FOUR-DIMENSIONAL KAKEYA SETS HAVE FULL HAUSDORFF DIMENSION
A version of the proof written for a reader who has finished high
school mathematics. The theorem is the same. Every step of the
argument is here. The presentation is different: abstract objects
are introduced only after an intuitive explanation, and each
technical engine is preceded by a plain statement of what it
produces and why we want it.
SECTION 1. INTRODUCTION
1.1 The needle problem
Take a unit-length needle. Put it flat on a table. Can you turn it
all the way around, so that it points exactly the opposite way,
without ever lifting it off?
Yes: just spin it through a half circle of radius 1/2. The needle
sweeps out a solid disk of area pi/4. That is the obvious answer.
Kakeya's question, posed in 1917, was whether you could possibly do
better. He guessed not. The intuition is that turning the needle
in any direction forces it to sweep out area.
In 1919 Besicovitch answered, in a shocking way: yes you can do
better, and in fact you can make the swept area as small as you
like. Even zero. There is a set of area zero that contains a unit
segment in every direction. It is not anything you can draw, but it
exists, and the construction can be made fully explicit. This is
the Besicovitch set, or Kakeya set, or needle set. All three names
mean the same thing.
The Besicovitch construction is not just one clever trick. It is
the germ of a whole family of ideas. Cut the needle into short
pieces. Rotate a little bit, slide the center to a new clever
position, rotate another little bit, slide again. The short pieces
of the swept region overlap in a cunning way, and the total area
collapses to almost nothing. Perfected at every scale, it goes to
zero.
1.2 Kakeya sets in higher dimensions
The same question makes sense in dimension 3 and higher. A "unit
line segment" in R^n is a set of the form a + t*e for 0 <= t <= 1,
where a is a starting point in R^n and e is a unit vector (the
direction). A set K inside R^n is a Kakeya set if, for every unit
direction e, K contains a unit segment pointing in that direction.
That is, for every e there is some a (allowed to depend on e) with
{ a + t*e : 0 <= t <= 1 } inside K.
The set is not required to be closed, or open, or measurable, or
compact, or even nice in any way. It is a purely geometric
condition: the set must touch every possible direction with a
whole unit of line.
Two observations. First, such a set must be "big" in the obvious
sense that it has no preferred direction missing. Second, thanks to
Besicovitch, it can have Lebesgue measure zero. So ordinary volume
is the wrong way to measure size here. We need a finer notion.
1.3 Hausdorff dimension, on the ground floor
Think of covering a set by tiny balls. Give a number s >= 0, and
for a cover of the set by balls of radii r_1, r_2, r_3, ..., look
at the sum
r_1^s + r_2^s + r_3^s + ...
For small s, this sum tends to be large. For big s, it tends to be
small. Try a few examples.
A single point. You can cover it with one tiny ball of any radius r,
so the sum is just r^s. As the ball radius r shrinks to zero, r^s
goes to zero for every s > 0. The critical value is s = 0. We say
the point has Hausdorff dimension 0.
A unit line segment. Cover it by roughly 1/r balls of radius r. For
any s, the sum is about (1/r) * r^s = r^(s-1). If s > 1, this goes
to zero as r shrinks; if s < 1, it blows up. The critical value is
s = 1. Dimension 1.
A unit square. Cover by 1/r^2 balls of radius r. The sum is about
r^(s-2), finite at s = 2 and infinite below. Dimension 2. Same for
a solid cube: dimension 3.
The Cantor set, obtained by repeatedly removing the middle third of
a unit interval, is stranger. At step n you have 2^n intervals of
length 3^(-n). Cover by balls of radius 3^(-n): the sum is 2^n *
3^(-n*s). This is bounded exactly when 2 <= 3^s, i.e. when s >=
log 2 / log 3, about 0.63. So the Cantor set has Hausdorff
dimension log 2 / log 3, strictly between 0 and 1. It is bigger
than a point and smaller than a line. Its ordinary Lebesgue
measure is zero, but its Hausdorff dimension knows something.
So Hausdorff dimension is a number in [0, n] that measures how a
set fills space, in a way that survives sets with zero volume. It
is defined carefully as the value where a certain cover sum jumps
from infinity to zero, but the intuition is what we said: "at
scale r, about r^(-s) balls are needed."
1.4 The Kakeya dimension conjecture
Kakeya sets contain a unit segment in every direction. Given they
already "fill every direction", it feels like they should be as big
as possible: dimension n. That is the Hausdorff-dimension Kakeya
conjecture.
Conjecture. Every Kakeya set in R^n has Hausdorff dimension n.
The case n = 2 is a theorem of Roy Davies, from 1971. The
three-dimensional case was proved by Hong Wang and Joshua Zahl in
early 2025, and then simplified by Larry Guth, Wang, and Zahl in
2026. The four-dimensional case is the content of this paper. The
dimension n = 5 and higher remain open.
1.5 What was known before this paper
Return to four dimensions. The trivial bounds are: Hausdorff
dimension at least 3 and at most 4. Why at least 3? Because a
Kakeya set in R^4 covers every direction in S^3, which is a
three-dimensional sphere of directions. Forcing a whole segment for
every direction already gives lots of line.
In 1995, Tom Wolff improved the lower bound using a "hairbrush"
argument: pick one tube of the set, look at all the tubes that hit
it, and use that to get dimension at least (n+2)/2. In R^4 that is
3. That was stuck for a while.
Then came a sequence of improvements toward 4:
3 -> 3 + 1/40 (Guth-Zahl, Zahl, Katz-Rogers, 2018)
3.049 (maximal) (Katz-Zahl, 2019)
3.0543 (maximal) (Borges-Chan-Chen-Liu-Xi-Zhan, 2025)
3.059 (Katz-Zahl planebrush, 2019)
13/4 = 3.25 (sticky) (Rai Choudhuri, 2026)
Here "maximal" means the bound comes from a Kakeya maximal-function
estimate, which is a stronger kind of analytic result. "Sticky"
means the case where the family of witnessing lines is not
arbitrarily small: it has its own dimensional size. The general
Kakeya set, with no stickiness assumption, was stuck around
3.059 in 2026. This paper closes the gap.
