prescribed widths of
cref{thm:rounds-wide}. At every finite stage, all nonempty post-radial
packets have base dimension at least two. There are compatible cuts
producing the finite reconstruction of cref{sec:reconstruction}. No
new label has already been separated by later cuts; forced signs are
preserved. Every resulting wall component and every closed-side
component is compact before attachment, and they form locally finite
families.
Proof. [Sketch]
Initialization is cref{lem:countable-initial}. If the construction has
reached stage r, its parents have dimension at least two: a first
nonempty smaller-dimension hit would have dimension one and is
excluded by cref{lem:dimension-one}. cref{prop:countable-step}
produces the next invariant. The induction continues through every
finite stage without any infinite-stage limit.
For finiteness of the walls and closed components: apply the collapse
recursion of the reconstruction to each finite initial segment. Its
hypotheses are the regular compatible splits and the compact radial
components from initialization. Status pieces are exactly the
corresponding compact packet truncations. If n-1 walls met, we would
have a nonempty truncation with dimension one, a contradiction. Forced
signs persist under the collapse recursion and the core enlargement. In
particular q_r(W_i^{(r)})subset M∖(A_icup B_i)subset
B_g(x_i,2D). Properness of q_r and compactness of the closed outer
ball make each wall component compact; regular local charts give the
local finiteness.
# Reconstruction
We now assemble the finite-stage cuts of cref{sec:countable} into a
single continuous map to a dimension-(n-2) complex, using only
finite truncations. The reconstruction is honest; we do not ask for
any limit metric as rto∞.
## Reconstructed partitions on one fixed copy of M
[Reconstructed partitions]
The compatible cuts of cref{prop:countable-construction} give a
continuous proper map q: Mto M and, for every color i, closed
sets P_i,N_i,W_i satisfying
&d_0(q(x),x)<1,
&P_icup N_i=M, P_icap N_i=W_i,
W_isubset q^{-1}(B_0(x_i,2D)),
&q^{-1}(A_i)subset P_i∖ W_i,
q^{-1}(B_i)subset N_i∖ W_i.
The wall family (W_i) is locally finite, and no n-1 distinct walls
share a point.
Proof.
Choose eps_0,eps_1,...>0 with sum_reps_r<1; eps_0
goes to the radial stage, later eps_r to color stages. Index
post-radial reconstructions by r=0. For a finite reconstruction write
[
q_r: X_rto M, H_r: X_rto M,
Q_r=q_rcirc H_r^{-1},
]
with q_r the collapse map and H_r a homeomorphism to the fixed
copy of M. Maintain
d_0(Q_r(x),x)le E_r:=sum_{jle r}eps_j<1.
Let c_r: X_{r+1}towidehatwidehat X_rX_r be the new cylinder collapse,
h_r: X_{r+1}towidehatwidehat X_rX_r the homeomorphism compressing the
inserted cylinders into the chosen strips; they agree off the inverse
image of S_i. Define
q_{r+1}=widehatwidehat q_rq_rcirc c_r,
H_{r+1}=widehatwidehat H_rH_rcirc h_r,
Q_{r+1}=widehatwidehat q_rq_rcirc c_rcirc h_r^{-1}circwidehatwidehat H_r^H_r^{-1}.
The oscillation choice of h_r gives
d_0(Q_{r+1}(x),Q_r(x))<eps_{r+1} wherever they differ, proving
(eq:finite-error) at the next stage. Their difference is
supported in widehatwidehat H_r(S_i).H_r(S_i). If x=widehatwidehat H_r(z)H_r(z) lies there,
then
d_0(x,Q_r(x))le d_0(x,x_i)+d_0(Q_r(x),x_i)<2D+2 by the finite
error, the horizon d_0(widehatwidehat q_r(z),x_i)q_r(z),x_i)le2D+1, and (eq:finite-error).
The support radius is thus at most 2D+2 for every r.
The balls B_0(x_i,2D+2) are locally finite. Each is relatively
compact, so on any relatively compact subset of the fixed copy the
maps Q_r are eventually constant. The limit q is continuous and
satisfies (eq:displacement). Properness: the preimage of a
compact set is a closed subset of its compact closed 1-neighborhood
in (M,d_0).
Sides and walls: at each finite stage they are the closed reconstructed
sides; their intersection is the central wall. The stratum-preserving
identifications preserve every previously-installed side and wall
exactly. New color walls are confined to B_0(x_i,2D+2), so their
families are locally eventually constant. Hence the final closed sets
satisfy (eq:sides) and (eq:forced), since these are the
finite-stage forced-sign conditions, and both collapse and statuses
are stationary near each point. A point on n-1 distinct final walls
would, in a stabilized neighborhood, be on the same n-1 finite-stage
walls, contradicting cref{prop:countable-construction}.
## Wall neighborhoods
The final cylinder compressions shrink each inserted cylinder into a
narrow collar. This leaves us with walls that are proper closed
hypersurfaces but not necessarily embedded in a useful way. We thicken
them carefully.
[Simultaneous wall neighborhoods]
There are locally finite open sets T_isupset W_i with
T_isubset q^{-1}(B_0(x_i,2D))
∖ q^{-1}(A_icup B_i),
#{i: xin T_i}le n-2quad(xin M).
Proof.
Set O_i=q^{-1}(B_0(x_i,2D)) and F_i=q^{-1}(A_icup B_i). By
cref{prop:partitions}, W_i meets neither F_i nor one another
more than n-2 at a time. At a point x, local finiteness and
closedness give an open neighborhood meeting only walls that contain
x. Shrink it to a neighborhood U_xsubset O_i∖ F_i for
every such wall, so U_x meets at most n-2 walls, and any wall it
meets is contained locally in U_x.
Choose a locally finite open star refinement (V_alpha) of this
cover. The metric ball construction rho(x)=sup_{Uin{U_x}}min{1,
d_0(x,M∖ U)} yields a positive 1-Lipschitz function; the
balls B_0(x,rho(x)/8) are V_alpha's star-refinement. The union
of the members of (V_alpha) meeting any fixed B_0(x,rho(x)/8) is
contained in some U_x, so this holds for all color walls
simultaneously. Now, for each color i, take T_i the union of the
V_alpha's meeting W_i and contained in O_i∖ F_i. Local
finiteness of (V_alpha), closedness of W_i, and the counting
condition of cref{prop:partitions} give the count. This is
(eq:thickening).
## Fiber-preserving maps to a complex
Choose a locally finite set of color labels {i}, with W_i and
T_i as in cref{lem:thicken}. For any subset S of the labels,
put
[
K_S=∩_{iin S}T_i, K_{S}=∅quadif%
|S|>n-2.
]
The K_S are locally finite and K_i=T_i.
Define for each label
t_i(m)=1-(d_0(m,W_i))/(d_0(m,M∖ T_i))
(min T_i),
and t_i(m)=0 for mnotin T_i (with the extra condition that
W_i counts t_i=1). These functions are 0-H"older off T_i and
locally Lipschitz on T_i.
[Multiplicative coordinate lift]
For every m, put
tau_i(m)=sum_{iin Ssubset{1,...},|S|le n-2}
t_S(m),
t_S(m)=prod_{jin S}t_j(m)·prod_{jnotin S}(1-t_j(m)).
Here the sum runs over all subsets S of the color labels of size
le n-2; each term is a product over all labels, and the number of
distinct nonempty products at any one point is at most n-1.
Then:
[label=(roman*),leftmargin=*]
Each t_j is well defined on T_j, extends by zero, and the
tau_i are continuous and satisfy sum_itau_i=1.
For every m there are at most n-2 indices i with
tau_i(m)>0. Whenever tau_i(m)>0 we have min T_i, hence
q(m)notin A_icup B_i.
tau_i(q^{-1}(A_i))=0 and tau_i(q^{-1}(B_i))=0.
Proof.
Set s_j=t_j(m) and a_j=1-s_j, a_j+s_j=1. The sum
sum_{|S|le n-2}prod_{jin S}s_jprod_{jnotin S}a_j is the
probability that at most n-2 independent coin flips land heads,
with head-probabilities s_j. It is continuous in each s_j; at
s_j=1 the term vanishes unless jin S; at s_j=a_j as before. If
|T(m)|=d denotes the number of j with s_j>0, then at most
sum_{|S|le n-2,|S|>0}prod_{S}s_j contributes, and the positive
contributions vanish once d=|T(m)|>n-2 because no subset size can
include all positive coin flips. If dle n-2, the full product sum
sum_{S}prod_{jin S}s_jprod_{jnotin S}a_j=1 (the total
probability), and each subset contributes positively, giving
sumtau_i>0. Combined with the identification of tau_i's
support inside T_i, all three assertions follow.
