THE NONMAXIMAL ANTI-DE SITTER PENROSE INEQUALITY
=================================================
and original-data rigidity
A readable presentation of the paper by OpenAI, October 2026.
Original source: ads_penrose_nonmaximal_original.txt
=================================================
ABOUT THIS VERSION
=================================================
The original paper is a long, technical document in
relativity and geometry. It proves two theorems about
black holes in an anti-de Sitter universe: an inequality
relating the total mass of the spacetime to the size of
its horizon, and a characterisation of exactly when
equality holds.
This version explains what was proved and why in plain
prose. It is aimed at a reader with high-school algebra
and some exposure to physics who wants to understand
not only the statements, but also the shape of the
proof, without going through every page of estimates.
All the main formulas are numbered. Every theorem from
the original paper is stated here at least once, with a
short explanation of what it means.
=================================================
ABSTRACT
=================================================
This paper proves two theorems about black holes in an
anti-de Sitter universe.
Theorem A (Inequality). For any initial data that is
asymptotically anti-de Sitter in the sense of the paper,
satisfies the dominant energy condition, has a future-
trapped boundary, and whose mass flux is finite and
future-timelike, the anti-de Sitter metric mass m_AH
satisfies the numerical bound:
m_AH >= ( r_A + r_A^3 ) / 2 (0.1)
where
r_A = sqrt( A_min / (4 pi) ), (0.2)
and A_min is the infimum of the areas of all enclosing
surfaces in the chosen exterior.
Theorem B (Equality and rigidity). Suppose equality holds
in (0.1), and the two-sided boundary S is a connected,
outermost, outer-area-minimizing future-marginally-outer-
trapped surface. Then the original pair (g, K) must be
realised as a smooth, proper spacelike slice in the
smooth maximally extended Schwarzschild-anti-de Sitter
spacetime of parameter exactly m_AH(g), with the interior
sitting in one black-hole exterior, the boundary mapping
onto a full spacelike section of the future horizon (the
bifurcation sphere is allowed), and the end reaching
conformal infinity. The reconstruction recovers not only
g but also the second fundamental form K.
Sharpness. For every M > 0, the Schwarzschild-AdS
exterior of mass M attains equality in (0.1). Moreover,
infinitely many compactly supported non-static time-graph
perturbations of the exterior also attain equality, as
long as they satisfy the initial-data hypotheses.
=================================================
TABLE OF CONTENTS
=================================================
Part I. Setting and statements
1. What this paper proves
2. What is the Penrose inequality
3. Anti-de Sitter space and its mass
4. The full setting and main statements
Part II. How it is proved
5. Proof architecture
6. Step 1: auxiliary stress and the one-sign case
7. Step 2: the flat deformation
8. Step 3: local regularity for the rank-one
trace equation
9. Step 4: the hyperbolic deformation
10. Step 5: the causal adjoint
11. Step 6: the static quotient
12. Step 7: compactification and the smooth end
13. Step 8: static rigidity
14. Step 9: reconstructing the original slice
15. Sharpness: non-static examples
Part III. Technical appendix
A. Key lemmas
B. References
=================================================
Part I. Setting and statements
=================================================
1. What this paper proves
-------------------------------------------------
Imagine you have a black hole. In Newtonian physics, a
black hole is just a region of space where gravity is so
strong nothing can escape. In Einstein's general
relativity, "black hole" is a geometric notion: a place
where spacetime curves so severely that light cannot
travel from the inside to the outside.
Two physical quantities describe a black hole:
- its MASS M, which measures how much it bends
spacetime far away;
- its HORIZON AREA A, which measures how big the
black-hole surface is.
For an isolated black hole, of course there is a
relationship between M and A - you can look it up in
any textbook. But what if the "black hole" you are
looking at is only an approximation to some genuinely
complicated, non-spherical, time-dependent
configuration? Surely M and A still have to satisfy
SOME inequality. That inequality is what Penrose
conjectured (in the asymptotically flat case) and what
this paper proves (in the anti-de Sitter case).
Here is the inequality, in the simplest form this paper
proves. Let M be the mass of the whole configuration,
and let A be the least area of a surface that encloses
the "would-be black hole" - the boundary of the region
you are calling the exterior. Then:
M >= sqrt( A / (16 pi) ) . (1)
If you are familiar with the Schwarzschild black hole,
you will recognize the right side: for a Schwarzschild
black hole of mass M, the horizon area is A = 16 pi M^2,
so sqrt(A / (16 pi)) = M. So the Schwarzschild black
hole SATURATES (1): mass exactly equals the right side.
The Penrose inequality says no other configuration can
have less mass for its horizon area than Schwarzschild
does.
Now add a cosmological constant. The cosmological
constant Lambda is a parameter in Einstein's equations
that describes how empty space "curves itself", even
without any mass around. Observations of the real
universe suggest Lambda > 0 (accelerating expansion),
but a NEGATIVE Lambda is also physically interesting and
mathematically cleaner. When Lambda < 0, the natural
"empty" background is anti-de Sitter space, and the
analogue of Schwarzschild is Schwarzschild-anti-de
Sitter (Schwarzschild-AdS).