1.6 The theorem
Theorem 1.1. Let K be a subset of R^4. Suppose that for every
unit vector e in S^3 there is a point a_e in R^4 such that
{ a_e + t*e : 0 <= t <= 1 } is contained in K.
Then K has Hausdorff dimension 4.
There is no measurability assumption, no compactness assumption,
and no stickiness assumption. The choice of witnessing segment may
be completely arbitrary, nonmeasurable, and wild. The proof has to
work with no regularity at all, and it does.
1.7 How the proof works, in outline
The plan has three stages.
First, reduce. A Kakeya counterexample -- a set with Hausdorff
dimension below 4 that still has a segment in every direction --
gives a certain weighted problem about many line segments and
their incidences. If the original set had dimension 4 - sigma, with
sigma > 0, the resulting problem has a quantity called the horizon
that sits above zero. The horizon measures the smallest "time cost"
of any approximate local model. Positive horizon means the problem
is genuinely resisting compression, exactly as a dimension deficit
would predict.
Second, contradict. Show that a weighted problem with positive
horizon cannot exist. The proof follows a nearly optimal sequence
of local models, called charts, and extracts a "critical path":
nested charts whose density profile is linear and whose time rate
is a positive number k. This critical path has a single extra
number, its narrowness ell, which measures how much thinner the
smaller spatial direction is than the time scale. Narrowness is 0
when the charts are nearly isotropic, is infinite when they are
degenerately skinny.
Third, case analysis. There are three cases:
ell = 0 the chart is isotropic
0 < ell < infinity the chart has a definite aspect ratio
ell = infinity the chart is degenerately narrow
In every case, the paper constructs a strictly better chart: one
with smaller time cost, or with the same time cost but extra
narrowness, on a definite fraction of the original weight. "Better"
contradicts the extremality used to extract the critical path. This
proves the contradiction, and with it the theorem.
The construction of the better chart is where the hard work lives.
The two main analytic engines are:
polynomial fits, and
multilinear Kakeya estimates.
A polynomial fit is a low-degree polynomial in the spacetime
variables that matches the incidence data to within a small
tolerance. If it can be shown to exist, it is a compact
representation of a huge number of line-segment constraints. The
polynomial fit machinery in Sections 5 and 6 is where the argument
does most of its heavy lifting.
A multilinear Kakeya estimate is a strong analytic bound on how
many tubes in independent directions can simultaneously cover a
region. It is used to control how the segments concentrate in
space. The endpoint of the multilinear Kakeya theorem was proved by
Larry Guth in 2010 and appears in Section 6 as the determinant
Kakeya estimate.
Put together, the proof looks like this:
- Build a hypothetical Kakeya counterexample.
- Reduce to a weighted problem with positive horizon.
- Extract a critical path with positive time rate k.
- Classify the path by its narrowness.
- In every case, construct a better chart on a substantial
residual (this is the hard analytic part).
- The better chart contradicts extremality.
- Therefore the hypothetical counterexample cannot exist.
Each piece of the reduction is a careful statement about how mass,
density, and geometric hypotheses behave under normalization,
composition, and restriction. Sections 2 and 3 build those
statements. Sections 4 and 5 set up the critical path and the
scalar tools. Sections 6 through 9 do the three case analyses.
1.8 Reading guide
For a reader meeting this material for the first time, here is how
the sections stack up.
Section 2, "Weighted models and the reduction from arbitrary
Kakeya sets," is where the geometric problem becomes an
incidence-counting problem. It is technical but each definition
has an intuitive meaning that you can keep in mind. Read the
statements of Theorems 1.1 (main theorem) and 2.1 (reduction)
carefully. The rest of the section is bookkeeping.
Section 3, "Comparison of algebraic sheets," proves a technical
lemma about polynomial roots. You can skim this on a first read.
Its purpose is to say: two polynomial-root-shaped functions that
agree at many sampled points must agree on a whole neighborhood.
Section 4, "Critical paths," sets up the extremal argument: the
maximizing path and the narrowness classification. This is the
structural heart of the proof.
Section 5, "Scalar concentration and polynomial fitting," is where
most of the analytic work happens. It develops the polynomial fit
machinery. It is dense. But its two top-level results --
Lemma 5.1 (Scalar coefficient clustering) and Proposition 5.10
(Higher-jet polynomial fitting) -- say something geometric: an
incidence family with little scalar support must cluster into a
small number of coefficient bins, and from the coefficients you
can build a low-degree polynomial fit.
Section 6, "The isotropic improvement," handles the narrowness
case ell = 0. It is where multilinear Kakeya first appears. It is
long but its structure is: build a velocity field, interpolate it,
make it approximately conservative, and restore the degree
required for a chart.
Sections 7, 8, and 9 handle the case where narrowness is finite
and positive. Section 7 is a set of planar rigidity estimates
(imported additive-combinatorics machinery). Section 8 reduces
the finite-narrowness case to a horizontal (scalar) model.
Section 9 is the case analysis by critical time rate; there are
three sub-cases, below / at / above the critical rate.
Section 10 handles ell = infinity, using a scalar projection. It
also assembles the entire proof for all narrowness values.
Section 11 collects the geometric consequences: bounds on other
notions of dimension, extensions to other direction sets, Nikodym
sets, curved Kakeya sets.
Every section has a short paragraph at the start saying what the
section produces and why it is needed. If you get lost, read those
paragraphs and the theorem statements -- that is the skeleton of
the argument. The proofs are where you go when you want details.
SECTION 2. WEIGHTED MODELS AND THE REDUCTION FROM ARBITRARY
KAKEYA SETS
2.0 What this section does
This is where we replace geometry with a random-sampling problem.
A Kakeya set is continuum-many line segments; we cannot handle
that directly. So we do three things.
First, we replace K by a family of probability distributions over
line segments, sampled from K by a random process.
Second, we sample a random time on each selected segment, and ask
whether the resulting spacetime point landed in one of the tiny
balls that almost-covers K.