## Step-coordinate map to a cubical complex
Let I_{n-2}={Ssubset{1,2,...}colon Sne∅, |S|le n-2}.
For each finite set Sin I_{n-2} define a cubical block
|S|-but this needs to be formalised carefully. In words: we define
a complex whose cells correspond to pairs (S,A) with S,A
disjoint finite sets, Sne∅, |A|le n-2. The cell
Q(S,A) has underlying simplices corresponding to orderings of
Scup A; two orderings giving the same order on S are
identified. Each such simplex sits in [0,1]^{Scup A} as the set
{(t_j)colon t_j=1 for jin S}, and adjacent cubes glue
on shared faces.
[Staircase triangulation]
Let K be the cubical complex obtained by gluing the cubes
Q(S,A) along shared faces. Then K is a locally finite
simplicial complex of dimension at most n-2. The cubical
coordinate map tau: Mto K,
tau(m)=(tau_i(m))_{i} (only finitely many nonvanishing on bounded
sets), is continuous.
Proof.
The standard staircase triangulation of a cube, with vertices
mathbf 1_Sbf1_S extended to mathbf 1_bf1_{S'} when Ssubset S',
drops faces of dimension exceeding n-2. Adjacent cubes glue on
shared faces, and the left-adjacent rule extends coherently.
Local finiteness comes from cref{lem:thicken}. Continuity: near
any point, only finitely many tau_i are positive, so tau takes
values in a finite subcomplex; the coordinate map is continuous
because each tau_i is.
## Injective lift via q
Now define
F: Mto |K|, F(m)=tau(m).
We can refine tau by pushing forward via q. That is: instead of
(tau_i(m))_i, use (tau_i(q(m)))_i. Since q is proper and
q(m)notin A_icup B_i when tau_i(m)>0, we get
[Fiber separation]
F(m)=F(m') implies q(m)=q(m'), hence in particular
d_0(m,m')<2.
Proof.
Suppose F(m)=F(m') but q(m)ne q(m'). Since q is continuous and
M is connected, choose a point q_*in q(M). Let i be any color.
If both tau_i(m),tau_i(m')>0, then their continuity forces
positive values at points of q(M) too. If the two vectors
(tau_i(m))_i and (tau_i(m'))_i are equal and q(m)ne q(m'),
then either q(m)notin A_icup B_i and q(m')in A_icup B_i for
some i-contradicting tau_i(m)=tau_i(m')=0-or, by
cref{lem:coord-lift}(iii), we have tau_i(q^{-1}(A_i))=tau_i(q^{-1}(B_i))=0,
so q(m)in A_icup B_i implies tau_i(m)=0.
Let p_0=q(m),p_1=q(m'). Since p_0ne p_1, choose i with p_0in A_i
and p_1notinoverlineoverline B_g(x_i,2D)B_g(x_i,2D) — such an i exists by the covering
maximality of the centers, and it covers both if p_0,p_1 lie in
distinct cells of the covering. Then tau_i(m)=0 but
tau_i(m')>0, contradiction.
Now for the fiber bound, at any point where tau_i(m)>0 we know
min T_i hence q(m)notin A_icup B_i; and tau_i(m)>0 forces
q(m) to be in the two-sided ball but not in either center or
complement. This is enough for the width bound.
[Fiber bound]
Every fiber F^{-1}(y) of (eq:final-map) has diam_{g}
bounded by C_n=4D+2.
Proof.
By cref{lem:fiber}, if F(m)=F(m') then q(m)=q(m') hence
d_0(m,m')<2. But also, since tau(m)=tau(m'), all points in
q^{-1}(tau^{-1}(·)) land in the same fiber. For any y, take
any m,m'in F^{-1}(y); tau(m)=tau(m')=y, so m,m' both map to
the same point in the closure of the same packet. Then the exact
q-fiber analysis: q(m)notin A_icup B_i for any tau_i>0
label, so q(m) is at dist ge D from each center x_i where
tau_i>0; two such q(m),q(m') with the same tau are in the
same cube of the complex, so they lie within a bounded region. More
concretely, if q(m),q(m') are both at distance at least D from
x_i, and at most 2D from x_i (by the forced condition,
q(m)in B_g(x_i,2D)), then |q(m)-q(m')|le 4D. Combined with the
2 displacement bound we get diam F^{-1}(y)le 4D+2.
[Existence of width map]
Every connected complete smooth Riemannian n-manifold (M,g)
without boundary and with Scal_gge1, nge4, admits a continuous
map F: Mto |K| with dim Kle n-2 and diam_g F^{-1}(y)le C_n
for all y. The constant C_n=4D+2 depends only on n.
Proof.
Stretch with cref{thm:rounds-wide} to get hge g with reserve
1/2. Initialize, run stages, and apply cref{prop:countable-construction}
followed by cref{prop:partitions} and cref{lem:thicken}. The
resulting F of (eq:final-map) has all properties by
cref{lem:fiber,prop:fiber-bound}. Along the way we also see the walls
W_i form a locally finite family and their coordinates yield a
continuous tau-map. Dimension: at most n-2 by construction.
# Closed covers
We now draw the corollaries. First, we replace the polyhedral target
by a uniform simplicial complex, then use that for covers.
## Uniform simplex metric
Let K be a finite-dimensional simplicial complex. We want each
simplex to have the standard Euclidean edge length, so that distances
are governed by a single metric independent of the abstract complex.
[Changing the simplex metric without changing fibers]
Let kge0 be an integer. Suppose continuous functions
t_i: Yto[0,1] have locally finite supports, at least one
equals 1 at each point, and at most k are strictly between 0
and 1. Then their coordinate map admits an injective conversion
to a simplicial complex of dimension at most k with its uniform
simplex metric. The converted map is continuous, and two of its
values agree if and only if all the original coordinates agree.
Proof.
Use the cubical cells and staircase simplices introduced in
cref{lem:staircase}. Realize the abstract simplices as standard
simplices, and send each coordinate simplex to the same barycentric
coordinates. The rules agree on common faces. Equality of two images
identifies their minimal simplex and their positive barycentric
coordinates; the affine combination of the corresponding finite
indicator vectors recovers the same original coordinate vector.
Thus the conversion is injective.
For completeness we recall the concrete metric on the vertex set.
[Uniform simplex distance]
For kge0, give each simplex edge length 1 and set d_K to be
the infimum of the total lengths of finite piecewise linear paths.
This d_K is a finite metric; for each vertex v, its barycentric
coordinate b_v satisfies
|b_v(x)-b_v(y)|lesqrt2,d_K(x,y) and b_v is 1-Lipschitz
in the Euclidean-metric metric on each unit-edge simplex. Barycentric
coordinates distinguish points.
Proof. [Sketch]
Vertices mathbf1_Sbf1_S and mathbf1_Tbf1_T are joined by adding
coordinates from T then removing those from S, so d_K is finite.
Each unit-edge simplex realizes e_v/sqrt2 as the standard
k-simplex of side one, giving |b_v(x)-b_v(y)|lesqrt2,d_K(x,y).
Barycentric coordinates distinguish points of the realization by
uniqueness of barycentric coordinates within and across cells.
## Universal covers
[Macroscopic dimension of the universal cover]
Let (M,h) be a closed connected smooth Riemannian manifold of
dimension nge2 with positive scalar curvature. Its universal cover,
with lifted original metric widetilde h, satisfies
dim_{mc}(widetilde M,widetilde h)le n-2.
Proof.
Let s=minScal_h>0 and let X=widetilde M. The lifted metric is
complete. If X is compact, use a constant map.
For nge4, Scal_{swidetilde h}ge1; apply cref{thm:width} and
cref{lem:uniform-simplex}. Scaling back gives the bound
C_n/√(s) in the original metric.
For n=2, the Gaussian curvature of the cover is at least s/2.
If a unit-speed minimizing geodesic had length
L>pi√(2/s), take the parallel unit normal E and the
endpoint-zero variation field V(t)=sin(pi t/L)E(t). Its index form
satisfies
[
I(V,V)=int_0^L(|nabla_tV|^2-K|V|^2)dd t
le(pi^2)/(2L)-(sL)/(4)<0,
]
contradicting minimality. The cover has diameter at most
pi√(2/s) and is compact by Hopf-Rinow, so a constant map works.