In the anti-de Sitter setting, the analogous statement
takes the form (0.1):
M_AdS >= ( r + r^3 ) / 2 .
where r = sqrt(A / 4 pi), M_AdS is the anti-de Sitter
mass, and the factor (r + r^3)/2 comes from evaluating
the mass of a Schwarzschild-AdS black hole at its
horizon. If Lambda = 0, r^3 disappears from the formula
and you are back to (1). If Lambda < 0, the r^3 term
becomes important at large radius.
So the inequality has the right limit at Lambda = 0. But
of course, that is not enough. This paper actually
proves it, under the right technical hypotheses. And it
proves a second, complementary fact: that equality in
(0.1) can happen ONLY for the Schwarzschild-AdS
configuration, or for extremely specific perturbations
of it.
The rest of this section spells out what all these
words mean.
=================================================
2. What is the Penrose inequality
-------------------------------------------------
The Penrose inequality has a story going back to the
1960s.
Penrose's original argument (1973). Suppose you have a
spacetime containing a trapped surface - a surface from
which both expansions of null geodesics are negative, so
that even light cannot escape from the "outgoing"
direction. Under cosmic censorship, that trapped
surface must be inside a black hole. Now assume for
contradiction that the mass M of the whole spacetime is
LESS than sqrt(A / (16 pi)) where A is the area of the
trapped surface. Consider "evolving" the trapped surface
forward in time along a suitable null flow. Under the
assumption M < sqrt(A/(16 pi)), the resulting "horizon"
would have to eventually become a naked singularity -
violating cosmic censorship. Therefore M must be at
least sqrt(A/(16 pi)).
This is a beautiful argument, but it "proves" the
inequality only conditionally: it uses cosmic
censorship, which is not itself proven. The challenge
was to prove the inequality directly from the dominant
energy condition, without needing cosmic censorship.
The Riemannian case. The first clean result was for
time-symmetric initial data, where the second fundamental
form K vanishes. In that case the constraint equations
reduce to a single scalar equation, R_g >= 0, and the
inequality becomes the pure geometry statement:
M_ADM >= sqrt( A_min / (16 pi) ) . (2.1)
Bray proved this fully in 1997 using a conformal-flow
argument. Huisken and Ilmanen proved a slightly weaker
version (for a single outermost horizon component) in
2001 using the inverse mean curvature flow. Both proofs
have since been generalized to higher dimensions and to
charged black holes.
The Lorentzian case. The fully general inequality, with
K allowed to be nonzero, is still open in complete
generality. It is proven in important special cases:
- Spherical symmetry (several authors, 1970s-1990s).
- A single connected outermost future-trapped surface
with rotational symmetry (various results).
- Near Schwarzschild initial data (many authors,
using perturbation theory).
- In parts of the "maximal" class where tr K = 0
(where the argument reduces to the Riemannian case).
The original paper's contribution. The paper proves the
Lorentzian Penrose inequality in full generality, in the
anti-de Sitter setting, WITHOUT assuming maximal gauge
or rotational symmetry. This is the first such result.
The key technical novelty is the "auxiliary stress"
construction that reduces the fully general case to
something tractable.
=================================================
3. Anti-de Sitter space and its mass
-------------------------------------------------
What is anti-de Sitter space? It is the analogue of
Minkowski space (flat spacetime) but with a NEGATIVE
cosmological constant. Instead of being flat, it curves
so that "large" triangles are thinner than in flat
space, and balls of large radius grow exponentially in
circumference rather than polynomially.
In the spatial picture (a single moment of time), the
natural empty space is the hyperbolic metric. In three
spatial dimensions, using a radial coordinate r and the
round metric sigma on the two-sphere:
b = dr^2 / (1 + r^2) + r^2 sigma . (3.1)
The end lives at r -> infinity, where the radius grows
but the metric begins to look like a cylinder. The
scalar curvature of b is exactly -6, which is the
traces of the cosmological constant Lambda = -3 in
Einstein's equations:
G_ab + Lambda g_ab = 8 pi T_ab, Lambda = -3.
WHAT IS A BLACK HOLE IN THIS SETTING? The simplest
black-hole spacetime is Schwarzschild-AdS. Its metric in
spherical coordinates (t, r, theta, phi):
g_M = - F_M(r) dt^2 + dr^2 / F_M(r) + r^2 sigma ,
F_M(r) = 1 + r^2 - 2M/r . (3.2)
This is a static, spherically symmetric solution of
Einstein's vacuum equations. The function F_M is
strictly increasing on (0, infinity): it starts at
-infinity (because of -2M/r) and goes to +infinity
(because of +r^2). So it has exactly one positive
zero for every M > 0, and this zero r_h is the black
hole horizon.
WHY IS (r + r^3)/2 THE MASS OF SCHWARZSCHILD-ADS?
Set F_M(r_h) = 0:
0 = 1 + r_h^2 - 2M / r_h ,
M = (r_h + r_h^3) / 2 . (3.3)
So given the horizon area A = 4 pi r_h^2 (or
equivalently r_h = sqrt(A/(4 pi))), the mass is
completely determined. This is the mass formula that
appears on the right-hand side of the Penrose
inequality.