Third, we count. The density of selected points at each resolution
scale gives us a "density exponent" f(r). The geometry of K
imposes inequalities between these exponents. The result is a
class of probabilistic problems -- the "cap class C(d)" for some
d < 1 -- with positive "horizon".
The horizon is the smallest time cost of any local model. The rest
of the paper shows this class of problems is empty. Section 2 is
just the setup.
2.1 Goal: the reduction theorem
Theorem 2.1 (Reduction). Let K be a Kakeya set in R^4 with
Hausdorff dimension below 4. Then there exist d in (0, 1) and
S, eta > 0 such that some exact problem in the cap class
C(d; S, eta) has positive horizon.
The proof occupies Subsections 2.2 through 2.8. We build the
machinery first, then assemble the proof at the end.
2.2 Exact problems
An exact problem is a sequence of probability laws indexed by a
resolution parameter N_0 that tends to infinity.
A "trajectory" is one of two kinds.
Version 2 (main model): y(t) = y_0 + t V, w(t) = w_0 + t W,
with y in R^2, w, W in R, and |y| + |V| bounded by K_0.
Version 1 (scalar model): y(t) = y_0 + t V, and
w(t) = w_0 + w_1 t + w_2 t^2, with y in R and
sup_{0 <= t <= 1} |w| bounded by K_0.
Both models come with a "native signal" X(t) = P_*(t, y(t), w(t)),
a degree-2 polynomial in t. In version 2, the polynomial P_*
satisfies: |X| <= K_0 everywhere on the trajectory; |d_w P_*| is
at most K_0 B; at selected incidences |d_w P_*| is at least B/K_0;
and |d_ww P_*| is at most K_0 B^2. The parameter B is the
"derivative scale". In version 1, the test is simply w - F(t, y)
for a quadratic F, and B = 1.
The "worlds" are finitely many auxiliary parameters gamma, with
weights omega_gamma adding up to one. The trajectory prior in
world gamma is nu_gamma. An "incidence event" E_gamma is a subset
of trajectory-times-t space.
The selected incidence measure is
mu = sum over gamma of omega_gamma * 1_{E_gamma}(l, t)
* nu_gamma(l) * dt.
We assume the total mass of mu is at least K_0^{-1}. If the mass
is lighter than that, we throw the world away; the problem is not
considered.
The "observation" at depth r is the tuple
p_r = ( gamma, t, y(t), [w(t)]_{a_r/B} ),
where a_r = N_0^{-r} and [.]_h denotes a dyadic grid of width h.
So at depth r we know the world, the time, the base y exactly, and
the hidden w only up to width a_r/B. Increasing r refines the
hidden observation.
The "density" of observations at depth r is measured per unit
dt dy and per hidden bin, and is denoted n(r) = N_0^{-f(r)} up to
subpower factors. The function f is called the density profile.
A phrase that keeps appearing is "subpower". A number K_0 >= 1
is subpower if log K_0 / log N_0 -> 0 as N_0 -> infinity. So K_0
grows slower than any fixed power of N_0. The point is that
subpower factors are invisible to power counting: if something
is bounded by K_0 * N_0^alpha, then, for any fixed eps > 0 and
N_0 large enough, it is also bounded by N_0^{alpha + eps}. We
are sloppy about them everywhere. When two quantities are equal
up to subpower factors, we write A ~ B.
The "cap class C(d; S, eta)" is defined by two inequalities. Fix
d in (0, 1) and S, eta > 0. The problem is in the class if, for
all 0 <= r <= L,
f(L) - f(r) <= d (L - r),
and the joint density of (p_L, [V]_{a_u}), at angular depth u in
(0, eta L], is at most n(L) N_0^{-Su+o(1)}. The first inequality
is the one that matters. It says: refining from depth r up to
terminal depth L loses at most d per unit of depth. The factor
d < 1 is the "compression budget" the problem allows. The second
inequality is an angular cap: the velocity V of a terminal
observation is confined to balls of angular width N_0^{-u}, with
density at most N_0^{-Su+o(1)} times the terminal density. This
restrains how much incidence mass is concentrated near any one
direction, at depth u.
What are these inequalities for, intuitively? The cap inequality
f(L) - f(r) <= d (L - r) says the density profiles are almost
linear. As we refine the hidden observation, we lose at most d
per unit of refinement. Since d < 1, the loss is strictly smaller
than the refinement scale. That gives room to compress.
2.3 Homogenization: why we can fix a density profile
Right after defining exact problems, we want to work with one
fixed function f(r) rather than a family that changes with N_0.
The "homogenization lemma" says this is fine: on a subsequence,
and after restricting to a subpower fraction of the total mass,
the density profile can be replaced by a fixed function f(r),
which is monotone nondecreasing in r and 1-Lipschitz.
Lemma 2.3 (Homogenization). On a subsequence and after a
restriction that retains inverse-subpower mass, an exact
problem has a homogeneous profile f(r) on every compact depth
interval. It is nondecreasing and 1-Lipschitz.
Why monotone? Because a coarser observation is a union of finer
ones; its density is a sum, so it is at least as large as any one
finer bin's density.
Why 1-Lipschitz? Because refinement from depth r' to depth r
subdivides a bin into at most N_0^{r'-r} sub-bins (up to subpower
factors), and the density over the union can grow by at most that
many times. So f(r') - f(r) <= r' - r, i.e. 1-Lipschitz.
The homogenization argument is a pigeonhole. At each resolution
N_0 we quantize the log-densities at finitely many tested depths
and pick the most common "type" (the vector of quantized
densities). There are N_0^{o(1)} types, so the winning type
carries inverse-subpower mass. Passing to a subsequence makes the
quantized log-densities converge, and the limit is f(r). The
monotonicity and 1-Lipschitz bounds pass to the limit.
Joint refinements by velocity are handled by the same argument
on a two-parameter family f(r, u).
The proof also gives an important structural fact. If a piece of
the problem has become light under a restriction (say, its mass
dropped below an inverse-subpower threshold), we can remove it
using a purely deterministic table of predicates. This is how we
"prepare" a problem: remove all the exceptional pieces once, then
work with the rest.