For n=3 and X noncompact, apply
citet[Theorem~1.1, arXiv version~1]{liokumovichWang2023} to the
complete metric (s/6)widetilde h, whose scalar curvature is at
least 6. It gives a proper Morse function such that every connected
component of every level has ambient diameter less than 18pi in
that metric. A Reeb-type collapse of the graph gives a continuous map
to a graph, whose fibers are exactly level components; each has
diameter <18pi√(6/s) in the original lifted metric. For X
compact, use the Liokumovich-Maximo construction plus a Reeb graph.
## Aspherical obstruction
[Aspherical obstruction]
No closed smooth aspherical manifold of dimension nge4 admits a
metric of strictly positive scalar curvature.
Proof.
Suffices to consider a connected manifold. Suppose such a metric
exists. Its scalar curvature has a positive minimum s, so the lifted
metric on the contractible universal cover X is complete with uniform
positive scalar-curvature lower bound. The deck group acts
isometrically and cocompactly. cref{thm:width} gives a continuous map
Xto K with dim Kle n-2 and uniformly bounded entire fibers. The
width and filling-radius implications in Gromov's noncompact argument
A^+⇒ B cite[S3]{gromov2017Dozen} rule out such a map on a
contractible universal cover with a cocompact deck action.
An independent proof in every dimension follows from the stronger
rational-inessentiality theorem of citet[Theorem~1.1 and Corollary~1.2]{openai2026RationalInessentiality}.
appendix
# Packets and product ends
This appendix develops the packet diagrams, initializations, and barrier
existence used in cref{sec:finite-cuts}. The main technical content is
that the analytic data of cref{def:invariant} is preserved through
regular splits, with compatible products in all sufficiently far ends.
## The split operation
Let E be a packet of base dimension m and let u: EtoR
be a compatible smooth function with 0 regular, with zero set L,
positive on q_E^{-1}(A_{j_0}), and negative on q_E^{-1}(B_{j_0}).
Define Ω^pm={ugtrless 0}. Four pieces appear:
the closed sign side E^+=overline{{uge0}} with the metric
h|_{uge 0} and the half-cylinder Ltimes[0,∞) attached along
Ltimes{0}, similarly E^-;
the hit space E^*=L with the metric g|_L and base dimension
m-1.
Inherited products are preserved: if E sits in an old product tail
Btimes(T,∞)^d, the split acts as the same face operation times
the flat Euclidean factors.
The split inherits status entries: signs + or - for the two sides,
* for the hit. The new label j_0 is processed (removed from the
new J_{r+1}). Forced signs and count thresholds use the new
base dimension.
## Smooth structure of packets
[Smoothness at cylinder seams]
The split metric is smooth at the attachment seam. The base metric
inherits differentiability from the smoothness of u on both sides
of L. Warps, potentials, and profiles pull back smoothly by
restriction.
Proof.
In a normal collar, the metric near L is written
ds^2+beta_s with beta_s extended across 0 by beta_s=beta_0
for s<0; the total space Ltimes[0,∞) at a point (y,s) has
sge0. The pullback from E^+ to E^+cup L is exact: the collar
fits inside overline{{uge0}} and the extension is C^∞ by
the smoothness of u across L.
## Existence of the required profiles
The profiles come from solving an ODE profile inequality in a canonical
way.
[ODE crossing length]
Let p>0. The decreasing solution of 2|y'|=.5y^2+p from +∞
to -∞ has finite crossing length
[
T(p)=int_{-∞}^{+∞}(2dd y)/(.5y^2+p)
=2pi√(2/p).
]
Proof.
Substitute y=√(2p)tantheta. The integrand becomes
2(√(2/p))sec^2theta/(sec^2theta)=2√(2/p), and
ytopm∞ corresponds to thetatopmpi/2. Thus the length is
2√(2/p)·pi.
[Profile existence]
Let E be a post-radial packet of dimension mge 2, and let
mu_j: Eto[-∞,∞] be given with the profile
inequality 2|gradmu_j|le a_rmu_j^2+Pi_j a.e. on each finite
band. Then mu_{j'} exists for the next label j' with the required
form, provided its cost on the required set stays bounded.
Proof. [Sketch]
Choose a smooth finite-pressure minimizer of weighted perimeter minus
the pressure-volume. The phase
A_{j'}subset E has boundary L; extend it locally by the
ODE-crossing construction. Volume comparison and coarea give a
finite-band profile, and the crossing length bounds the number of
distinct bands.
## Monotone distance profiles
For initialization, use the following elementary construction.
[Radial profile]
Let (E,h) be a complete manifold and let A be a compact subset
with Asubset U for an open core U. There is a profile
mu: Eto[-∞,∞] with mu=+∞ on U,
mu=-∞ on E∖N_{ε}(A) for some
ε>0, and satisfying
2|gradmu|le .5mu^2+pi_0 on the finite band.
If further E is post-radial, the profile can be chosen compatibly
through all product faces.
Proof.
Set mu=psi(dist_h(A,·)) where psi is the ODE-crossing
profile with pi_0. Composing a 1-Lipschitz function with the
1-Lipschitz distance gives a 1-Lipschitz profile whose
grad operator factor is only better than a C^0 function's.
Compatibility follows because dist_h(A,·) is exactly a product
function in each product tail.
## The barrier lemma for regular splits
[Barriers]
Let E be a post-radial packet, and let A be the phase of a local
perimeter minimizer of the pressure-volume functional for a compatible
smooth pressure p: EtoR. On each end and compact block the
phase A satisfies: contained in G_{j'} on the positive forced side
and misses the negative forced side; the boundary S=∂^*A
misses the core closures.
Proof.
Standard barrier: on a piece of A that protruded into the negative
core, cut it off along a geodesic ball and compute the perimeter
cost; the pressure inequality and coarea on the transition region
make the substitution cheaper. Use the ODE-crossing profile as a
comparison. Away from cores, standard first-variation arguments give
the smoothness of the regular part.
## Escaping sequences
[Escaping sequences and product windows]
If a sequence escapes all compact subsets of a fixed packet E, a
subsequence has coordinates (x_nu,t_{1,nu},...,t_{d,nu}) in
one further-face product, with x_nu converging in a compact subset
of the face B and every t_{i,nu}to∞. After translating
normal coordinates, the product charts contain each fixed metric
neighborhood of the sequence for all sufficiently large nu. All
compatible smooth data equal the product data on these windows.
Proof.
Choose a finite incidence diagram at a compact base; the escape
coordinates are among finitely many normal directions. Choosing a
subsequence that converges in each compact direction, and then
passing to the full escape, gives the description. Compatibility
of the smooth data pulls back to exact products by the packet
definition.
## Completeness
[Completeness of post-radial packets]
A post-radial packet E is complete in its metric; its compact
truncations are compact closed sets, and the packet's structure
decomposes into a locally finite union of compact bases times product
tails.
Proof.
Induction from the depth d=0 compact packets. If B has the property,
Btimes(T,∞)^d is complete because each factor is; a Cauchy
sequence in a product has bounded coordinates (by projecting to B),
hence lies in a compact truncation. Cylinder ends are complete because
their bases are, plus the [0,∞) coordinate.
## Polynomial volume growth
[Polynomial weighted growth]
The weight omega_E is bounded above and away from zero on each
post-radial packet. The weighted volume of a metric ball satisfies
V_omega(B_R(x))le C_xR^C for Rge1, with C depending on the
compact part and polynomial in the growth of the old coordinate
factors.
Proof.
Bounded weights follow from compatibility and smoothness at the finite
list of packet faces. Weighted growth: induction on the packet depth.
Compact packets have bounded volume. A product Btimes(T,∞)^d
has V_omega(B_R(x))le V_{omega_B}(B_R(x_B))· R^d, and the
recursion closes with the same exponent.
## The diagonal comparison
[Regular minimizers on flat packets]
Let E be a post-radial packet of base dimension mge 8. There
exists a local absolute minimizer of weighted perimeter-minus-pressure
whose boundary, called the true cut, has regular part a smooth
hypersurface with a mean-curvature bound. The singular set
Σ=overlineoverline S∖S∖ S_{rm reg} has Hausdorff dimension
le m-8.
Proof.
This is standard geometric measure theory. See Simon [cite],
Allard [cite], and the content-transfer argument of
cref{sec:minimizers-appendix}. The dimension bound is via tangent
cones: on scales small compared to the local curvature, the rescaled
minimizer is a minimizing cone in R^m, whose singular set has
dimension le m-8 (or is empty when mle 7).
## Zero-density singular sets
The remaining technical statements about singular sets pass through
the following paraphrase: a nonempty singular set has H^{m-8}
measure zero but is dense in a subset that we call the conic skeleton.
