================================================================================
ENDPOINT POINTWISE CONVERGENCE FOR THE SCHRODINGER EQUATION IN HIGHER DIMENSIONS
================================================================================
================================================================================
PART I: THE QUESTION
================================================================================
1. A SMALL STORY ABOUT WAVES
--------------------------------------------------------------------------------
Drop a stone into a still pond. A ripple spreads out. If you knew the exact
shape of the water at the instant of impact, you would in principle know the
shape at every later instant; the wave simply propagates. The Schrodinger
equation describes the same kind of thing for waves in n-dimensional space
(the physical case is n = 3; n = 1 and n = 2 are the training wheels).
The subtlety is at time zero. Suppose the surface at t = 0 is a messy,
sloshy jumble. At any time t > 0 the wave is smooth, because it has evolved.
As t shrinks to 0, does the smooth wave "return" to the messy data at each
point of space? Usually yes, except possibly at some unlucky points.
Carleson's pointwise convergence problem asks exactly this: for which initial
data does the smooth-at-t>0 evolution return to the raw data at almost every
point x, as time goes to zero? "Almost every x" (almost everywhere, a.e.)
means: at every x except possibly a set of measure zero. This is the precise
sense in which "most points" recover the data while an exceptional "thin"
set may fail.
The answer depends on how rough the initial data is allowed to be. Rougher
data means wilder high-frequency oscillations, a more violent interaction
near t = 0, and a larger chance some point is unlucky. The whole subject is
about locating the exact threshold of roughness at which convergence holds.
2. SCHRODINGER'S EQUATION, CONCRETELY
--------------------------------------------------------------------------------
The free Schrodinger equation in n space dimensions is
i (d/dt) u(x,t) = -Laplacian u(x,t) x in R^n, t in R,
where u(x,t) is a complex-valued wave, x = (x_1,...,x_n), and
Laplacian = (d/dx_1)^2 + ... + (d/dx_n)^2
is the sum of the second derivatives. The initial condition is
u(x,0) = f(x). For nice enough f (say, smooth and decaying fast at infinity)
there is a unique solution, and it is given by a Fourier integral. Write xi =
(xi_1,...,xi_n) for the frequency (momentum) variable, x.xi = x_1 xi_1 + ... +
x_n xi_n for the dot product, |xi|^2 = xi.xi, and f_hat for the Fourier
transform of f (defined in Chapter 11). Then
S(t) f(x) = (2 pi)^{-n} INT_{R^n} e^{ i x.xi - i t |xi|^2 }
f_hat(xi) d xi.
Do not worry if this looks scary; Chapter 11 explains every symbol. For now
read it as: the data f is a superposition of plane waves e^{i x.xi} indexed by
frequency xi, and the plane wave of frequency xi steadily picks up the phase
factor e^{-i t |xi|^2} as time advances. The operator S(t) is "run the wave
for time t". It is exactly the evolution for initial data f, and it preserves
the L^2 size of the data (unitarity, Chapter 11).
The question of this paper is then, in one line:
Q. For which f does S(t) f(x) -> f(x), as t -> 0+, at almost every x?
The answer is given in terms of how much "roughness" f is allowed, measured
by a Sobolev exponent (next chapter).
3. SMOOTHNESS AND THE SOBOLEV EXPONENT
--------------------------------------------------------------------------------
Roughness is measured in Sobolev spaces. For a real number s, the H^s norm
of a function f on R^n is
||f||_{H^s}^2 = (2 pi)^{-n} INT_{R^n} ( 1 + |xi|^2 )^s
|f_hat(xi)|^2 d xi.
In words: multiply the energy at frequency xi by (1+|xi|^2)^s, so high
frequencies are weighted more heavily. A large s means high frequencies are
strongly penalized, i.e. f is very smooth. s = 0 is just the L^2 norm
(up to the same constant). Larger s means the allowed data is smoother; as
s drops, rougher and rougher data is admitted.
The main result of this paper is that convergence holds down to the exact
value
s_n = n / ( 2 (n+1) ), for every n >= 3.
For n = 3 this is 3/8, for n = 4 it is 2/5, and so on. This number is not
arbitrary; it is forced from both sides (Chapter 9). Two features of s_n
matter for the whole narrative:
- It sits strictly between 0 and 1/2. It is a "rough" regime. In
particular the data is not assumed continuous: functions in H^{s_n}
can be unbounded on sets of positive measure.
- It is an "endpoint". The paper proves convergence for every f in
H^{s_n}, not merely for every f in H^s with s > s_n. The endpoint is
where the estimates have zero slack, and it is exactly the place the
harder parts of the proof are engineered to reach.
One more word: the paper proves the theorem for complex-valued f. Real-valued
is the special case everyone pictures, but complex data is what the Fourier
method handles cleanly, and it is strictly more general.
4. WHY ONE DIMENSION IS A THEOREM AND HIGHER DIMENSIONS ARE HARD
--------------------------------------------------------------------------------
In one dimension the story is complete and old. For n = 1 the endpoint is
s_1 = 1/4. Carleson (1980) proved convergence for H^{1/4}(R), and Dahlberg-
Kenig (1982) proved that below 1/4 there are counterexamples. So in one
dimension, convergence holds exactly for s >= 1/4, endpoint included.
For higher n, convergence at a given s is a strictly harder question. The
reason is dispersion. The free evolution spreads a sharp packet of waves out
over wider and wider regions as time grows, and it does so more slowly in
higher dimensions in a certain precise sense. Near t = 0 the high-frequency
pieces can arrange themselves to concentrate again, and the geometry of
"where can many frequency pieces pile up at time t" becomes genuinely n-
dimensional.
Historically, for n >= 2, progress came in stages:
- Sjolin (1987) and Vega (1988): convergence for s > 1/2 in every
dimension. Half a derivative was the first natural threshold.
- The plane n = 2 was attacked with bilinear and restriction methods:
Bourgain (1995), Moyua-Vargas-Vega (1996), Tao-Vargas (2000),
Tao (2003), culminating in Lee (2006) with s > 3/8.
- Bourgain (2013) reached s > 1/2 - 1/(4n) in dimensions n >= 3.
- Bourgain (2016) produced the counterexample that rules out every
s < n/(2(n+1)) in every dimension n >= 2. This made
s_n = n/(2(n+1)) the universal conjectured threshold.
- Du, Guth, Li (2017) reached the endpoint 1/3 for n = 2.
- Du, Guth, Li, Zhang (2018) reached s > (n+1)/(2(n+2)) in higher
dimensions, the first dent below 1/2.
- Du, Zhang (2019) reached s > n/(2(n+1)) in higher dimensions. This
is the immediate predecessor: every exponent strictly larger than the
conjectured endpoint.
Until this paper, for n >= 3 the endpoint s = s_n itself was open. The
conjecture was that it should hold. This paper proves it. The plane n = 2
endpoint is the subject of the companion planar paper, not this one.
5. THE STATE OF THE ART BEFORE THIS PAPER
--------------------------------------------------------------------------------
Read this chapter as the map before the journey.
Bourgain's counterexample (2016) gives a lower bound: if s < n/(2(n+1)),
there is initial data f in H^s for which the evolution does not converge at
points of a set of positive measure. Hence no exponent below s_n can work.
Du and Zhang (2019) give a one-sided upper bound: for every s > n/(2(n+1)),
convergence holds. Their proof introduced (among other things) a sharp higher-
dimensional maximal estimate, and it is the point of departure for this paper.
But a sharp inequality for every s > s_n does not, by itself, give the
endpoint s = s_n, because at s = s_n all the epsilon-slack is gone.
The gap between "every s > s_n" and "s = s_n" is not just a nuisance. Many
techniques lose an arbitrarily small power of the large scale N. If a proof
loses N^eps for every eps > 0, it can prove "every s > s_n" but never the
endpoint. The main novelty of this paper is a collection of estimates with
zero loss at the endpoint: the near-Cauchy theorem, the buffered refined
decoupling, the transverse frame estimate, and the codimension-two strip
saving (all explained in Part IV). These are the tools that let the endpoint
close.
The companion paper does the same job in the plane n = 2. This paper is for
n >= 3, and the n-dimensional geometry is genuinely harder: the "transverse"
configuration of wave packets is a simplex of n+1 pieces, not 3.
================================================================================
PART II: THE MAIN THEOREM, AND WHAT IT REALLY SAYS
================================================================================
6. THE MAIN THEOREM, STATED EXACTLY
--------------------------------------------------------------------------------
As explained in Chapter 2, for rough f the Fourier integral defining S(t) f
need not converge at a point. So we cannot simply write "S(t) f(x) -> f(x)";
we must say how the expression is given a meaning. The paper's choice is:
- First smooth the data slightly. For a > 0 define the Gaussian-regularized
evolution
U_a f(x,t) = (2 pi)^{-n} INT_{R^n}
e^{ i x.xi - i t |xi|^2 - a |xi|^2 }
f_hat(xi) d xi.
The extra factor e^{-a |xi|^2} cuts off frequencies above |xi| ~ a^{-1/2}.
For every a > 0 the integral converges absolutely, U_a f is smooth, and
U_a f is the exact evolution of the smoothed data U_a f(.,0).
- Then let a -> 0, and see whether the regularized evolutions settle down to
a single limit function u.
- Finally let t -> 0 and see whether u(.,t) returns to the data.
The theorem is that this procedure works, for every f in H^{s_n}, on one
fixed full-measure set of x, uniformly in t in the open interval (0,1).
To state the size of the exceptional set precisely: a subset E of R^n has
full measure if its complement has Lebesgue measure zero. "For almost every
x" is shorthand for "for every x in some full-measure set E".
There is one more piece of bookkeeping. For f in L^1_loc(R^n) and a point x,
the Lebesgue value of f at x is
f*(x) = lim_{r -> 0} (average of f over the ball B(x,r))
provided the limit exists. By Lebesgue's differentiation theorem it exists
for almost every x, and f* = f almost everywhere. So f* is a "canonical
representative" of f, defined pointwise, that equals f almost everywhere.
THEOREM (main theorem of the paper). Fix n >= 3 and put
s_n = n / (2(n+1)). For every f in H^{s_n}(R^n) there exist a set
E subset R^n of full measure and a complex-valued function u(x,t) on
E x (0,1) such that:
(a) For every x in E,
sup_{0 < t < 1} | U_a f(x,t) - u(x,t) | -> 0
as a -> 0+. In words, the regularized evolutions converge to u
uniformly in t on the open interval 0 < t < 1.
(b) For every x in E,
u(x,t) -> f*(x) as t -> 0+.
In words, u returns to the Lebesgue value of the data at time 0.
(c) Consequently, for every x in E and every path t_a -> 0+,
lim_{a -> 0} U_a f(x, t_a) = f*(x) = f(x) (a.e. x).
In particular, convergence holds simultaneously for all times.
This is the sharp form for which the paper is written. The abstract sums it
up as: for initial data in H^{s_n}, the free Schrodinger evolution converges
almost everywhere to the initial data, the evolution being defined by
removing Gaussian regularization on one full-measure set, uniformly over
0 < t < 1.
Three remarks on the statement.
- Part (a) is where all the analysis lives. It is a maximal, uniform,
"on every point of E simultaneously" statement, and the paper ultimately
reduces it to a maximal inequality for a single frequency-localized
problem, with a constant that does not depend on the frequency.
- Part (b) is essentially a Lebesgue differentiation argument, but it needs
part (a) plus a small amount of extra care about how the regularizer a
and the time variable t interact. Chapter 30 handles this.
- Part (c) is the reason the theorem is stated with "for every path
t_a -> 0". Because (a) is uniform in t in (0,1), the limit as a -> 0 of
U_a f(x, t_a) does not depend on how t_a varies with a.
7. WHY THE ORDER OF LIMITS MATTERS
--------------------------------------------------------------------------------
You might ask: why not define the evolution as the limit of S(t) f(x) as
t -> 0 for each fixed x, using the Fourier integral directly? The reason is
that for f in H^{s_n}, with s_n > 0 but very small, the Fourier integral for
S(t) f(x) at a fixed t > 0 is not absolutely convergent. The data f is only
L^2, and the multiplier e^{-i t |xi|^2} does nothing at infinity. So S(t) f
itself is an L^2 object, not a function, and "evaluate at x" is meaningless
at the point level without extra structure.
The Gaussian factor e^{-a |xi|^2} fixes both problems at once:
- It makes the frequency integral absolutely convergent, so U_a f is an
honest function of (x,t), smooth in both variables.
- It smooths the data at scale a^{1/2}. At frequencies |xi| << a^{-1/2}
the factor e^{-a |xi|^2} is ~ 1. So U_a f is the true evolution of the
data with frequencies above ~ a^{-1/2} gently removed.
The order of limits that the theorem uses is: remove the smoothing first
(a -> 0, uniformly in t), then take time to zero. If instead you tried to
interchange a -> 0 and t -> 0, you would not get a well-defined limit,
because at time t the plane wave of frequency xi carries the factor
e^{-i t |xi|^2}, which for |xi| ~ a^{-1/2} sees phase ~ t/a; altering the
relative rate of a and t can pick out any phase at all. So the order of
limits is not an aesthetic choice. It is what makes the statement true and
canonical.
The reader who is used to convergence-almost-everywhere statements from
elliptic or heat-flow theory should note: the reason the paper uses the
Gaussian version is not because the unregularized claim is known to be
false, but because it is the version that the additive combinatorics and
packet analysis actually delivers. The unregularized pointwise convergence at
s = s_n is a stronger statement that is not attempted here. For the plane,
the analogous unregularized statement at s = 1/3 is what the companion
planar paper proves. (See Chapter 32.)
8. SHARPNESS: WHY s_n = n/(2(n+1)) IS EXACTLY RIGHT
--------------------------------------------------------------------------------
There are two halves to "exact".
(a) Sharpness from below. Bourgain (2016) constructed, for every
s < n/(2(n+1)), data f in H^s and a set of x of positive measure
on which the evolution does not converge to f. So no s < s_n can
work. This bound holds in every dimension n >= 2.
(b) Sharpness from above. This paper proves convergence at s = s_n,
hence a fortiori at every s > s_n.
Together: the endpoint of convergence is exactly s_n, and the present paper
reaches it.
It is worth understanding the shape of Bourgain's counterexample, because it
is the reason the number n/(2(n+1)) appears. The construction is a
concentration of high-frequency mass near a codimension-2 set (a point in the
plane, a line in three dimensions). Near such a set, wave packets can be
arranged so that the phases align at every short time, but at each fixed
point the phases do not align once the regularization is removed. This is a
"fractal" obstruction. Its scale is dictated by the codimension of the
concentration set, which in turn produces the exponent n/(2(n+1)): the
denominator n+1 is the sum of the spatial dimension n and the codimension 1
of the "time direction" as a coordinate, and the powers in the counterexample
come from the volume growth of n-dimensional balls inside R^{n+1} spacetime.
There is no stronger counterexample. The conjecture from Bourgain 2016 onward
has been: convergence holds at s = s_n. This paper proves the conjecture.
One caution about the word "endpoint". An endpoint statement is qualitatively
harder than the corresponding "every s > s_n" statement. If you have a family
of estimates whose constants blow up as s -> s_n from above, you have no
endpoint. The entire architecture of this paper is engineered to produce
estimates whose constant is finite at s = s_n. That is the deep content.
9. THE NON-ENDPOINT THEOREM, AND WHY THE ENDPOINT RESISTS STANDARD ARGUMENTS
--------------------------------------------------------------------------------
It is worth stating the Du-Zhang (2019) result cleanly, because it is the
closest point of comparison.
THEOREM (Du-Zhang 2019). Fix n >= 3. For every s > n/(2(n+1)) and every
f in H^s(R^n), the free evolution S(t) f converges to f almost everywhere
as t -> 0+.
The proof of this theorem has the following shape. Let N be a large
frequency cutoff, and let P_N be the frequency projector onto |xi| ~ N. The
goal is a maximal estimate
|| sup_{0<t<1} | P_N S(t) f | ||_{L^p(B(0,1))} <= C_eps N^{s + eps}
|| P_N f ||_2
for some finite p and every eps > 0. This is a "frequency-local maximal
estimate with an epsilon loss". Given such an estimate, one localizes to a
ball, sums over dyadic N via a Borel-Cantelli argument, and concludes
convergence a.e. on the ball. The epsilon loss is absorbed by the
summability in N, because (at exponent s strictly bigger than s_n) the
N^{s - s_n} decay dominates.
At the endpoint s = s_n, this scheme fails for two separate reasons:
1. The epsilon loss is fatal. If you lose N^eps at the endpoint, then the
N^{s_n} growth and the N^eps loss cannot be summed, and the
Borel-Cantelli argument gives nothing. So an endpoint proof needs an
estimate with NO epsilon loss.
2. The Borel-Cantelli argument must hold on EVERY unit ball, not just a
fixed one, or the conclusion "a.e. convergence" fails at the boundary
between balls. This is a localization issue, and it has to be handled
uniform in the location of the ball.
The present paper's architecture solves (1) with four lossless estimates
(Chapter 10 lists them), and (2) with a common time tree and a careful
Gaussian regularization argument (Chapters 29 and 30). The lossless estimates
are where the depth lives; the localization and the time-tree argument are
delicate but essentially bookkeeping once the lossless estimates are in hand.
10. STRATEGY AT 30,000 FEET: THE THREE BIG MOVES
--------------------------------------------------------------------------------
The proof splits into three big moves, and it is useful to hold all three in
your head before diving into any of them.
MOVE A. Reduce the theorem to a frequency-local maximal estimate at the
endpoint, with no epsilon loss.
This is a statement about the solution of the Schrodinger equation whose
initial data has frequency only near |xi| ~ N, and whose time interval is
of length 1/N (much shorter than the natural frequency-local time). The
reduction lives in Chapters 29 and 30. Statement:
For every f with frequency near |xi| ~ N, every unit ball Omega,
and every time interval I of length 1/N,
|| sup_{t in I} | S(t) f | ||_{ L^{p, infinity}(Omega) }
<= C N^{s_n} ||f||_2,
where p > 2 is some fixed exponent, and the constant C does not
depend on N, Omega, or I. (Here L^{p, infinity} is weak L^p,
defined in Chapter 27.)
MOVE B. Prove Move A, by descending into a hierarchy of packet estimates
and a geometric induction. This is the entire middle of the paper, and
it is where the real content is. It splits into three further parts:
B1. Fractal and transverse-saving estimates. Chapters 21-23. These
give the input: high density of packets at a point forces their
frequencies to be spread out (transverse), and transversality buys
a quantitative saving.
B2. Sparse iteration. Chapters 23 and 24. Combined with fractal and
decoupling inequalities, the transverse saving becomes a genuine
power saving D^{-c_s} in a "sparsity parameter" D. This is the
sparse estimate (Chapter 24).
B3. Geometric induction. Chapters 25-28. This upgrades the sparse
estimate (sparsity level D a small positive power of N) to the
full range 1 <= D <= N, by reducing a general maximal estimate to
the frame estimate applied in dimension n-1. This is the hardest
and most novel part.
MOVE C. Assemble Move A across frequencies on a single "time tree", and
convert the result to pointwise convergence a.e. via a Gaussian
approximate identity. Chapters 29 and 30.
The three moves fit together as follows. Move C needs Move A as a black box.
Move A is proved by Move B. Move B decomposes into B1, B2, B3, each of which
relies on a family of packet, fractal, and geometric inequalities proved in
Chapters 19-23. Everything upstream of Move B is standard in spirit (it is
the Du-Zhang engine with lossless replacements); almost all of the novel
content is in B3, i.e. in the transverse frame estimate and its use.
11. WHAT IS NEW: THE FOUR INGREDIENTS THAT BUY THE ENDPOINT
--------------------------------------------------------------------------------
The whole point of the paper is to reach the exact endpoint s = s_n. Every
place a proof normally loses N^eps has to be replaced by an estimate with no
loss. The paper introduces four such replacements. Memorize these; they are
the four reasons the paper exists.
(i) THE NEAR-CAUCHY THEOREM [paper Sec 2, packets.tex]. A point count
lemma for frequency-localized wave packets, at the sharp exponent q_k
= 2(k+2)/k, with no N^eps. It says: if you have many high points of a
sum of packets, then their frequencies are nearly orthogonal, and you
can count them. This is the "packet algebra" of the paper, and it
replaces the epsilon-losing point counts in prior work.
(ii) BUFFERED REFINED DECOUPLING [paper Sec 2, packets.tex]. A version of
Bourgain-Demeter decoupling that pays a fixed small power of the
buffering scale, not N^eps. The reason it is needed: in the paper's
iterative arguments, decoupling has to be applied inside itself, and
each inner application must be lossless relative to the outer one.
This repairs a "collar gap" in the Singh-Singh-Parmar paper, and is
proved in every dimension in the present paper (Lemma p:buffered-
refined).