2.4 Time charts and the true floor
A "chart" is a local model of the problem: a collection of
trajectories that live in a moving box, on a common time interval
I, and that all pass a common degree-2 polynomial test.
Take a depth r in (0, L]. Fix a dyadic partition of the time axis
into intervals of length
q = N_0^{-h + o(1)}, h >= 0.
The exponent h is the "time cost" of the chart; smaller h means
longer time intervals, which is what we want.
Inside each interval I and each world, a chart system partitions
the surviving trajectories into disjoint label sets A_c. It
requires:
- an affine spatial center y_c(t);
- an invertible matrix D_c, with ||D_c|| bounded by q K_0 and
Jacobian J_c = |det D_c|;
- the bound sup_{t in I} |D_c^{-1}(y(t) - y_c(t))| <= K_0 on
the assigned trajectories;
- a degree-2 polynomial P(t, y, w) that is small on them.
"Small" here means:
|P(t, y(t), w(t))| <= K_0 a_r for t in I,
|d_w P(t, y(t), w(t))| <= K_0 B for t in I,
|d_w P(t, y(t), w(t))| >= B / K_0 at selected incidences,
|d_ww P(t, y, w)| <= K_0 B^2 / a_r globally.
The local box condition lets the chart's center follow the
trajectories over the interval; the box is allowed to be long (up
to q) in every direction. The Jacobian J_c measures the spatial
volume of this box.
Two numbers summarize a chart:
nu_c = nu_gamma(A_c), the prior mass of the assigned set;
m_c = mu_gamma(selected incidences in this chart).
There is an unavoidable budget: the assigned sets are disjoint
within one time interval, so
m_c <= q nu_c, and sum_c omega_(gamma(c)) q nu_c <= 1.
The basic "goodness" condition for a chart is that its incidence
mass is not too small compared to what the ambient density would
suggest. That is the "true floor":
m_c / q >= K_0^{-1} n(r) J_c.
A chart satisfying this is called true. A chart system is true if
every chart is true.
The key structural lemma is that a true chart can be assumed to
saturate: its two masses are comparable.
Lemma 2.4 (Saturation). A true system may be restricted,
retaining inverse-subpower mass, so that every chart satisfies
m_c / q ~ nu_c ~ n(r) J_c, (up to subpower),
and its label is recoverable from p_r in a subpower list.
Idea of proof. The trivial budget already gives m_c/q <= nu_c.
The true floor gives the other direction. For the middle term,
one uses a sublevel-set bound (a degree-2 polynomial with bounded
first derivative and small value has bounded sublevel set)
combined with the density cap and the trajectory prior. The
label-knowledge part follows from analogous reasoning applied
to the observation p_r itself.
2.5 Composition and the horizon
The chart machinery is useless unless it composes: a chart at
shallow depth, followed by a chart in the normalized coordinates
of the first chart, should produce a chart in the original
problem.
That is what the "chart normalization" lemma (Lemma 2.5) says.
The correspondence is:
q_(cd) = q * q_d,
J_(cd) = J_c * J_d,
m_(cd) = q nu_c * m'_d,
m_(cd) / q_(cd) = nu_c * m'_d / q_d.
So masses and Jacobians "multiply through" cleanly. The cap class
is preserved: if the original problem is in C(d; S, eta), so is
the normalized problem, with the same d and (up to subpower
changes in the tolerance) the same S, eta. The new true floor
follows from the old true floor and the normalization formula.
Lifting a true terminal chart in the normalized problem gives a
true terminal chart in the original, with composed time cost.
The "horizon" H(E) of an exact problem E is the minimum, over all
true terminal systems of E, of the exponent h. Equivalently, the
smallest time cost any true terminal chart system can achieve.
Two universal facts bound the horizon for every exact problem, no
matter how nasty. They are discussed next.
2.6 Two universal horizon bounds
Proposition 2.6 (Initial horizon bound). Every homogeneous
exact problem satisfies H(E) <= L.
The proof uses the native test P_* directly. Choose the time
scale q ~ a_L, so that the native quadratic polynomial is small
on the scale of the box. The value bounds and derivative bounds
of the exact problem already give the chart conditions. What
remains is to find spatial boxes with enough incidence mass. The
weighted plank estimate (Subsection 2.7) supplies those boxes.
The result is a true terminal chart system with h = L.
Theorem 2.1 (Restated). A hypothetical Kakeya set with
Hausdorff dimension below 4 produces an exact problem in the
cap class C(d) for some d < 1 with positive horizon.
We prove this in Subsection 2.8. The two facts together say:
every problem has H(E) <= L (always), and any counterexample has
H(E) > 0. So a counterexample would force the horizon to sit in
the open interval (0, L]. The rest of the paper shows that no
problem occupies this interval, so no counterexample exists.
2.7 The weighted plank estimate: importing 3D Kakeya
The paper has exactly one imported black box from the analytic
theory of Kakeya sets. It is the discretized, weighted,
full-time version of the 3D Kakeya maximal estimate of Wang and
Zahl, as restated in their streamlined proof by Guth, Wang, and
Zahl.
Lemma 2.7 (Weighted full-time plank estimate). Fix a bounded
affine-line chart in R^2. For every eps > 0 there is eta_eps
> 0 and C_eps < infinity with the following property. Given
finitely many affine traces M_i(t) = b_i + t u_i in the
chart with positive weights omega_i, and put rho = D^{-1}.
Suppose that every full-time plank P (of width a rho by
b rho, 1 <= a <= b <= D, with arbitrary orthonormal frame
and center/velocity) satisfies
sum_{i : M_i inside P} omega_i <= A a b.
Suppose each trace i is marked for at least D^{1-eta_eps}
time bins of size rho. Then the total weighted marks obey
sum_i omega_i #J_i <= C_eps A D^eps #E_vis,
where E_vis is the number of visited time-space mesh cells.