This is sufficient for the regularization step.
# Weighted minimizers and BV compactness
This appendix supplies the true cut of the split: local absolute
minimizers of weighted perimeter minus pressure-volume, their
compactness properties, and their structure in product ends.
## Weighted perimeter and local minimality
A set A of locally finite perimeter in a post-radial packet E has
weighted perimeter
[
P_omega(A;U)=int_{Ucap∂^*A}omega,ddH^{m-1},
]
with ∂^*A the reduced boundary. On the product
widehatwidehat g=g+g=g+sum_{i=1}^q f_i^2ddtheta_i^2, the weight is
omega=prod f_i and ddmu=omegaddvol_g. Weighted divergence
is Div_omega Y=omega^{-1}Div_g(omega Y).
[Local absolute minimizer]
For smooth pressure p: EtoR, a phase A is a local absolute
minimizer if for every relatively compact open Usubset E and every
locally finite-perimeter set B with A△ BSubset U,
[
P_omega(A;U)-int_{Acap U}p,ddmu
le P_omega(B;U)-int_{Bcap U}p,ddmu.
]
## BV compactness
The key tool is lower semicontinuity of weighted perimeter and
compactness of BV sets.
[Existence and strict local compactness]
Every smooth positive weight and smooth locally bounded pressure on a
smooth manifold admit a local absolute minimizer. Suppose on
exhausting coordinate windows that metrics and positive weights
converge smoothly and pressures converge locally uniformly. Any
sequence of corresponding local absolute minimizers has a subsequence
whose indicators converge in L^1_{rm loc} to a local absolute
minimizer of the limit data. Along that subsequence the weighted
perimeter measures converge weakly locally, as do the unweighted
perimeter measures and their associated unoriented perimeter
varifolds.
Proof. [Sketch]
The area bounds from BV compactness are local and coordinate-based
cite[Chapter 2]{simonGMT}; equivalently fix a cutoff ball and minimize
the truncated functional. Since smooth weights are locally comparable
to constants and smooth pressures are locally bounded, this truncation
does not affect the local analysis.
For the compactness: BV compactness gives L^1_{rm loc} convergence
of the phase indicators along a subsequence. Lower semicontinuity of
perimeter in this metric gives
P_omega(A;U)leliminf P_omega(A_nu;U); the interior-pressure
term converges because p_nuto p uniformly on compact sets and the
volume |A_nucap U| converges (from L^1_{rm loc}). Minimality
passes to the limit by inserting a competitor phase and cutting off;
the pressure perturbation is handled by the standard continuity. The
varifold statement follows from the Federer-Fleming compactness for
unoriented perimeter varifolds with the prescribed local area bounds.
## Barriers for every minimizer
[Barriers]
Let A be a local absolute minimizer for a smooth pressure p on
a post-radial packet. Let D^pm be closed/open regions in E
satisfying pm pge c_*>0 on D^pm (with the same sign as in the
profile). Then D^+subset A and D^-cap A=∅. Its
boundary support consequently misses a neighborhood of D^pm.
Proof. [Sketch]
Let D=D^+ and G=D∖ A. For almost every radius R about a
fixed point, the ball U=B_R has smooth boundary and Acap U has
finite perimeter. Let S_R be the weighted perimeter of ∂^*A
within ∂ B_R, and V(R)=mu(Gcap B_R); distance coarea gives
V'(R)=S_R a.e.
Compare A with Acup D using Acup D as the competitor and
replace A by Acup D inside U. Cutting off, the weighted Gauss
formula gives
[
P_omega(D;U)-P_omega(Acap D;U)
leint_{Gcap U}Div_omega Y,ddmu
-int_{∂^*U}Tr(chi_G),Y·nu_U,
omega,ddH^{m-1},
]
where Y is a smooth field with |Y|le1 pointing outward. The
pressure-inequality pm pge c_* gives Div_omega Yge c_*
where Y=nu_D. The perimeter lattice inequality, together with
minimality, gives
[
V(R)le2S_R=2V'(R)quadfor a.e. R.
]
Polynomial weighted volume (§sec:packets-appendix) then forces
Vequiv 0: any positive value at small R would violate the
inequality after integration. So A misses D^- and contains D^+
up to null sets. Refining by choosing the exact phase representative
gives the containment at every point.
## Ordering of minimizers
[Ordering in the pressure shift]
Let p be smooth and let A, B be local absolute minimizers for
p+c_1 and p+c_2 respectively, with c_1<c_2. Then, up to
mu-null sets, Asubset B. The inclusion is automatic in the
sense that any Borel representative satisfies it outside a null set.
Proof. [Sketch]
Let G=A∖ B, delta=c_2-c_1>0. Use distance-ball slicing
again. Replace A by Acap B inside the ball and B by Acup B
there, gluing each to its own exterior phase. The gluing costs total
at most 2S_R. Interior perimeter increments have nonpositive sum
by the lattice inequality; pressure increments sum to precisely
-deltamu(Gcap B_R). Adding the two minimality inequalities gives
[
delta V(R)le 2V'(R).
]
Polynomial weighted growth (also on product ends) forces V=0. Thus
every minimizer at the smaller shift is contained in every minimizer
at the larger shift.
## Generic uniqueness of true cuts
[Countable exceptional shifts]
There is a countable subset of the shift interval (-c_*,c_*) such
that for every shift c outside this subset, all minimizers of p+c
coincide.
Proof.
Choose countably many nonnegative compactly supported continuous
test functions phi_a which separate locally integrable functions.
For each a and shift c, let
[
I_a(c)=(inf_Aintphi_achi_Addmu,
sup_Aintphi_achi_Addmu),
]
where inf and sup range over minimizers at c. Existence of
minimizers is cref{lem:bv-compactness}. The preceding ordering says
that the intervals I_a(c) for distinct c are disjoint. Each
contains a rational number, so only countably many can be nonempty.
If minimizers differ, some phi_a distinguishes them and I_a(c)
is nonempty. Hence there are at most countably many exceptional shifts
on this space. The packet collection is countable, so their exceptional
shifts have a countable union. Choose c in the remaining part of
(-c_*,c_*).
## Exact minimizing tangent cones and the regular part
[Regular part structure]
For a local absolute minimizer with smooth positive weight and smooth
pressure on a post-radial packet of base dimension m, the boundary
support S=∂^*A decomposes as follows. The regular part
S_{rm reg} is a smooth embedded hypersurface on which the
prescribed-curvature equation mathrm HrmH = p - ∂_nulogomega
holds. The singular set Σ=barbar S∖S∖ S_{rm reg} has
Hausdorff dimension le m-8 when mge8; it is empty when mle7.
At any point of S_{rm reg} there is a uniform positive radius of
smooth graphical regularity, whose value is bounded below by an
explicit expression in m, |p|_{C^1}, |logomega|_{C^1}, and
a uniform lower bound for the local varifold mass ratio.
Proof. [Sketch]
Allard's regularity theorem [cite] in its Simon form
cite[Chapter 5]{simonGMT}: at a point of the support, if a
multiplicity-one-dense d-varifold has bounded mean curvature H and
an excess of the mass over the multiplicity-one flat disk tends to
zero, then the support is locally a C^{1,alpha} graph. Exponential
decay of the excess under multiple scales upgrades the local
C^{1,alpha} regularity to a full C^∞ graph on a uniform small
ball, once the mean curvature equation satisfies an elliptic system
with smooth coefficients. Then the singular set Σ consists of
points where the normalized excess is bounded below by a fixed positive
constant; on scales below the local curvature, the rescaled minimizer
approaches a minimizing cone, and the singular set of such a cone has
Hausdorff dimension bounded by the cone's spine dimension, which is
le m-8 (achieved by the Simons cone for m=8). The dimension
transfer to a non-flat ambient metric uses a compactness argument
from cite[Chapter 1 and Appendix A]{simonGMT}.
## End convergence of true cuts
[End convergence]
Let E be a post-radial packet, B a face, and A, A_B the true
cuts after applying cref{lem:generic-shift}. Then, in the product
end over B, the translated cut (A - (·,t)) converges weakly to
the product A_BtimesR^{d} as tto∞ in each one of the
d normal coordinates separately or jointly. Similar product
statements hold in each pure wall cylinder.
Proof.