For example, if the "black hole" is small (r_h ->
0), M -> 0 and the black hole is close to the flat
space limit. If the black hole is large (r_h ->
infinity), M grows like r_h^3 / 2, much faster than
the asymptotically flat analogue.
WHAT IS THE "MASS" OF A GENERAL (NON-SCHWARZSCHILD)
SPACETIME? In general relativity, extract mass from a
metric by looking at the far-field behavior at infinity.
But anti-de Sitter space is curved even in the absence
of matter - so the definition of "mass" is not the same
as in asymptotically flat space.
The standard definition (Wang, Chrusciel-Herzlich) for
the anti-de Sitter setting involves four components:
p_0, p_1, p_2, p_3 , (3.4)
which are computed by integrating (g - b) (the
difference between the metric and the background
hyperbolic metric) against four "static potentials":
V_0(x) = sqrt(1 + r^2) (the time-direction) ,
V_i(x) = r x_i (i = 1, 2, 3) . (3.5)
The precise formula is a flux integral over spheres at
infinity. Roughly:
p_a = (1 / (16 pi)) lim_{R -> inf}
integral_{r=R} U_b(V_a, g - b)(nu) dA . (3.6)
The details of the integrand U_b are technical, but
its physical meaning is: p_0 is the total energy, and
p_1, p_2, p_3 are the total momentum components.
For a future-timelike four-vector p, we have
p_0 > sqrt( p_1^2 + p_2^2 + p_3^2 ) , (3.7)
and we define the metric mass as the Lorentz length:
m_AH = sqrt( p_0^2 - p_1^2 - p_2^2 - p_3^2 ) . (3.8)
For Schwarzschild-AdS, only p_0 is nonzero, and it
equals M. So m_AH(Schwarzschild-AdS) = M = (r_h+r_h^3)/2,
matching (3.3).
WHY "m_AH"? The subscript AH stands for "asymptotically
hyperbolic". It distinguishes this mass from the more
familiar anti-de Sitter mass (sometimes written m_ADM),
which is defined with different boundary conditions.
The AH mass is invariant under changes of the spatial
chart at infinity (that is, changes of coordinate
system that respect the decay hypotheses) but NOT
invariant under changes of which time slice of the
spacetime you look at. In other words: the mass depends
on the SPATIAL slice, not just on the spacetime. This
is why the paper's theorem is stated for a specific
initial-data slice, not for the full spacetime.
This is subtle, but important. In asymptotically flat
space, you can pick any spatial slice and compute the
ADM mass and always get the same answer. In anti-de
Sitter space, changing the slice changes the answer.
The metric m_AH correctly captures this.
=================================================
4. The full setting and main statements
-------------------------------------------------
Now we pin down all the hypotheses precisely. The
definitions below are the ones that appear in the
original paper, translated into readable form.
MANIFOLD AND METRIC. Omega is a connected, smooth,
orientable three-dimensional manifold with nonempty
compact boundary S, exactly one end. Outside a compact
set, there is a chart (0, infinity) x S^2 in which the
metric g and the second fundamental form K decay to the
hyperbolic background b and to zero at rate r^{-tau}:
g - b = O(r^{-tau}),
K = O(r^{-tau}), (4.1)
for some 3/2 < tau < 3, in a specific weighted Holder
norm. The lower bound tau > 3/2 is needed for the mass
flux to converge; the upper bound tau < 3 is a technical
convenience that rules out a degenerate limit.
CONSTRAINT DENSITIES. Any pair (g, K) on Omega satisfies
the constraint equations that come from projecting the
Einstein equations onto the slice. Define
mu = (1/(16 pi)) ( R_g + 6 + (tr_g K)^2 - |K|_g^2 ),
J = (1/(8 pi)) div_g ( K - (tr_g K) g ) . (4.2)
Here R_g is the scalar curvature of g, tr_g K is the
trace, |K|_g^2 is the squared norm, and div_g is the
divergence of a symmetric two-tensor. The sign convention
has Lambda = -3.
DOMINANT ENERGY CONDITION (DEC). This is the assumption
that matter has nonnegative energy density in every
frame. In initial-data terms:
mu >= |J|_g , pointwise on Omega. (4.3)
WEIGHTED INTEGRABILITY. The paper also requires that
the "mass flux integrand" is integrable at infinity:
integral_Omega sqrt(1 + r^2) (mu + |J|_g + |K|_g^2)
dV_g < infinity. (4.4)
Without this, the four mass components p_a would not
converge.
TRAPPED BOUNDARY. On each connected component S_i of
S, the future null expansion must be nonpositive:
theta_+ = H_{S_i} + tr_{TS_i} K <= 0 . (4.5)
Here H_{S_i} is the mean curvature of S_i with respect
to the outward normal nu, and tr_{TS_i} K is the
tangential trace of K along S_i. This is the "weakly
future-trapped" condition: S_i has nonpositive future
expansion, which is what makes it a candidate horizon.