The statement is technical, but the intuition is short. If the
traces are geometrically spread out -- not concentrated in planks
-- then their weighted marks distribute across many cells. That
is the whole content. Compare with the case where a whole cluster
of traces sits in one short-wide box: then their marks can be
tiny in number, and the estimate would be false. The plank cap is
exactly the hypothesis that rules out this concentration.
The estimate is the "endpoint" of the multilinear Kakeya theorem
of Bennett-Carbery-Tao, proved by Guth in 2010 and made into a
weighted form by Bourgain-Guth. It descends to the discretized
full-time statement above by a standard rounding argument.
Why is this the specific input we need? Because building a true
chart at terminal depth L amounts to building a polynomial fit
whose incidence mass is not confined to a small spatial region.
That is exactly what the plank estimate gives you. The chart's
2D slices satisfy the plank cap -- that is precisely the 3D
Kakeya conclusion, and it is why the 4D theorem can be proved by
using 3D as a black box.
The same estimate is used twice more. Once in Proposition 2.6,
to build the trivial terminal charts with h = L. And again in
Section 6 as the "determinant Kakeya estimate," which controls
Kakeya concentrations directly.
2.8 Proof of the reduction theorem
We sketch the four stages. Full details are in the original.
Stage 1. Finite selection from outer measure.
Fix a bounded chart of directions in S^3 (a small open box in
the three-dimensional sphere of directions). For directions e in
this chart, K has a unit segment a_e + [0, 1] e. Take a small-ball
cover of K at scale delta = 2^{-k}. Because dim_H K < 4 - sigma,
we may assume the number of balls is at most C delta^{-(4 -
sigma)}.
Count the "productivity" of each direction: how much time of its
unit segment falls inside the cover balls at scale delta. Since
the segment has length one and the cover is total, at least c/k^2
of its time falls in this scale for some k depending on the
direction. Outer subadditivity gives: over countably many k and
balls, one combination has positive outer angular measure a. This
is the "productive directions and productive scale" selection.
Take delta-separated directions in the selected set; there are
between c k^{-2} delta^{-3} and C delta^{-3} of them. For each,
choose one witnessing segment. This is the only witness choice,
and it is finite.
Stage 2. Smoothing and three density bounds.
Perturb each selected segment's endpoints by a small multiple of
delta; this "smoothes" it into a small box. Summing smoothed
witnesses over the selected directions gives a version-2 exact
problem at resolution N_0 = delta^{-1}, with
L = 1, P_* = w, B = 1.
The selected incidences are those whose smoothed time lies inside
the enlarged cover balls. Selected mass is at least c k^{-2}, up
to subpower. Nothing is lost yet in a serious way.
Let f(r, u) be the joint density exponent of (observation at
depth r, velocity at depth u), and put g = r - f. Then three
inequalities hold:
g(1, 0) >= sigma,
g(0, 0) <= 0,
g(r, u) <= r + 2 - 2u.
First inequality. An enlarged ball uses O(1) hidden w-bins over
base (t, y)-volume O(delta^3). The total state volume is at most
C delta^{-(4 - sigma)} * delta^3 = C delta^{-(1 - sigma)}. Since
mass is inverse-subpower, n(1, 0) is at least delta^{1 - sigma}
up to subpower, so f(1, 0) <= -(1 - sigma) and g(1, 0) >= sigma.
Second inequality. The projected y-lines satisfy a full-time
plank cap at every dyadic mesh: any spatial plank of half-widths
a rho, b rho contains projected y-traces of total prior mass at
most K_0 a b rho^2 (this is the "planar" fact observed in the
original). If n(0, 0) grew by a fixed positive power, the weighted
plank estimate (Lemma 2.7) would break this cap, so g(0, 0) <= 0.
Third inequality. The velocity marginal has bounded range and
prior density at most K_0. At fixed t, the conditional y-density
is at most C delta^{-2}, because the smoothed segments are of
width delta. So the joint density is at most K_0 N_0^{2 - 2u};
hence g(r, u) <= r + 2 - 2u.
Stage 3. Endpoint selection.
Choose small fixed R, S > 0 (for example R + 4S < sigma/2) and
maximize g(r, u) - Rr + Su on [0, 1] x [0, 4]. The maximizer
(r_0, u_0) satisfies r_0 > 0, u_0 < 4, g(r_0, u_0) >= sigma/2.
Restrict the problem to the maximizer's scales. The maximization
inequality acts like this: the local problem lies in C(1 - R) at
angular parameters S, eta = (4 - u_0) / (2 r_0). Refining from
depth r up to r_0 loses at most (1 - R)(r_0 - r), and the angular
cap holds. Length becomes r_0.
Stage 4. Positive horizon.
Take any true terminal chart in this restricted problem. Its mass
in the ORIGINAL unnormalized problem is at least
m_old,c / q >= K_0^{-1} m(r_0, u_0) a_{r_0} J_phys,
where J_phys is the chart's physical-area Jacobian.
Now compare with the total mass accessible via the "slope volume".
For a chart on a time interval of length q, the slopes that could
possibly account for its incidence mass -- slopes whose lines fit
the chart's spatial box AND pass the small-value and first-
derivative tests -- form a set of "Euclidean slope" measure at
most
|Sigma| <= K_0 a_{r_0} J_phys / q^3.
The reason is that the chart's tests confine the endpoints of
unit-time traces to a small region, and a Jacobian argument
(quadratic in the derivative scale, cubic in the time scale)
bounds the accessible slope volume.
Then the original slope-density ceiling (from Stage 2, giving the
prior an upper bound K_0 in slope) bounds the prior mass of the
chart by K_0 |Sigma|. Comparing with the lower mass floor gives
m(r_0, u_0) <= K_0 q^{-3},
that is, q^3 <= K_0 / m(r_0, u_0). Since m(r_0, u_0) >= N_0^{sigma/2}
(from Stage 3), we get
q <= N_0^{-sigma/6 + o(1)},
i.e. h >= sigma/6 > 0. The horizon of this problem is positive.
This closes the reduction. Every Kakeya counterexample in R^4
yields a homogeneous exact problem in the cap class C(d) for some
d < 1, with positive horizon H(E) >= sigma/6. The rest of the
paper shows no such problem exists.