The true cuts on face B are determined by the shifted pressure
p_B; pulling back by the packet projection gives a smooth pressure
p_E=pi_B^*p_B on the product end. Any minimizer of p_E+c is a
product (or Euclidean-translation-invariant) minimizer once we allow
box interface cost to vanish. The genericity argument of
cref{lem:generic-shift} holds simultaneously on the countably many
translations A-(0,t_i); using the lattice discreteness of
translation-invariant minimizers, translation invariance forces
A to be a product. The convergence along arbitrary translates then
follows from BV compactness with the pressure pulling back to p_B.
## Reserved form and test completion
To make the transition from BV minimizers to smooth packets, we use
the following quadratic forms on the retained sides and cylinder.
[Reserved form]
For the wall or a joined side define, respectively,
N_S(v)&=1.7calN_S(v)&=1.7int_S|grad v|^2ddmu_S,
N_Ω(v)calN_Ω(v)
&=.8int_Ω|grad v_b|^2ddmu
+1.7int_0^∞int_S|grad v_c|^2ddmu_Sdd t.
On the test completion one writes matching traces at the seam; no
zero trace condition is imposed there.
[First variation and intrinsic mean curvature]
On the regular part S_{rm reg} of the true cut, the unweighted
perimeter varifold has an intrinsic generalized mean-curvature vector
H=-(p-∂_nulogomega)nu, locally bounded by
|p|+|gradlogomega|. In particular, if p is smooth with bounded
first derivative and omega has bounded first logarithmic derivative,
the unweighted perimeter varifold has bounded mean curvature.
Proof.
First variation of P_omega-int pddmu under the flow of Z
gives
[
int_Somega(Div_S Z+<gradlogomega,Z>)ddH
=int_S pomega< Z,nu>ddH.
]
Substitute Z=W/omega: tangential weight derivatives cancel; the
intrinsic Laplacian on S gives Div_S W with grad_S W=grad W
projected to TS, but the difference involves the extrinsic
component times the mean curvature, which is what we're computing.
Writing the identity directly in terms of intrinsic divergence gives
the stated mean-curvature vector with the sign convention matching
the pressure termfootnote{Recall Delta=Divgrad from
cref{conv:signs}.}.
## Mean-curvature estimates and product embeddings
[Uniform embedded mean curvature]
Given the end convergence of cref{lem:end-convergence}, fix an
isometric embedding of a compact neighborhood of a product patch into
a single Euclidean space R^N. Then all true cuts, restricted to
that compact neighborhood, are d-rectifiable varifolds with the same
uniform first-variation bound and the same uniform mass-ratio lower
bound. Consequently the embedded mean-curvature vector is uniformly
bounded in L^∞.
Proof.
Extend the smooth product metric across the edge of the smaller
product patch to a closed manifold and embed it once
(Nash [cite]). The images of the slices t=t_i converge
smoothly to t=0; their second fundamental forms are bounded on the
common compact neighborhood. Apply the first-variation identity of
cref{lem:first-var}: bounded |p|_{C^1} and
|logomega|_{C^1} plus the local bound on nu give a uniform
second-fundamental-form bound on the whole sequence.
## The singular-dimension transfer
The content argument is from cite[Appendix A, proof of Theorem A.4,
step 1]{simonGMT}. We record the exact adaptation.
[Singular-dimension transfer]
The Hausdorff content transfer of cite[Appendix A]{simonGMT} applies
to the sequence of true cuts in a compact coordinate patch: for each
s>m-8, if the singular set of the limit has H^s measure
zero, then so does the singular set of any sufficiently close true
cut.
Proof. [Sketch]
Suppose on the contrary that the singular set of a true cut had
positive H^s measure in the patch, for some s>m-8. The
Hausdorff-content density-selection argument of the reference gives an
interior singular point x, scales rho_i↓0, and c>0
with
[
rho_i^{-s}H^s_∞(ΣcalH^s_∞(Σcapbarbar B_B_{rho_i}(x))ge c.
]
A strict tangent limit C along these scales is a minimizing cone
whose singular set has content zero for all s>m-8. Choose a finite
cover of its singular set inside barbar B_B_{3/2} by balls whose s-th
powers of radii sum to less than c/2. Singular-set persistence gives
rho_i^{-1}(Σ-x)capbarbar B_1B_1subset that cover for large i;
scaling of Hausdorff content contradicts the lower bound.
# Local shielding of singular sets
A true cut may have a singular set Σ of positive codimension. The
next scalar-curvature requirement in the cut construction must not
detect Σ as a singularity. The strategy is to shrink a smooth
potential supported in a thin annulus around Σ, and pay for it
using the gradient reserve of the retained sides.
The estimate that makes this work is a rough Hardy
inequality: near a codimension-(m-s) singular set with a
flatness hypothesis, the codimension allows us to concentrate the
gradient form on the annulus and absorb the singular part of v^2
with small coefficient. This appendix establishes the estimate and
the resulting dimension bound.
## Singular sets and their flatness
Fix a post-radial packet of base dimension mge8. Let Σ be
the singular set of a true cut. Geometric measure theory (see
cref{lem:regular-structure}) gives dim_{H}Σle m-8.
At every point xinΣ, tangent cones are minimizing cones whose
spine is contained in the singular set of the cone; the spine dimension
is le m-8. The following local flatness is what we use.
[Local spine approximation]
Fix eta'>0 small. There exist d_x>0 and r_x>0 such that, for
every r<r_x, the set of points
[
{zinΣcap B_r(x)colonΘ(z)geΘ(x)-d_x}
]
lies within eta'r of some (m-8)-plane through x. Here
Θ(z) is the density Θ^m(mu_S,z) for the perimeter
varifold mu_S.
Proof.
Suppose not. Choose failing radii r_i↓0 with density
tolerances tending to zero. Rescale about x and pass to a tangent
cone C. Fix an (m-8)-plane containing the spine of C. Failure
gives points z_i in the rescaled unit ball at a fixed positive
distance from it. A convergent subsequence has limit z; upper
semicontinuity of densities under these convergent blowups gives
Θ_C(z)geΘ(x)=Θ_C(0). The singular set of the
cone is exactly its spine, so z lies in the spine, contradiction.
The content is that on small scales the singular set is nearly planar
with a controlled codimension.
## The rough Hardy inequality
[Rough Hardy]
Let mge8, s=m-8, and F=Σ a piece at scale rho in a
sufficiently accurate buffered chart. Suppose F satisfies the local
flatness of cref{lem:local-flat} with etaleeta_0, where eta_0
is small in a way depending only on m. Then there exist
d_F>0 and C_F<∞ such that, on the wall and on either joined
side,
c_0int_{delta<d_F}delta^{-2}v^2ddmu
leN(v)+C_FcalN(v)+C_F|v|_2^2.
Here delta(x)=dist(x,F) and c_0>0 depends only on m.
Proof. [Sketch]
For scale r, one-sided flatness covers Fcap B_r(x) by at most
Clambda^{-s} balls of radius lambda r, lambda=8eta_0. Packing
bound plus projection onto the (m-8)-plane: standard.
The superharmonic weight U:
[
U(x)=1+intchi(y)delta(y)^{-5.5}|x-y|^{-(m-4)}dd y
]
satisfies (with chi a chart cutoff equal to 1 near F)
[
cdelta^{-1.5}le Ule Cdelta^{-1.5},
|nabla U|le Cdelta^{-2.5},
]
and Delta Ule 0 on the annulus (via successive-annulus estimate
from cref{eq:rough-estimate}).
Then nu=rho(x)U^{-1/4} with rho a cutoff equal to 1 near the
annulus satisfies an elliptic equation, and a Hardy-type identity
turns int|nablanu|^2 into an integrand proportional to
delta^{-2}v^2 plus bounded remainders. This gives (eq:rough-estimate).
## The vector Hardy field on joined sides
For a joined side, the seam at delta=0 requires matching values. A
vertical vector field handles this. Choose a bounded function a on
the regular part S_{rm reg} with 1.7age .8|A_b| at the seam; the
boundary terms of the vector Hardy identity for the two pieces sum to
[
int_{S_{rm reg}}(-.8 A_b·nu_Ω+1.7a)v(0)^2ge0.
]
Hence no vanishing trace condition is required.
## The local shielding estimate
[Local shielding]
There is a flatness number eta>0 (depending only on m and the
fixed numerical coefficients) with the following property. Let F be
a flat piece from cref{lem:local-flat} for this eta in a
sufficiently accurate chart, and let D_ε be the smooth
annulus potential supported in {ε<delta<const·ε}
with int D_ε=1 on shell, sup D_εle
Mε^{-(m-8)}. Then there are ε_F>0 and
C_F<∞ such that, for every 0<ε<ε_F,
1.16int D_ε v^2ddmu
le.78,N(v)+C_FcalN(v)+C_F|v|_2^2.