FUTURE-TIMELIKE MASS. The four mass components p_a are
required to form a future-timelike four-vector:
p_0 > 0, p_0^2 > p_1^2 + p_2^2 + p_3^2 . (4.6)
This is a positivity assumption on the mass, automatic
in reasonable physical situations. Without it, m_AH
could be imaginary.
ENCLOSING CUTS AND A_MIN. A full enclosing cut is the
entire intrinsic boundary Gamma = partial D of a
connected codimension-zero smooth submanifold D with
boundary in Omega, which contains the entire end and
whose interior lies in the interior of Omega. Gamma may
coincide with S, may have multiple components, and need
not be minimal or trapped. Define
A_min = inf Area_g(Gamma) , (4.7)
then
r_A = sqrt( A_min / (4 pi) ) . (4.8)
MAIN NUMERICAL THEOREM (Theorem A). Let (Omega, g, K) be
an admissible initial-data exterior in the sense above.
Then:
m_AH(g) >= ( r_A + r_A^3 ) / 2 . (4.9)
Equivalently, substituting (4.8):
m_AH >= sqrt( A_min / (16 pi) ) * ( 1 + A_min / (4 pi) ) , (4.10)
which is the same statement written with A_min rather
than r_A.
MAIN EQUALITY THEOREM (Theorem B). Suppose in addition:
(1) S is a single connected component;
(2) S is outermost: no weakly future trapped enclosing
surface exists outside S;
(3) S is outer area-minimizing: for every enclosing cut
Gamma outside S, Area_g(Gamma) >= Area_g(S).
Then:
(a) A_min = Area_g(S);
(b) equality holds in (4.9);
(c) the pair (g, K) is realised as a proper spacelike
slice of the maximally extended Schwarzschild-AdS
spacetime of parameter M = m_AH, with the interior in
one black hole exterior, S mapping onto a full spacelike
cross section of the future horizon, and the end
reaching conformal infinity;
(d) the reconstruction recovers both g and K.
Outside the connected outermost outer-area-minimizing
subclass, the paper does NOT claim that equality implies
Schwarzschild-AdS. There may exist other equality
configurations, and the paper does not rule them out.
SHARPNESS. For every M > 0, the Schwarzschild-AdS
exterior with mass M attains equality in (4.9). Its
mass is M, its boundary area is 4 pi r_h^2, r_A = r_h,
and M = (r_h + r_h^3)/2. Infinitely many further
examples (compactly supported non-static time-graph
perturbations) are described in Section 15 of this
document.
This closes Part I. Part II explains how the inequality
and equality theorem are proved.
=================================================
PART II. HOW IT IS PROVED
=================================================
5. Proof architecture
-------------------------------------------------
The proof of the numerical theorem has four stages. They build on
each other, but each is a technically self-contained problem.
Stage 1 (Section 6). Auxiliary stress. Given general K, we
construct a scalar potential f solving a nonlinear Dirichlet
problem with a "saturating flux". Its derived stress tensor L
makes (g, K+L) satisfy strict dominant energy and strict
future-trapping, at the cost of a small controlled mass shift.
In the "one-sign" case (K - (tr K) g has a definite sign), the
mass bound follows by further reducing to asymptotically flat
data.
Stage 2 (Section 7). Flat deformation estimate. We need the
Riemannian Penrose inequality in a specific form: for data
with one asymptotically flat end, DEC, and trapped boundary,
the ADM energy dominates the radius of the area-minimizing cut.
This is proven by a coupled elliptic deformation of the flat
constraint equations, the "flat deformation estimate". It is
proved in the paper, not assumed from external results.
Stage 3 (Section 8). Rank-one regularity. Both the flat and
the hyperbolic deformations need the auxiliary trace equation
to have a bounded gradient and a Holder Hessian, uniformly
over measurable coefficients. The principal matrix of the
trace equation differs from the identity by a rank-one term
pointing along the gradient of f. A dedicated regularity
theorem supplies the estimates.
Stage 4 (Section 9). Hyperbolic deformation. For a general K
(not one-sign), deforming the hyperbolic end directly with a
scalar profile removes K at lowest order and gives a strict
DEC in the limit. The mass deficit from this deformation is
controlled by a weighted divergence estimate.
The numerical theorem is then the combination of these four
stages.
For equality, five more stages:
Stage 5 (Section 10). Causal adjoint. Varying (g, K) in all
mass-preserving directions and plugging in the mass-area
stationarity that equality provides yields a causal stationary
field (u, X). This field satisfies the static Killing
equations in the sense of an orthogonal splitting.
Stage 6 (Section 11). Static quotient. On the timelike region
N = u^2 - |X|^2 > 0, take the quotient by the stationary
vector field. The resulting Riemannian metric h and lapse
lambda satisfy the static vacuum equations with Lambda = -3.
A twist argument shows the closed one-form X^flat/N is exact.
A flat-leaf exclusion shows the null set N = 0 is empty.
Stage 7 (Section 12). Compactification. The static base has a
normalized form near its conformal boundary, from which the
third-order trace coefficient (encoding the mass) and the
third-order lambda coefficient are computed explicitly.