SECTION 3. COMPARING ALGEBRAIC SHEETS AND EXTENDING TERMINAL
TIME
3.0 What this section does
The charts in Section 2 use polynomial tests of degree 2. When we
build a better chart, one common strategy is: "extend the fit at
terminal depth L backward through a longer time interval."
Concretely, we want to replace a chart on a short time interval
with a chart on a longer time interval, using the SAME polynomial
test (up to small rescaling and composition with a change of
coordinates).
For this to work, we need a statement of the following form:
If two polynomial-fit solutions agree to small error on
many sampled incidences, then they agree everywhere, up to
a slightly larger error.
That is the content of this section.
The key objects are "algebraic sheets": scalar functions f(u)
that satisfy a polynomial equation Q(u, f(u)) = 0 for a nonzero
polynomial Q of bounded total degree. The roots of a degree-2
polynomial in w, for example, are algebraic sheets of degree 2.
More general sheets arise when we pass to complex extension, or
when we differentiate, iterate, or compose the fits.
Everything is done in the complex plane, because that is where
polynomial roots behave nicely. The main lemma is a "Bernstein-
Remez"-type inequality for algebraic functions: it bounds a
sheet's norm on a comparison ball by its norm on a smaller piece
of it. That is exactly the rigidity we need: agreement on a
large set forces agreement everywhere.
Read this section on a first pass as:
3.1 -- an analytic fact about algebraic sheets (bounded
extension, derivative bounds).
3.2 -- a geometric fact about extending a sheet along a
path (uniform cost even with tiny leading
coefficients).
3.3 -- an entropy fact about two lists of sheets agreeing at
many points (impossible if the lists are too short).
3.4 -- the combination used later: a scalar-fit comparison
on a real box.
3.5 -- the actual terminal extension: turn a terminal chart
into an atlas at slightly smaller time cost.
Only 3.5 (and implicitly 3.3 and 3.4) is used in the rest of the
paper; 3.1 and 3.2 are the technical inputs.
3.1 Algebraic sheets and the norm comparison lemma
Say f is an "algebraic sheet of degree at most D" if there is a
nonzero polynomial Q(u, Z), of total degree at most D in the
complex variables (u, Z), such that Q(u, f(u)) = 0 for all u in
the domain of f.
Examples: a constant function (Q = Z - c), a square root
(Q = Z^2 - g(u)), any root of any low-degree polynomial in Z
whose coefficients are low-degree polynomials in u.
The degree is bounded but the coefficients are allowed to be
arbitrary real numbers, including very small or very large ones.
That last point matters: we cannot rule out the relevant equations
having tiny leading coefficients. The rest of the section is
about being careful when they do.
The main lemma is a quantative "value + derivative bound" and a
"Remez-style" lower bound. It comes from the classical Bernstein-
Remez theory of algebraic functions, but the specific form we
need is:
Lemma 3.1 (Algebraic analytic norm comparison). Let Omega
be a fixed ball or polydisc in C^n, B a fixed closed real
ball inside Omega, and K a compact subset of Omega. For
algebraic sheets f of degree at most D, normalized so that
||f||_B = 1, the following bounds are uniform:
(a) for every fixed integer q,
max_{|alpha| <= q} ||d^alpha f||_K <= C_q,
(b) at every z in K,
max_{|alpha| <= D} |d^alpha f(z)| >= c_K > 0.
If B_1 is another fixed real ball in Omega, and E is a
measurable subset of B with positive relative measure
theta = |E|/|B|, then
||f||_{B_1}
+ max_{|alpha| <= q} ||d^alpha f||_{B_1}
<= C_q theta^{-C} ||f||_E.
These bounds are homogeneous (no normalization needed).
They are unchanged, up to rescaling derivatives, by affine
changes of coordinates.
The first two parts say: a normalized sheet has bounded
derivatives (of any fixed order) and cannot be locally
flat-to-all-orders -- some derivative of order at most D is
nonzero at each point. The final estimate is the Remez
inequality: a sheet's norm on a smaller ball is at least a
polynomial fraction of its norm on a bigger ball, provided the
smaller ball has positive measure.
What does that mean concretely? Take a degree-2 sheet, e.g. the
square root of a bounded analytic function. Such a sheet cannot
have a "spike": if it is order 1 in norm on a full ball, then it
must be order 1 everywhere, up to a polynomial loss in the
measured fraction of the ball. If it were tiny at a positive-
measure set of points, it would have to be tiny everywhere,
because the relation Q(u, f(u)) = 0 would force it.
Proof idea. Compactness: take a sequence of normalized sheets
with no holomorphic limit of the required type. Normalize
relation coefficients. On a subsequence they converge to a
limiting relation Q. That relation must actually involve Z
(otherwise f would escape), so the roots of Q(u, .) lie in a
bounded set on most of Omega minus {coefficient zeros}. A
continuation argument along a path avoiding the zeros shows
that the bounded root cannot escape; complex extension and
maximum principle finish the job. The quantitative Remez-style
part is a standard sublevel estimate on the resulting
algebraic function.
3.2 Polynomial cost of continuation
The next question is: if we have a stable branch on a small
complex ball, and we want to continue it along a path to a
distant ball, how much can its norm and derivatives change?
The answer, for the specific sheets we need, is that the change
is bounded by a fixed power of the "clearance" along the path.
Lemma 3.2 (Polynomial cost of continuation). Fix base dimension,
polynomial degree, and the number of sheets. Work in a fixed
bounded complex base region. Let A >= 2. Suppose the linear
coefficients or discriminants of the relevant equations stay
nonzero throughout an A^{-1}-neighborhood of a piecewise-linear
path. On endpoint balls of radius c/A, the norms of a continued
function compare in both directions with loss A^C, where C
depends only on the fixed degree and dimension. Fixed derivatives
are controlled by these norms.
The point is that the loss A^C is independent of the numerical
values of the coefficients. Even if the discriminant of a
quadratic is A^{-100}, as long as it is nonzero on the tube, the
continued root is well-defined and its norm ratio is at most a
fixed power of A.