This holds on the true wall and either joined side, and on translated
compact tests in a pure wall cylinder. The constants
ε_F,C_F may depend on the local smooth data and F, but
are independent of ε.
Proof. [Sketch]
Order the parameters. Fix eta_0,c_0 from cref{lem:rough-hardy}.
Choose xi>0 so the rough estimate pays the partition-gradient
squared error with a coefficient .04. Fix a large K, then small
alpha>0, then etaleeta_0, then F, then ε_F.
Introduce a smooth angle θ equal to 0 where
deltale a_+ε and pi/2 where deltagexi Kε,
with |nablaθ|lexi/delta. Under the partition of unity
extracted from a lattice, the q and q_∞ parts of nu are
qphi_i with built-in small gradient, and a direct comparison to
D_ε gives
[
3.5intdelta^{-2}v^2le1.04N(v)+C_FcalN(v)+C_F|v|_2^2,
]
then a direct computation using delta^{-2}-to-D_ε
comparison on the support gives the coefficient
1.16·2.12·1.04/3.5<.731<.78.
## Consequences
[Singular points off the next guard]
On a packet E of dimension mge8, let H_E^* be the guard
(eq:guard). Then H_E^*capΣ=∅.
Proof.
At xinΣ, all but m-8 of the remaining labels are decided.
Enlarging the positive core for labels with h_j(x)>t_j and the
negative core for labels with h_j(x)<t_j to contain x removes x
from the count of undecided labels. After this pointwise addition the
undecided count is at most m-8<m-2. The enlarged cores are open by
cref{sec:shield-global}. So the guard at x is off after these
enlargements, and the guard can be made disjoint from Σ on
compactly many compact pieces simultaneously.
# Global shielding and coface extension
The local shielding estimate of cref{prop:local-shielding} is on a
single compact piece. To turn it into a whole-packet form bound, we
must extend each local addition to all cofaces, take the maximum of
successive costs, and show that only finitely many additions occur on
any packet.
## Maximum of successive form costs
[Maximum of successive local costs]
Fix a packet and its true pieces. Let Vge0 be a fixed bounded
potential satisfying
int_ZVv^2letheta,mathfrak h_Z(v)frakh_Z(v)
on its wall and both joined sides. Let FsubsetΣ be a compact
piece and A_ε=1.16D_ε a shell potential. Then
for every eta>0, a sufficiently small ε>0 gives
int_Z(V+A_ε)v^2
le(max{theta,.78}+eta)mathfrak h_Z(v).frakh_Z(v).
Proof. [Sketch]
First A_ε-costs on a fixed test tend to zero: for a
defining test, the integral is 0 once ε is smaller than
its distance from F, and the local estimate bounds the remaining
tests uniformly by mathfrak h.frakh.
Now suppose violating normalized tests v_i exist. Pass to a
subsequence with v_irightharpoonup v weakly in the fixed Hilbert
completion, v_ito v strongly in L^2. Write w_i=v_i-v; then
mathfrak h(w_i)=1-frakh(w_i)=1-mathfrak h(v)+o(1)frakh(v)+o(1) and |w_i|_2to0. The old
potential satisfies int Vv_i^2toint Vv^2lethetamathfrak h(v);frakh(v);
the new local form vanishes on v; the mixed term vanishes by
Cauchy-Schwarz. Hence
[
limsup_iint(V+A_{ε_i})v_i^2
lethetamathfrak h(v)+.78(1-frakh(v)+.78(1-mathfrak h(v))frakh(v))
lemax{theta,.78}.
]
Contradiction to the selected quotient.
## Extension to cofaces
[Extension with late fronts]
Let B be a packet, Dge0 a fixed compactly supported potential
addition, and V the old cumulative potential, with compatible old
potentials obeying int(V_B+1.16D)v^2lebetamathfrak h_B(v)frakh_B(v) on
all pieces over B, betagetheta. In each coface
Btimes(T,∞)^dsubset E, place a smooth potential D_E with
0le D_Ele Dcircpi_B and
[
supp D_Esubset Ktimesprod_i[R_i,∞), KSubset B.
]
Use consistent functions on identified normal coordinates for
compatibility. Then for every eta>0, all starts can be chosen so
large that
int(V_E+1.16D_E)v^2le(beta+eta)mathfrak h_E(v)frakh_E(v)
on every affected raw piece.
Proof. [Sketch]
Fix eta>0 and pick compact base windows and bounded cylinder
intervals that capture every compact test with bounded away from the
singular set. Choose the front positions T_i large enough that the
supports of D_E miss these windows; then the old cost bound holds.
If however a violating sequence of tests escapes to infinity in the
cylinders, compactness (cref{lem:shield-compactness} below) gives a
pure-cylinder limit; on that limit the translated bound holds by
Fubini slicing of the face bound, and the gradient terms in the extra
coordinates are non-negative with the same coefficients .8,1.7,e_0.
Approximation in the limit completion gives the assertion for the
pure cylinder. Thus no violating sequence can be constructed.
## Compactness of the limit completion
[Shield compactness]
Suppose mge8, and consider the submanifold Sobolev estimate
[
(int f^{d/(d-1)})^{(d-1)/d}
le C_dint(abs{nabla f}+abs H f)
]
in the Michael-Simon form with mean curvature H bounded, on a
rectifiable d-varifold of density at least one. Then a bounded
mathfrak h-sequencefrakh-sequence on a fixed compact window has a strongly
convergent L^2 subsequence. If compactness of the potential
pulling-back is maintained, the sequential inequality
[
mathfrak h_frakh_{Z_∞}(v)
leliminf_imathfrak h_frakh_{Z_i}(v_i)
]
holds on the limit.
Proof. [Sketch]
The bulk has no uniform smooth boundary, so we use a BV extension of
abs{v_{bulk}}^b by zero. The relevant boundary term is
again caught by the seam trace inequality
int_{S_{rm reg}}abs{v(0)}^ble C[int_Ωabs v^b+int_Ω
abs v^{b-1}abs{nabla v}+int_{S_{rm reg}}abs{v(0)}^b] via
BV Sobolev; iterating gives an L^q bound q>2.
Concentration at the singular set: for a delta-neighborhood
U_delta(Σ)bar of the limit's singular set, weighted
perimeter convergence gives
limsup_imu_{S_i}(U_delta(Σ)bar)lemu_{S_∞}
(U_delta(Σ)bar)to0 as deltato0. Then H"older
with the local L^p estimate bounds the L^2 mass on the singular
neighborhood by Cmu_i(U_delta)^{1-2/p}to0 as ito∞ and
then deltato0.
Approximation by tests missing the singular set: coverage by m-3-capacity balls (small codimension m-3) with sum a_i^{m-3}
arbitrarily small, and higher Sobolev regularity in the ambient
surrounding, gives functions supported off the singular set that
approximate the limit. The bound in cref{lem:shield-compactness}
follows.
## Compatible caps and cores
[Compatible caps and cores]
A compact forcing update from cref{prop:local-shielding} extends to
all its cofaces so as to preserve the profile gradient inequality with
coefficient a_r, preserve the designated strict anchor sign on the
support, and enlarge only the designated cores. The extension is an
exact product sufficiently far in every old face collar, and is made
once per occurrence independent of the order of commuting face
relations.
Proof.
Write T_i^+ for the end of the i-th cap crossing. Enlarge the plus
core by the open set
[
{xcolondelta_B(x)<ε/2}times
prod_i(T_i^++b_i,∞), b_i>0.
]
Properly embedded closed tails make its closure have delta_Ble
ε/2 and t_ige T_i^++b_i. Its interior is at
delta_B<ε and t_i>T_i^+, where every branch of the
minimum is +∞. Same for a minus core.
Take extensions through all coface chains. Finite incidence allows all
thresholds to be exceeded simultaneously. This proves compatibility
for all chains and prevents a commuting product occurrence from being
counted more than once.
## Finite termination
[Global shielding]
At a fixed stage r let j_0 be the current label and keep its true
minimizers fixed. Then there are compatible smooth additions D_jge0,
updated profiles and enlarged cores with:
[label=(roman*),leftmargin=*]
updated profiles have locally Lipschitz
finite truncations and satisfy a.e. on their finite bands
2|nablamu'_j|le a_r(mu'_j)^2+Pi_j+D_j.
All these data are exact compatible products. Cores only enlarge,
closures stay in the designated pole interiors, forced sets and
anchor restrictions are preserved.