(iii) THE TRANSVERSE FRAME ESTIMATE [paper Sec 3, frame.tex]. The most
novel single ingredient. It is a transverse maximal estimate in k
dimensions, closely tied to the geometry of the paraboloid. Its main
content: k+1 transversally arranged wave packets cannot concentrate at
many points if each packet is individually large. This is the engine
of B3. Chapters 21 and 28.
(iv) THE CODIMENSION-TWO STRIP SAVING [paper Sec 4, fractal.tex]. A
quantitative gain available for wave packets whose frequencies lie in
a thin neighborhood of an affine (n-2)-plane (subspace of codimension
two in frequency space). It is the reason the geometric induction can
reduce dimension from n to n-1 (Chapters 22 and 28).
The rest of the proof is the Du-Zhang machinery, with two other technical
repairs (a general Du-Guth-Li-Zhang multilinear restriction estimate, and a
lossless endpoint version of the Du-Zhang fractal estimate). These are
standard in spirit, and are explained in Chapter 23.
With the four ingredients in hand, the proof is a long but structured
argument: packet hierarchy -> transverse saving -> sparse estimate ->
geometric induction -> short-time maximal estimate -> frequency summation ->
Gaussian representative -> theorem.
================================================================================
PART III: BACKGROUND
================================================================================
This part is the toolbox. The paper only
uses a handful of facts from classical analysis, but each fact has to be
available before Part IV can make sense.
12. THE FOURIER TRANSFORM: SPLITTING A WAVE INTO PURE FREQUENCIES
--------------------------------------------------------------------------------
Think of a musical note. A cellist bowing one string produces a single pitch,
but plucking the same string produces a mixture, and any mixture can be
written as a sum of pure tones. The Fourier transform is the same operation
for a function of position. It decomposes f into "pure frequencies" xi.
Formally, for f : R^n -> C which is smooth and decaying fast at infinity,
f_hat(xi) = INT_{R^n} f(x) e^{- i x . xi} d x.
The inverse operation is
f(x) = (2 pi)^{-n} INT_{R^n} f_hat(xi) e^{ i x . xi} d xi.
Two operations are important:
- The Fourier transform almost preserves energy: Plancherel's identity
says ||f||_{L^2} = (2 pi)^{-n/2} ||f_hat||_{L^2}, with the same
constant for f and f_hat. The whole "energy" picture of the paper is
built on this.
- Differentiation becomes multiplication. If D_j = d/d x_j, then
the Fourier transform of D_j f is i xi_j f_hat. Hence "smoothness" is
"decay of f_hat at infinity", and this is why the Sobolev norm in
Chapter 3 is a weighted L^2 norm of f_hat.
For all the fancy objects in the paper, this is the mental model: high
frequencies correspond to fast spatial oscillation, low frequencies to slow
spatial variation.
13. HOW A SINGLE FREQUENCY MOVES: GROUP VELOCITY AND WAVE PACKETS
--------------------------------------------------------------------------------
A single "pure frequency" xi, viewed as a function of (x,t), is
e^{ i x . xi - i t |xi|^2 }.
We can rewrite the phase as
x . xi - t |xi|^2 = (x - 2 t xi) . xi + t |xi|^2.
The time-dependent constant t |xi|^2 is a phase; the spatial argument is
x - 2 t xi. This says: at time t, the plane wave of frequency xi has the
same spatial profile as at time 0, translated by 2 t xi. So the wave moves
with velocity
v(xi) = 2 xi (this is called the group velocity).
Consequence: each frequency has a definite direction and speed. A packet
of waves with frequencies in a small ball around xi travels almost like a
single rigid object moving at speed 2 xi. The direction vector
v_vec(xi) = (2 xi, 1) in R^{n+1}
is important. Its last coordinate is 1 because the packet moves forward in
time at unit time-normalized rate; the first n coordinates give the
direction in space. This is the vector whose determinant conditions appear
in Chapter 16.
14. NATIVE COORDINATES: MAKING A PACKET LOOK THE SAME AT EVERY SCALE
--------------------------------------------------------------------------------
The Schrodinger equation has a scaling symmetry. If u(x,t) solves it, so
does a rescaled copy. The scaling is:
x -> x / N, t -> t / N^2,
together with a corresponding rescaling of the amplitude. This "parabolic
scaling" reflects the parabolic relation between space and time: even
though the equation looks like it has first order in t and second order in x,
the natural length scale for time is (length)^2.
A wave packet of "scale N" with central velocity v occupies, at a fixed
time, a spatial ball of radius ~ N, and its frequency support is a ball of
radius ~ 1/N centered at v. In time it survives for duration ~ N^2. This is
the "unit packet" of the paper.
The paper expresses this using "native coordinates": for a fixed scale N
and velocity v,
X_{N,v}(x,t) = ( (x - 2 v t) / N , t / N^2 ).
In these coordinates, a packet at scale N and velocity v occupies a unit
box (of side ~ 1) centered at the origin. The "packet" is then a single
fixed function of X, and every N-packet at every v is just that same
function translated in X.
Native coordinates are used everywhere in the paper. They are why the
statement "a packet is a compact array" makes sense; see Chapter 22.
15. MEASURING SIZE: LEBESGUE, BALL GROWTH, AND FRACTAL DIMENSION
--------------------------------------------------------------------------------
Physics would like to say "the set of points where the wave has size at
least lambda has size such-and-such". But this set may be a fractal: neither
empty nor full, with non-integer dimension. The paper measures sizes with a
positive Borel measure mu, and controls "which sets can mu give mass to" via
a "ball growth inequality":
mu(B(z,r)) <= m r^n for every z in R^{n+1} and every r >= 1.
Interpretation: mu assigns mass at most m r^n to any ball of radius r. This
says mu is n-dimensional inside the (n+1)-dimensional spacetime. If mu were
n-dimensional Lebesgue measure, this bound would hold with m = volume of
the unit ball. If mu were the counting measure on an n-dimensional set (for
example, a spacetime surface), the bound would hold with some m.
The parameter m is the "ball parameter". It is the first scalar that
controls every estimate.
There is a second important property: the "tube moment condition". Fix the
trajectory of a single packet of scale L, which is the set
{ (x,t) : |x - y_alpha - 2 t omega_alpha| <= L }
(isotropic in x, time bounded). The "tube moment" of mu on this trajectory
is
INT (1 + |x - y_alpha - 2 t omega_alpha| / L)^{-M} d mu(x,t),
where M is a fixed large constant. The paper requires this to be at most
m L^{n+1} / D, for a fixed sparsity parameter D (Chapter 22). When D is
large, mu "avoids" the packet's trajectory. When D = 1, the condition is
just a consequence of the ball growth inequality (up to a constant).
Together, m and D control: how much mass, how focused. Both have to be
tracked through the entire proof because each estimate produces a new m and
D.
16. TRANSVERSALITY: WHEN DIRECTIONS ARE GENUINELY DIFFERENT
--------------------------------------------------------------------------------
Two directions are "transverse" if they are genuinely different, not nearly
parallel. The precise notion: a collection of directions is transverse if
the determinant of the corresponding direction vectors is bounded away from
zero.
For a frequency xi in R^n, let v_vec(xi) = (2 xi, 1) as in Chapter 13. A
collection of n+1 frequency sets U_1, ..., U_{n+1} (each a subset of R^n) is
"delta-transverse" if for all choices of xi_j in U_j,
| det( v_vec(xi_1), ..., v_vec(xi_{n+1}) ) | >= delta > 0.
The determinant here is the (n+1)-by-(n+1) determinant of the matrix whose
jth column is v_vec(xi_j).
Geometric meaning. The vector v_vec(xi) is tangent to the parabola
xi |-> (2 xi, |xi|^2). The n+1 vectors v_vec(xi_1), ..., v_vec(xi_{n+1}) are
transverse iff they span R^{n+1}. The determinant is a "volume" of the
spanned simplex. It is small if some pair of directions nearly coincides or
if the directions lie near a lower-dimensional plane.
In every transverse-saving argument, the procedure is:
- Take n+1 families of wave packets, all of high amplitude on a common
dense set of points. Show that either (i) their directions cluster
into a lower-dimensional affine subspace (the "narrow" case), or (ii)
they are transverse.
- In case (ii), apply the transverse estimate (Kakeya or refined
Strichartz) and get a saving.
- In case (i), reduce to a lower-dimensional problem (by an inductive
argument on the dimension of the subspace) or use decoupling directly.
The "simplex lemma" (paper Lemma s:simplex) is the exact statement: given
n+1 sets of frequencies, either they lie in a neighbourhood of a hyperplane,
or a fixed small subset has bounded-below determinant. Its proof is a
Gram-Schmidt argument. It is a combinatorial fact, not an analytic one.
17. MULTILINEAR KAKEYA: MANY TUBES IN MANY DIRECTIONS
--------------------------------------------------------------------------------
Kakeya problems ask how much space can be covered by thin tubes in many
directions. The multilinear Kakeya theorem (Bennett-Carbery-Tao 2006, with
later weighted forms by Guth) is the fundamental estimate, and it is the
"soft" transverse estimate used everywhere in the paper.
Statement (simplest version). Let d >= 2. For j = 1, ..., d let T_j be a
family of infinite cylinders of radius h in R^d, with unit direction vector
e(T) for T in T_j. Suppose
| det( e(T_1), ..., e(T_d) ) | >= theta > 0
for every tuple (T_1, ..., T_d) in T_1 x ... x T_d. Then for any nonnegative
weights c_j(T),
INT_{R^d} prod_{j=1}^d
( sum_{T in T_j} c_j(T) 1_T )^{1/(d-1)} d z
<= C_d theta^{-1/(d-1)} h^d prod_{j=1}^d
( sum_{T in T_j} c_j(T) )^{1/(d-1)}.
What this says. On the left, we count the size of the "interaction region":
for each point z, we take the number of j-tubes of T_j covering z, raise each
to the power 1/(d-1), multiply the d factors, and integrate. On the right,
h^d is the volume of a tube cross-section times length, and the last factor
is the "cost" (sum of weights). The estimate says: the tubes cannot
pigeonhole into a small region -- the mass remains controlled.
The determinantal transversality condition is what makes the estimate true.
If some pair of direction families coincided, tubes in those families could
overlap completely and the estimate would fail.
The paper's weighted form (Proposition p:kakeya) is exactly the one shown
in the paper's Section 2. The "lattice version" of it (for unit cube sums
rather than integrals) is also needed and follows from the integral version
by summing over cubes.
The reason multilinear Kakeya is central: it is the one transverse estimate
that is entirely elementary in spirit (a combinatorial incidence argument),
which makes the whole edifice "close" without invoking deep restriction
theory in every step.
18. REFINED STRICHARTZ AND ITS BUFFERED VERSION
--------------------------------------------------------------------------------
Strichartz estimates are the L^p analogues of Plancherel for the Schrodinger
equation. The classical Strichartz estimate is
||U f||_{L^q(R^{n+1})} <= C ||f||_{L^2(R^n)} U f = S(t) f,
for the "admissible" pairs (q, p) satisfying q = 2(n+2)/n when p = 2. In the
paper the exponent is q_k = 2(k+2)/k with k = n; this is the endpoint of
admissibility.
Refined Strichartz. A refined estimate gains a factor when the packets have
been placed in a "transverse" configuration. The Du-Guth-Li-Zhang 2018
result proves a multilinear refined Strichartz estimate: given n+1
transverse families of packets, the number of high unit-cube incidences is
bounded by an appropriate product of coefficient energies times a
subpolynomial loss. The subpolynomial loss is "N^{eps}", and this is what
blocks the endpoint.
Buffered refined decoupling (paper Lemma p:buffered-refined). This is the
paper's lossless replacement. The estimate says: for packages buffered with
an extra scale Lambda = R^vartheta, the constant is bounded by R^{eps} times
a fixed power of the buffer, not by a generic N^{eps}. The gain is a fixed
positive power of the buffer, which is exactly what iterating decoupling
inside itself requires. The proof is a substantial multi-scale induction in
the paper's Section 2, modeled on Singh-Singh-Parmar and Brady, and it is
where the "collar gap" repair happens. In this companion we state the lemma
and use it as a black box; the reader who wants the full induction can
follow the paper's Section 2 line by line.
The key point to remember: without the buffered version, every inner use of
decoupling would lose N^{eps}, and after a few iterations you would be over
budget. With the buffered version, the loss is at most a fixed positive
power of the buffer (a much smaller quantity in the relevant range), and
it sums.
19. DECOUPLING (BOURGAIN-DEMETER)
--------------------------------------------------------------------------------
Decoupling is the fundamental estimate behind all modern restriction theory.
The idea: if a function H has Fourier support split into delta-caps J of a
surface, then the L^p norm of H behaves like the l^2 sum (in J) of the L^p
norms of the pieces, for p up to a critical exponent.
Precise statement for the paraboloid. Fix k >= 1 and let J range over
squares of side N^{-1/2} whose union covers the frequency box in R^k. For a
field H = sum_J H_J with each H_J supported in a fixed enlargement of the
cap J, and 2 <= p <= q_k where
q_k = 2(k+2)/k,
Bourgain-Demeter (2015) proves
|| H ||_p <= C_eps N^eps ( sum_J || H_J ||_p^2 )^{1/2}.
This is a "square-root" estimate: any overlap between the caps J and J' in
L^p is at most the l^2 sum, up to a tiny N^eps loss. The exponent q_k is
sharp; above it the estimate is false without an extra loss.
Important constant features:
- The loss is N^eps. For endpoint arguments this is fatal; it must be
replaced by something with better dependence on N. The paper's
"buffered refined decoupling" (Chapter 18) is exactly this replacement.
- There is a "local weighted" form useful in practice. Fix a lattice cube
Q_s of side s. Then for 2 <= p <= q_k,
|| H ||_{L^p(Q_s)}
<= D_{delta,L,k,p} s^delta ( sum_J || H_J ||_{L^p(w_{Q_s,L})}^2 )^{1/2}
with w_{Q_s,L}(z) = (1 + dist(z,Q_s)/s)^{-L}. The cost is s^delta, a
fixed small power of the cube side, not N^eps. For the paper's
purposes this is the relevant form (paper Lemma p:bd-local).
- The decoupling exponent q_k can be below 1 in the exponent of the norm
of the sum; that is why the paper repeatedly splits products into
bounded-overlap color classes and uses l^2 / Holder bookkeeping.
The phrase "decoupling inequality" describes the whole family of estimates
that let you pull the Fourier pieces of H apart, compute the pieces one at a
time, and reassemble at the end. It is a very powerful reduction: instead of
estimating a complicated sum H, you estimate each H_J separately. The cost
is the tiny N^eps (or s^delta) factor, and everything in the endpoint proof
is devoted to controlling that cost.
20. POLYNOMIAL PARTITIONING (GUTH-KATZ)
--------------------------------------------------------------------------------
Polynomial partitioning is the geometric engine behind all modern incidence
estimates. The idea: given a set of M points and an integer parameter D,
there is a nonzero polynomial P of degree at most D whose zero set divides
R^n into "sign cells", none of which contains too many points.
Precise statement (Guth 2016, Guth-Katz 2015). Fix dimensions n >= 1. Let
A subset R^n be a finite set of M distinct points and let D >= 1. Then
there exists a nonzero real polynomial P of degree <= D such that each
"sign cell" of P (each connected component of a set where P has a fixed
sign) contains at most
C_n M / D^n
of the points. The constant C_n depends only on n.
Complementary fact (Wongkew 1993). A real algebraic variety of codimension
c has a tubular neighbourhood of radius r that has volume at most a
polynomial in r, depending on the degree and codimension. This is used to
control how much mass a partition's zero set can hold.
The polynomial partitioning theorem is used in the same way as the
multilinear Kakeya: it gives a "soft" geometric estimate that the proof
uses whenever the frequencies cannot be further simplified. The paper needs
it in the transverse saving (Chapter 35) and in the wall decomposition
(Chapter 39). It is a deep theorem by itself (it required work by Guth,
Guth-Katz, and many others), but it is used at a fixed "level" in the
present proof: it is the classifier of the partition, not the source of the
saving.
The finite-set partition theorem is exactly what the paper cites from
Guth-Katz. The integrable-density continuous version is from Guth 2016
Restriction (Theorem 1.4).
21. PROJECTION THEOREMS: HOW A SET LOOKS FROM A RANDOM DIRECTION
--------------------------------------------------------------------------------
A projection theorem answers: for a set A in R^d and a direction v, how
many cells does A cover in the projection onto the hyperplane v^perp? If A
is spread out in R^d, then most projections of A should be "bijective" up
to a small loss. The theorem says this: unless the direction v is very
special, the projection does not collapse much.
The paper uses a specific robust form of this, due to He (2020). Let A be
a set in a unit ball in R^d, and let Delta > 0 parametrize a fine grid.
For a small eta > 0, suppose:
(i) The number of Delta-cells meeting A is between Delta^{-alpha+eta}
and Delta^{-alpha-eta} for some 0 < alpha < d.
(ii) The set A has non-concentration: A cap B(x,r) contains at most
Delta^{-eta} r^rho times as many Delta-cells as A itself, for all
Delta <= r <= 1 and some rho > 0.
(iii) nu is a probability law on unit directions (kernel directions) that
has the same nonconcentration nu{v : |v . w| < r} <= Delta^{-eta} r^rho
for every |w| = 1 and all Delta <= r <= 1.
Under these hypotheses, He's theorem guarantees that some direction v in
the support of nu projects A bijectively up to a fixed power:
there is gamma > 0 such that for every subset A' of A whose Delta-cell count
is at least Delta^eta of that of A,
N_Delta( pi_{v^perp} A' ) >= Delta^{-(d-1) alpha / d - gamma}.
This is the "robust" form: the conclusion holds uniformly for every
sufficiently large subset A' of A, not only for A itself. The price is the
extra slack eta in the hypotheses and gamma in the conclusion.
Why the paper needs it. The transverse saving's proof (Chapter 35) has to
rule out the case where the packet directions are all "almost contained" in
a hyperplane. The projection theorem is what makes this exclusion rigorous:
if they were all close to a hyperplane, their projection onto the normal
direction would be small, but the non-concentration hypothesis (Lemma t:rank)
violates that. It is a piece of "linear algebra plus Dirichlet principle",
not a deep restriction-theoretic input. It is quoted from He's paper and
used as a black box.
22. COMPACT ARRAYS, ENERGY AND SPARSITY
--------------------------------------------------------------------------------
The wave packets introduced in Chapter 14 are the core objects of the whole
proof. They are formalized as "compact arrays":
U(X,T) = sum_alpha
c_alpha e^{ i ( X . omega_alpha - T |omega_alpha|^2 ) }
phi( X - y_alpha - 2 T omega_alpha, T ).
Breakdown:
- alpha indexes the packets. The label alpha = (omega_alpha, y_alpha) is a
frequency omega_alpha in a fixed bounded frequency box, together with a
spatial intercept y_alpha.
- phi is a single fixed compact profile: a bump function that decays
rapidly away from the unit ball and has a fixed number of derivatives
bounded. It is the same profile for every alpha. It is what makes the
sum an array of packets rather than a generic superposition.
- c_alpha is a complex coefficient.
- The phase e^{i(X . omega_alpha - T |omega_alpha|^2)} is the carrier
wave; the profile phi(.) slides along with the group velocity as
expected (Chapter 13).
The coefficient energy of U is
E(U) = sum_alpha | c_alpha |^2.
This is the L^2-squared size of the input. It is constant under most of the
operations the paper performs. It plays the role of "how much total data
is here".
Two useful basic properties:
- Supremum bound. By Cauchy-Schwarz and bounded phase-space multiplicity,
|| U ||_{infinity} <= C E(U)^{1/2}.
So amplitude is controlled by energy.
- Energy summable under translation. Translating a packet by a constant
changes its phase but not its energy. This is used everywhere.
Now the sparsity parameter D. The measure mu on spacetime satisfies:
(a) Ball growth: mu(B(z,r)) <= m r^n for every z and r >= 1.
(b) Tube moment: for every label alpha,
INT (1 + |x - y_alpha - 2 t omega_alpha| / L)^{-M} d mu(x,t)
<= m L^{n+1} / D.
Interpretation. D measures how much mu avoids the packet trajectories.
When D is large (up to ~ L^tau for the sparse estimate, then up to ~ L for
the full weak estimate), each individual trajectory sees very little mass.
When D is small (near 1), mu is dense and the tube moment is just a
consequence of ball growth.
The compact array + measure + sparsity parameter (U, mu, D) is the principal
input to the sparse estimate (Chapter 36), and, with more structure, to the
weak estimate (Chapters 37-41). All the work of the proof goes into
controlling this "amount of concentration" (D) and turning it into a
saving.