Notice how the "admissible path" condition is used: it says the
path must stay away from the zeros of the relevant coefficients,
by a specific distance. In our applications, we can always choose
such a path, because the zeros are a dimension-1 complex set and a
generic broken path avoids them.
The proof is a two-step argument. First, cover the path with
O(A) comparison balls, use Lemma 3.1 on each ball to control
successive ratios, and multiply. Second, argue that the resulting
ratio is best-possible-described-by a semialgebraic function of
the parameter A, which by standard facts about one-variable
semialgebraic functions has at most polynomial growth. So the
constant is polynomial in A, independent of the coefficients.
3.3 Two-color matching: agreement forces closeness
Now we come to the main technical content of this section. The
setup is a "two-color" comparison: two separate lists of algebraic
sheets, of bounded length, whose root-pairs at a sampled point
agree TO VALUE within a small tolerance but OPPOSE in derivative.
Lemma 3.3 (Two-color matching). Fix D >= 1 and an integer
J > D. Fix d_0 < 1. There is no sequence N -> infinity with
the following properties.
-- Two lists of real C^{J+1} functions on a unit interval,
each of size at most N^{d+o(1)}, with d <= d_0. All
derivatives through order J+1 bounded by
K = N^{o(1)}.
-- A probability law on events (i_1, i_2, t), with (index,
time) marginal in each color bounded by K*pi_kappa(i)*dt,
where sum_i pi_kappa(i) = 1.
-- Every retained event satisfies
|f_{1, i_1}(t) - f_{2, i_2}(t)| <= K N^{-1},
max_{1 <= j <= D} |f_{1, i_1}^{(j)}(t)
- f_{2, i_2}^{(j)}(t)| >= K^{-1}.
The impossibility is uniform with sufficiently small fixed
power losses in place of all the subpower losses.
In plain words: two lists of size subpolynomial, agreeing in value
to accuracy N^{-1} on a substantial set of points, cannot
uniformly disagree in every derivative of positive order up to D,
unless the lists are essentially length N^d with d >= 1. But our
lists are length N^{d_0 + o(1)} with d_0 < 1. So the agreement
must fail on a small margin of derivative.
Why is this the right statement? Because for the terminal
extension argument, we need the "derivative data" of a chart to
DETERMINE it, up to a bounded list. If two charts agreed in value
but could disagree widely in derivative, they would be genuinely
different charts; we could not merge them. The lemma says: no, if
they agree in value on a substantial set, their derivatives must
agree too, up to a specific quantitative bound.
Proof idea. The proof uses an "entropy" argument.
(a) At each scale N^{-a}, a state records: the location of each
function's jet to precision N^{-a}, plus a "derivative"
family. There are at most N^{d + r(1) + o(1)} states, where
r(1) is a bookkeeping integral of the local jet widths.
(b) Along a "switching experiment," we sample alternating pairs
of points, jumping from color 1 to color 2 and back. Each
jump is nearly free (from the value-agreement hypothesis);
but its effect on the derivative accumulates. After m+1
switches, the accumulated derivative gap becomes a
Jacobian-controlled map from switch times to endpoint jets.
(c) That map's image has small measure but at least a
Jacobian-determined amount of "state mass." Comparing the
state count with the image size gives a strict inequality
on state counts. Summing over scales forces d >= 1.
The details involve information-theoretic entropy chains and
polynomial interpolation of jet trajectories, but the essence is
what (c) says: "you cannot have too small a list of functions be
pinned down to value N^{-1} while their derivatives range over
N^D independent directions."
3.4 The local matching corollary
The abstract two-color matching lemma has a quantitative form
that is more directly usable. Translate N-scales into a ratio R
= M/Delta of a "norm scale" M to a "value scale" Delta, and
require the two lists to be small: subpower in R.
Corollary 3.4 (Local matching of scalar graphs). Fix base
dimension, algebraic degrees, real comparison balls with fixed
holomorphic margins, and d_0 < 1. There is eps_* > 0 with the
following property. Let M, Delta > 0 and R = M/Delta. For
sufficiently large R, two lists of real-on-real algebraic
sheets cannot satisfy all of the following (with 0 < eps <=
eps_* and d <= d_0):
(a) each list has at most R^{d+eps} members, and each member
is within M R^eps in norm of one common algebraic
reference sheet;
(b) a subprobability law on (index, mass) pairs has total
mass at least R^{-eps}, with each index-base marginal
bounded by R^eps pi_kappa(i) dz, where pi_kappa are
probability weights, separately in each color;
(c) on events, the two values differ by at most Delta R^eps,
whereas the norm of their difference on the comparison
ball is at least M R^{-eps}.
In words: "two short lists of algebraic sheets, matched in value
to within Delta, but separated in norm by at least M, cannot
both be small." Since our lists will always be subpower, this
will be the contradiction mechanism.
The proof is a rescaling of Lemma 3.3: subtract the common
reference, divide by M to normalize, and then note that two
values that are Delta-close in value have derivative gap at
least M / Delta = R. Slice, normalize to a unit domain, and
apply Lemma 3.3.
3.5 Terminal time extension
Now we use the matching lemma in its final form: turn a
terminal-depth chart into an "atlas" at a slightly smaller time
cost (i.e. a longer time interval), at the price of a slightly
larger thickness.
Lemma 3.5 (Terminal time extension). Fix d_0 < 1 and angular
parameters S, eta > 0. There are constants C < infinity and
c > 0 (depending only on these fixed parameters and the
fixed model dimensions and test degrees) with the following
property. Suppose an exact problem in C(d), d <= d_0, of
length L, has a true terminal system at horizon exponent
h > 0. Then for every 0 < e < min(h, cL) there is, on a
retained subsequence, a known atlas at horizon exponent h - e,
whose labels are known at the old observation p_L, and whose
fitting thickness is a_L' with 0 < L' < L and L - L' <= Ce.
Read this as: "you can always trade a small amount of thickness
(loss of depth Ce) for a small amount of time localization
(horizon reduced by e)." Every loss is a fixed multiple of e.