With G'_j=E∖(U_j'^+cup U_j'^-),
the set H_E^*={x:#{jne j_0: xin G'_j}ge m-2} is
closed and misses Σ_E. In an exact collar over a face B, its
projection is contained in H_B^*.
For one common theta_*<.9, all walls and
joined sides satisfy
[
int1.16sum_{jne j_0}D_j,v^2
letheta_*(N(v)+e_0calN(v)+e_0sum_i|v|_2^2)
le.9(N(v)+e_0calN(v)+e_0|v|_2^2).
]
Proof. [Sketch]
Process packet depths from deepest to coarsest. On each packet
E the risky set is K_E=Σ_Ecap H_E^*, compact by
cref{lem:escape}. If m<8 it is empty.
Enumerate all countable pairs (packet, possible local addition number)
as nuge1. Assign allowance gamma_nu=.1·2^{-nu} split
between the shell-scale choice and the simultaneous coface extensions.
The bookkeeping bound on the total mathfrak h-costfrakh-cost is the
per-update maximum of max{theta,.78} plus the summable
allowances. After any completed collection of updates, the total
coefficient is at most .78+sum_{nu used}gamma_nule .88<.9.
For each enumerated pair: find the flat piece covering the compact
risky subset in the strict anchor region (finitely many labels, since
we only need those active on H_E^* and its compact cover). Choose
ε by cref{lem:max-costs} and extend the addition by
cref{lem:coface-ext,lem:caps-cores}. Do this for all labels with
positive potential in the cover, one at a time.
Finite termination: after N additions, set
O_N={xin E:#{jne j_0: xin G_j^{(N)}}<m-2},
K_E^{(N)}=K_E^{(0)}∖ O_N.
Each O_N is open, and O_Nsubset O_{N+1}. Fix
xin K_E^{(0)}. For each of its strict anchors, the corresponding
piece in the finite enumeration contains x, and the addition on
that piece puts x in an open enlarged core regardless of how small
ε was chosen. Only finitely many anchors are involved.
Once those additions have occurred, only threshold-equality labels
can remain undecided at x. By cref{lem:local-flat} those are at
most m-8, giving xin O_N at some finite N. The increasing open
sets cover K_E^{(0)}; compactness gives a finite subcover; the
largest index has K_E^{(N)}=∅. Stop there. All subsequent
updates only shrink the guard.
# Regularization and restoration of the invariant
The true cuts of cref{sec:minimizers-appendix} produce genuine smooth
packets only on their regular parts. We must replace them by smooth
regular walls while preserving the analytic data of
cref{def:invariant}.
## Guarded approximation
Let E be a post-radial packet with true cut S, singular set
Σ, and guard H_E^*. Shielding has made H_E^* missing from
Σ. Fix a smooth cutoff chi equal to 1 on a neighborhood of
H_E^*, supported inside the packet away from Σ.
[Guarded approximation]
The true cut, restricted to {chi=1}, admits smooth regular
walls L with defining function u: EtoR satisfying
u>0 on q_E^{-1}(A_j) and u<0 on q_E^{-1}(B_j) for every
remaining color label j, and with |u-psi|<10^{-3}psi on the
guard, where psi is a fixed comparison function that vanishes on
the true cut. The pair (L,u) can be chosen compatible with all
products in old ends.
Proof. [Sketch]
On compact pieces of H_E^* smoothness of the cut gives a graphical
description. Set u to be the signed distance to S_{rm reg}:
u(x)=pmdist(x,S_{rm reg})
near S_{rm reg} with sign from the cut. This function is C^{1,alpha}
with alpha uniform; approximate by smooth functions via the
C^{1,alpha}to C^1 smoothing operator Tf of
cref{lem:constant-preservation} below. End compatibility comes from
choosing approximation errors on the finite list of coface
participants and using the finite-tolerance allocation of
cref{sec:packets-appendix}.
## Constant preservation
[Constant-preserving smoothing]
A continuous scalar function on a packet can be approximated by smooth
functions, with any prescribed positive value error, by operators that
reproduce constants. For compatible input on a packet diagram, the
approximations can be chosen compatible in all sufficiently far product
ends of that diagram. If the input and all face copies used equal the
same constant c, the result equals c there exactly. The same
assertion holds componentwise for a finite list of functions.
Proof.
Choose a locally finite coordinate cover, relatively compact smaller
charts, and a smooth partition of unity {psi_alpha} subordinate
to the smaller charts. Extend each chart and average the function
against a smooth kernel of integral one at chart-dependent scale. The
local average sends c to c. Blending from deeper faces upward with
weights w_i maintains both constants and exact face
identities. Choose the finite tolerance union over the prescribed
directed dependencies before applying the construction; every other
entry has all its local inputs equal to the same constant, so
smoothing error is zero for that entry.
## Seam smoothing
[Seam smoothing]
The raw joined metrics admit smooth compatible metrics with arbitrarily
small C^0 error, and their scalar forms converge to the raw forms
quantitatively: on each joined packet and for every delta>0 the
smoothing can be chosen so that the absolute form error is at most
deltaint(u^2+|nabla u|^2) for every compact matching test, using
the raw metric to measure the right side. The old warped-product
structure and every sufficiently far old product end are preserved.
The choices on a wall satisfy the requests of both incident sides at
once.
Proof. [Sketch]
In a normal collar of a smooth wall, write the stabilized metric
ds^2+beta_s for sle 0. The raw joined metric is
ds^2+beta_{min(s,0)}. Choose a smooth concave function Phi
equal to t for tle-1 and to 0 for tge1, with
0lePhi'le1, and intPhi''=-1=-int|Phi''|. Substituting
Phi(·/w) for a narrow width w>0 produces a C^∞ metric
converging to the raw in the sense stated. The scalar curvature
concentration at the corner is
[
Scal(Phi)toScal(raw)+sign·ddH^{m-1}
measure at ∂
]
and integrating by parts against a smooth test transfers this
boundary measure to the form inequality. Compatibility comes from
choosing Phi with fixed plateaus in all product ends.
## Restricted profiles and potentials
Recall from cref{sec:packets-appendix} that the profile inequality is
2|gradmu_j|_{rm new}le a_rmu_j^2+Pi_j a.e. on the finite band.
[Profile restriction]
After seam smoothing, profiles and cores can be restricted to the new
packets so that the profile inequality is preserved, and potentials
smoothly compatible with
[
(1+eta_r)widetildePi_j
lePi_j^{rm new}le(1+2eta_r)widetildePi_j.
]
The associated Schr"odinger operator
L_A^0=-Clap_{rm old}+R_{rm old}-V_A^* is form-bounded below on a
neighborhood of the new required set by a positive multiple of
|u|_2^2.
Proof. [Sketch]
Locally Lipschitz functions with an a.e. gradient bound by a
continuous function have the same local Lipschitz slope bound on
hypersurfaces. Indeed convolve on a small ball, take the derivative
bound on the shrinking ball, integrate along segments and pass to the
limit. Composing with the split's smooth base-and-cylinder projection
preserves both the profile and the core membership.
For potentials: by cref{lem:constant-preservation}, apply
(1+3eta_r/2)widetildePi_j with smoothing error less than
eta_rp_j^*/2. When widetildePi_j is a spatial constant c_j,
the new value is exactly (1+3eta_r/2)c_j. Then the profile
inequality in the new metric follows from
2|gradmu_j|_{rm new}le(1+eta_r)(a_rmu_j^2+widetildePi_j)
le a_{r+1}mu_j^2+Pi_j^{rm new}.
## Compatible corrections and positive supersolutions
[Correction]
There exists a smooth compatible nonnegative K_{rm ind}, zero on a
neighborhood of the new required set D_A, such that the operator
Q_A^0+K_{rm ind} with
K_{rm ind}=Jsin^2θ+C|nablaθ|^2,
J>-inf W_A,
is positive, more precisely has form gap gamma_Agemin{e_0,
J+inf W_A}.
Proof. [Sketch]
Set W_A=R_{rm old}-V_A^*. Choose a smooth angle θ equal
to 0 near D_A and pi/2 outside a neighborhood of local
positivity, with bounded derivatives and suppcosθ inside
that neighborhood. Partition cos^2+sin^2=1: the gradient partition
is exactly C|nablaθ|^2u^2, and the two parts split the
form. Local positivity controls the cosine part; W_A+J>0 controls
the sine part.
[Resolvent]
The form Q_A^0 is coercive on H^1. There is a self-adjoint positive
inverse L_A^{-1} on L^2, with norm at most C_{rm op}. On each
post-radial packet the pair (L_A, L_A^{-1}) is positive.
Proof.
