THE AFFINE BERNSTEIN THEOREM IN DIMENSIONS THREE THROUGH NINE
a reader's guide
========================================================================
PURPOSE
This note explains, from scratch, what the "affine Bernstein theorem in
dimensions three through nine" says, why it is hard, and how the proof
works. It is written so a strong high-school student who knows vectors,
matrices, and a little calculus can follow the ideas. The full rigorous
proof is the accompanying LaTeX paper; here we state every definition
precisely, give the intuition, and lay out the logical skeleton so that
nothing is a mystery.
Notation: no bold. Write vectors as plain letters and use ^T for
transpose. sum_i means sum over i.
========================================================================
PART 1. THE QUESTION
========================================================================
1.1 A CONVEX FUNCTION
Let O be an open convex region in R^n, and let u : O -> R be a smooth
function. Say u is (strictly) convex if its Hessian matrix
H(x) = D^2 u(x), with entries H_{ij} = d^2 u / dx_i dx_j, is positive
definite at every point. "Locally uniformly convex" means H stays
positive definite with a bit of room on compact sets.
1.2 THE GRAPH
The graph of u is the set of points
Graph(u) = { (x, u(x)) : x in O } in R^{n+1}.
For example, u(x) = (1/2)|x|^2 gives the round paraboloid. If
u(x) = (1/2) x^T A x + b.x + c with A positive definite, the graph is an
"elliptic paraboloid" (a paraboloid with an elliptic cross-section).
1.3 TWO DIFFERENT "LENGTHS" ON THE GRAPH
A graph inherits a notion of length from the surrounding Euclidean
space R^{n+1}. Moving x by dx moves the graph point by
(dx, du) = (dx, Du . dx), so squared length is
g_{ij} = delta_{ij} + u_i u_j,
where u_i = du/dx_i. This is the INDUCED EUCLIDEAN METRIC. Call the
graph EUCLIDEAN-COMPLETE if every path that runs off to infinity has
infinite length in this metric (you cannot escape to the edge in finite
distance).
Affine geometry gives a SECOND metric, the AFFINE (or Berwald-Blaschke)
metric:
g^aff_{ij} = (det H)^{-1/(n+2)} H_{ij}.
Completeness for this metric is a different condition. The whole
subtlety of the problem is that "complete" can mean either of these
two, and they are not the same.
1.4 AFFINE AREA AND AFFINE MAXIMALITY
Ordinary area of the graph is the integral of sqrt(det g). Affine
geometry has its own area, invariant under volume-preserving affine
maps, called AFFINE AREA:
A(u) = integral_O (det D^2 u)^{1/(n+2)} dx.
Functions that are critical points of affine area (the "affine maximal"
graphs) satisfy a PDE. It is cleanest to write it with two objects:
H = D^2 u (the Hessian, positive definite)
U = det(H) H^{-1} (the cofactor matrix of H)
w = (det H)^{-(n+1)/(n+2)} (a weight; note (n+1)/(n+2) = 1 - delta
with delta = 1/(n+2))
The AFFINE MAXIMAL EQUATION is
sum_{i,j} U^{ij} w_{ij} = 0, (AM)
where w_{ij} = d^2 w / dx_i dx_j. This is exactly the Euler equation
of affine area A(u); a graph satisfying (AM) is "affine maximal".
========================================================================
PART 2. THE THEOREM
========================================================================
THEOREM (main). Let 3 <= n <= 9. Suppose u : O -> R is smooth, has
positive definite Hessian, and satisfies (AM). Suppose also that the
graph of u is complete for the induced Euclidean metric g. Then O = R^n
and
u(x) = (1/2) x^T A x + b.x + c
for a positive definite symmetric A. In other words the graph is an
elliptic paraboloid (an invertible affine image of the standard round
paraboloid).
In the language of the field: a Euclidean-complete, locally uniformly
convex, affine maximal graph in dimensions 3 to 9 must be a paraboloid.
This is the "affine Bernstein theorem" in those dimensions.