23. INFORMATION THEORY: BITS OF A POINT, BITS OF A DIRECTION
--------------------------------------------------------------------------------
Chapter 5 of the paper uses a language that is unusual in harmonic analysis:
entropy and conditional information. The reason is that the transverse
saving's proof needs to compare, at a fixed scale, how much information a
sampled point gives about the direction of a wave through it. This is
exactly an information-theoretic question, and the language makes the
argument clean.
Shannon entropy. For a finite set A with probability p_A(a), the entropy in
base Lambda > 1 is
I_Lambda(A) = - ( 1 / log Lambda ) sum_a p_A(a) log p_A(a).
In words: I_Lambda(A) "counts" the number of Lambda-ary possibilities for A.
If A has at most Lambda^{C+o(1)} values, then I_Lambda(A) <= C + o(1). If A
is deterministic, I_Lambda(A) = 0.
Conditional information. For two finite labels A and B,
I_Lambda(A | B) = I_Lambda(A, B) - I_Lambda(B).
Interpretation: I_Lambda(A|B) counts how many additional Lambda-ary choices
A has, given B. If B determines A up to Lambda^{o(1)} labels,
I_Lambda(A|B) = o(1). If A is independent of B, I_Lambda(A|B) = I_Lambda(A).
The two facts the paper uses constantly:
- Chain rule. I_Lambda(A,B | C) = I_Lambda(A|C) + I_Lambda(B | A,C).
- Conditioning reduces information. I_Lambda(A|B) <= I_Lambda(A).
The "profiles". The transverse saving proof uses a "countersequence": a
sequence of configurations parameterized by a large scale L, in which the
quantities of interest take values near their limiting extremal profiles.
For each exponent s in [0,1], the resolution Lambda^{-s} defines a level of
scale, and:
- F(s) is the limiting (base-L) information of the sampled spatial cell
X_s at that resolution.
- G_i(s) is the same information, but conditioned on one of the n+1
incident tube directions T_i.
- G_i'(s) is the same, but conditioned on an alternative tube i'.
These functions are nondecreasing in s and Lipschitz with constants that are
dimension-dependent but fixed. The "profile increments"
F(s+t) - F(s), G_i(s+t) - G_i(s)
are the marginal bits of position at the finer scale, and by the chain rule
they are exactly the information that the finer position provides about the
coarse position. Controlling them is the whole content of the rigidity part
of the transverse saving (Chapters 35-37).
The reason the paper uses base-L information (not natural log): the constants
become clean. If a quantity has L^{C+o(1)} labels, its base-L information is
C + o(1), and no epsilon bookkeeping is needed.
================================================================================
PART IV: THE PROOF
================================================================================
One chapter per paper section, plus three extra chapters for the three pieces
of the hardest section (paper Sec 7), plus one extra chapter for the two
"smoothing" pieces of the frame estimate. Each chapter has the same shape:
what the step proves, the big idea, the precise statement, and where it is
used. We are now in the main body of the proof.
24. SECTION 2: THE PACKET TOOLBOX, AND THE NEAR-CAUCHY THEOREM
--------------------------------------------------------------------------------
The paper's Section 2 is where the packet-level estimates live, and every
later section depends on it. We have already introduced compact arrays
(Chapter 22) and the central exponents; here is the cast of characters.
- Fixed dimensions. Write
m = k+1, p_k = 2(k+1)/k = 2m/k, q_k = 2(k+2)/k = 2(m+1)/k.
The exponent p_k is the linear Strichartz end for the Schrodinger
equation in k dimensions. The exponent q_k is the endpoint of the
Bourgain-Demeter decoupling inequality in k dimensions (Chapter 19).
They differ by a constant in the exponent.
- A packet sum V(x,t) = sum_T s_T f_T(x,t), where T ranges over a finite
index set of N-scale packets (fixed profile, fixed frequency box). The
envelope of V at a point z is
B(z)^2 = sum_T s_T^2 w_T(z),
where w_T(z) = (1 + d_T(z))^{-P} with P a fixed large envelope exponent
and d_T the native distance to the packet's cell (Chapter 22).
- The "high set" of V at level lambda is the set of unit cubes Q where
|V| is >= lambda * (its envelope) on some point of Q. The count of such
cubes is what the packet estimates control.
The claims in Section 2 that we quote as black boxes:
- The multilinear Kakeya estimate (Chapter 17, Proposition p:kakeya).
- The robust hyperplane projection theorem (Chapter 21, Proposition
p:projection, due to He 2020).
- The buffered refined decoupling lemma (Chapter 18, Lemma
p:buffered-refined, proved in the paper).
- The weighted packet estimates and their hierarchical versions, which
we state in the next chapter (Propositions p:weighted and p:hierarchy).
- The local linear and local multilinear estimates (Chapter 25).
- The near-Cauchy theorem (below).
THEOREM (near-Cauchy; paper Theorem p:near-cauchy). Fix k >= 3, fix the
frequency box and packet profile. There is a finite p_* such that for every
p > p_*, there is a constant C_p with: for every N-scale packet sum V on a
ball B_R of radius R, and every set Y of unit cubes on which V equals its
envelope at some point up to a fixed constant factor,
| Y | <= C_p R^p ( sum_T s_T^2 ).
The infimum p_* is at most q_k = 2(k+2)/k.
So for every eps > 0 the exponent p = q_k + eps works. In words: even
though a packet sum can pile up coherently on many unit cubes, the number
of high cubes is at most R^{q_k + eps} times the square of the input L^2
size. Non-coherent Cauchy would give |Y| ~ sum s_T^2; coherent packet sums
can be much worse than Cauchy, but the worsening is bounded by a fixed power
of R.
The name. If V is a single packet, |Y| <= S trivially. The point of the
theorem is that even adding many packets coherently, the count of high unit
cubes is at most a fixed power of R bigger than the obvious one. In this
sense the theorem says "packet sums are close to Cauchy in a point count".
It is the point count that lets the paper run induction on packet scales.
Sketch of the proof. We indicate the strategy; the details are a long
induction in the paper's Section 2.
1. By the "polynomial start" lemma (paper p:polynomial-start), p_* is
finite: some finite power of R suffices trivially.
2. Suppose p_* > q_k. Then there is a sequence of configurations with
ratio |Y| / S approaching R^{p_* - o(1)}; this is a "counterexample" to
any exponent below p_*.
3. From the counterexamples, choose coarse and fine scales K, N as in
the proof (Chapter 25) and assemble the fine packets into coarse ones.
Use the weighted packet estimates to control how the fine amplitudes
relate to the coarse amplitudes.
4. At a typical point of Y, get a lower bound on the number of incident
coarse packets (Cauchy-Schwarz). Compare with the point count for
coarse packets and get a contradiction unless p_* <= q_k.
5. The precise constants follow from the Cauchy estimate (Chapter 25) and
the near-Cauchy's use of the "occupancy" p:occupancy-upper.
The reason the near-Cauchy theorem is important: it is the point-count input
the sparse estimate (Chapter 37) uses to control how many high cubes a
packet sum can have at a given scale. Without it, the sparse iteration
cannot close.
25. SECTION 2: WEIGHTED PACKET ESTIMATES AND LOCAL CAUCHY
--------------------------------------------------------------------------------
The last pieces of Section 2 that the paper uses throughout are a family of
weighted L^p estimates for packet sums. They are the "hard analysis" of the
paper, and they quantify how the envelope B controls the full sum V.
Weighted packet estimate (paper p:weighted). Fix 2 < p <= q_k, weights
b_T >= 0 with the convention that b_T = 0 forces s_T = 0, and put
B_b(z)^2 = sum_T b_T^2 w_T(z). Then for every eps > 0,
INT B_b(z)^{2-p} | sum_T f_T(z) |^p d z
<= C_eps N^eps N^{k+2} sum_T s_T^p b_T^{2-p}.
Interpretation: if we "pay" the packet sum by its weighted L^p mass with a
positive-definite weight B_b^{2-p}, the cost is bounded by the sum of
individual packet contributions. When p = 2 this is trivially an L^2
identity. When p > 2, the extra power of p is paid by N^eps and the
frequency-localization volume N^{k+2}.
Hierarchical form (paper p:hierarchy). Take nested scales K < K' with bounded
ratio. Assemble K'-scale packets into K-scale packets. If a_T is the
amplitude of a K-scale parent packet and b_T is its weight, and similarly
for K'-scale packets a_{T'} and b_{T'}, then for every eps > 0,
sum_T K^{k+2} a_T^p b_T^{2-p}
<= C_eps (K'/K)^eps
sum_{T'} (K')^{k+2} a_{T'}^p b_{T'}^{2-p}.
The weighted p-energy is preserved when coarsening, up to (K'/K)^eps.
This is the "energy transport" estimate. It says you can move up the
packet hierarchy without losing.
Two local estimates. These are the "Cauchy at a point" bounds:
Local linear (paper p:local-linear). Suppose V(x,t) = sum_T f_T(x,t) is
a sum of N-scale packets, with all frequencies in a fixed bounded box.
Let B(z) be the envelope. Then for every p > p_k = 2m/k, and every ball D
of radius N centered at a fixed point, and every witness z in D,
|| V ||_{L^p(D)} <= C_p N^{k/2} B(z).
So the pointwise envelope at one point controls the L^p norm on the whole
packet neighborhood. This is the "local Cauchy" bound.
Local multilinear (paper p:local-multilinear). With m = k+1 fields
V_1, ..., V_m transverse with determinant >= delta > 0, and with any
choice of witness points z_j in D,
|| prod_{j=1}^m |V_j|^{1/m} ||_{L^p(D)}
<= C_p delta^{-C_p} N^{k/2} prod_{j=1}^m B_j(z_j)^{1/m}.
The transversality costs a polynomial factor in delta^{-1}, but the
bound is otherwise pointwise in the envelopes.
These local estimates are used in the sparse estimate (Chapter 37) to
bound how much mass can pile up on a "broad" collection of packets. They
are also used in the frame estimate (Chapter 26) exactly once, at a key
step. The powers N^{k/2} are the natural "packet size" factor.
The reader should think of these estimates as the "Cauchy-Schwarz at every
scale" that makes the packet algebra close. The buffered refined decoupling
of Section 2 is what supplies the missing N^eps absorption to reach the
exponents p = p_k and q_k without loss. We quote both as black boxes and
use them as they are used in the paper.
26. SECTION 3: THE TRANSVERSE FRAME ESTIMATE (STATEMENT AND IDEA)
--------------------------------------------------------------------------------
This is the deepest single estimate of the paper. It is also the estimate
that most visibly distinguishes the proof from its predecessors, and it is
what makes the endpoint close.
Physical idea. Suppose we have m = k+1 transverse families of wave packets,
all at the same scale h, and suppose at many points of spacetime each family
has a "big" packet through it (amplitude >= A_i, envelope <= 1). The claim
is: this cannot happen too often. The number of such points is bounded by a
factor reciprocal to the "geometric mean amplitude" of the fields.
In symbols, the theorem is:
THEOREM (transverse frame; paper Theorem f:frame). Fix k >= 2, c > 0, and
fixed packet bounds C_0. There is a finite integer P_0 such that for every
fixed P >= P_0, there is a finite integer J_0 and a constant C (all
depending only on k, c, C_0, P and the fixed conventions) with the
following property.
Let h >= 2, let m = k+1, and let V_1, ..., V_m be scale-h packet sums on a
common time interval of length at most C_0 h^2. Suppose their normalized
compact profiles have derivatives through order J_0 bounded by C_0, and let
Xi_i denote the frequency-label sets of the i-th family. Assume:
(i) bounded phase-space counting of the packet family;
(ii) transverse determinant: for all xi_i in Xi_i (i = 1, ..., m),
| det( v_vec(xi_1), ..., v_vec(xi_m) ) | >= delta > 0,
with v_vec(xi) = (2 xi, 1) in R^{k+1} (Chapter 13);
(iii) envelope exponent P (so w_T(z) = (1 + d_T(z))^{-P}).
Let Y be a set of distinct unit cubes in spacetime, and for every Q in Y
and every i, let z_{i,Q} be a witness point in Q with
| V_i(z_{i,Q}) | >= A_i >= h^c,
B_i(z_{i,Q})^2 = sum_T s_{i,T}^2 w_T(z_{i,Q}) <= 1.
(Here the packets s_{i,T} are the packe-ts of the i-th family.)
Define
H_i = A_i^{2/k}, H_g = ( prod_{i=1}^m H_i )^{1/m},
D_i = h^{k+2} sum_T s_{i,T}^2.
Then
|Y| <= C delta^{-C} H_g^{-2} sum_{i=1}^m D_i / A_i^2.
What this says. The number of high points |Y| is bounded by a "gain" factor
H_g^{-2}, multiplied by the total energy of the families (sum D_i / A_i^2).
The bigger the amplitudes A_i, the smaller H_g (since H_g = geometric mean
of A_i^{2/k}), and hence the smaller the bound. If some fields are small,
H_g is small; if all fields are big, H_g is big, but the transverse
determinant assumption still makes the fields "spread out" and there is a
contradiction unless the number of points is small.
Why k+1 fields, not fewer? The number k+1 is forced by the geometry of the
paraboloid in R^{k+1}. Fewer than k+1 transverse fields can concentrate;
k+1 is the minimum needed for the "simplex volume" to be positive. This is
why the frame estimate is in dimension k (or, in the geometric induction of
Section 7, in dimension k = n-1, with m = n).
Where the frame estimate is used. Chapters 39-41 will use the frame estimate
in dimension k = n-1 to control a "plate case" of the geometric induction.
That is, the frame estimate is the engine that makes the codimension-one
reduction work: packets that are concentrated near a codimension-one
"plate" are, after flattening along the plate, a k = n-1 dimensional frame
configuration.
Strategy of proof. We cannot reproduce the full proof; it is 25 pages of
the paper. Instead, we sketch the four logical steps. They are:
A. REDUCTION TO EXACT PACKETS AND h-BALANCE (Chapter 27). One may assume
the packets are exact, smooth solutions of the Schrodinger equation,
with fixed derivative orders. From the failed estimate, one derives
that all H_i are within h^{o(1)} of H_g and all |Y| H_i^{k+2} / D_i
are within h^{o(1)} of 1. This "h-balance" is a strong equality.
B. AMPLITUDE-RATIO PROFILES AND BINARY RIGIDITY (Chapter 28). Decompose
each field down to a fine packet scale. Define a real-valued "amplitude
ratio profile" f_i(s) along the packet hierarchy, and prove a binary
rigidity: f_i' is either 0 or -k/2 at almost every slope. This is the
most technical part.
C. ANGULAR PRUNING (Chapter 29). Use the codimension-two strip saving
(Chapter 32) to prove that for a specific parameter l (a scale in the
hierarchy), the direction law of the packets satisfies
nonconcentration in every affine hyperplane. Then use He's projection
theorem (Chapter 21) to extract a direction that survives.
D. INCIDENCE CONTRADICTION (Chapter 30). Use the strong nonconcentration
from step C to contradict the h-balance of step A. This completes the
proof.
The four steps are precisely where each input is used, and each one is a
small paper of its own. We will go through them at the level of "which
estimate is invoked and why".
27. SECTION 3: REDUCTION TO EXACT PACKETS, AND THE h-BALANCE
--------------------------------------------------------------------------------
Step A of the frame estimate proof has two pieces.
First piece: exact-packet reduction (paper Lemma f:exact). The frame estimate
is stated for "compact arrays of scale h". The reduction says: it suffices to
prove the estimate for exact paraboloid extensions. That is, replace each
packet f_T by the exact solution S(t)(of the appropriate native profile).
The difference between the compact packet and its exact counterpart is at
most a fixed number of derivatives times a rapid decay factor, which the
envelope power P absorbs. The reduction thus loses only constants depending
on P and the fixed profile order.
Second piece: suppose the estimate fails. Then there are counterexamples
with |Y| much larger than the asserted bound. By picking a subsequence
that violates every fixed constant and every fixed inverse power of delta,
one obtains a sequence with h -> infinity and delta = h^{-o(1)}. The
"absolute Cauchy" inequality gives H_i <= C h (each amplitude is bounded by
the total energy squared, and total energy is bounded by the input squared).
Then Proposition p:weighted (Chapter 25) gives, for each i,
|Y| <= h^{o(1)} D_i H_i^{-k-2}.
Combining this with the failure of the frame estimate gives the h-balance:
(HB1) H_i / H_g = h^{o(1)} for every i, (all fields equal)
(HB2) |Y| H_i^{k+2} / D_i = h^{o(1)} for every i, (all balanced)
Read the h-balance as: any counterexample has all fields with essentially the
same H_i and each contributing equally to the point count. This is a strong
constraint: no field can be much bigger or smaller than the others. The rest
of the proof of the frame estimate shows that the h-balance is inconsistent
with the amplitude-ratio structure of the fields.
The h-balance is derived purely from the weighted packet estimates, so it
inherits the endpoint strength: the constants are exact, no eps loss.
The reader can think of the h-balance as the "invariant" the rest of the
proof must destroy. Each subsequent step (profile rigidity, angular
pruning, incidence) develops a stronger consequence of transversality and
shows that at least one of (HB1) and (HB2) fails in a specific way.
28. SECTION 3: AMPLITUDE-RATIO PROFILES, AND THEIR ONE-DIMENSIONAL STRUCTURE
--------------------------------------------------------------------------------
Step B of the frame estimate proof is the extraction of a one-dimensional
"amplitude-ratio profile" from each field, and the proof that this profile
has rigid structure.
Setup. Fix a field V and its scale-h packets s_T. For a relative scale
s in [0,1], consider the packets of scale h^s (a coarser scale inside the
original scale-h packet decomposition). Write kappa(s) for that width
(kappa(0) = 1 is the top scale, kappa(1) = h is the finest). For each
s, define:
- r_s = the "local amplitude" of the field on the relevant packet at
that scale, after removing the natural factor kappa(s)^{-k/2} coming
from packet normalization.
- The amplitude-ratio profile
f(s) = log_Lambda ( r_s / (A * kappa(s)^{-k/2}) )
where A is the original witness amplitude at scale h and Lambda is the
fixed base (Lambda = h in the "power case", Lambda = S in the
"subpower case" -- see below).
The profile f(s) is real-valued and can be positive or negative. It is
defined by the packet decomposition, and its values are recorded at every
dyadic scale.
The two cases. Exactly one of the following is proved to hold (paper
forced_profiles.tex has a picture):
POWER CASE: Lambda = h, kappa(s) = h^s.
SUBPOWER CASE: Lambda = S, kappa(s) = H_i S^s,
for a fixed parameter S with S = h^{o(1)}, and for a parameter H_i that
encodes the field's amplitude at scale 1. In the power case, the scale
hierarchy is dyadic in h. In the subpower case, the packet widths grow
more slowly, with exponent S^{o(1)}. The dichotomy is forced by the
constraints on the profiles below; there is nothing to choose.
Basic properties. From the definitions and the packet algebra (Chapter 25):
- Lipschitz: f(s) - f(s') <= (n+1)|s - s'| + o(|s - s'|) for
s' < s. The constant n+1 comes from the "bounding" of profiles by the
ambient dimension.
- Endpoint bounds: f(1) >= 0 and f(0) <= k e_i / 2 (in the subpower
case), with e_i the "cutoff" parameter. The profiles are zero at
s = 0 in the power case.
- Limits. In the power case, f(s) -> 0 as s -> 0 (the profile has a
"left limit" 0 at the top of the hierarchy). In the subpower case, the
left limit is -ks/2 as s -> -infinity for an appropriate extension.
Classification of points. At a differentiability point s of f (which exists
almost everywhere by the Lipschitz property):
- s is a FULL-SLOPE point if f'(s) = -k/2.
- s is a ZERO-SLOPE point if f'(s) = 0.
The "log-derivative of M_s" is zero at full-slope points and k+2 at
zero-slope points (where M_s is the local occupancy count associated to the
profile at scale s; it measures the number of high unit cubes the packets
of this profile contribute around a point).
THEOREM (midpoint closure; paper Lemma f:midpoint). For almost every pair
of full-slope points l < a in one field, the midpoint (l+a)/2 is also a
full-slope point. As a consequence, the full-slope set of each field is an
interval (possibly empty, possibly the whole [0,1]) modulo null sets.
The proof is a clever Lagrangian-interpolation argument: if the midpoint
were a zero-slope point, one could build a small chain of scale changes
that would force a "free" gain of information, contradicting the chain
rule for the sampled profiles. The reader who wants the details can look
at the paper's Section 3, "The location of the full-slope interval".
THEOREM (left limit; paper Lemma f:left-limit). In the subpower case, the
leftmost differentiability point of any full-slope interval is not a
"boundary" point of a zero-slope region: the profile must actually achieve
slope -k/2 at the left endpoint of any full-slope region. In particular,
subpower profiles with 0 < e_i < 1 cannot have a full-slope region that
"starts later" than s = 0; they must start at s = 0.