Stage 8 (Section 13). Static rigidity. A static Heintze-
Karcher deficit on the static base is shown to have nonnegative
integral over any bounded domain. Evaluating it in the limit
yields the Schwarzschild-AdS mass, and the vanishing of the
deficit identifies the whole base with Schwarzschild-AdS.
Stage 9 (Section 14). Reconstruction. The static quotient is
lifted back to a spacelike slice in Schwarzschild-AdS, and its
future normal is shown to reproduce the original K.
Each of Section 6 through Section 14 elaborates one stage.
Section 15 verifies sharpness with explicit non-static
examples.
=================================================
6. Step 1: auxiliary stress and the one-sign case
-------------------------------------------------
The technical obstacle for the numerical theorem is that a
general K couples badly with the deformation. If we could
replace K by K + L where L is a "good" correction, we might
reduce the general case to a much nicer one.
The correction L used in the paper comes from a scalar
potential f. Its construction is beautiful and I want to
give it in detail.
6.1 The saturating flux function.
Define a scalar function t: [0, infinity) -> [0, infinity) by
t(s) = a(s) s^2,
where a(s) is the unique positive solution of
a + s^3 a^{3/2} = 1. (17)
That is, at s = 0 we have a(0) = 1 and t(0) = 0; as s grows,
a decreases. Some useful limits:
a(s) = 1 - s^3 + O(s^6) as s -> 0,
t(s) = s^2 - s^5 + O(s^8)
a(s) ~ s^{-2} as s -> infinity,
t(s) -> 1 as s -> infinity,
and in fact
b(s) = s t'(s) / t(s) satisfies b(0) = 2, b -> 4/3 as s -> infinity.
The function t is strictly increasing, satisfies t(s) <= s^2,
and has t(s) ~ s^2 at zero and t(s) ~ 1 at infinity. In other
words, t behaves like the square of its argument near zero and
like a constant near infinity.
This is the "saturating flux". The adjective is because t(s)
"flattens out" at large s: it does not grow past 1.
6.2 The Dirichlet problem.
Let (X, g) be an exterior with compact boundary T and one end,
and let D > 0 be a constant. Consider the Dirichlet problem
for a function f:
div_g ( t(|df|_g) grad_g f / |df|_g ) = 0,
f|_{T} = D,
lim_{r -> infinity} f = 0. (18)
The vector field t(|df|)/|df| * grad f is smooth in df, even at
zero: the quotient t(s)/s has a finite limit as s -> 0 because
t(s)/s = a(s) * s -> 0.
Theorem (essentially Theorem 3.15 of the original). For every
D > 0 and every exterior (X, g) with the hyperbolic decay
hypotheses, this Dirichlet problem has a unique weak solution
0 < f < D. It has bounded physical gradient, is C^{1, theta}
on compact sets, and is smooth wherever df != 0. On the
conformal boundary at infinity, it has an asymptotic
expansion
f = z C_f(omega) + O(z^2), (19)
where z is the standard defining function (z = 1/r roughly) and
C_f is a smooth positive function on S^2.
The last piece is the important part. Define
F_D = integral_{S^2} C_f(omega)^2 dA_sigma. (20)
Then a computation shows
-lim_{r -> infinity} integral_{r = const} t(|df|) <grad f/|df|, nu> dA = F_D, (21)
so F_D is the flux of the saturated vector field across any
large sphere.
6.3 The stress tensor.
Given f solving (18), define sigma = |df|_g, let e = df/sigma
be the unit gradient direction (near a noncritical point), and
define
k(sigma) = sqrt( 1 - t(sigma)^{3/2} ), (22)
W(sigma) = ( 3 + k^2 ) / ( 2 k ).
Then the stress tensor is
P = W g - sigma W'(sigma) e tensor e. (23)
Its trace-reversal
L = P - (1/2) (tr_g P) g (24)
is the actual correction we will add to K.
Here are its key properties, each proven in Lemma 3.19 of the
original:
(a) Critical points and boundaries. At a critical point (df=0),
we have t(0) = 0, k(0) = 1, W(0) = 2. Consequently P = 2g
and L = -g, extending (23)-(24) to critical points.
(b) Conservation. If f solves (18), then div_g P = 0 in the
sense of distributions. This follows from the fact that f
is a critical point of the variational problem
integral W(|df|) dV = integral t(|df|)|df| dV,
and P is the associated Euler-Lagrange stress-energy tensor
of the (non-standard) variational integrand.
(c) Constraint on L. Everywhere the following identity holds:
(tr_g L)^2 - |L|_g^2 = 6. (25)
At a critical point this reads (-3)^2 - 3 = 6, as expected.
(d) Negative definiteness. L is negative definite: its
eigenvalues are L_e < 0 (along e) and L_omega < 0 (all
other directions).
These four facts make L a perfect "stress correction". Adding
L to K increases the constraint Hamiltonian, decreases the
boundary expansion, and is conserved.
Specifically, if K - (tr K) g >= 0 (positive semidefinite
"variance tensor"), then
P : K <= 0. (26)
This single inequality is the engine. It implies that the
flat constraint Hamiltonian of the new pair (g, K+L) is:
H(g, K+L) = R_g + (tr(K+L))^2 - |K+L|^2
= R_g + (tr K)^2 - |K|^2 - 2 P : K (27)
>= H(g, K).