2.1 COMPANION RESULT (no graph assumed)
The same conclusion holds for a connected smooth open (noncompact, no
boundary) locally uniformly convex immersed hypersurface that is
affine-complete (complete in the affine metric) and classically affine
maximal, in dimensions 3 to 9. The bridge is a theorem of Trudinger and
Wang: for locally uniformly convex hypersurfaces in dimension >= 2,
affine completeness IMPLIES Euclidean completeness. So the graph
theorem, once proved, upgrades to the abstract hypersurface statement.
2.2 WHY THE DIMENSIONS 3 TO 9
The number 9 is not decoration. Deep in the proof there is a quadratic
form whose positivity is required ("coercivity"). That form has
determinant
(n-2)(10-n)/16.
It is positive exactly when 2 < n < 10, i.e. 3 <= n <= 9. At n = 10 it
vanishes, and in dimension 10 there is a known singular (nonsmooth)
affine maximal graph. So the method naturally stops at 9; the boundary
case n = 10 is where the smooth statement is not reachable this way.
========================================================================
PART 3. HISTORY AND THE REAL OBSTACLE
========================================================================
3.1 THE CONJECTURE
Chern (1979) asked whether an "entire" (defined on all of R^n) locally
uniformly convex affine maximal surface must be a paraboloid. Calabi
(1982) built the affine variational framework and settled the surface
case (n = 2) under both Euclidean and affine completeness. Trudinger and
Wang (2000) proved the Euclidean-complete n = 2 theorem. Li and Jia
(2001) gave another n = 2 proof under affine completeness. The higher-
dimensional case was the open problem.
3.2 THE OBSTACLE: A MODULUS OF CONVEXITY
Trudinger and Wang reduced the high-dimensional problem to a clean
analytic question. Their interior estimate (Theorem 4.2 in their 2000
paper) says, roughly:
IF a convex solution of (AM) has a UNIFORM MODULUS OF STRICT
CONVEXITY -- a single function that controls how "round" the graph
is at every point and every scale -- THEN the higher derivatives are
controlled uniformly, and a rescaling argument forces the graph to
be flat in the third derivatives, i.e. to be a quadratic polynomial.
So everything hinges on GETTING that uniform modulus in dimensions 3 to
9 without extra hypotheses. In dimension 2 Trudinger and Wang got it by
studying convex limits and their contact sets. This paper gets it in
higher dimensions by a new geometric-plus-analytic combination.
3.3 SECTIONS
The key geometric objects are the TANGENT SECTIONS. At a base point a
with supporting plane l_a(x) = u(a) + Du(a).(x-a), define, for height
t > 0,
S_u(a,t) = { x in O : u(x) < l_a(x) + t }.
This is the "footprint" of the part of the graph between the tangent
plane and a parallel copy raised by t. The graph is well-behaved (has a
uniform modulus of convexity) exactly when these sections are UNIFORMLY
BALANCED: there is a fixed number 0 < rho <= 1, the SAME at every base
point and every height, such that reflecting a section about its base
point and shrinking it by rho still lands inside the section:
-rho * (section - a) is contained in (section - a). (BAL)
That is the target. Everything else is machinery to prove (BAL) with a
rho that does not depend on the point or the height.
========================================================================
PART 4. THE STRATEGY IN FIVE STEPS
========================================================================
The proof has two halves. The first half (geometry plus analysis, steps
A through D) proves the uniform balance (BAL). The second half (step E)
turns balance into rigidity with the Trudinger-Wang interior estimate.
STEP A (geometry: convex limits and centering).
Take the epigraph D = {(x,z): x in O, z >= u(x)}, a closed convex body.
Look at all "local limits" of invertible affine images of D. These are
the AFFINE MODELS. Organize a model by the number k of "nonnegative flat
directions" it has: a model with k directions splits coordinates as
(s,y) in R^k x R^{n+1-k}, is supported in s_i >= 0, contains the whole
nonnegative orthant s_i >= 0 at y = 0, and has compact caps
sum_i s_i <= T. If k could keep growing, fibers would be badly centered.