The proof is a "bottom-up" analysis: from the fact that the profile is
bounded by 0 at s = 0 in the subpower case, and by the Lipschitz condition,
the slope at the left endpoint of any negative-slope region has a forced
value related to the derivative of M_s.
What these two lemmas give. The profile f is now known to be:
- a Lipschitz function that is zero to the left of any full-slope region;
- a function whose derivative is -k/2 on an interval [s_0, s_1]
(possibly with s_0 = 0 and/or s_1 = 1), and 0 outside;
- in the power case, f is either identically zero or is of the form
f(s) = (k/2)(s_H - s)_+ for a "transition scale" s_H in (0, 1] (the
forced profile in forced_profiles.tex).
The one-dimensional structure is total: an amplitude profile is basically
a "ramp" from 0 down to some negative value, with a single slope limit
-k/2 and two flat ends. This is very restrictive, and it is exactly what
the incidence contradiction in Chapter 30 exploits.
We have deferred the proof that "the profile is either entirely zero-slope,
or has a full-slope interval." That is the content of the binary rigidity
theorem, which uses the angular test -- the next chapter.
29. SECTION 3: ANGULAR PRUNING AND THE BINARY RIGIDITY
--------------------------------------------------------------------------------
Step C of the frame estimate proof is where the codimension-two strip
saving enters, and where the profile slope is forced to be exactly -k/2.
The angular test (paper Lemma f:angular). Fix a field, a scale l in the
hierarchy, and a finite mesh of 0 < v <= u with l+Cu < e_i. Suppose that
on this mesh, the profile drops by at least xv:
f(l) - f(l+v) >= x v, for some fixed x > 0.
Then, for a small eta > 0 depending only on x and k, one may restrict the
hierarchy so that:
- At typical (Z, tau_l), the new conditional direction law assigns mass
at most Lambda^{-eta v + o(1)} to every affine hyperplane strip of
width Lambda^{-v} in the native frequency coordinates of tau_l.
- At a much smaller precision v_0 << v, the sampled active l-cells also
have "broad witnesses" for their original pre-modification fields:
k+1 cap groups whose directions are transverse throughout their
supports (determinant >= Lambda^{-C v_0}) and each of magnitude at
least Lambda^{-C v_0} r_l b_{tau_l}(Z) at a point within native
distance Lambda^{C v_0} of the cell.
The two assertions are not independent: the first is the "quantitative
angular information" (the direction law avoids hyperplanes), the second is
a "pre-modification certificate" that lets later arguments use the original
packets without the pruning.
Sketch of the proof. Fix a cap, its parent box, and a mesh value v. Do the
following, from coarsest to finest v:
(i) Choose a macro box of side Lambda^{3u}. Since all descendant widths
are <= Lambda^{2u + o(1)}, the ratios of descendant weights to the
current weight are bounded above and below by Lambda^{o(1)} on such
a box. So the direction law is "almost invariant" across the macro
box.
(ii) For each mesh v, greedily remove all (l+v)-caps intersecting an
enlarged hyperplane neighborhood of width Lambda^{-2 eta v} whose
reference measure is at least Lambda^{-2 eta v}. The cost:
a hyperplane neighborhood of width Lambda^{-2 eta v} can hold at
most Lambda^{2 eta v} caps, and this is a fixed saving.
(iii) Show that the omitted hyperplane contribution is cheap at level l.
This uses:
- the profile assumption (that the profile drops at rate x);
- the recorded ratio class of f:selection;
- Hölder with the measure b_T^2 |T|;
- the "plate-supported" version of Proposition p:weighted at
exponent q' = 2(k+1)/(k-1) > q_k;
and obtains a bound of the form
( omitted planes contribution ) <= Lambda^{-c v}.
The key inequality is a negative power budget: the omitted planes
contribute a fixed inverse power of Lambda to the total, and the
fixed positive exponent c is chosen after the other parameters.
The codimension-two strip saving (Chapter 32 of this companion) is used
implicitly here: it is what makes the "plate-supported" packet estimate
sharp enough to give the negative power. Without it, the cost of a
hyperplane contribution would be too big by a power.
Binary rigidity (paper Lemma f:binary). Under the same hypotheses, the
profile f has only two possible slopes: at almost every differentiability
point s, either f'(s) = 0 or f'(s) = -k/2.
The proof structure is:
(a) The angular test, applied at every mesh value of v, shows that the
"profile slope" of any scaling interval [s, s+v] is at least the
"gain" x = -f'(s) as far as the direction nonconcentration is
concerned. Since nonconcentration has a fixed upper bound (the
number of lambda-caps in a hyperplane strip is at most Lambda^{something}),
a positive slope x > 0 forces the profile increment to be
dominated by the "full" increment k v / 2 (the maximal possible
drop of the local amplitude by the packet normalization).
(b) Conversely, midpoint closure (Chapter 28) forces full-slope sets to
be intervals: once a "partial" slope 0 < x < k/2 holds on an
interval, one can find full-slope points on either side that
contradict the intersection structure.
(c) Combining, the slope is either 0 (well inside the "no-drop" region)
or k/2 (well inside the "full-drop" region). Intermediate slopes are
impossible.
We compress the proof substantially. The full details take the paper about
8 pages, and the reader who wants to follow them should read the paper's
Section 3, "Angular pruning inside one profile block".
What this gives. Now every profile f is a "two-slope" function: zero to the
left, then a linear ramp of slope -k/2, then zero again. The transition
points s_0 < s_1 are the only interesting scale parameters. This is much
stronger than the Lipschitz-plus-midpoint statement of Chapter 28.
30. SECTION 3: THE INCIDENCE CONTRADICTION, AND THE END OF THE FRAME ESTIMATE
--------------------------------------------------------------------------------
Step D of the frame estimate proof. We now have, from Chapters 27-29:
- h-balance: H_i / H_g = h^{o(1)} and |Y| H_i^{k+2}/D_i = h^{o(1)}
(all fields "equal and balanced").
- profiles: each field f_i is a two-slope function with slope 0 or -k/2,
a full-slope interval [s_0, s_1], and endpoints 0 / -(k/2)(1-s_1).
- angular pruning: for each field at the appropriate scale (in the
transition region of the profile), the direction law is strongly
non-concentrated in every affine hyperplane of width Lambda^{-v}.
- broad witnesses: at a much finer precision v_0, the l-cells have
k+1 transverse cap groups with comparable amplitude.
Contradiction. The end of the frame proof is a direct counting argument
using the angular pruning and the properties of the profile at the
"incidence" level.
The key step is to extract, from the surviving masks (the ones left after
the angular pruning of Chapter 29), a set of original scale-h packets with
two properties:
(a) Each surviving mask has an "original amplitude" lower bound
a_T^{orig} >= c S H^{-k/2},
where S is a scale parameter related to the transition point of the
field profile, and H is the local amplitude bound. The constant c
is universal.
(b) The total original integrated square energy of all surviving masks
is at most
sum_{T} |a_T^{orig}|^2 <= C H^{k+2} |Y| Lambda^{-o(1)},
by the "original-energy" bound at the top of the frame estimate.
Combining (a) and (b), each surviving mask contributes at most
(1/|Y|) S^{-2} to the total, so summing over all masks and using the
original-amplitude lower bound gives
D_i = sum_T |a_T|^2 >= c |Y| S^2 H_i^{k+2} Lambda^{-o(1)}.
On the other hand, the h-balance (from Chapter 27) says
|Y| H_i^{k+2} / D_i = h^{o(1)},
i.e., D_i <= |Y| H_i^{k+2} h^{o(1)}. So we have the two inequalities:
D_i >= c |Y| S^2 H_i^{k+2} Lambda^{-o(1)}
D_i <= |Y| H_i^{k+2} h^{o(1)}.
Together:
c S^2 Lambda^{-o(1)} <= h^{o(1)}.
In the POWER case, Lambda = h and S = h^{s_0} for the transition scale
s_0 in (0, 1]. Since s_0 > 0, S^2 = h^{2 s_0}, which grows like h^{c_0}
for a constant c_0 > 0; the inequality fails for h sufficiently large.
In the SUBPOWER case, Lambda = S and S^2 = Lambda^2. The inequality becomes
Lambda^2 Lambda^{-o(1)} <= Lambda^{o(1)}, which fails for any base
Lambda > 1 as h grows (with the subpower relation between h and S).
This contradiction proves the frame estimate.
The whole frame estimate is now proved except for the technical estimates
(the fixed constants, the profile derivative orders, the phase-space
counting, the sharp derivative bounds) that enter every step. Those are
finite and the paper handles them in the "tooling" parts of Section 3.
Where does the frame estimate go next? Into Section 7 (Chapters 40-44),
where it is applied in dimension k = n-1 to control the "plate case" of
the geometric induction. The frame estimate, plus the codimension-two
strip saving, is Section 7's engine.
31. SECTION 4: FRACTAL NORMS AND THE DU-ZHANG INPUT
--------------------------------------------------------------------------------
We now begin the "fractal" machinery that converts a sparse measure into a
square function estimate. It is the descending step that gives Sections 6
and 7 the estimates they need. The main input is Du-Zhang 2019's fractal
estimate; the paper's contribution in Section 4 is a weighted version and
a strip saving.
Exponents. Define
beta = 2 / (n+1),
e = n^2 / (n+1),
gamma_n = 2n / ( (n+1)(n+2) ).
These are the canonical exponents of the fractal machinery. The exponent
beta is the homogeneity of the fractal norm in the measure; e is the
critical smoothness in the Du-Zhang paper; gamma_n is the "gain per
density" exponent in the weighted fractal estimate.
The Du-Zhang input in squared form (paper Theorem d:dz-input). Let Q be a
collection of distinct unit cubes in a ball of radius S >= 1. Suppose their
counting function in every ball of radius r >= 1 is bounded by Gamma r^n,
and that their number in any ball of radius sqrt(S) is at most lambda_0.
Then for every eps > 0,
sum_{q in Q} sup_q | E g |^2
<= C_eps S^eps
Gamma^{4/((n+1)(n+2))}
lambda_0^{gamma_n}
S^{gamma_n}
|| g ||_2^2.
Interpretation: if the set Q is "n-dimensional" (counting bounded by
Gamma r^n) and "not too concentrated" (bounded lambda_0 cubes per
sqrt(S)-ball), then the L^2 mass of the wave on Q is bounded by its
L^2 size, with a small overcount that depends on Gamma and lambda_0. This
is the refined Strichartz estimate.
Squares and occupancy. The exponent gamma_n is exactly the gain from
"fewer cubes per sqrt(S)-ball". If lambda_0 were 1 (one cube per
sqrt(S)-ball), the bound would be much better; the gamma_n factor captures
this quantitatively.
Weighted version (paper Proposition d:fractal). Suppose mu is supported in
a ball of radius CS, satisfies the ball growth inequality mu(B(z, r)) <= m
r^n, and additionally
mu(B(z, sqrt(S))) <= d m S^{n/2}, S^{-A_0} <= d <= 1,
for a fixed A_0. Then
sum_q mu(q)^{beta} sup_q | E g |^2
<= C_{eps, A_0} m^beta
S^{n/(n+1) + eps} d^{gamma_n} || g ||_2^2.
With d = 1 the estimate reduces to the unweighted Du-Zhang bound. The
factor d^{gamma_n} is the "occupancy gain": the smaller d is (i.e., the
more evenly mu is distributed at scale sqrt(S)), the better the bound is.
Setting d = 1 recovers no gain; setting d small recovers a fixed power
since gamma_n > 0.
Why this exponent pattern. The exponent n/(n+1) is the "sharp" fractal
exponent that the maximal estimate has to beat; the exponent gamma_n =
2n/((n+1)(n+2)) is the gain from lambda_0; the ratio Gamma / lambda_0 is
what the counting hypotheses encode. The formula is not arbitrary: it is
what the Du-Zhang refined Strichartz estimate produces, and it cannot be
improved (a known sharpness example).
Instantiation for compact arrays (paper Corollary d:power-loss). If U is a
compact array of scale L, mu satisfies the ball growth inequality on a
spacetime box of side O(L^2), then for every eps > 0 and sufficiently many
fixed profile derivatives,
sum_q mu(q)^{beta} sup_q | U |^2
<= C_eps m^beta L^{2n/(n+1) + eps} E(U).
This is the "compact array" version of the Du-Zhang estimate. It differs
from the unweighted version by L^{eps} (an unavoidable epsilon loss in the
compact array form) and no lambda_0 factor; the lambda_0 is encoded in
the compact array structure. The exponent 2n/(n+1) is the "endpoint of the
fractal estimate for compact arrays".
The reader may notice: the fractal exponent here (2n/(n+1)) is the same as
2 - e (since e = n^2/(n+1) and 2n/(n+1) = 2 - n/(n+1); and n/(n+1) =
(2n)/(2(n+1))... let me not belabor). It is the fractal exponent that
comes out of Du-Zhang's theorem and is what the paper calls "sharp at
the endpoint".
Where these are used. The compact-array estimate (Corollary d:power-loss)
is used in Section 6 (Chapter 38) to close the "terminal" step of the
sparse estimate: when the broad/narrow iteration reaches its final scale,
the remaining packet exceeds the Du-Zhang bound by a fixed power, and the
sparse estimate is closed. The weighted version (Proposition d:fractal)
is used in Section 4's strip saving (Chapter 33) and in Section 7's
geometric induction (Chapters 40-44).
32. SECTION 4: COMPACT LAYERS AND EXACT LAYERS
--------------------------------------------------------------------------------
The Du-Zhang fractal estimate (Chapter 31) is a statement about
Eg -- the exact extension of a function g on R^n. The paper's packet
framework, however, works with compact arrays U (Chapter 22). We need a
bridge between the two: given an exact extension Eg, decompose it into
compact arrays; given a compact array, realize it as a sum of exact layers.
The bridge is the following.
THEOREM (compact layers; paper Lemma e:compact-layers). Let g be a
function on R^n with bounded frequency support and let L >= 1. On the
time range |T| <= C L^2, the extension Eg(x, T) can be written as a sum
Eg = sum_{q in Z^n} U_q
of compact arrays U_q of scale L. For any prescribed finite profile
derivative order J and any A > 0, the arrays can be chosen with uniformly
bounded C^J native profiles, and their coefficient energies satisfy
E_q <= C_{A,J} <q>^{-2A} ||g||_2^2,
where <q> = (1 + |q|^2)^{1/2}.
Key consequence: the layer energies are summable, sum_q E_q <= C ||g||_2^2
(up to a constant depending on A, J). Moreover, each fixed q has uniformly
bounded phase-space multiplicity, so the packet-algebra bookkeeping can be
applied to the whole sum.
Interpretation. The exact extension is a "global" object: at each
spacetime point, it is determined by all of g. To fit it inside the
compact-array box, we have to allow the sum over q. The q = 0 layer
captures most of the mass (where g is closest to the spacetime box), and
higher q layers account for far-away contributions. The decay <q>^{-2A}
pays for those.
Why the decomposition exists. In native variables (X, T) with the
L-scaling from Chapter 14, the exact field Eg is a smooth function of a
vector (eta, zeta) in Fourier space. Split eta = q + theta with q in Z^n
and theta in the unit box; split zeta = u - |eta|^2 with u in R. The
change of variables is exact, and the Fourier inversion assembles U_q as a
sum over q of compact-array terms:
U_q(X,T) = sum_alpha c_{q, alpha} e^{i(X . omega_q - T |omega_q|^2)}
phi( X - y_{q,alpha} - 2 T omega_q ).
The weights in the Fourier split are chosen to be Schwartz-decaying, so
each E_q gets a decay factor <q>^{-2A}.
Reverse direction, for local arguments. The paper also uses the reverse:
a compact array U = sum_alpha c_alpha phi_alpha may be treated as a sum of
exact layers, each supported in a small neighborhood, so that the fractal
machinery of Chapter 31 applies to each layer separately. This is
essentially the same decomposition run in reverse, with the roles of U and
Eg interchanged.
Both directions are used in Section 4 and thereafter. The reason is
simple: fractal estimates want exact extensions, packet estimates want
compact arrays, and the layer decomposition lets us pass freely between
the two at summable cost.
33. SECTION 4: THE CODIMENSION-TWO STRIP SAVING
--------------------------------------------------------------------------------
This is the fourth of the four new ingredients introduced in Chapter 11.
It is what lets the geometric induction reduce dimension from n to n-1,
by exploiting the case where packet frequencies concentrate near a
codimension-two affine plane.
THEOREM (codimension-two strip saving; paper Proposition d:strip). Fix
0 < alpha < 1/8. Suppose g has bounded frequency support contained in an
O(S^{-alpha})-neighborhood of an affine (n-2)-plane in R^n, and let Q be a
collection of unit cubes in a ball of radius S >= 1, with
#(Q cap B(z, r)) <= m_0 r^n for every ball B(z, r), r >= 1.
Fix any b with beta < b < 2/n (recall beta = 2/(n+1)). Then there is
c_0 > 0, depending only on the fixed dimension and b (and NOT on alpha),
such that for every eps > 0,
sum_{q in Q} sup_q | E g |^2
<= C_{alpha, eps} m_0^b S^{ n/(n+1) - c_0 alpha + eps }
|| g ||_2^2.
Compare with the unweighted Du-Zhang bound (Chapter 31), which would
be m_0^beta S^{n/(n+1)+eps} ||g||_2^2. The strip saving has:
- m_0^b replacing m_0^beta (b > beta, so a more generous measure budget);
- S^{-c_0 alpha} replacing 1 in the exponent (a genuine power saving).
The saving is the point: as long as m_0 is not too large (say m_0 ~
S^{o(1)} or a fixed power small enough), the S^{-c_0 alpha} gain is a net
win. The m_0^b allows for a slightly larger measure parameter on the left
side, which is the price.
Idea of the proof. The (n-2)-plane is a codimension-two affine subspace.
Its orthogonal complement has dimension two. In frequency space, the
support of g is a thin slab of width H = S^alpha around this plane.
Step 1. Decouple the frequencies into "transverse caps" of size K^{-1}
for a dyadic parameter K, chosen later. At the finest relevant level,
the transverse frequency support has size H = S^alpha.
Step 2. On each K^2-cube, apply the strip decoupling estimate
(d:strip-decoupling). This is a refined-decoupling estimate tailored to
strip-supported fields; it bounds the L^P norm on a K^2-cube by an l^2
sum over the caps, at the cost of K^eps. Its content is a version of
Bourgain-Demeter for strip-supported fields, in the spirit of refined
Strichartz.
Step 3. Transport. Under parabolic rescaling (Chapter 14), the strip
condition is preserved (Chapter 34 of this companion has the transport
lemma d:transport). Iterate.
Step 4. Close. Each rescaling step replaces the exponent by beta - b,
which is strictly negative. The number of rescaling steps is J, and each
step also incurs the K^{C_*} of the terminal estimate; choosing J large
makes the net factor K^{-(b-beta)/2}, hence a power saving S^{-c_0 alpha}
in the original scale.
The precise exponent arithmetic is this. Suppose we rescale by H = K^J
(the strip width) in J steps. The transport factors per step are:
- Jacobian of parabolic rescaling: K^{2 - (n+2) b} (from the map Phi)
- measure parameter inflation: K^{(n+1) b} (measure growth)
- terminal estimate at scale T = S/H^2: K^{-2n/(n+1)}
Multiplying:
K^{2 - (n+2)b} * K^{(n+1)b} * K^{-2n/(n+1)}
= K^{2 - b - 2n/(n+1)}
= K^{beta - b},
where in the last equality we used beta = 2/(n+1). Since b > beta, this
factor is strictly less than 1. Iterating J times and using the K^{C_*}
of the terminal estimate gives K^{-(b-beta)/2} when J is large enough.
Since K^J = H = S^alpha, this is S^{-(b-beta)alpha/2}. Setting
c_0 = (b-beta)/2 proves the theorem.
One subtlety: the K^{C_*} from the terminal estimate must be beaten by J
large enough. This is possible because C_* is independent of J (it comes
from the fixed dimensional exponents of the fractal estimate at scale T).
The reader who wants to see the constant-tracking should read the paper's
Section 4, "Stopping and the terminal square-root-scale gain".
Why this theorem is not in Du-Zhang. The Du-Zhang paper uses the
codimension-two structure in its final argument, but not as a black-box
theorem of the form stated above. The present paper needs the quantitative
version because it will apply the saving inside a recursion. That is why
the paper proves it explicitly in Section 4.
Where it is used. In Section 3, the strip saving supplies the "negative
power budget" of the angular test (Chapter 29). In Section 7, it supplies
the strip case of the geometric induction (Chapter 43). It is the "second
transverse" saving, complementary to the frame estimate.
34. SECTION 4: COARSE FUNCTIONALS, THE TERMINAL ESTIMATE, AND THE STRIP
DECOUPLING LEMMA
--------------------------------------------------------------------------------
The last three pieces of Section 4 are the bookkeeping for the strip
saving. They are technical but all have the same shape: sell the strip
structure, buy a fixed power saving.