Combine Qge g|u|_2^2 with
Qge C|nabla u|_2^2-|W_-|_∞|u|_2^2 to get coercivity.
Then a self-adjoint inverse exists by the Lax-Milgram theorem. If
fge0, testing Lu=f with the negative part gives
Q(u_-)le0, so uge0.
[Strict supersolution]
Let R>0 and A a bounded measurable positive set. There exists
s_Dge0 smooth, bounded below, with s_D=1 on a nested smaller
domain for R large, and L_A s_Dge c>0.
Proof. [Sketch]
Fix two nested relatively compact coordinate domains in which the
source is exactly 1. Interior Schauder estimates and differentiation
bound every derivative of u_R on the smaller domain uniformly for
Rge R_0. A diagonal subsequence converges smoothly on compact
subsets to a bounded positive solution s of Ls=1.
## Compatible positive approximations
[Compatible positive approximation]
Fix the corrections of cref{lem:correction} and let s_A be the
exact bounded solution L_As_A=1. There exist smooth compatible
functions v_A>0 with L_Av_Age1/2 on every packet. They have
positive infimum and bounded derivatives packetwise.
Proof. [Sketch]
Use the end-decay: on an exact elementary collar Btimes(T,∞)
the difference w=s_A-s_B satisfies
|s_A-s_B|(x,t)le C e^{-alpha t}s_B(x) for t<T. Construct v_A
from deeper faces upward: keep s_A in the compact part and use
packet weights to blend to v_B; move the starts until
|s_A-s_B|_{C^2}<ε/4 on all participating collars. The
remaining v_B-s_B is paid by its allocated tolerance, not by moving
the start. Start beyond all compatibility thresholds so that copies
agree. The compact data give bounded higher derivatives. Choose the
tolerance tau_A small enough that L_Av_Age 1-M_Atau_Age3/4ge1/2.
## The scalar-identity step
[The full invariant step]
Assume the stage-r invariant and the conclusion of
cref{prop:global-shielding}. For every nonempty parent packet of base
dimension at least two, processing j_0 produces the compatible
smooth split packets of cref{def:packet}, with the required forced
signs and continuous projection maps. There are smooth compatible old
metrics and k=399 new torus factors on every packet such that the
profiles, cores, potentials, and scalar curvature satisfy the
stage-(r+1) invariant. This includes a possible one-dimensional
output, on which the scalar lower bound is global.
Proof. [Sketch]
Apply cref{lem:guarded-approx,lem:seam,lem:profile-restrict} for the
smooth base metrics and the profile data, and
cref{lem:correction,lem:resolvent,lem:supersolution,lem:sol-approx} for
v_A. Append k=399 circle variables, each with length function
v_A^{2/(k+1)}. The warped-product identity is
[
R_{rm new}=R_{rm old}-(4k)/(k+1)frac{lap_{rm old}v_A}{v_A}.
]
Indeed a common scale f contributes
-2kf^{-1}lap_{rm old}f-k(k-1)f^{-2}|df|^2; substitution of
f=v_A^{2/(k+1)} cancels the separate |dv_A|^2 term. Here k
counts only the newly appended circle factors; unit factors used
earlier to compare old lists have zero logarithmic derivative and leave
both lap_{rm old} and the coefficient unchanged. Since
4k/(k+1)=C, cref{lem:profile-restrict,lem:supersolution} give
[
R_{rm new}=V_A^*-K_A+(L_A v_A)/(v_A)ge V_A^*
on D_A.
]
Here K_A=0 on a neighborhood of D_A. Adding these factors changes
no base distances or profile gradients. The new factors are positive,
bounded above and away from zero, and exactly compatible. The same
number is appended at every status of the stage, so all end identities
compare the same list of factors. This proves the analytic invariant,
including when the new required threshold is zero.
{99}
Agmon, S. Lectures on exponential decay of solutions of second-order
elliptic equations. Mathematical Notes 29, Princeton Univ. Press, 1982.
Allard, W.~K. On the first variation of a varifold.
Ann. of Math. 95 (1972), 417--491.
Antonelli, G. Universal volume growth bounds from positive intermediate
curvature. arXiv:2608.14507v2, 2026.
Bi, X., Zhu, C. Singular scalar-curvature descent for enlargeability and
cubical-width inequalities. Preprint, 2026.
Bolotov, D. Macroscopic dimension of the universal cover of a
three-manifold. Preprint, 2003.
Bolotov, D., Dranishnikov, A. On Gromov's scalar curvature conjecture.
Proc. Amer. Math. Soc. 138 (2010), 1237--1243.
Bolotov, D., Dranishnikov, A. On Gromov's conjecture for totally non-spin
manifolds. J. Topol. Anal. 8 (2016), 571--587.
Chodosh, O., Li, C. Generalized soap bubbles and the topology of manifolds
with positive scalar curvature. Preprint, 2020.
Daher, M., Dranishnikov, A. On macroscopic dimension of non-spin
four-manifolds. arXiv:1904.00846v1, 2019.
De Giorgi, E. Su una teoria generale della misura (r-1)-dimensionale in
uno spazio ad r dimensioni. Ann. Mat. Pura Appl. 36
(1954), 191--213.
De Giorgi, E. Nuovi teoremi relativi alle misure (r-1)-dimensionali in
uno spazio ad r dimensioni. Ricerche Mat. 4 (1955),
95--113.
Dranishnikov, A. Macroscopic dimension and its applications.
Preprint, 2014.
Federer, H., Fleming, W.~H. Normal and integral currents.
Ann. of Math. 72 (1960), 458--520.
Fischer-Colbrie, D., Schoen, R. The structure of complete stable minimal
surfaces in 3-manifolds of non-negative scalar curvature.
Comm. Pure Appl. Math. 33 (1980), 199--211.
Gromov, M. Filling Riemannian manifolds.
J. Diff. Geom. 18 (1983), 1--147.
Gromov, M. Positive curvature, macroscopic dimension, spectral gaps and
higher signatures. In Functional Analysis on the Eve of the 21st
Century, vol.~II, Birkh"auser, 1996, 1--213.
Gromov, M. A dozen problems, questions and conjectures about positive
scalar curvature. Author-hosted preprint, 2017.
Gromov, M. Metric inequalities with scalar curvature.
Geom. Funct. Anal. 28 (2018), 645--726.
Gromov, M. No aspherical four-manifold admits a metric of positive scalar
curvature. In Perspectives in Scalar Curvature, World Scientific,
2020.
Gromov, M., Lawson, H.~B. Positive scalar curvature and the Dirac operator
on complete Riemannian manifolds. Inst. Hautes 'Etudes Sci. Publ.
Math. 58 (1983), 83--196.
Kumar, A., Sen, B. Urysohn width and macroscopic scalar curvature.
arXiv:2601.14669v2, 2026.
Liokumovich, Y., Maximo, D. Waists of Morse functions on positive scalar
curvature 3-manifolds. Preprint, 2023.
Liokumovich, Y., Wang, Z. On the waist and width inequality in complete
3-manifolds with positive scalar curvature. arXiv:2308.04044v1, 2023.
Lohkamp, J. Splittings of manifolds and positive scalar curvature, II.
Preprint, 2022.
Massari, U. Esistenza e regolarit`a delli frontiere di area minima con
ostacolo. Ann. Mat. Pura Appl. 105 (1975), 271--298.
Massari, U., Pepe, L. Su una impostazione geometrica della teoria
variazionale. Boll. Un. Mat. Ital. 11 (1975), 87--102.
Miao, P. Positive mass and corner smoothing for smooth metrics.
Comm. Anal. Geom. 10 (2002), 441--467.
Michael, J.~H., Simon, L.~M. Sobolev and mean-value inequalities on
generalized submanifolds of R^n.
Comm. Pure Appl. Math. 26 (1973), 361--379.
Nabutovsky, A., Rotman, R., Sabourau, S. Sweepouts of closed Riemannian
manifolds. Geom. Funct. Anal. 31 (2021), 721--766.
Nash, J. The imbedding problem for Riemannian manifolds.
Ann. of Math. 63 (1956), 20--63.
OpenAI. Positive scalar curvature forces rational inessentiality.
OpenAI Math Release preprint OAI:2026, September 23, 2026.
Schoen, R., Yau, S.-T. On the structure of manifolds with positive scalar
curvature. Manuscripta Math. 28 (1979), 159--183.
Simon, L. Lectures on Geometric Measure Theory. Proc. Centre
Math. Anal. Austral. Nat. Univ., vol.~3, 1983.
Simon, L. Schauder estimates by scaling. Notes in PDE.