The maximal possible k governs everything. PROPOSITION: if k is maximal,
the fibers K_s = {y : (s,y) in model} are UNIFORMLY CENTERED:
-K_s is contained in C_* K_s for one constant C_*. If centering failed
along a sequence, a rescaling would build a model with k+1 directions,
contradicting maximality.
STEP B (affine area plus a cap identity).
Using that the graph solves (AM), one shows affine area is STATIONARY
under compactly supported variations, and one derives a "cap identity":
an exact integral identity over any cap (region below a level of an
affine function), balancing a second-derivative term against first-
derivative and constant terms. This identity later bounds an inverse-
Hessian weight on each cell.
STEP C (an integral inequality for convex tubes).
Over a fibered region (the "tube": s in a box, y on the sphere of
directions), write the affine area and an affine-invariant volume in
terms of the support function h(s,e) of the fiber. Stationarity of
affine area becomes a PDE. Combining a few integrated identities gives a
COERCIVE inequality:
integral v^2 (1 + |p|^2 + |d|^2 + |d_e|^2) dmu
<= C_n integral |Dv|^2 dmu,
where p and d are gradients of log h and of the area/volume ratio, and
d_e is the angular gradient of d. The constant C_n depends only on n.
THIS inequality needs the quadratic form to be positive, and that is
where 3 <= n <= 9 enters (determinant (n-2)(10-n)/16 > 0). A log-change
of variables turns the coercive inequality into a Poincare-type
logarithmic inequality for a measure sigma = (1 + sum_i |D_s log s_i|^2)
dmu.
STEP D (growth contradiction, excluding k >= 2).
If a model had k >= 2 directions, the logarithmic inequality (step C)
would force the measure sigma of the log-box
Q_L = {s > 0 : |log s|_inf <= L}
to grow EXPONENTIALLY in L: sigma(Q_L) >= c0 q^L with q > 1. But the cap
identity (step B) plus bounds on Hessian minors gives a uniform bound on
each log cell, and finitely many cells cover Q_L, so
sigma(Q_L) <= C_box (L+1)^k, which is POLYNOMIAL growth. Exponential
cannot beat polynomial. Contradiction. Hence the maximal k is 1.
STEP E (balance, then rigidity).
With k = 1 for every model, step A's centering gives (BAL): one uniform
rho works everywhere. (BAL) yields a uniform modulus of strict
convexity. Normalize each section by a linear map (John's ellipsoid
theorem) so the section sits between two fixed balls. Feed this fixed
modulus to the Trudinger-Wang interior estimate: the normalized third
derivatives are uniformly bounded. Un-normalizing, the third derivatives
at the base point are bounded by a constant times t^{-1/2}, which tends
to 0 as the height t -> infinity. Since the base point was arbitrary,
all third derivatives vanish. A function with zero third derivatives is
a quadratic polynomial. Done.
========================================================================
PART 5. WHAT EACH STEP ACTUALLY PROVES
========================================================================
5.1 STEP A: CONVEX LIMITS AND CENTERING
- Lemma (proper sections). Because the graph is Euclidean-complete, at
every finite boundary point of O the height u -> +infinity. So the
epigraph D is closed and full-dimensional, its boundary is exactly the
graph, and each sublevel {u - l_a <= t} is compact inside O. In
particular caps of D are compact.
- Local limits. Closed convex sets converge "locally" if their distance
functions converge uniformly on compact sets. The family F(D) of all
full-dimensional local limits of affine images of D is affine-invariant
and closed under further such limits.
- Definition (model with k directions). A member C of F(D) with
coordinates (s,y) in R^k x R^{n+1-k} such that the nonnegative orthant
{(s,0): s_i >= 0} sits inside C, C sits inside {s_i >= 0}, and every
cap {sum_i s_i <= T} is compact. Its fibers are K_s = {y: (s,y) in C}.
- Lemma (model fibers). Adding a vector (v,0) with v_i >= 0 keeps you in
C; each positive fiber K_s is compact, full-dimensional, contains 0,
and fibers are monotone: K_s is contained in K_t when s_i <= t_i, and
K_{lambda s} is contained in lambda K_s for lambda >= 1.