1. Coarse functional. For dyadic K >= 2, 0 < b < 1, and P = 2/(1-b),
define
T_b(V, nu) = sum_{Q in Q_{K^2}} nu(Q)^b || V ||_{L^P(Q)}^2,
where Q_{K^2} ranges over the lattice of K^2-cubes in spacetime. This is
the "K^2-cube version" of the fractal norm.
2. Terminal estimate (paper d:terminal-functional). On one such K^2-cube,
T_b(V', nu') <= C_eps K^{C_*} T^{n/(n+1) + eps}
(m')^b || child input ||_2^2,
where C_* is a fixed exponent independent of J, and m' is the ball growth
parameter of nu'.
Why is this true. We replace the field norm on each K^2-cube by the
"volume factor" times a unit-cube witness for the supremum, and move the
mass of the K^2-cube to that witness. The resulting measure has ball
parameter at most C K^{2n} m'. Since b >= beta, its b-th power is at most
(C K^{2n} m')^{b - beta} times its beta-th power. Applying the weighted
fractal estimate (Chapter 31) at scale T to the moved masses, absorbing
the K factors into C_*, gives the terminal estimate.
The important feature: C_* does not depend on J (the number of rescaling
levels). It depends only on dimension and on b. This is what makes the
strip saving "iterable" J times without losing.
3. Strip decoupling lemma (paper d:strip-decoupling). Group a
strip-supported field V = sum_theta V_theta by frequency-label bins of
side K^{-1}. Suppose the field's Fourier support lies in a fixed
enlargement of a strip near an affine (n-2)-plane, and lies within a
vertical distance O(K^{-2}) of the paraboloid. Then for each K^2-cube Q
and every 2 <= P <= q_{n-2} = 2n/(n-2),
|| V ||_{L^P(Q)}^2
<= C_eps K^eps sum_theta sum_{a in Z^{n+1}}
<a>^{-A} || V_theta ||_{L^P(Q + K^2 a)}^2.
This is decoupling "along the strip": the transverse factor K^{-1} is
small enough that the caps decouple with only K^eps loss. The RHS is a
bounded-overlap sum of translated cubes; it is what lets us move to the
child scale in the transport lemma (d:transport, in the paper).
Why "codimension two"? A strip near an (n-2)-plane has two transverse
directions. The decoupling is in those two directions. The transverse
paraboloid structure (Chapter 14) is exactly what Bourgain-Demeter
decouples, up to the q_{n-2} exponent. That is the numerical reason for
the exponent 2n/(n-2) in the lemma.
4. Packet expansion (paper d:packet-expansion). An exact extension can be
localized to spatial boxes of side T with summable input energies. This
lemma (used inside the strip saving) says that "cutting the extension to a
spatial box of side T" costs an exponentially small tail; the local norm
is controlled by incident input. It is bookkeeping but important: without
it, the sum over K^2-cubes in the terminal estimate would not close.
Together these four pieces turn the strip saving into a proved theorem.
Section 4 is, in effect, the technical heart of the "transverse savings"
side of the paper, and the frame estimate (Section 3) is the technical
heart of the "frame" side. The two meet in Section 7.
35. SECTION 5: THE TRANSVERSE SAVING IN EVERY DIMENSION
--------------------------------------------------------------------------------
This is the second of the paper's two big internal savings (the first was
the frame estimate of Section 3). The frame estimate proves a saving in
dimension k. The transverse saving proves a saving in dimension n,
exploiting the same transversality idea but in a "pruned tube model"
setting, and it does not require the frame estimate. Together the two
savings power the sparse estimate (Chapter 37) and the geometric induction
(Chapters 40-44).
Statements and form.
THEOREM (transverse saving; paper Proposition t:transverse). Fix a bounded
frequency region in R^n. There are c > 0 and L_0 < infinity (depending only
on the region and n) such that the following configuration is impossible
for L >= L_0.
Let Q be distinct unit lattice cubes in a ball of radius L^2 in R^{n+1},
with
# Q >= L^{2n - c}, #(Q cap B(z, r)) <= L^c r^n (r >= 1).
Let f_1, ..., f_{n+1} have ||f_i||_2 <= 1 and frequency support in the
fixed region, with
sup_q | E f_i | >= L^{-e - c} (q in Q, 1 <= i <= n+1),
recall e = n^2/(n+1). Then it is IMPOSSIBLE that every choice of one
frequency from each of the n+1 supports has direction determinant at least
L^{-c} in absolute value.
Read it as: no matter how you choose f_1, ..., f_{n+1}, if they are that
high on a very dense set Q of cubes, then some tuple of frequencies is
degenerately close to a lower-dimensional plane (determinant L^{-c}). The
transversality is not compatible with the concentration.
Why the exponent e shows up. L^{-e} = L^{-n^2/(n+1)} is the amplitude at
which a unit-L^2 wave is "naturally" as high as it can be on density "L^{2n}
cubes" (the volume of Q). It is the critical threshold where the
concentration problem is at its hardest. Below L^{-e-c} the fields are
smaller, and the theorem is easier (smaller levels trivially spread out);
above L^{-e+c} the Du-Zhang-type estimates already forbid the
concentration. The case at the boundary is where the theorem bites.
The proof, in outline. We prove the theorem by contradiction.
Step 1. Multilinear Kakeya with weights. Given a hypothetical
configuration, apply Proposition p:kakeya (Chapter 17) to the tube
families corresponding to the f_i. This forces the configuration to have
a "pruned tube model": there is a probability distribution on tubes
(each tube a packet trajectory of some f_i), and a random variable
(X, Y) in the tube model, such that:
- X is roughly uniform on the "body" and Y is a
"tube-distance" coordinate.
- Every tuple (T_1, ..., T_{n+1}) with independent sampled pairs
(X_i, Y_i) has |det(v_vec direction)| >= L^{-c}.
- Each tube T_i has length ~ L^2 and width ~ L^2 / (fill).
Step 2. Random variable sampling. From the tube model, define a family
of informative labels X_s (position at scale L^{-s}) and T_i (the tube
parameter). For each s in [0, 1], consider the information content
F(s), G_i(s), G_i'(s) defined in Chapter 23. These are the "profiles"
of the transverse problem.
Step 3. Rigidity. The information profiles obey:
- F(s+t) - F(s) = sum_i [G_i(s+t) - G_i(s)] + o(t) + o(1),
- the same for G_i' versus G_i,
- endpoint bounds: F(1) = n, G_i(1) = G_i'(1) = n/(n+1).
The second equality is the "single common profile" (Chapter 36).
Step 4. Angular restriction. Using the codimension-two strip saving
(Chapter 33), prove that the sum
sum_i [G_i(s+t) - G_i(s)]
obeys a codimension-two upper bound c_0 t. This gives a lower bound on
the "angular increment" of the profiles.
Step 5. Binary derivative + alternation. Combine step 4 with the
well-known "block comparison" lemma (Chapter 36) to show that at any
differentiation point, either no angular increment happens (a "full
slope point") or all angular increment happens (a "zero slope point"),
with no intermediate values. This is the binary derivative.
The full-slope set of any coordinate profile is an interval.
Step 6. Full-to-empty jump. The binary derivative plus the endpoint
inequalities forces a positive-probability event: there is some block of
scales [s, s+b] on which the coordinate profile is "full" everywhere at
the top (u) and "empty" everywhere at the bottom (v). This is the
"full-to-empty jump" (t:jump).
Step 7. Contradiction via partitioning. The full-to-empty jump is
impossible. The proof uses polynomial partitioning (Chapter 20) to
cover the point set with sign cells; a detailed incidence argument
(t:no-jump) shows that under the jump hypothesis, the sign cells cannot
cover the mass. This is a nontrivial incidence geometry argument.
Steps 1-2 and 4 are the "transverse saving" proper; steps 3 and 5-6 are
information theory bookkeeping; step 7 is the final combinatorial
contradiction. The next two chapters go through these steps.
Why the theorem is important. The transverse saving is used in two ways:
- In Section 6 (sparse estimate), it supplies the "broad gain" bound
that controls what happens when a collection of high points carries
a genuinely transverse set of packet directions.
- In Section 7 (geometric induction), it is the input to the "wall"
and "plate" arguments, alongside the frame estimate and the strip
saving.
Both uses are quantitative: they need the shape of the exponents, not just
the truth of the theorem.
36. SECTION 5: INFORMATION PROFILES AND THE COMMON-PROFILE THEOREM
--------------------------------------------------------------------------------
We continue the transverse saving's proof. Step 3 (profiles) is the
information-theoretic core. Chapter 23 introduced the machinery; here we
see it applied.
Setup. Fix a random sample (X, Y, T_1, ..., T_{n+1}) from the pruned tube
model. The labels are:
- C = parent cube of X_t, a size-L^{-t} cube in the ambient position
space.
- X_s = position of X at resolution L^{-s}, for 0 <= s <= t.
- T_i = the tube parameter of the i-th field, i = 1, ..., n+1.
- T_i' = an "alternative tube" (a reference tube independent of the
original choice; used to define the conditional profile G_i').
The profile functions are:
F(s) = I(X_s), conditional information on nothing
G_i(s) = I(X_s | T_i), conditional on tube i
G_i'(s) = I(X_s | T_i'), conditional on alternative i
All information is base-L. The natural "volume" bound on labels (L^{ns}
possible positions at resolution L^{-s}) gives F(s) <= ns for all s, and
an endpoint F(1) = n.
Endpoints and Lipschitz.
- F(s) >= ns (Chapter 23, from the growth of the model).
- F(1) = n and G_i(1), G_i'(1) >= n/(n+1) (from the positivity of the
tube marginal probability and the density assumption).
- All profiles F, G_i, G_i' are nondecreasing and Lipschitz with
constants n and 1 respectively. So they have derivatives almost
everywhere.
Block comparison (paper Lemma t:additivity). For any scale block
[s, s+t] and any reference list of tubes,
sum_{i=1}^{n+1} [G_i(s+t) - G_i(s) - ell_i]
<= F(s+t) - F(s) - ell,
where ell and ell_i are the "conditioning costs" of including a whole
block of reference bins. The block comparison is the "chain rule at
block level":
F(s+t) - F(s) >= sum_i [G_i(s+t) - G_i(s)] - cost.
Making the block size b_* small, the cost becomes small and we get the
"information-increment replacement":
F(s+t) - F(s) = sum_i [G_i(s+t) - G_i(s)] + negligible.
Cardinality coincidence. The profiles satisfy a further "coordinate
subset" fact: for any subset A of coordinates, the sum over i in A of
G_i increments is at most F increment minus remaining, up to exchange
costs. This is what lets us localize to a single direction without fixing
the other n-1.
THEOREM (single common profile; paper Theorem t:common-profile). At almost
every s, the profiles agree:
G_i(s) = G_j(s) = G(s) for all i, j,
F(s) = (n+1) G(s).
Proof strategy. Use the block comparison inequality, the coordinate subset
identity, and the fact that at the endpoint the two sides are equal (both
equal n/(n+1) per coordinate). Combining:
- If on some block, the j-th coordinate profile G_j has strictly
smaller increment than G_i for i != j, then the coordinate subset
identity is violated (the subset {j} increment would need to be
larger than the total). This forces G_i - G_j <= o(1) uniformly.
- A standard "sum of nonnegative" argument closes to exact equality.
The details use several intermediate lemmas:
- The "restricted determination" (paper t:block-comparison): for a
scale increment [s, s+t], the extra information in X_{s+t} given X_s
and the tube direction is bounded by the corresponding coordinate,
up to a conditioning cost.
- The "block comparison and telescoping costs" (paper t:additivity):
conditioning on individual reference bins costs only O_n(b_*) per
block, so summing over blocks loses only O_n(b_*) total.
These are, in effect, elementary estimates about chain rules on finite
partitions. The reason they have to be tracked carefully is that they
hold with constants that are dimension-dependent but uniform in the
sample size, and the limiting information profile has to be derived from
the finite-scale version carefully.
What the common-profile theorem gives. Now we have a SINGLE function G
such that F = (n+1) G, and G(1) = n/(n+1), F(1) = n, both achieved at
s = 1. So G is a nondecreasing Lipschitz function with G(1) = n/(n+1) and
G(0) = 0. The whole transverse saving's information content is captured
by this single one-dimensional function G.
The geometrical content of G. By construction, G(s) is the base-L
information of position given one of the tube directions. An increment
G(s+t) - G(s) measures how much more information position carries at the
finer scale s+t given the tube T_i. Because the tube is constrained to
pass through the sampled point, the increment at resolution L^{-(s+t)} is
exactly the "extra" position information contributed by the resolving of
scale.
The endpoint inequality from Chapter 35 (from codimension-two) will say
that sum_i (G_i(s+t) - G_i(s)) <= c_0 t, i.e. the total conditional
increment is bounded by a constant c_0 < 1. But F increment = (n+1) times
G increment, so sum_i increment * (n+1)/... -- this forces a constraint
that is violated unless something specific happens. That is the jump,
which the next chapter rules out.
37. SECTION 5: ANGULAR RESTRICTION, BINARY DERIVATIVE, AND THE JUMP
--------------------------------------------------------------------------------
We now finish the transverse saving proof.
Step 4: angular restriction. Fix a scale block [s, s+t]. The common-profile
theorem (Chapter 36) says F increment = (n+1) times G increment. The
codimension-two strip saving (Chapter 33) supplies a constraint in the
opposite direction:
LEMMA (angular sum bound). In the pruned tube model, for every block
[s, s+t],
sum_{i=1}^{n+1} [G_i(s+t) - G_i(s)] <= c_0 t + o(t),
where c_0 < 1 is a fixed constant (related to the factors in Chapter 33).
The reason the codimension-two strip saving enters here: any
configuration in which the sum on the left has a bigger derivative than
c_0 would give a quantitative saving in the tube model that contradicts
the Du-Zhang machinery. That is, if the profile increments were large,
we would be in the strip saving regime (packets concentrating near an
(n-2)-plane in the direction space), and the strip saving provides a
strict improvement.
Sketch. Suppose the sum had larger derivative. Then there is a
macroscopic coordinate j such that the "direction variable" concentrates
near a hyperplane. By the codimension-two structure of the (n+1)-tuple,
this puts the direction into a "codim-2 stratum", and the strip saving
applies with a fixed gain c_0 t. Contradiction. (The precise reduction
uses the projection theorem of He 2020, which is why Chapter 21 was
included.)
Step 5: binary derivative. Combine the angular sum bound with the block
comparison and the "single common profile" theorem to show:
LEMMA (binary derivative; paper Lemma t:binary). At almost every s,
G'(s) is either 0 (a "zero slope point") or G'(s) = 1/(n+1) (a "full
slope point" of G).
Proof sketch. If G'(s) were strictly between 0 and 1/(n+1), then the
angular sum bound would be violated by a small bi-linear interpolation
argument (choose two scales inside the block, average the two angular
bounds, and use the chain rule to force an intermediate value). More
concretely, the projection theorem lets us extract a "good direction" on
any subset of the point set of sufficiently large cardinality; applied
to the profile, this forces a specific binary structure.
COROLLARY (binary alternation). The set of full-slope points of G is an
interval; G is a "ramp" that is 0 on the left, increases linearly at rate
1/(n+1) on an interval [s_0, s_1], and is 0 again on the right. In
particular, F = (n+1) G is a ramp of the same shape, scaled by n+1.
This corollary is exactly the "profile rigidity" of the frame-estimate
world (Chapter 28), now obtained in the tube setting. The two worlds
(frame, transverse) both eventually force profiles to be two-slope
functions.
Step 6: full-to-empty jump (paper t:jump). The endpoints F(1) = n and
G(1) = n/(n+1) together with the binary alternation force a specific
event: there is a block [s, s+b] of scales such that at the coarse
scale s the profile is at "full" level (F(s) = ns, G(s) = n/(n+1) s),
and at the fine scale s+b it is at "empty" level (F(s+b) = (n+1)*0, G(s+b)
= 0). This is not quite literally "F(s+b) = 0"; the "empty" of the jump is
a specific quantitative statement at a scale b < 1/8. The details are in
the paper's Section 5, "Block comparison, additivity, and coordinate
subsets" (the jump lemma t:jump).
Step 7: partitioning contradiction (paper t:no-jump). The key claim is
that the full-to-empty jump is IMPOSSIBLE. The proof is an incidence
geometry argument using polynomial partitioning.
Sketch. Assume a full-to-empty jump on a small block. Choose the partition
parameter P = K^{1/2 - eta} (a specific power of the ambient K). By
polynomial partitioning with parameter P, cover the point set with sign
cells of the partition. Each cell contains few points by construction.
- "Good" endpoints: at scales s and s+b, most of the point set has
profile in the "full" or "empty" range. Cells that contain good
endpoints have bounded mass.
- "Central line crossing": the tubes cross each sign cell along a
central line, and each central line meets O_n(P) cells. So the mass
that is "seen" by the tubes as they traverse the cells is bounded by
the cell structure.
- Combine: the number of good endpoints in cells the tubes visit is at
most (mass per cell) * O_n(P) * (number of cells), and the mass
per cell is bounded by the partition. This gives a bound on the pairs
of "good off-wall" endpoints in the same cell. The bound is too
small to be compatible with the jump's lower bound, giving a
contradiction.
The full proof of the no-jump lemma is a 5-page incidence argument in the
paper. We have indicated the structure and the role of the partition.
What this completes. Steps 1-7 together prove the transverse saving
(Theorem t:transverse). The four pieces (multilinear Kakeya, information
theory, codimension-two strip saving, polynomial partitioning) are all
invoked at their sharp endpoints, and the exponents track without loss.
The transverse saving does not need the frame estimate. It uses the tube
geometry of n+1 transverse families directly. It is a coarser result than
the frame estimate (it is about D = 1 configurations in n dimensions, not
about packets at scale h in k = n-1), and where it applies it is simpler.
In Section 6 (next chapters), it provides the "broad" gain of the sparse
estimate, and in Section 7 it is one of the two engines of the geometric
induction.
38. SECTION 6: THE SPARSE ESTIMATE
--------------------------------------------------------------------------------
Section 6 (sparse.tex) is the bridge between the geometric savings of
Sections 3-5 and the geometric induction of Section 7. It converts the
transverse savings (which give a gain when the packet frequencies are
transverse) into a single, quantitative estimate over the whole packet
array. The name "sparse" is that D, the sparsity parameter (Chapter 22),
governs the size of the saving.
Statement.
PROPOSITION (sparse estimate; paper Proposition s:sparse). Fix the
spacetime dimension n+1. There is c_s > 0, depending only on dimension
and the fixed frequency box, with the following property.
Fix tau > 0 and D with
L^tau <= D <= L.
Let U be a full-dimensional compact array of scale L with coefficient
energy E = sum_alpha |c_alpha|^2, and let mu be a nonnegative measure on
the native time range of U satisfying:
- ball growth: mu(B(z, r)) <= m r^n for all z and r >= 1;
- tube moment: for every nonzero label alpha,
INT (1 + |x - y_alpha - 2 t omega_alpha|/L)^{-M} d mu(x, t)
<= m L^{n+1} / D.
(Recall M is a fixed large constant from Chapter 22.)
Then, for sufficiently many bounded profile derivatives,
sum_{q unit cube} mu(q)^{beta} sup_q |U|^2
<= C_tau m^beta L^{2n/(n+1)} D^{-c_s} E.
Here beta = 2/(n+1) (Chapter 31). The constant C_tau may depend on tau
(the lower bound on D) but not on L, D, U, or mu.
The same estimate holds with U replaced by any fixed finite number of its
derivatives, with the same exponent.
What it says. Given a measure that avoids each packet trajectory by a
factor D (the tube moment condition), the beta-fractal sum of U is
strictly smaller than the "ambient" bound m^beta L^{2n/(n+1)} E by a
fixed power of D. The saving D^{-c_s} is the point of the sparse estimate,
and it is what Section 7 uses to close the induction, with D being the
scale parameter along a hierarchy.
Where the estimate is used. The sparse estimate is the main tool in
Section 7. In that section, configurations with D >= L^tau are "already
handled" (the sparse estimate applies); configurations with D < L^tau
need the deeper geometric induction (Chapters 40-44). So the sparse
estimate's role is to handle the "large D" region of the problem, so that
only 1 <= D <= L^tau needs the hard work.
Structure of the proof. The proof has three parts, and we describe them
here at the level of what each part does.
Part (i): localization and Fourier layers. On the time range of interest,
localize U to boxes of side L^2 (there are boundedly many by the growth
assumption). Decompose U into exact Fourier layers (Chapter 32), one for
each integer frequency-index q in Z^n and each time-frequency u in R,
with summable weights. Each layer is a compact array that "lives" on the
same spacetime box.