Similarly, the boundary expansion of K+L is strictly smaller
than that of K, because L is negative definite.
The "one-sign case" of the Penrose theorem is exactly the
case where K - (tr K) g has a definite sign. By flipping K
to -K if necessary (which preserves DEC and preserves the
mass p_0, and strictly decreases the boundary expansion), we
can assume the sign is nonnegative.
6.4 Flux saturation.
Here is the final piece of the auxiliary stress construction.
Solve (18) at height D, and let the stress L_D and its flux
F_D be as above. Then (Lemma 3.18 of the original):
0 <= F_D <= Area_g(T), lim_{D -> infinity} F_D = Area_g(T). (28)
That is: as we raise the auxiliary height D, the flux from L_D
saturates exactly at the horizon area. And by Jensen's
inequality applied on S^2,
( F_D / (4 pi) )^{3/2} <= (1/(4 pi)) integral_{S^2} C_f^3 dA_sigma. (29)
So the volume integral of the cube of C_f is at least the
(3/2)-power of the flux.
Why does this matter? The flat deformation estimate applied
to (g+auxiliary, K+L) gives, in the limit inside the same
paper:
p_0(g) >= sqrt(A_*/(16 pi)) + (1/2)(F_D/(4 pi))^{3/2}, (30)
where A_* is the flat-end horizon area. Sending the auxiliary
field height D to infinity and using (28) gives
p_0(g) >= sqrt(A_*/(16 pi)) + (1/2)(Area_g(T)/(4 pi))^{3/2}. (31)
Then adding a small inward collar to (X, g) makes T the actual
horizon S and lets A_* be replaced by A_min(S;g). This yields
p_0(g) >= sqrt(A_min(S;g)/(16 pi)) + (1/2)(A_min(S;g)/(4 pi))^{3/2}. (32)
And the right side is exactly (r_A + r_A^3)/2 with
r_A = sqrt(A_min(S;g)/(4 pi)).
That completes the one-sign case.
6.5 General K: the preparation lemma.
For a general K (not one-sign), we cannot directly apply the
auxiliary stress. Instead:
(i) Lemma 3.11 ("preparation") reduces the general K to
strict DEC and strict future-trapping with a conformal
perturbation. Concretely, on a sequence of approximating
data (g_i, K_i), we can arrange:
R_{g_i} + (tr K_i)^2 - |K_i|^2 - 2 |J_i| >= c_i (1+r)^{-3-delta},
H_{g_i} + tr_{T S} K_i < 0,
and the approximating metrics converge to the original in
the physical weighted norms. The mass and area limits are
preserved.
(ii) Lemma 3.12 ("a collar with almost unchanged cut infimum")
adds a short inward collar of mean-curvature zero, so
the new metric is defined on a slightly larger exterior
than the original without changing A_min up to o(1).
(iii) The auxiliary Dirichlet problem (18) is posed on the
enlarged exterior, so the auxiliary stress sits in the
added collar as a free parameter.
After (i)-(iii), we have a well-behaved setup that fits the
one-sign theorem. Letting the deformation parameters and the
collar width tend to zero gives the original result.
6.6 What this section accomplished.
The one-sign theorem (Theorem 3.1 of the original) states:
p_0(g) >= sqrt(A_g(S)/(16 pi)) + (1/2) (A_g(S)/(4 pi))^{3/2} (33)
for every (Omega, g, K) satisfying the initial-data hypotheses,
with compactly-supported K, and with a definite sign on the
variance tensor K - (tr K) g.
The proof, which we have just sketched, uses only the auxiliary
stress and the flat deformation estimate (which we will describe
in the next section). It is the foundation on which the general
case is built.
The remaining task: prove the flat deformation estimate. That's
Section 7. It is itself the technical heart of the paper.
=================================================
7. Step 2: the flat deformation
-------------------------------------------------
After the auxiliary-stress trick of Section 6, we are
in the "one-sign" case: K - (tr K) g has a definite
sign. Even in this favorable case, we are not done.
The Penrose inequality we want (4.9) is a Lorentzian
statement about the pair (g, K). The Riemmanian
Penrose inequality (Bray, Huisken-Ilmanen) is a purely
metric statement about g alone, with R_g >= 0. To
bridge the two, we would like to convert the Lorentzian
inequality into a Riemannian one.
The idea of the flat deformation: construct a NEW
metric hat_g, on the SAME three-manifold, which:
(a) is close to g near the horizon S;
(b) is asymptotically FLAT at infinity (not
anti-de Sitter);
(c) has nonnegative scalar curvature: R_hat_g >= 0;
(d) has mass m_hat which differs from m_AH by a
controlled positive correction.
If all four hold, then the classical Riemannian Penrose
inequality applies:
m_hat >= sqrt( A_min(hat_g) / (16 pi) ) . (7.1)
If m_hat is close to m_AH + (positive correction) and
A_min(hat_g) is close to A_min(g), we obtain exactly
the AdS Penrose inequality (4.9). In fact the
correction turns out to be (1/2) (A_min(g)/(4 pi))^{3/2},
which is precisely the extra r^3 term on the right
side of (4.9).