- Proposition (centering at maximal k). If m = n+1-k > 0, one constant
C_* >= 1 (depending only on the family F(D)) gives
-K_s is contained in C_* K_s for every k-model, every positive fiber,
every admissible coordinate choice. Proof by contradiction: if the
inclusion with constant j failed along a sequence, normalize each
failing fiber to sit in a fixed pair of balls, then change coordinates
to build a (k+1)-model, contradicting maximality. The "change" is the
thin-slice rescaling: cut the big model at level epsilon in one new
direction r = y_1/epsilon, then rescale the remaining transverse
coordinates by a linear map A_epsilon chosen from John's theorem.
5.2 STEP B: AFFINE AREA AND THE CAP IDENTITY
- Lemma (stationarity). (AM) implies affine area is stationary under
every smooth compactly supported variation, and every invertible
affine image of the graph is also stationary. Uses the cofactor
identity sum_i d_i U^{ij} = 0 (differentiate the alternating
determinant formula; terms cancel by symmetry of third derivatives).
- Lemma (area in a ball). The affine area of a smooth positively curved
convex piece inside a ball of radius R is bounded by a constant
depending only on n and R. Uses injectivity of the Gauss map, plus
projection estimates, plus Holder's inequality.
- Lemma (cap identity). For an affine image M of the graph bounding its
convex body, a center o inside, an affine function l, and a smooth
test function psi whose support is compact on M, the exact identity
integral_M [ nu(o-X) |dl|_{H^{-1}}^2 psi''(l)
+ n (l(o) - l(X)) psi'(l)
- n(n+1) psi(l) ] dA = 0
holds. Here nu is the inward conormal, |dl|_{H^{-1}} the dual norm,
H the second form. It is derived by testing stationarity with
Z*psi(q), where Z = nu(o-X) and q = l restricted to the graph, and
integrating by parts once. The delta = 1/(n+2) exponent turns the
log-det term into the -n(n+1) psi term. This is the workhorse that
later bounds inverse-Hessian weights.
5.3 STEP C: THE TUBE AND THE COERCIVE INEQUALITY
- Setup. Split coordinates (s,e) with e a direction on the sphere
S^{m-1}; the boundary part is parametrized by X(s,e) = (s, Y(s,e)) with
Y = h e + nabla_S h, where h(s,e) is the support function of the fiber
K_s. Two matrices appear:
B = -D_s^2 h (base block)
R = nabla_S^2 h + h Id (angular block)
Both are positive definite. The second form is the block-diagonal
diag(B,R).
- Densities. With delta = 1/(n+2),
dA = (det B)^delta (det R)^{1-delta} ds domega (affine area)
dmu = h^{-n/2} sqrt(det B det R) ds domega (invariant vol)
dmu is invariant under separate linear changes in s and y; that is
what makes the whole argument scale-free.
- Cofactor identities (sum_i d_i (cof B)^{ij} = 0, and the angular
analogue) give integration by parts in both blocks. Writing
a = (det B / det R)^delta and f = log(a/h), and setting
p = D_s log h, d = D_s f, p_e = nabla_S log h, d_e = nabla_S f, with
T_s, T_e the trace operators in the two blocks, stationarity of affine
area becomes the PDE
n - 2k + T_s f - |p|^2 - (n+1)|p+d|^2
+ T_e f + 2 + |d_e|^2 = 0.
- Two integrated identities (a p-identity and a d-identity) combine into
a cancellation, then into a coercive identity with a parameter lambda.
The quadratic form in (p,d) has, per pair of components, the matrix
[ n+1-2 lambda (n+2)/4 ]
[ (n+2)/4 lambda ].
At lambda = 1 its determinant is (n-2)(10-n)/16, positive for
3 <= n <= 9. Fix lambda = 15/16. Then the scalar term
k - 2 lambda >= 1/8 whenever k >= 2, and the form's determinant is
(-8n^2 + 88n - 137)/128, which is at least 7/128 on 3 <= n <= 9 (a
concave quadratic with positive endpoint values). So there is a
constant kappa_n > 0 with
F_lambda >= kappa_n (1 + |p|^2 + |d|^2 + |d_e|^2).