Part (ii): broad/narrow iteration. At each iteration level j, we have a
compact array U_j of scale L/K^j (where K is a dyadic scale chosen at the
end), and a measure nu_j with ball growth m_j and sparsity D_j. We split
U_j into broad and narrow parts using the simplex lemma (Chapter 16):
- broad part: contains n+1 subfields whose directions are transverse
with determinant at least c K^{-n} (the simplex bound for K-caps);
- narrow part: the rest of the frequency content.
For the broad part, apply the transverse saving (Theorem t:transverse,
Chapter 35) at the current scale. This gives a fixed inverse-power gain
of L^{-c_b/2} for the broad contribution at this level.
For the narrow part, iterate to level j+1 by applying the transport
(parabolic rescaling, Chapter 4). The narrow prefactor is at most
C h^{-beta} per level, where h = K^j (the amount of "narrowing").
Part (iii): terminal and summation. At a terminal node, the current
scale is small enough that the fractal estimate (Chapter 31) applies
directly, giving a gain of (l/D_j)^{theta_n} (where l is the current
scale and D_j the current sparsity). The terminal scale is where the
measure's tube moment condition kicks in fully.
Summation over the tree. Because each leaf has a positive power saving in
either c_b or theta_n, and the branching structure loses only the fixed
prefactor C^j (with j bounded by log(L) / log(K) = O(1/tau)), the leaves
sum to at most D^{-c_s} times the top.
We describe the summation more precisely in the next chapter. The point
to remember: the broad/narrow iteration is a strict adaptation of
Bourgain-Guth 2011 and Du-Zhang 2019; the innovation is the terminal
gain and the constant tracking from the sparse parameter D.
Relation to other estimates. The sparse estimate does not use the
frame estimate (Section 3). It uses:
- the transverse saving (Section 5),
- the exact-layer decomposition (Section 4),
- the compact-array fractal estimate (Section 4),
- the parabolic transport / narrow decoupling (Section 4).
So Section 6's proof rests entirely on Sections 4 and 5, plus the packet
machinery of Section 2. The frame estimate (Section 3) and the strip
saving (Section 4) are used only in Section 7, not here. This is why the
paper develops the frame estimate early and the sparse estimate later:
they are logically independent sections that meet in Section 7.
39. SECTION 6: THE NARROW PREFACTOR, THE TERMINAL GAIN, AND THE SUMMATION
--------------------------------------------------------------------------------
We continue the sparse estimate's proof. This chapter is the bookkeeping.
The narrow prefactor. Each narrow step costs a factor h^{-beta}, where
h = K^j is the total rescaling factor up to level j. Recalling
beta = 2/(n+1), this is h^{-2/(n+1)}. For h large (k many levels of
narrowing), this is a real loss.
The broad gain. Each broad step gains at least l^{-c_b} for some fixed
c_b > 0 (from the transverse saving; the constant depends only on
dimension and the simplex bound K^{-n}). And l = current scale, so at the
bottom of the tree the gain is substantial.
The terminal gain. At the terminal scale l, the measure mu_j has tube
moment parameter D_j (from the direct pushforward, Chapter 34). The
fractal estimate at scale l gives
sum_q mu_j(q)^{beta} sup_q |V|^2
<= C l^{n/(n+1) + eps} (min(1, D_j))^{gamma_n} m_j^beta E_j.
The "terminal gain" comes from D_j being at least D_0 = L^tau for our
starting D, and growing as we iterate (each step multiplies by a factor
from K^{n+1}). For the fractal-estimate bound at terminal, we get a
gain at least (l/D)^{theta_n} for some fixed theta_n > 0 (this is proved
by organizing the local mass at the terminal mass bound
m_j l^{n+1}/D).
Combining. Define
a_* = min(c_b/2, theta_n/2),
c_s = a_* / 8.
Then a leaf has either
- a broad gain of l^{-c_b} times the narrow prefix h^{-beta}, which is
l^{-c_b} * h^{-beta} <= (L/l)^{-c_b/2} * L^{-a_*}, after adjusting the
exponents by the choice of the dyadic K in the tree; or
- a terminal gain of (l/D)^{theta_n} times the narrow prefix, which is
(l/D)^{theta_n} * L^{-a_*} in the same way.
In both cases the net gain is at least L^{-a_*}. With c_s = a_*/8, we
have a uniform slack of a factor of 8, which absorbs every fixed
polynomial constant (from the branching factor, from the localization,
from the Sobolev reduction).
The exact exponent arithmetic (paper proof of s:sparse). At each level,
the relevant factors are:
- Jacobian of parabolic rescaling: K^{2 - (n+2)beta} (loss)
- measure parameter inflation: K^{(n+1)beta} (loss)
- narrow decoupling prefactor: K^{-beta} (loss)
- fractal-estimate factor: K^{-2n/(n+1)} (gain)
Multiplying gives
K^{2 - (n+2)beta} * K^{(n+1)beta} * K^{-beta} * K^{-2n/(n+1)}
= K^{2 - 2beta - 2n/(n+1)} (using (n+2)beta = (n+1)beta + beta)
and with beta = 2/(n+1), we get
= K^{2 - 4/(n+1) - 2n/(n+1)}
= K^{2 - (2n+4)/(n+1)}
= K^{2 - 2 - 2/(n+1)}
= K^{-2/(n+1)}
= K^{-beta}.
So the product at each step is exactly K^{-beta}, i.e. a fixed inverse
power per step. Multiplied by the number of steps (= log(L)/log(K) =
1/v), the cumulative is K^{-beta * 1/v} = L^{-beta}, i.e. a fixed power of
L, if v is chosen to make beta/v = a positive constant. That is the "clean"
case; in the actual proof the four factors are not all applied at every
step (they apply to different parts of the tree), and the exponent
tracking is more delicate. But the cancellation
K^{2 - (n+2)beta} K^{(n+1)beta} K^{-beta} K^{-2n/(n+1)} = K^{-beta} is the
core "algebraic miracle" that makes the sparse iteration close.
The precise statement of the sparse estimate at the end (from the paper's
proof of s:sparse):
a_* = min{ c_b/2, theta_n/2 }, c_s = a_* / 8.
Every leaf gains D^{-a_*} before its small losses, and the terminal K
factors absorb into a fixed polynomial. The exponent c_s is the
paper's "sparse exponent", and it is what Section 7 uses.
Where the sparse estimate "ends". For D >= L^tau, the sparse estimate is
strictly better than the "trivial" bound (D^{-c_s} is a genuine saving
for L large). For 1 <= D <= L^tau, the sparse estimate's D^{-c_s} is
cancelled by the constant C_tau and the estimate becomes useless. This is
why Section 7 must handle the "small D" case directly via the geometric
induction.
The three sections (3, 5, 6) interlock like this:
- Section 5 (transverse saving) supplies the broad gain.
- Section 4 (fractal + strip saving) supplies the terminal gain and the
narrow transport.
- Section 6 (sparse estimate) packages them into a single D^{-c_s}.
- Section 3 (frame estimate) is not used in Section 6. It is used in
Section 7, specifically in the "plate" case of the geometric
induction (Chapters 40-44).
The reader who wants to follow the details: the paper's Section 6 is a
straightforward adaptation of Bourgain-Guth 2011 and Du-Zhang 2019, with
the exponents and the terminal gain tracked. The novel content is the
terminal estimate (which uses the Du-Zhang input in the squared form of
Chapter 31) and the exponent bookkeeping (which is where the constants
c_s, c_b, theta_n are pinned down).
40. SECTION 7: THE WEAK ESTIMATE: STATEMENT AND EASY RANGE
--------------------------------------------------------------------------------
Section 7 (weak.tex) is the longest and heaviest section. It proves the
frequency-local weak L^p maximal estimate at the endpoint, which is then
used in Section 8 to get pointwise convergence. This is where the frame
estimate (Section 3), the strip saving (Section 4), and the sparse
estimate (Section 6) are all needed.
Statement.
THEOREM (weak estimate; paper Theorem w:weak). Fix the spacetime
dimension n+1 >= 4 (so n >= 3). There is a finite p_0 (depending only on
n) and fixed exponents a, c_0 > 0 with the following property.
For every p with p_0 < p < infinity, there is a constant C_B = C_B(p),
such that for every compact array U of scale L with coefficient energy
E(U) = sum |c_alpha|^2, every measure mu on the native time range
satisfying:
(i) ball growth: mu(B(z, r)) <= m r^n for all z and r >= 1;
(ii) unit density: mu(B(z, 1)) <= m chi for a fixed chi in (0, 1];
(iii) tube moment with sparsity D: for every nonzero label alpha,
INT (1 + |x - y_alpha - 2 t omega_alpha|/L)^{-M} d mu
<= m L^{n+1} / D,
and every g > 0,
mu{ |U| >= chi^a g L^{-e} }
<= C_B m L^{2n} g^{-p} D^{-c_0} E(U)^{p/2},
where e = n^2/(n+1) (Chapter 31).
The theorem has four independent parameters: L (the packet scale), D
(tube moment), chi (unit density at scale 1), and g (the level). p is
free (must be > p_0); a and c_0 are fixed by n. The point is that the
constants do not depend on L, D, chi, g.
What it says. If a packet field U is larger than the critical amplitude
chi^a g L^{-e}, then the measure of the level set is strictly smaller
than the "ambient" bound m^beta L^{2n} times a fixed power of g and a
fixed power of D. The factor D^{-c_0} is the whole point of the sparse
estimate. The factor g^{-p} is the "level gain".
The easy range. Split by the values of chi and D:
- SMALL CHI. If chi <= L^{-tau_chi} for a fixed small tau_chi > 0, then
the amplitude chi^a g L^{-e} is small compared to g L^{-e}, and the
"clean" bounds for small amplitude suffice. In fact the trivial
bound mu{category} <= mu(all) contributes nothing for small enough
chi, and one can use the sparse estimate (Chapter 38) directly with
the effective chi-small improvement.
- LARGE D. If D >= L^tau for a fixed small tau > 0, the sparse
estimate (Chapter 38) applies and gives D^{-c_s}. So for these
parameter ranges, Theorem w:weak follows from Chapter 38.
So the difficult range is:
L^{-tau_chi} <= chi <= 1 and 1 <= D <= L^tau.
The "geometric" or "many-point" case: the measure is neither very sparse
nor very concentrated away from the tube trajectories. This is where the
geometric induction runs.
Reduced form. Theorem w:weak reduces to the following "geometric level
estimate":
For chi in [L^{-tau_chi}, 1], D in [1, L^tau], and g bounded
below by a fixed power (chi-dependent) and above by L^{C_g}:
mu{|U| >= chi^a g L^{-e}} <= C_B m L^{2n} g^{-p} D^{-c_0} E^{p/2}.
This is what the next four chapters prove.
Exponent order. The constants and the order in which they are chosen
matter, because the proof uses each exponent at a specific step. The
order (from the paper, w:eq-ab):
(i) Fix b with beta < b < 2/n. Fix a small delta > 0, and set
2a = 1 - b + delta, zeta = 1 - beta - 2a = b - beta - delta.
(So zeta > 0 and 2a + beta + delta = 1 - ... well, zeta is positive
by the choice b > beta and delta small.)
Then choose J large enough that the angular lemma (Chapter 41)
holds with a fixed gain exponent c_2 > 0.
(ii) Fix a small unit-density cutoff v_0 > 0. The margin
n - 1/2 - e = (n-1)/2(n+1) > 0
allows the "frame cutoff" power (the boundary between small-density
cases and the frame saturation bound) to be fixed here.
(iii) Fix sigma > 0 much smaller than uc_2 and c_s (with u determined
below). This is the "sparse slack" used to absorb the smaller
factors in the summation.
(iv) Fix the frame estimate's parameters (profile orders, extension
constants, and so on). Choose the "u" small enough.
(v) Reduce u further if the frame estimate's costs require it.
(vi) Choose p (the level exponent) after all of the above, taking
p-2 small compared with the exponents c_2, sigma, c_s.
Finally, choose the induction constant C_ind > C'_B.
The pivotal inequality is:
C_g (p-2) + sigma < c_s / 2.
Since C_g is fixed by geometry, p-2 can be made small enough for this to
hold. This is where the "easy range" (small chi) and "hard range"
(geometric) are glued: the same inequality must hold in both.
Strategy. The geometric induction proceeds as follows:
1. Reduce to a "heavy" subset of the level set where at each point the
Fourier angular sectors ("bins") carry most of the coefficient mass
in a small number of bins. (Chapter 42.)
2. At each such point, either the heavy bins are transverse (a genuine
(n-1)-simplex), or they lie near an affine (n-2)-plane (a "wall").
(Chapter 42.)
3. Transverse case: apply the frame estimate (Section 3) in dimension
n-1. This is the deepest step and where the frame estimate is
finally used. (Chapter 43.)
4. Wall case: apply the strip saving (Section 4) to sell the (n-2)
concentration, and then the angular lemma (Section 7-angular) to
bound the curvature of the wall. Either the wall has large second
fundamental form (angular), or it is locally a plate (flattening).
(Chapters 41 and 43.)
5. Sum the contributions against the measure mu, using two energy
budgets (one per time bin, one global with density weights). Close
with the induction constant. (Chapter 44.)
The four pieces are interleaved: the plate-case frame estimate (step 3)
needs the heavy-bin reduction (step 1), and the wall-case strip saving
(step 4) needs the angular lemma (step 4) which uses the strip saving.
41. SECTION 7: THE ANGULAR INDUCTION AND THE DETECTION OF CURVATURE
--------------------------------------------------------------------------------
The key piece of the wall case is a self-contained lemma about "how much
does the heavy bin field curve", which uses the strip saving quantitatively.
This is the angular induction.
Setup. Suppose at each point of Y, a field V (a "heavy" component of U)
has magnitude >= lambda, and V is supported in a small frequency cap
around a fixed vector. So V is a "narrow-band field". The level set Y is
a subset of spacetime.
Tangent plane. For the frequency cap, there is a natural "affine (n-2)-
plane" H, roughly the tangent to the paraboloid at the frequency of the
cap. Locally, the field is (approximately) a function of the "transverse"
coordinates to H. The "wall" is the codimension-1 surface where it is
as big as lambda.
Second fundamental form. The derivative of the heavy-bin directions
across H is the "shape" of the wall. Its second fundamental form (Hessian
of the surface in the "shape" coordinates) either is large or is small.
Large Hessian. If the second fundamental form is large, the wall curves
sharply. Along a curve crossing the wall, the second derivative of the
heavy-bin amplitude with respect to the transverse parameter is at least
a threshold. This forces the level set Y to be small: the curvature
(effectively, the phase modulation from the parabolic structure) rapidly
decorrelates the wavelets.
Small Hessian. If the second fundamental form is small, the wall is
locally close to a hyperplane (a "plate"). Along the wall, the
configuration resembles an (n-1)-dimensional problem, and the frame
estimate applies after flattening.
Quantitative form (paper w:angular). There is a fixed exponent c_2 > 0
such that: if the heavy field is not contained in a single quadratic
neighborhood (radius R^{-1} around the plate; see below) at scale R, then
the measure of Y is reduced by a factor R^{-c_2} compared to its natural
size.
Setup of the "quadratic neighborhood". The reference quadratic
neighborhood is defined by a specific bounded polynomial of degree 2
with coefficients controlled by the "shape" of the wall. A configuration
"inside" the quadratic neighborhood is exactly the plate case (small
Hessian). A configuration "outside" is the "large second fundamental
form" case.
Why the strip saving enters. To prove w:angular, one must detect a
curvature in the wall. The detection uses the codimension-two strip
saving (Chapter 33) in the following way. Compact the heavy field at scale
R. If it stays inside a quadratic neighborhood, it is "plate-like"; if
not, the second derivative is quantitatively large. Then project to the
codim-2 subspace (transverse to the wall) and use the strip saving to
bound the projection: the projected configuration is a "strip-saved"
configuration with saving R^{-c_0}, and the angular gain c_2 is a fraction
of that strip-saving exponent.
The sublevel estimate (paper w:wall). With the strip saving, the "large
curvature" case gives a direct measure bound:
mu({|V| >= lambda, heavy bins not in quadratic neighborhood})
<= R^{-c_2} mu({|V| >= lambda}).
This is what allows the induction to keep making progress: at each step,
either the field is well-approximated by a quadratic neighborhood (plate
case), or its measure is reduced by R^{-c_2} (angular case).
Structure of the argument (sketch).
- Fix the "reference hyperplane" H (from the tangent cap). Choose a
basis of the transverse directions and the "shape" parameter.
- Suppose the heavy field is NOT in the quadratic neighborhood. Then
the second fundamental form at some point of Y is at least a
threshold (in the shape coordinates).
- Use the parabolic rescaling (Chapter 4) and the strip saving
(Chapter 33) in the "transverse" directions to convert the large
second derivative into a strict measure saving.
- The saving is at least R^{-c_2}, where c_2 depends only on the
dimensions and on the given simplex bounds.
The proof is about 8 pages in the paper. The reader who wants the full
argument should look at the paper's Section 7, "Angular induction".
Two important things to know:
- The angular lemma is stated at a FIXED scale R (with R = K^J for
the fixed dyadic parameters). It is applied once at each level of
the induction, at the corresponding scale. So the induction combines
angular lemmas at multiple scales.
- The curvature dichotomy (large Hessian vs. plate) is what forces the
wall to be "piecewise affine". Either curvature is large (angular
handles), or it is small (plate handles). There is no intermediate
case, because the transition is sharp (bounded below by 1/R and
bounded above by 1/R, with a gap of R^{small}).
This is the geometric heart of the paper: the "curved ruled wall" that
defeated naive codimension-two arguments in earlier work is now handled by
combining angular (for curvature) and flattening (for plates).
42. SECTION 7: HEAVY BINS AND THE HEAVY-BIN REDUCTION
--------------------------------------------------------------------------------
We continue Section 7. This chapter reduces the level set to a "heavy"
portion on which a few Fourier angular bins carry most of the mass, and
then presents the transverse/plate dichotomy for the heavy bins.
Heavy bins. Fix the current compact array U (of scale L) and the level
lambda. Split the frequency support of U into R^{-1}-caps (angular
sectors), with R = K^J the fixed dyadic parameter from Chapter 40's
exponent choice.
For each cap b, let E_b be the coefficient energy in bin b, and let b_T
be the effective envelope. Define the "heavy bins" at a point z:
the caps b with |V_b(z)| >= W(z) * (sum of amplitudes over b, times a
threshold). Here W(z) is the local weighted square mass (Chapter 22).
Roughly, a heavy bin is one that supplies at least a fixed fraction of
the mass of U at z. Because the total weighted mass is 1 by normalization,
the number of heavy bins is bounded by a fixed constant (like 1/eta with
eta the fraction threshold).
Heavy-bin reduction (paper Lemma w:heavy-reduction). There is a subset
Y' of Y with mu(Y') >= c M_Y such that, at each point of Y':
(i) the heavy-bin field is >= lambda/2 in magnitude;
(ii) the heavy-bin centers are not contained in a strip of any fixed
large constant times R^{-1} about an affine (n-2)-plane, nor in
a single quadratic neighborhood of w:angular's form;
(iii) every heavy bin supplies >= W/2 of its weighted coefficient
square mass from trajectories within distance X^{C_w} L.
In particular, some n heavy bins have direction vectors (2 omega, 1)
whose n-volume is at least R^{-C} throughout the bins.
What each clause does.
(i) is a "clean" reduction: at most half the points can lose half the
magnitude, and we take the surviving set.
(ii) is the important geometric clause. It says the heavy bins are
"not all almost in a codim-2 plane". If they WERE in a codim-2 plane,
we would be in the wall case (Chapter 41's angular lemma handles it
separately); the reduction removes this case, so we may assume from
now on that the heavy bins are genuinely transverse (contain an
n-simplex of positive volume).
(iii) is a "local mass" clause: most of the heavy bin's weighted mass
comes from trajectories within distance X^{C_w} L of the point. This
is used to apply the frame estimate, which requires the envelopes
to be concentrated near the point.
Sketch of the proof. For each bin, restrict mu|_Y to the macros where
that bin is "light" (contributes less than a threshold). Because R^4 is
much smaller than L, the weight is comparable at the macro center and at
each point of the macro. So the total contribution of the "light" part
of any bin is at most a fixed fraction of the mass. Summing over the
finitely many "light" cases and taking the complement gives the "heavy"
subset.
Then clause (ii): apply the quadratic-neighborhood removal (the angular
lemma's alternative). The subset excluded by clause (ii) is exactly the
"quadratic / plate" case, which will be handled separately. So we can
assume the good clause (ii) without loss.
Clause (iii): similar local restricted-measure argument, at the scale
X^{C_w} L.
Consequence. At each point of Y' (the heavy subset), there are n+1 heavy
bins whose direction vectors (2 omega, 1) have full (n+1)-volume. In
particular, they are transverse with determinant >= c R^{-C} (the fixed
simplex bound for R-caps). This is the configuration to which the frame
estimate will be applied.