AUXILIARY METRIC. The deformation is implemented
through a new metric
bar_g = g + l^2 df tensor df , (7.2)
where l is a smooth positive "profile" function (a
length scale) and f is a scalar to be chosen. Then a
conformal rescaling
hat_g = e^{4t} bar_g (7.3)
is applied, where t is another scalar. The pair (f, t)
is characterized by two coupled PDEs.
TRACE EQUATION (rank-one). The scalar f satisfies:
A_b : Hess_g f = F , (7.4)
where
A_b = I + (b - 1) e tensor e ,
e = df / |df| ,
b = b(x) in [c_0, 1] , (7.5)
and F is a source term determined by the deformation
stage. The notation A_b : Hess_g f means the contraction
of A_b against the covariant Hessian of f, taken in
the nondivergence sense.
Equation (7.4) is called "rank-one" because, in the
eigendirection e (the gradient of f), the coefficient
is b < 1; in the two directions orthogonal to e, the
coefficient is 1. So the operator A_b is like the
identity on a two-dimensional subspace and like b on
the remaining one-dimensional subspace. Crucially, the
distinguished direction e depends on f itself, so the
equation is genuinely nonlinear even though its
structure looks very simple.
A subtlety: at points where df = 0 (critical points of
f), the direction e is undefined and the equation
becomes formally ambiguous. The paper shows (Section 8)
that weak solutions exist and are unique even when b is
merely measurable, so the equation's meaning at
critical points is not a genuine obstacle.
SCALAR EQUATION. The variable
Z = t + (1/8) log(1 + l^2 |df|^2) (7.6)
satisfies a divergence-form equation:
div_g ( u A_chi grad Z + u K(w, .) ) = u Xi , (7.7)
where
u = exp(2t) l sqrt(D) ,
w = l df / sqrt(D) ,
D = 1 + l^2 |df|^2 ,
A_chi = a projection operator ,
Xi = a source chosen so that a positivity
property ("floor") holds. (7.8)
The "floor" condition is essentially:
t(x) >= -epsilon for a small constant epsilon. (7.9)
If the floor holds, then the constructed metric hat_g
has R_hat_g >= 0 and appropriate boundary behavior on
S, so the Riemannian Penrose inequality applies.
THE FOUR-STAGE CONSTRUCTION. The coupled system (7.4)-
(7.7) is far too complicated to solve directly. The
paper solves it in four stages, each a homotopy of
deformations:
Stage 1: Short range in the profile l, small positive
epsilon. This is the foundational stage where
the deformation is small everywhere.
Stage 2: Increase l from a short value to a large
value N, keeping everything controlled.
Stage 3: Fix l at N, run t upward from the floor,
pushing into the domain where floor holds.
Stage 4: Set f = 0 (removing the auxiliary scalar
entirely), and run t alone to recover the
conjugate Riemannian metric hat_g = e^{4t} g.
At the end of Stage 4:
- hat_g = e^{4t} g is the metric we searched for;
- m_hat = p_0(hat_g) equals the original p_0(g) up to
a controlled error;
- A_min(hat_g) is comparable to A_min(g);
- R_hat_g >= 0.
EXISTENCE THEORY. The existence result for (7.4)-(7.7)
at each stage is the bulk of the technical work. It
uses:
- Maximum principles (to bound t below and above);
- A gradient bound on f (to control df everywhere);
- Rank-one regularity (Section 8) for the trace
equation;
- Classical local estimates from linear elliptic
theory (Schauder, W^{2,p});
- Degree theory to get existence under the homotopy.
MASS COMPLETION (Theorem 5.2 of the original). Once
hat_g is constructed with all the needed properties,
the Riemannian Penrose inequality gives:
m_hat >= sqrt( A_min(hat_g) / (16 pi) ) . (7.10)
Computing m_hat in terms of the original data, one
finds:
m_hat = m_AH + (1/2) (A_min(g)/(4 pi))^{3/2}
+ O( l + l^2/tau ) . (7.11)
Letting l -> 0 and A_min(hat_g) -> A_min(g), one obtains:
m_AH >= sqrt( A_min(g)/(16 pi) )
+ (1/2) (A_min(g)/(4 pi))^{3/2} (7.12)
= (r_A + r_A^3) / 2 . (7.13)
That is the numerical Penrose inequality (4.9).
SUMMARY. The flat deformation converts the Lorentzian
(g, K) data into a Riemannian hat_g with R_hat_g >= 0,
at the cost of an explicit positive correction to the
mass. The correction, when compared to the classical
Riemannian bound, produces the extra r^3 term that
distinguishes the AdS Penrose inequality from its
asymptotically flat predecessor.
The technical machinery that makes this work - the
coupled trace and scalar equations, the four-stage
homotopy, the gradient estimates, the boundary
barriers, and the mass completion - is what makes
this paper (and the linear estimates in its appendix)
so long. But the underlying idea is simple: reduce a
hard Lorentzian problem to a hard (but known-to-be-
true) Riemannian one.
The next section explains the deepest ingredient: the
local regularity for the rank-one trace equation (7.4).