Young's inequality then gives the coercive bound
integral v^2 (1 + |p|^2 + |d|^2 + |d_e|^2) dmu
<= C_n integral |Dv|^2 dmu,
with C_n depending only on n.
- The logarithmic inequality. Put sigma = (1 + sum_i |D_s log s_i|^2)
dmu. Substituting v(s) = tilde v(log s) and using Cauchy-Schwarz in
the positive metric turns the coercive bound into
integral |tilde v(log s)|^2 dsigma
<= C_n integral |grad tilde v(log s)|^2 dsigma,
for every compactly supported tilde v on log of the box. No
completeness of the metric is required.
5.4 STEP D: THE GROWTH CONTRADICTION
- Positive mass. The area-in-a-ball bound and a cap with bounded
geometry give a lower bound on affine area of a slab of caps; Holder
then gives sigma(Q_{L0}) >= c0 on one fixed log box Q_{L0}.
- Cell bound. After normalizing a log cell by rescaling s_i and a linear
change in y, the cap identity bounds the inverse-Hessian weight, a
change-of-variables (principal-minor gradient) bound controls det B,
and Cauchy-Schwarz plus a weighted interpolation (exponent
theta = (n+2)/(2(n+1))) bound the whole cell mass by one constant
C_cell, independent of the cell's location. Finitely many cells cover
Q_L, so
sigma(Q_L) <= C_box (L+1)^k (polynomial growth).
- Growth. Applying the logarithmic inequality to a cutoff that is one
on |log s|_inf <= l and zero beyond l+1 forces
sigma(Q_l) <= C_grow ( sigma(Q_{l+1}) - sigma(Q_l) ),
i.e. each box is at most a fixed fraction of the next. Iterating gives
sigma(Q_L) >= c0 q^L, q = 1 + 1/C_grow > 1 (exponential growth).
- Contradiction. Choose L large enough that the exponential lower bound
exceeds the polynomial upper bound; then choose one approximant large
enough for all finitely many cell estimates. Both bounds hold at once,
which is impossible. Hence k = 1.
========================================================================
PART 6. THE RIGIDITY FINISH AND THE COROLLARY
========================================================================
6.1 FROM k = 1 TO BALANCE
Fix a base point a and use the coordinates s = z - l_a(x), y = x - a.
The epigraph in these coordinates is a 1-model, whose fiber at height t
is exactly the tangent section K_a(t). Step A's centering, with its
single constant C_*, applies to EVERY 1-model and hence to EVERY section,
giving (BAL) with rho = min(1, 1/C_*).
6.2 DOUBLING AND THE ENTIRE DOMAIN
Balance (BAL) implies that along any ray from a the tangent-subtracted
gap g(r) = u(a+re) - u(a) - r Du(a).e satisfies a DOUBLING bound
g(beta r) <= A g(r)
for fixed beta > 1 and A > 1 depending only on rho (explicitly
beta = 1 + rho/2 and A = (2+rho)(1+rho)/rho). Iterating the same
argument also shows every ray from a stays in the domain, so O = R^n.
6.3 NORMALIZED SECTIONS AND THE INTERIOR ESTIMATE
Fix a and subtract its tangent so u(0) = 0, Du(0) = 0, u >= 0. Normalize
each section by a linear map (John's ellipsoid theorem) to sit between
two fixed balls, producing rescaled functions
v_t(y) = t^{-1} u(A_t y), with B_1 subset S_{v_t}(0,1) subset B_R.
The doubling bound yields ONE lower modulus b(r), the same for all t,
with
S_{v_t}(x, b(r)) is contained in x + B_r (x in B_{1/4}, any r>0).
That is precisely the uniform strict-convexity hypothesis of the
Trudinger-Wang interior estimate (Theorem 4.2 of their 2000 paper) on
the ball B_{1/4}. Applying it, there are constants c0, C0 independent
of t with
D^2 v_t(0) >= c0 Id, |D^3 v_t(0)| <= C0.