The "wall" (Chapter 41) case. What if, contrarily, the heavy bins WERE in
a codim-2 plane? Then the angular lemma applies and either:
- the angular gain R^{-c_2} reduces the measure, or
- the heavy field is confined to a quadratic neighborhood (a "plate"),
and the plate case is handled by flattening (Chapter 43).
So the heavy-bin reduction removes the wall; either the heavy subset
survives and is transverse, or we are in a wall/plate configuration. The
geometric induction handles both cases at each step.
The "budget" bookkeeping. The paper tracks two things:
- The MEASURE budget: mu(Y) can only decrease along the induction, so
we can sum contributions at each level without double counting.
- The ENERGY budget: the coefficient energy E(U) is preserved by
projection and localization (with a fixed constant factor), and is
the "numerator" of the bound.
- The HEAVY-BIN budget: the heavy-bin mass condition ensures that the
weight W(z) is comparable to the sum over heavy bins, so when we
discard light bins we lose only a fixed fraction.
Combining, the heavy-bin reduction preserves both budgets up to fixed
constants, so the estimate we get on Y' transfers to Y with only a
constant loss.
43. SECTION 7: FLATTENING AND THE FRAME ESTIMATE IN DIMENSION n-1
--------------------------------------------------------------------------------
We continue Section 7. This is where the frame estimate (Section 3) is
finally plugged into the geometric induction.
The plate case. From the heavy-bin reduction (Chapter 42), one of two
things holds at each point of the heavy subset Y':
(a) The heavy bins are contained in a thin strip about an affine
(n-2)-plane (a "plate" of frequency codimension 2).
(b) The heavy bins are NOT in such a strip; their direction vectors
(2 omega, 1) span a full (n+1)-simplex (determinant at least
c R^{-C}).
In case (b), the frame estimate (Chapter 26) applies directly in
dimension n, with m = n+1 transverse fields. That is the "generic"
case, and it is the only place the full frame estimate is applied as-is.
In case (a), the field is trapped near a lower-dimensional surface: the
heavy bins, though individually of bounded volume, have collapsed into a
codim-2 subspace. The generic transversality condition fails. The
geometric induction must "flatten" along the plate, reducing the problem
by one dimension, and reapply the frame estimate in dimension k = n-1.
Flattening.
Let V be the codim-2 affine subspace. Its normal space N has dimension
2. Choose coordinates:
u in V: an (n-1)-vector;
v, w in R: the two normal coordinates.
The heavy field is approximately u-independent near the plate (that is
what "plate" means: the field is roughly constant along V). Parabolic
rescaling along the plate direction:
u -> u / b_1, t -> t / b_1^2,
with b_1 = L/h (hence b_1 h = L), turns the plateau into a parabolic
rescaled problem in (n-1)+1 dimensions, with the two normal directions
as separate "fields".
Why k = n-1 and m = n. The plate is (n-1)-dimensional. The (n+1) heavy
bins satisfy one linear constraint (their simplex lies in the normal
space, which is 2-dimensional, so the sum of their projections onto any
common hyperplane in N is constrained). This leaves (n+1) - 1 = n
independent transverse directions. The frame estimate in dimension k
requires m = k+1 fields. Setting k = n-1 gives m = n. Exactly right.
Why the frame estimate applies after flattening. Its hypotheses are:
- common scale h. After rescaling, all packets have the same native
scale in the flattened coordinates.
- transverse determinant >= c R^{-C}. This is heavy-bin clause (ii)
verbatim, in the flattened coordinates.
- envelopes bounded at witnesses. This is heavy-bin clause (iii).
So the frame estimate applies, and gives a bound on the number of high
cells in the flattened problem.
The tuple bound.
A "tuple" is a local unit configuration in the flattened problem: one
witness per heavy bin, one normal slab of width h (an n-dimensional slab
in the original), one time bin of duration L^2/N (equivalently b_1^2 h^2
in the flattened coordinates).
LEMMA (tuple bound; paper Lemma w:tuple). For each tuple, the level-set
measure satisfies
lambda^2 M_0
<= C R^C (1 + sum_i ell_i)^C m chi^{1-beta}
h^{n + 2n/(n+1)}
( prod_i (d_{ell_i} E_i) )^{beta/n}
( sum_i E_i )^{1-beta}.
Where:
- M_0 is the tuple's cell measure (the mass of mu on the tuple);
- d_{ell_i} is a density parameter for the i-th direction, bounded by
ell_i in a dyadic hierarchy;
- E_i is the coefficient energy of the i-th bin;
- beta = 2/(n+1) is the fractal homogeneity.
Idea of the tuple bound. Partition the two normal directions into slabs
of width h and depth b_1 h = L. For each slab, the mass of mu on the
slab is bounded by the unit-density condition (Chapter 40):
mu(slab) <= m chi^{1-beta} L^n (up to bounded constants).
Then apply the frame estimate in the flattened coordinates:
|high cells in the slab| <= H_g^{-2} sum_i D_i / A_i^2,
where the flattened field has common scale h, and:
A_i = lambda d_{ell_i}^{1/2} (amplitude, augmented by
direction density),
D_i = h^n E_i (coefficient energy, up
to h^n per packet),
H_g = ( prod_i A_i^{2/(n-1)} )^{1/n} (frame-estimate "geometric
mean" of amplitudes).
The (prod_i (d_{ell_i} E_i))^{beta/n} (sum_i E_i)^{1-beta} energy factor
appears when the amplitudes are reweighted by the fractal measure on the
tuple. The h^{n + 2n/(n+1)} factor is the product of:
- h^n: unit-cube count in the (n+1)-dim ambient;
- h^{2n/(n+1)}: the fractal dimension factor of the tuple.
The (1 + sum_i ell_i)^C cost is the number of dyadic bands in the
summation over "narrow" regions; bounded overlap per band. The R^C is the
accumulated frame-estimate cost (R^delta factors, simplex counts, band
counts). Both are absorbed in the h^{small} slack and the D^{-c_s/2}
margin at the end (next chapter).
Where the "saving" comes from. The tuple bound alone does not yet give a
saving. The saving comes from combining:
- the frame estimate's H_g^{-2} factor, which is a "transverse" gain;
- the density-weighting of the tuple's mass, which is bounded by the
ball growth condition;
- the D-sparsity of the underlying measure (mu's tube moment).
The product of these three gives the D^{-c_s/2} factor that appears in
the final w:eq-finalL2 (Chapter 44).
We do not reproduce the full proof of the tuple bound here. The reader
can find it in the paper's Section 7, "Flattening and the frame estimate
in the lower dimension". The key point for the narrative: the frame
estimate is not used directly in n dimensions in the plate case, but in
n-1 dimensions after the (n-2)-plate has been "flattened out".
44. SECTION 7: TWO ENERGY BUDGETS AND THE CLOSURE OF BOTH INDUCTIONS
--------------------------------------------------------------------------------
Step 7 of Section 7. We have the tuple bound (Chapter 43) and must now
sum it over all tuples, all time bins, and all slabs, against two
independent energy budgets. This gives the final w:eq-finalL2 estimate,
which closes both inductions of Theorem w:weak.
The two energy budgets.
BUDGET 1 (per time bin). Fix a time bin of length T = L^2/N (or the
flattened analogue). The total coefficient energy of all bins active
in this time bin is bounded by a fixed power of N times E(U):
sum_{i : bin i active at this time} E_i <= C N^{C_1} E(U).
BUDGET 2 (density weighted). Weighted by the measure mu and by the
density parameter chi, the total coefficient sum across all bins is
bounded by the same E(U) times a fixed power of R:
sum_{i, slabs} mu(slab_i) chi^{-1} E_i <= C R^{C_2} E(U).
Why two. The frame estimate's right-hand side is
(prod_i E_i)^{beta/n} (sum_i E_i)^{1-beta}.
When the E_i are all comparable, budget 1 (a simple (sum E_i) bound)
suffices. When they vary, budget 2 (a weighted bound) is needed.
Together, the two budgets give, uniformly in the energy distribution,
(prod_i E_i)^{beta/n} (sum_i E_i)^{1-beta} <= C E(U) (bounded).
Summation over tuples.
The tuple class is partitioned by the dyadic "ell" values (density) into
O(log L)^n classes. In each class:
- Budget 1 bounds the energy sum.
- The frame-estimate factor H_g^{-2} bounds the "spread" cost.
- mu-mass per tuple is at most m chi^{1-beta} L^n.
Summing over the (polynomially many) tuples in the class:
lambda^2 M_class <= C_B m chi^{1-beta} h^{n + 2n/(n+1)}
E(U) * h^{-n} b_1^{2n/(n+1)} * (slack).
Slab summation. The two normal directions partition into b_1^{-1} slices
each, giving b_1^{-n} slabs per unit normal volume. Density rescaling
contributes b_1^{2n/(n+1)} (fractal-norm transformation from flattened to
original).
Scale-factor collapse. The three scale factors combine:
h^{n + 2n/(n+1)} * b_1^{-n} * b_1^{2n/(n+1)}
= h^{n + 2n/(n+1) - n} * b_1^{2n/(n+1)}
= h^{2n/(n+1)} * b_1^{2n/(n+1)}
= (h b_1)^{2n/(n+1)}
= L^{2n/(n+1)}, since h b_1 = L.
This is the same "algebraic miracle" as in Chapter 39's exponent
bookkeeping, manifesting again in the summation.
Final assembly (w:eq-finalL2).
Summing over classes, using the two budgets, and applying the
scale-factor collapse:
w:eq-finalL2: lambda^2 M_Y <= C_B m chi^{1-beta}
L^{2n/(n+1)}
D^{-c_s/2}.
The D^{-c_s/2} is EXACTLY half of the sparse estimate's D^{-c_s}
(Chapter 38). The factor 1/2 is there because not all of Y survives the
heavy-bin reduction with full mass; clause (i) preserves only a fraction,
so we lose a factor of 2 in the exponent and then again at Chebyshev.
Recalling that a fixed fraction of M_Y remained from the heavy-bin
reduction, and adjusting chi^{...} accordingly, gives w:eq-finalL2.
Closing the outer induction (on level).
The strong hypothesis of w:weak involves the parameters p, m, chi, D, g,
and the induction constant C_ind. The outer induction is on the level
parameter gamma = lambda L^e / chi^a.
- Base case: gamma <= 1. Then the level set is empty or trivial; the
bound holds trivially.
- Inductive step: assume w:weak holds at level gamma' = gamma - delta,
and prove at level gamma. Use w:eq-finalL2 (via frame estimate,
sparse estimate, two-budget summation) as the engine. Suppose the
theorem fails at level gamma. Then w:eq-counterlevel (the
"counterlevel" statement) yields a configuration whose gamma-parameter
is at least gamma. Apply the geometric induction of Section 7 (this
chapter): the flatten/wall/frame cases reduce to level gamma - delta,
contradicting the induction hypothesis, unless C_ind is huge.
- Closing: choose C_ind (the induction constant) larger than C_B in
w:eq-finalL2. Then the counterlevel is impossible, and the theorem
holds at level gamma.
Closing the inner induction (on sparsity).
The inner induction is on the sparsity parameter D. In the sparse
estimate (Chapter 38), the case D >= L^tau is handled directly. For
D < L^tau, the geometric induction reduces to smaller sparsity. This is
the sequence of "reduce D, recurse":
- Start with D_0 = L^tau (initial sparsity).
- Each level of the induction reduces D by a factor (from heavy-bin
reduction).
- At the terminal level (D ~ L^{C_s sqrt(log L)} ~ L^{o(1)}), apply the
Du-Zhang fractal estimate directly (Chapter 31) with the square-root
occupancy, closing the recursion.
The exponent sigma = c_s/2 in w:weak is fixed by the inner induction's
final gain.
The choice of p. Finally, p must satisfy:
- p > max(p_0, q_{n-1}) (so the p-th moment is available);
- p large enough that Chebyshev's inequality at the very end of the
"summation against mu" step gives a good bound.
The exponent order (Chapter 40's "order of choices") fixes p > 2 after
all other exponents. This ensures the Chebyshev argument at the very end
is tight.
Theorem w:weak is proved.
================================================================================
PART V: THE ENDPOINT (SECTION 8)
================================================================================
45. SECTION 8: THE COMMON TIME TREE AND THE SHORT-TIME ESTIMATE
--------------------------------------------------------------------------------
Section 8 (endpoint.tex) uses the weak estimate (Chapter 40) to prove the
main theorem (Chapter 6). There are three pieces:
- A "common time tree" that sums the weak estimate over frequencies
and time intervals.
- An L^1 endpoint maximal estimate that follows.
- A Gaussian representative construction that finishes the theorem.
This chapter is the time tree.
The frequency-local short-time estimate. Let N >= 1, let f_N be a
function on R^n whose Fourier transform is supported in a fixed annulus
around |xi| ~ N (say dilated by N of a fixed annulus). The Schrodinger
evolution S(t) f_N is frequency-localized to scale N.
The weak estimate (Chapter 40) applied at scale L = N^2 (choosing the
measure mu as Lebesgue restricted to a unit ball) gives:
PROPOSITION (short-time estimate; paper Proposition e:short-time). Let
p be the exponent from Theorem w:weak, and let s_n = n/(2(n+1)). Then
for every unit ball Omega subset R^n, every time interval I of length
1/N (centered anywhere), and every N-frequency-localized f_N,
|| sup_{t in I} | S(t) f_N | ||_{L^{p, infinity}(Omega)}
<= C N^{s_n} || f_N ||_2,
where the constant C does not depend on N, Omega, or I.
What makes it "short-time". The time interval has length 1/N, not 1.
Why the short interval: at scale N, the wave packet of frequency N moves
a distance ~ N over time 1/N. So over time 1/N, the packet has "just
barely" displaced; on a unit ball, the wave is not fully "spread".
Summing time intervals. To go from time intervals of length 1/N to the
full interval [0, 1], we partition [0, 1] into N intervals of length 1/N.
For each interval, we have the short-time estimate. Setting I = [j/N,
(j+1)/N], summing over j = 0, ..., N-1, using the subadditivity of
supremum and the "many intervals" bound, we get the "long-time" version:
|| sup_{0 < t < 1} | S(t) f_N | ||_{L^{p,infty}(Omega)}
<= C N^{1/p + s_n} || f_N ||_2 (or N^{s_n} * N^{1/p} * ...).
Actually, the "long-time" version is: for every unit ball Omega and every
f_N,
|| sup_{0 < t < 1} | S(t) f_N | ||_{L^{p,infty}(Omega)}
<= C N^{1 + s_n} || f_N ||_2 / N = C N^{s_n} || f_N ||_2,
where the "1/N" piece comes from sup_{t in I} being <= a convex
combination of |S(t) f_N|, via the short-time estimate at each interval.
The exact bookkeeping: sum over N intervals of length 1/N of the
short-time bound is at most N * (short-time bound) / |I|^{1/p}? Actually,
the right formulation is an L^p estimate over a partition of [0,1]; the
short-time estimate is on each interval, and "sum" in L^p is bounded by
(subadditivity) sum_j ||sup_{t in I_j}||_p. So:
||sup_{0<t<1}|S(t) f_N|||_p <= sum_j ||sup_{t in I_j}|S(t) f_N|||_p
<= N * C N^{s_n} ||f_N||_2.
Hmm, the N factor is bad. But we get to use weak L^p on Omega, not over the
time interval. So the summation over time intervals gives a factor, but
the localization of sup over t within each dyadic interval at scale N is
"interpolated". Actually, the paper's Section 8 does not use this crude
summation; it uses a "time tree".
The common time tree. The idea: use ONE common dyadic tree for all
frequencies, so that summing over N does not multiply costs. Specifically:
- Dyadic tree: at each level j = 0, 1, ..., log_2 N, split each time
interval into two. The tree has depth O(log N).
- For each frequency N, associate N to one specific "leaf" of the tree
(via a measurable function from R to {0, 1}^N).
- At each node of the tree, sum the frequency-localized sup over the
frequencies assigned to that node.
This is the "single measurable time graph for all frequencies". The
result (paper m:eq-roadmap-maximal in introduction) is:
int_Omega sup_{0 <= t < 1} | S(t) f(x) | dx <= C || f ||_{H^{s_n}},
initially for compactly supported f, uniformly in translated unit balls
Omega.
This is the L^1 maximal estimate at endpoint.
Why it is at L^1 (not L^p). The exponent s_n is the critical one. The
maximal operator with this Sobolev input is bounded from H^{s_n} to
weak L^1. The L^1 (or weak L^1) version is what we need for the
frequency summation.
The "summation by the time tree" is the main content of Section 8's
first part. Its details are about the correct exponents and avoiding
N^eps losses: the tree assigns frequencies to dyadic levels in a way that
is "measurable" in t, so that the summation sup_t is controlled by the
disjoint dyadic-tree structure.
What is "common" about the tree. If each frequency N had its own dyadic
tree (e.g., aligned per-frequency), the summation over N would lose a
factor that grows with N. Using a COMMON tree (fixed a priori) means that
the summation over frequencies at each node uses the SAME dyadic interval
for all frequencies in that node, and the estimates add up. This is why
the proposal in Du-Zhang (individual trees) does not give the endpoint
and the paper's common tree does.
This completes the frequency summation. The next chapter states the L^1
endpoint estimate it gives.
46. SECTION 8: THE L^1 ENDPOINT MAXIMAL ESTIMATE
--------------------------------------------------------------------------------
We continue Section 8. The time tree sums frequencies and time intervals,
returning the L^1 maximal estimate.
THEOREM (endpoint maximal estimate; paper e:local-endpoint or m:eq-
roadmap-maximal). Let s_n = n/(2(n+1)). For every compactly supported
f in H^{s_n}(R^n), and every unit ball Omega subset R^n,
int_Omega sup_{0 <= t < 1} | S(t) f(x) | dx <= C || f ||_{H^{s_n}},
with C independent of f and Omega.
The endpoint L^1 maximal estimate is the crux of Section 8. It is
"endpoint" because (a) it uses the exact exponent s_n (no epsilon slack),
and (b) it is at L^1 (the sharpest integrability, since s_n is exactly
the threshold). The Du-Zhang proof at s > s_n would give L^p with p > 1
at exponent s + eps; here we get L^1 at exactly s_n.
Proof structure.
From the time tree (Chapter 45), a maximal estimate at frequency N:
||sup_{0 < t < 1} |P_N S(t) f| ||_{L^p, infinity(Omega)}
<= C N^{s_n} ||f||_2.
for p = p from Theorem w:weak (a fixed finite p > 2), uniform in Omega.
Then use the "square function" or "Littlewood-Paley" structure:
||f||_{H^{s_n}}^2 = sum_N N^{2 s_n} ||P_N f||_2^2.
The maximal operator sup_t |S(t) f| roughly splits as a sum over N of
the maximal operators sup_t |P_N S(t) f|, but summing L^p weak bounds
across N is nontrivial (a "square function" or "Littlewood-Paley"
argument is needed).
Actually, the paper uses a "square function" argument: the maximal
operator sup_t |S(t) f| is bounded in L^1 by the L^2 norm of its
"frequency-averaged" version, and the frequency-averaged version is
summable over N via the maximal estimate. The bookkeeping uses the
"square function" bound:
||sup_t |S(t) f| ||_{L^1} <= C || (sum_N N^{2 s_n} |P_N S(t) f|^2 )^{1/2} ||_{L^1}
with some appropriate time-integral version, and then the L^2 sum over N
is controlled by the maximal estimate.
Details. The reader who wants the precise exponents should read the
paper's Section 8, "Short-time estimate, time tree, and local L^1
endpoint maximal estimate". The key estimates are:
- The short-time weak L^p estimate at frequency N (Chapter 45).
- The "time tree" summation over frequencies and time intervals,
avoiding N^eps losses (the tree is common).
- The "square function" bookkeeping to convert the L^p maximal
estimate over frequencies to L^1.
- The exact constant tracking to have no eps loss.
What the L^1 endpoint estimate gives. It is the input to the Gaussian
representative construction (Chapter 47). From the L^1 estimate and
Borel-Cantelli, one gets that S(t) f converges to f a.e. on a full-measure
set. But the L^1 estimate alone is not quite enough for the theorem's
"Gaussian regularization on one full-measure set uniformly in t" form;
that requires extra care, which the paper handles in Chapter 47.
47. SECTION 8: THE GAUSSIAN REPRESENTATIVE AND THE MAIN THEOREM
--------------------------------------------------------------------------------
We finish Section 8 and the paper. This is where the main theorem
(Chapter 6) is actually proved.
Setup. Let f in H^{s_n}(R^n). We want to define a representative u(x, t)
of the evolution, on a full-measure set E subset R^n, such that:
(a) sup_{0 < t < 1} | U_a f(x, t) - u(x, t) | -> 0 as a -> 0+;
(b) u(x, t) -> f*(x) as t -> 0+;
for every x in E.