=================================================
8. Step 3: local regularity for the rank-one
trace equation
-------------------------------------------------
The previous section's deformation depends entirely on
the trace equation:
A_b : Hess_g f = F , (8.1)
where A_b = I + (b-1) e tensor e, e = df/|df|, b is
bounded between c_0 and 1. This section states and
sketches the deepest input to the whole construction:
a genuinely strong gradient bound for (8.1) which holds
without any regularity of b beyond measurability and
boundedness.
THE RESULT (Theorem 6.1 of the original):
||Df||_{C^{0,alpha}(B_{1/2})}
<= C (L + Q) ,
integral_{B_r} |Hess f|^2 dV_g
<= C (L+Q)^2 r^{1+2 alpha} , (8.2)
where L is a Lipschitz bound for f at the boundary, Q
bounds the source F, and the constants C, alpha are
UNIFORM over all measurable b(x) with c_0 <= b <= 1
and all bounded F. The word "uniform" is essential:
the estimate does not degrade if b oscillates
violently at every point.
WHY IS THIS HARD? Equation (8.1) has a principal part
A_b whose distinguished axis is precisely the gradient
direction e = df/|df|. The axis depends on the
solution itself. Standard Schauder theory for
nondivergence equations assumes something like Hölder
continuity of the principal coefficients; here the
coefficients are only measurable in b AND depend on
df, which is not even continuous a priori. At critical
points of f (df = 0), the axis is not even defined.
The proof has five main steps.
STEP 1 (W^{2,p_0} bounds, Lemma 6.2). First show that
f is bounded in W^{2,p_0} on a slightly smaller ball,
for some p_0 > 2. The key observation: the "defect"
I - A_b has norm at most (1 - c_0) < 1, and the Riesz
transform D^2 Delta^{-1} is bounded on L^p for
1 < p < infinity. So f is controlled at essentially
the "Laplacian plus small perturbation" level:
||D^2 f||_{L^p(B_{1/2})}
<= C_p ( ||F||_{L^p(B_1)} + ||Df||_{L^2(B_1)}
+ lower order ) . (8.3)
STEP 2 (flux identity). Introduce w = |df|^2 / 2. A
direct computation gives the exact identity:
div_g ( A_b grad w - F grad f )
= |Hess f|^2 - (Delta f)^2
+ Ric_g(df, df) , (8.4)
see (6.4) of the original. This "flux identity" for w
shows that w satisfies a SUPERSOLUTION inequality:
-div_g ( A_b grad w ) >= c |Hess f|^2
- C F^2 , (8.5)
uniformly in b. The right side is nonnegative whenever
|Hess f| is large relative to |F|. Combined with the
Hessian L^2 bound from STEP 1, this forces the
gradient to be locally bounded.
STEP 3 (dichotomy, Lemma 6.11). Take a unit ball B_1
where |df| is large at the center. One of two things
must happen:
(a) |df| drops by a FIXED factor on B_{1/2}, or
(b) f is well approximated by an affine function on
B_{1/2}.
If (a), we have "used up" some of the gradient and
can iterate at a smaller scale. If (b), the equation
is nearly the Laplacian on B_{1/2}, which is far more
regular.
STEP 4 (affine improvement, Lemma 6.12). In case (b),
the affine approximation on B_{1/4} is better than on
B_{1/2} by a fixed geometric factor. Iterating, either
the gradient stays bounded throughout, or f becomes
exactly affine on a small ball - a limiting case
easily dispatched.
STEP 5 (Newton potential). Given the Hessian L^2 bound
from STEP 1 and the gradient bound from STEPS 3-4,
write f = harmonic + (Newton potential of Delta f -
Hess f * curvature correction). A standard singular-
integral argument applied to the Newton potential
gives the Hölder gradient bound (8.2). (Lemma 6.8 of
the original.)
BOUNDARY HANDLING. Interior estimates have been the
focus so far. To handle the compact boundary of Omega,
extend the domain by odd reflection of f across the
boundary, and even reflection of g and b. The extended
f is C^1, its Hessian has no atom on the reflection
plane (because the normal derivative matches from both
sides), and the extended equation holds on both sides
of the plane. The interior estimate then applies to the
extended function on the doubled manifold.
WHY IT MATTERS FOR THE REST OF THE PAPER. The trace
equation (8.1) is the single place in the flat
deformation where the construction could genuinely
fail. It is a nonlinear nondivergence equation with a
degenerate axis, and standard PDE technology does not
suffice. Without the uniform gradient estimate (8.2),
the subsequent scalar equation (7.7) cannot even be
set up, and the four-stage homotopy of Section 7 would
collapse at Stage 1.
BEYOND THIS PAPER. Equation (8.1) is also a model case
for a broader question: what is the regularity of a
nondivergence equation whose principal part is rank-
one in a direction depending on the gradient of the
solution? The methods in this section - flux identity,
dichotomy, affine approximation improvement, Newton
potential - suggest that such equations have strong
gradient estimates without any regularity of the
distinguished coefficient. This is a genuinely new
result with probable uses beyond the AdS Penrose
inequality.