Un-normalize. Relating D^2 v_t(0) to H = D^2 u(0) bounds the inverse of
A_t by a constant times t^{-1/2}. Relating D^3 v_t(0) to D^3 u(0) gives
|D^3 u(0)| <= C t^{-1/2} -> 0 as t -> infinity.
Since a was arbitrary, D^3 u = 0 everywhere, so u is a quadratic
polynomial with positive definite Hessian. That is the theorem.
6.4 THE AFFINE-COMPLETE COROLLARY
If M is instead a connected open (noncompact, no boundary) locally
uniformly convex immersed hypersurface that is affine-complete and
classically affine maximal, then by Trudinger-Wang Theorem A it is
Euclidean-complete, and by Trudinger-Wang Corollary 1 an invertible
affine change makes it a global convex graph over some open convex O
(not assumed all of R^n). That affine change is bi-Lipschitz (bounded
singular values), so it preserves Euclidean completeness, and it scales
affine area by the fixed positive factor |det L|^{n/(n+2)}, so it
preserves stationarity. The graph theorem then forces O = R^n and a
quadratic u. Undo the affine change: M is an elliptic paraboloid.
========================================================================
PART 7. DEFINITIONS AND SYMBOLS (quick reference)
========================================================================
n dimension of the domain O (parameter space of the graph)
u convex function, graph z = u(x)
H = D^2 u Hessian, positive definite
U = det(H) H^{-1}
cofactor matrix; divergence free, sum_i d_i U^{ij} = 0
w = (det H)^{-(n+1)/(n+2)}
weight in the affine maximal equation
delta 1/(n+2); the affine area exponent
g_{ij} delta_{ij} + u_i u_j; induced Euclidean metric on the graph
g^aff (det H)^{-1/(n+2)} H_{ij}; affine (Berwald-Blaschke) metric
A(u) integral (det D^2 u)^{1/(n+2)} dx; affine area
(AM) sum_{i,j} U^{ij} w_{ij} = 0; the affine maximal equation
l_a(x) u(a) + Du(a).(x-a); supporting tangent plane at a
S_u(a,t) { x in O : u(x) < l_a(x) + t }; tangent section, height t
(BAL) -rho (section - a) is contained in (section - a);
uniform balance of sections, the goal
D epigraph {(x,z): x in O, z >= u(x)}
F(D) family of full-dimensional local limits of affine images of D
k-model member of F(D) with k nonnegative flat directions (s in R^k)
K_s transverse fiber {y : (s,y) in model}
h(s,e) support function of the fiber K_s in direction e
B, R -D_s^2 h and (nabla_S^2 h + h Id); the two positive blocks
dmu h^{-n/2} sqrt(det B det R) ds domega; affine-invariant volume
sigma (1 + sum_i |D_s log s_i|^2) dmu; the measure in the log bound
Q_L { s > 0 : |log s|_inf <= L }; a logarithmic box
C_* uniform centering constant for fibers of a maximal k-model
rho min(1, 1/C_*); the uniform balance constant
========================================================================
PART 8. WHAT TO REMEMBER
========================================================================
- The question: must a complete, round-shaped, affine-area-critical graph
be a paraboloid?
- "Complete" is ambiguous between the Euclidean metric and the affine
metric. This theorem is the Euclidean-complete version; a separate
Trudinger-Wang result turns affine-complete hypersurfaces into
Euclidean-complete graphs.
- The core is to get ONE strict-convexity modulus that is the same at
every point and every scale. Everything is engineered to produce it.
- The engine: study all affine limits of the epigraph. If fibers are not
uniformly centered, you can manufacture one more flat direction, which
is impossible once the number of directions is maximal.
- The analysis: a coercive integral inequality on fibers kills models
with two or more directions, via the exponential-versus-polynomial
growth contradiction. Its positivity is exactly why dimensions 3 to 9.
- The finish: the uniform modulus plus the Trudinger-Wang interior
estimate makes third derivatives tend to zero under rescaling; zero
third derivatives means quadratic; quadratic means paraboloid.
That is the whole story.