Approximation by compactly supported smooth data. Let f_k be a sequence
of compactly supported smooth functions in H^{s_n} with f_k -> f in
H^{s_n}. For each k, S(t) f_k is a smooth function of (x, t) on R^n x
(0, 1), because f_k is compactly supported.
For each k, define:
E_k = { x : sup_{0 < t < 1} | S(t) f_k(x) | < infinity }
= R^n (for smooth f_k).
But we need a "good set" for each k, on which convergence to f_k is
controlled. The L^1 endpoint estimate (Chapter 46) provides this: the
operator sup_t |S(t)(f - f_k)| has L^1 bound C ||f - f_k||_{H^{s_n}}.
Chebyshev / Borel-Cantelli on the operator sup_t |S(t)(f_{k+1} - f_k)|
gives that, on a full-measure set E, the series of sup_t |S(t)(f_{k+1} -
f_k)| is summable. Hence S(t) f_k(x) converges uniformly in t for each x
in E, to a limit u(x, t).
The set E. Construct E as follows: for each integer m (a "level" of the
Borel-Cantelli argument), define
A_m = { x : sup_t |S(t)(f_{m+1} - f_m)(x)| > 2^{-m} }.
By the L^1 endpoint maximal estimate, |A_m| <= C 2^m ||f_{m+1} - f_m||_{H^{s_n}}.
Choose f_k so that ||f_{k+1} - f_k||_{H^{s_n}} <= 2^{-k} after taking
||f - f_k||_{H^{s_n}} <= 2^{-k} in a suitable summation scheme. Then
sum_m |A_m| < infinity, so by Borel-Cantelli, A_m occurs finitely often
for a.e. x. On the complement E (a full-measure set), we have
sup_t |S(t)(f_{m+1} - f_m)(x)| <= 2^{-m} for all large m,
so the sequence S(t) f_k(x) is uniformly convergent in t. Define
u(x, t) = lim_k S(t) f_k(x) for x in E, t in (0, 1). Then (a) holds
with U_a f replaced by S(t) f (and by extension for U_a).
The uniform-in-t conclusion. Because the convergence is uniform in t,
u(x, t) is continuous in t on (0, 1) for each x in E.
Gaussian regularization. The paper also needs to show that U_a f -> u
as a -> 0, uniformly in t, on the same full-measure set E. This is
NOT an immediate consequence of S(t) f_k convergence, because U_a f is
defined via Gaussian smoothing.
The key input is the "Gaussian approximate identity" (paper Lemma
e:gaussian-local): for a measurable function M and any a > 0, the
convolution K_a * M is finite everywhere, and at every Lebesgue point x
of M, K_a * M(x) -> M(x) as a -> 0. The proof is just the standard
Lebesgue differentiation theorem, applied to the symmetric Gaussian kernel.
Apply this lemma to M = sup_{0 < t < 1} |u(x, t)|. The result: for a.e. x
on E, we have
U_a u(x, .) -> u(x, .) as a -> 0, uniformly in t.
Combined with U_a f being a smooth approximation of u (from (a) for the
compact-supported approximation), this completes part (a) of the
theorem.
The Lebesgue value part (b). Part (b) needs u(x, t) -> f*(x) as t -> 0.
For the approximate sequence f_k, S(t) f_k(x) -> f_k(x) as t -> 0 for
smooth f_k. So u(x, t) -> f(x) as t -> 0, provided u is the "correct"
limit. The Lebesgue value f*(x) appears because the full-measure set E
on which convergence holds is best described in terms of Lebesgue
differentiation. The paper's Section 8, "One representative and all real
times", is where this identification is made.
The endpoint theorem. The main theorem (Chapter 6) is now proved:
For f in H^{s_n}(R^n), n >= 3, there is a full-measure set E and a
function u(x, t) on E x (0, 1) such that (a), (b), (c) of Chapter 6
hold.
The order of limits in the theorem (a -> 0 uniformly, then t -> 0) is
exactly what the construction gives.
Why the construction is "Gaussian". The Gaussian e^{-a|xi|^2} is chosen
because it is a bounded smooth cutoff that commutes with the evolution
in a clean way, and its approximate-identity lemma (e:gaussian-local) is
elementary. Any other smooth cutoff would work equally well; the Gaussian
is picked for exponential decay and clean constants.
Why the theorem is "uniform in t". Uniformity is what makes U_a f -> u
hold simultaneously in t, so that the order of limits in the theorem is
a -> 0 first, then t -> 0, and the result is deterministic. Without
uniformity, one could get different limits along different paths.
Why "one full-measure set". The set E is constructed once (Borel-Cantelli
on the sum over k of A_m), and all three parts (a), (b), (c) hold on the
SAME E. This is stronger than "for a.e. x there exist times t where
convergence holds"; it is "for a.e. x, convergence holds in t uniformly
on (0, 1)".
This completes the proof of the main theorem. Section 8 is the
culmination of the paper. Its inputs are:
- The weak estimate (Section 7, Chapter 40).
- The time tree (Chapter 45).
- The L^1 endpoint maximal estimate (Chapter 46).
- The Gaussian approximate-identity lemma (e:gaussian-local).
Section 8 is, in a sense, the most "classical" section of the paper:
once the weak estimate is in hand, the analysis is standard (Borel-Cantelli,
Lebesgue differentiation, approximate identities). The hard work is all in
Sections 2-7.
================================================================================
PART VI: WHAT IT MEANS: SUMMARY AND AFTERMATH
================================================================================
48. WHAT THE PAPER PROVES, IN ONE PARAGRAPH
--------------------------------------------------------------------------------
For every spatial dimension n >= 3, and every initial data f in the
Sobolev space H^{s_n}(R^n) with s_n = n/(2(n+1)), the free Schrodinger
evolution of f converges to f almost everywhere as time goes to zero.
The precise statement: there is a set E subset R^n of full measure and a
complex-valued function u(x, t) on E x (0, 1) such that the Gaussian-
regularized evolutions U_a f(x, t) converge to u(x, t) uniformly in t as
a -> 0, and u(x, t) converges to the Lebesgue value of f as t -> 0, for
every x in E. Equivalently, lim_{a -> 0} U_a f(x, t_a) = f*(x) for every
path t_a -> 0+ and every x in E. This is the sharp endpoint of Carleson's
problem in higher dimensions; the exponent s_n = n/(2(n+1)) cannot be
improved (Bourgain 2016). In dimensions n = 1 and n = 2 the corresponding
endpoint theorems are due to Carleson (1980) and Du, Guth, Li (2017)
respectively; the planar case is treated in the companion paper.
Read more compactly: the conjecture that the Sobolev exponent for
almost-everywhere pointwise convergence is exactly s_n = n/(2(n+1)) is
now a theorem in every dimension n >= 1, with n = 1 classical, n = 2
the planar companion paper, and n >= 3 the present paper.
49. WHAT REMAINS OPEN
--------------------------------------------------------------------------------
A few natural questions are not settled by the paper.
1. The unregularized version. The main theorem is stated for the Gaussian-
regularized evolution U_a f, with the limit a -> 0 taken uniformly in
t. The stronger statement, that the unregularized S(t) f also converges
pointwise a.e. at the endpoint (without any smoothing), is not
claimed here for n >= 3. In the plane, the unregularized version is
what the companion paper proves; but the higher-dimensional statement
is open as far as this companion knows. This is one of the most natural
remaining questions and would require more structure than the maximal
estimate alone.
2. The order of limits and admissible classes. The paper takes the limit
a -> 0 first, then t -> 0. Other natural orders (t -> 0 first for each
fixed a, then a -> 0) might give different limit points; the sharpest
admissible version of the problem is precisely what the paper settles.
Relatedly, it is natural to ask whether the theorem extends from
H^{s_n}(R^n) to other function spaces (Besov, Triebel-Lizorkin, and so
on). The maximal estimate does have Sobolev-differentiability
parameters, and the paper's weak estimate can be extended with minor
modifications, but a systematic treatment is not attempted here.
3. Quantitative endpoint. The paper gives a qualitative a.e. convergence
statement; the "quantitative endpoint" version (some measure of the
rate of convergence or the size of the exceptional set in terms of the
size of the Sobolev norm) is a more delicate question. For the
non-endpoint case some quantitative statements are available in the
literature; the endpoint is not yet similarly understood.
4. Lower dimensional and other equations. Endpoints in n = 1 (classical)
and n = 2 (planar companion) are settled. Analogues for other
dispersive equations (wave equation, Klein-Gordon, KdV, and so on) are
partially known but not in the same general form. The endpoint
phenomenon is specific to the parabolic scaling of Schrodinger; each
equation has its own geometric issues.
5. The frame estimate in more general settings. The transverse frame
estimate (Section 3) is tied to the paraboloid. Generalizations to
other hypersurfaces of positive curvature would be an interesting
project, and could be relevant to endpoint problems for other
dispersive equations.
6. Simpler proofs. The present proof is long. It is natural to ask
whether the same conclusion can be reached by a shorter route, perhaps
via a refined version of the Du-Zhang argument combined with new
insight. No shorter proof is known to this companion.
7. Sharp constants. The constants in the paper are effectively
computable but not sharp. Optimal constants for the endpoint maximal
estimate are likely related to the sharp form of the refined
Strichartz estimate (Du-Guth-Li-Zhang 2018), which is itself an open
problem in some dimensions.
50. GLOSSARY
--------------------------------------------------------------------------------
Notation from the paper, listed alphabetically by symbol.
Almost everywhere (a.e.):
For every point except possibly a set of measure zero.
Ball growth / ball parameter (m):
A measure mu has ball parameter m if mu(B(z, r)) <= m r^n for all
centers z and all r >= 1. This says mu "does not concentrate more
than n-dimensional Lebesgue measure would".
beta:
beta = 2/(n+1). The homogeneity of the fractal norm in the measure.
Bourgain-Demeter decoupling:
A family of estimates (Decoupling, Chapter 19) at exponent
q_k = 2(k+2)/k, bounding the L^p norm of a field by the l^2 sum
over its frequency caps.
Buffered refined decoupling (paper Sec 2):
A lossless version of decoupling that pays only a fixed power of a
buffering scale, instead of N^eps. Proved in the paper, replaces
the epsilon-losing decoupling needed for the endpoint.
Carleson's theorem (1D):
For f in H^{1/4}(R), the free Schrodinger evolution converges to f
a.e. (Carleson 1980). Sharp by Dahlberg-Kenig (1982).
Common time tree (paper Sec 8):
A single dyadic tree of time intervals, used for the frequency
summation in the endpoint argument, so that summing over frequencies
does not lose an N^eps factor.
Compact array / packet:
A piece of a wave localized in space and frequency. A packet of
scale N with frequency label v lives in a native unit cell of the
parabolic rescaling X_{N,v}(x, t) = ((x - 2vt)/N, t/N^2). The
paper uses compact arrays (Chapter 22) to organize wave packets.
D (sparsity parameter):
The tube moment parameter appearing in the sparse estimate
(Chapters 22, 38). Large D means mu "avoids" each packet
trajectory; small D means mu is dense on packets.
Decoupling:
The inequality ||H||_p <= N^eps (sum_J ||H_J||_p^2)^{1/2} for fields
with Fourier support split into caps J, valid for p up to
q_k = 2(k+2)/k (Bourgain-Demeter 2015).
Du-Zhang fractal estimate:
A refined Strichartz estimate at each compact scale, with the
square-root-scale occupancy gain built in (Chapter 31). The
starting point for the paper.
e:
e = n^2/(n+1). The critical amplitude exponent in the transverse
saving and in the weak estimate.
Envelope (B = B_b):
The "square-root sum" of a packet family, B(z)^2 = sum_T s_T^2
w_T(z). Controls the size of a packet sum at every point.
Fourier layer:
A component of a compact array after a Fourier split; used to
apply the Du-Zhang estimate (Chapter 32).
Frame estimate (paper Sec 3):
The main "transverse packet" saving: k+1 transverse fields of
packets cannot concentrate at many unit cubes if each is large.
Proved by the profile/incidence argument (Chapters 26-30).
gamma_n:
gamma_n = 2n/((n+1)(n+2)). The Du-Zhang "density gain" exponent.
Heavy bin:
A Fourier angular sector that carries at least a fixed fraction
of the weighted coefficient mass of the field at a point
(Chapter 42).
H^s (Sobolev space):
The space of functions f whose Fourier transform satisfies
INT (1+|xi|^2)^s |f_hat(xi)|^2 d xi < infinity.
Iitaka (no; that is a different paper):
(A note: the reader who saw the Iitaka companion in the same
directory should know that this companion is about Schrodinger's
equation, not Iitaka fibration. The corpus contains many such
companions; this one is self-contained.)
Information (base L):
I_Lambda(A|B) = H(A,B) - H(B), in base Lambda. Counts Lambda-ary
possibilities for A given B. Used in Section 5 (transverse saving).
Lebesgue value (f*):
f*(x) = lim_{r -> 0} average of f over B(x, r). Equals f(x) for
a.e. x (Lebesgue differentiation theorem).
Littlewood-Paley:
The decomposition of a function into dyadic frequency annuli, and
the corresponding square function. Used in the L^1 endpoint
maximal estimate (Chapter 46).
Multilinear Kakeya (BCT/Guth):
The estimate integral of product of tube weights at exponent
1/(d-1) bounded by a polynomial in the weights, subject to a
determinant transversality condition (Chapter 17).
Native coordinates:
X_{N,v}(x, t) = ((x - 2 v t)/N, t/N^2). The parabolic rescaling
that makes a scale-N packet at velocity v a unit box.
near-Cauchy theorem (paper Sec 2):
A point count at exponent q_k for compact packet sums (Chapter 24).
p_k, q_k:
p_k = 2(k+1)/k = 2m/k (linear Strichartz end);
q_k = 2(k+2)/k (Bourgain-Demeter decoupling end). Here m = k+1.
Polynomial partitioning (Guth-Katz):
The decomposition of a point set by a degree-D polynomial into
sign cells of mass at most C_n * (points) / D^n. Used in the
transverse saving's no-jump lemma (Chapter 37).
s_n:
s_n = n/(2(n+1)). The critical Sobolev exponent for a.e. pointwise
convergence of the Schrodinger equation in R^n.
Sparse estimate (paper Sec 6):
sum_q mu(q)^beta sup_q |U|^2 <= C m^beta L^{2n/(n+1)} D^{-c_s} E
for D >= L^tau (Chapter 38). Converts transverse savings into a
power of D^{-c_s}.
Strip saving (paper Sec 4):
For fields supported near an affine (n-2)-plane, a power saving
S^{-c_0 alpha} beyond the unweighted Du-Zhang estimate
(Chapter 33).
Transversal / transverse:
A configuration of direction vectors with determinant bounded
away from zero (Chapter 16).
Transverse saving (paper Sec 5):
The theorem that n+1 transverse unit amplitude fields cannot
concentrate on a dense set of cubes at the amplitude L^{-e}
(Chapter 35).
Tube moment condition:
The condition INT (1+|x - y_alpha - 2 t omega_alpha|/L)^{-M}
d mu(x, t) <= m L^{n+1}/D for every packet trajectory. This is
the "sparsity" of mu at scale L (Chapter 22).
U_a f:
The Gaussian-regularized evolution:
U_a f(x, t) = (2 pi)^{-n} INT e^{i x xi - i t|xi|^2 - a|xi|^2}
f_hat(xi) d xi.
Unit density (chi):
The parameter chi in mu(B(z, 1)) <= m chi. chi = 1 is "dense",
chi << 1 is "sparse". Controls the amplitude threshold chi^a g L^{-e}.
Wall:
A codimension-1 hypersurface on which a heavy packet field is in
"transition" — approximately as big as lambda — and which is not
locally affine. Either the wall has large second fundamental form
(angular case) or it is locally a plate (frame case). See Section 7.
Weak estimate (paper Sec 7):
The endpoint frequency-local maximal estimate at L^p: the main
output of Sections 3-7 (Chapter 40).
51. SOURCES
--------------------------------------------------------------------------------
Primary.
The paper.
"Endpoint pointwise convergence for the Schrodinger equation in
higher dimensions", September 24, 2026.
Predecessors on which the paper is built.
Carleson 1980.
Some analytic problems related to statistical mechanics. In
Euclidean Harmonic Analysis (Proc. Sem. Univ. Maryland 1979),
Lecture Notes in Math 779, Springer.
Dahlberg and Kenig 1982.
A note on the almost everywhere behavior of solutions to the
Schrodinger equation. In Harmonic Analysis (Minneapolis 1981),
Lecture Notes in Math 908, Springer.
Sjolin 1987.
Regularity of solutions to the Schrodinger equation. Duke Math
Journal 55(3): 699-715.
Vega 1988.
Schrodinger equations: pointwise convergence to the initial data.
Proc. AMS 102(4): 874-878.
Bourgain 2013.
On the Schrodinger maximal function in higher dimension. Proceedings
of the Steklov Institute 280(1): 46-60.
Bourgain 2016.
A note on the Schrodinger maximal function. Journal d'Analyse
Math (2016). The counterexample for s < n/(2(n+1)).
Du, Guth, Li 2017.
A sharp Schrodinger maximal estimate in R^2. Annals of Math,
Theorems 1.1-1.2. The planar endpoint (s > 1/3, plus endpoint).
Du, Guth, Li, Zhang 2018.
Pointwise convergence of Schrodinger solutions and multilinear
refined Strichartz estimates. Forum Math Sigma, Theorems 1.1 and 4.2.
Du and Zhang 2019.
Sharp L^2 estimate of the Schrodinger maximal function in higher
dimensions. Annals of Math, Theorem 1.1. The non-endpoint predecessor
s > n/(2(n+1)).
Du, Li 2024.
Recent improvements in higher-dimensional maximal estimates
(Theorems 1.1-1.2), for other L^p ranges.
Pan, Sun, Tan 2026.
Bessel-potential extension, still assuming s > n/(2(n+1)) at p = 2
(Theorems 1.3-1.4).
Technical inputs.
Bennett, Carbery, Tao 2006.
On the multilinear restriction and Kakeya conjectures. Acta Math.
Multilinear Kakeya.
Bourgain and Demeter 2015.
The proof of the l^2 decoupling conjecture. Annals of Math.
Decoupling.
Bourgain and Guth 2011.
Bounds on oscillatory integral operators based on multilinear
estimates. GAFA. The broad-narrow paradigm underlying Chapter 38.
Guth 2010.
The endpoint case of the Bennett-Carbery-Tao multilinear Kakeya
conjecture. Acta Math. Weighted multilinear Kakeya.
Guth 2016 (Restriction).
A restriction estimate using polynomial partitioning. JAMS.
Guth and Katz 2015.
On the Erdos distinct distances problem in the plane. Annals.
Polynomial partitioning, finite-set version.
He 2020.
Robust hyperplane projection estimate. Cited in Chapter 21.
Singh, Singh, Parmar 2026.
Refined decoupling, cited for the "collar gap" repair in Section 2.
Wongkew 1993.
Volumes of tubular neighbourhoods of real algebraic varieties.
Pacific J. Math. Cited for the tubular volume bound.
Basu 2014.
Algorithms in real algebraic geometry: a survey. arXiv:1409.1534v1.
Cited for real algebraic background.
Tao 2020.
Multilinear estimate for the paraboloid patch, Theorem 1.7.
Used for the multilinear Strichartz exponent.
Companion.
Planar version: "planar_schrodinger_endpoint_v2.txt". The companion
paper for n = 2. Uses many of the same ideas (frame estimate, transverse
saving) adapted to the planar case, and covers the endpoint 1/3.
Sibling docs in the same directory.
The file "Uniform-Pluricanonical-Iitaka-Fibrations-complete-standalone-
equivalent-explainer-for-16-year-olds-with-nothing-missing-ascii.txt"
uses a different format ("Exercise: question / answer" pairs) at the same
level of accessibility. The present companion is longer, uses linear
prose, and does not follow the Q&A format. A reader who likes one format
may like the other; they are independent.
A final note.
This companion is meant to be read alongside the paper, not instead of
it. Every step in Part IV corresponds to a proof in the paper, and in
each case the paper is the authoritative version. Where this companion
is more explicit than the paper, that is a bonus, not a substitute.
If you are the reader who has read all the way:
- The single most technical input is the frame estimate (Section 3,
Chapters 26-30). If you have followed Chapters 26-30, you have
understood the hardest part of the paper.
- The two savings (frame, transverse) are logically independent and
meet only in Section 7.
- The endpoint is bought by the absence of N^eps factor in each of the
four "buying" ingredients (near-Cauchy, buffered refined decoupling,
frame estimate, strip saving). Each one replaces a lossy standard
inequality with a lossless version.
- The final theorem (Chapter 47) is, once the weak estimate is in
hand, essentially classical analysis (Borel-Cantelli, Lebesgue
differentiation, approximate identities).