The constants depend only on the fixed geometric parameters, not
on the particular problem.
Idea of proof. Take the terminal chart system at horizon h. Pick
a coarser time scale q_1 ~ q_0 * N_0^e, i.e. slightly longer time
blocks than the charts themselves. Sample a pair of times
t_0, t on the same trajectory, both selected incidences. At t_0
we have a source chart label A; at t we have a second label A'
from the same true terminal system (actually from the observation
list at t).
The core construction is:
1. Use A to build a coarse moving box D. This is what the
source chart's geometry looks like when extended to the wider
time scale.
2. In each D, pass to a common complex "anchor" alpha_D and
record the source branch (an algebraic sheet) and its jets at
that anchor.
3. Given a second sampled observable A', the source D is in a
subpower list. The corresponding anchor bin is in a subpower
list. Enumerate them and pick one representative.
4. That representative becomes a fixed test at scale q_1. A
Remez-type argument on the source trajectory extends the fit
through the longer time block; the sheet-matching corollary
prevents the second color from disagreeing (in the aggregate
over trajectories).
5. On the retained subset, the new charts have time length
q_1 ~ q_0 N_0^e (horizon reduced by e), thickness a_L' with
L' <= L - ce for a fixed c.
The "common complex anchor" step is why we spent time on complex
algebraic sheets: it fixes the coordinates on which two branch
representatives can be directly compared. Without this step,
extending a chart through a gap without a bridge is
uncontrolled.
This closes Section 3. The takeaway: charts can be slowed down
in time at a small cost in thickness, and the bridge used is
complex-sheet matching.
SECTION 3 wrap-up: what to remember going forward.
-- algebraic sheets + relative-extension (Lemma 3.1)
-- path continuation is polynomial-lossy but never singular
(Lemma 3.2)
-- two short value-close lists must have derivative-close
members (Lemma 3.3, Corollary 3.4)
-- terminal time extension: trade e of horizon for Ce of depth
(Lemma 3.5)
SECTION 4. CRITICAL PATHS AND THE NARROWNESS CLASSIFICATION
4.0 What this section does
By now we have: every Kakeya counterexample gives an exact problem
in the cap class C(d), for some d < 1, with positive horizon. The
horizon is a number in (0, L]. Every problem also has horizon at
most L (Proposition 2.6). So the interesting range is the open
interval.
This section sets up the extremal argument: we take a problem that
sits as far as possible from the trivial "horizon = L" bound within
C(d), and we extract from it a geometric object -- the critical
path -- whose parameters are so compressed that everything else in
the paper can be aimed at contradicting it.
There are four pieces:
4.1 Maximize. Among cap problems in C(d) for any d < 1, take
the smallest achievable d (call it d_*), and within those
take the problem with the largest normalized horizon (call
it k > 0). This problem is now extremal.
4.2 Extract a critical path. Take a hypothetical near-optimal
sequence of true terminal systems. Their density profiles
become linear (f(r) = d_* r) and their time rate becomes
k. Refining the terminals further reveals a nested chain of
true charts.
4.3 Classify by narrowness. The critical path has exactly one
structural degree of freedom left: how much thinner is the
smaller spatial direction than the time scale? Call this
narrowness ell. There are three cases: ell = 0 (isotropic),
0 < ell < infinity (finite aspect ratio), ell = infinity
(degenerately narrow).
4.4 State the criterion. In every case, we will construct a
"better chart" on a substantial residual. Section 4.4
states precisely what "better chart" means (Proposition
4.5, candidate-upgrade criterion), which is the bridge
from local constructions (Sections 5-9) to global
contradictions.
4.1 The extremal problem
The optimization is a two-step lexicographic one.
First, over all Kakeya counterexamples in R^4 (i.e. all Kakeya
sets with Hausdorff dimension below 4), and over all the exact
problems they produce via Theorem 2.1, minimize the cap dimension
d. That infimum is d_*. By construction, d_* < 1. Any exact
problem realizing the infimum is in the cap class C(d_*) and is
called "critical."
Second, over all critical problems, and over all their true
terminal systems, maximize the normalized horizon: that is, of
quantity k where the horizon h is written h = k * d_*. The
supremum is k.
Two easy observations.
-- k > 0. Otherwise there would be a critical problem with
horizon 0, contradicting positivity from the reduction.
-- d_* is achieved "on a subsequence." We may pass to a
subsequence and assume the cap dimension tends to d_* from
above. Similarly for the horizon ratio.
So from now on "critical problem" means: an exact problem in C(d)
with d tending to d_*, whose true terminal systems have horizon
tending to k d_* = k (writing k for the supremum horizon ratio,
and noting k > 0).
Nothing in this section is specific to R^4. The same setup would
work in any dimension n; the substance is in identifying what can
be improved in each geometry. For R^4 the three narrowness cases
are what we will handle.
4.2 Extracting a critical path
Given a critical problem, we want to read off finer structure than
"just an exact problem." The information we get is a "critical
path": a nested sequence of true systems at depths that are a
definite fixed fraction of L apart, whose density profiles are
approximately linear.
Proposition 4.1 (Critical path). Fix d_* < 1 and k > 0 as in
Section 4.1. Given a critical problem and any delta > 0, there
is a subsequence of exact problems, a nested sequence of true
chart systems, and a uniform constant c > 0, such that for
every pair of nested depths r_1 < r_2 the density profile
satisfies
f(r_2) - f(r_1) = d_* (r_2 - r_1) + O(delta),
and the associated time rates satisfy
h_2 - h_1 = k (r_2 - r_1) + O(delta),
on the same subsequence. In particular, at scales separated by
a bounded factor, the ratio of the two charts' time lengths is
N_0^{k(r_2 - r_1) + O(delta)}.
The key phrase is "linear profile." Since d_* is the maximum
allowed value of d, and since at the extremal problem the cap
constraint is tight, the density profile f has to drop at EXACTLY
rate d_* per unit of depth. If it dropped more slowly, we could
increase d_* (contradiction with criticality). If it dropped
faster, the problem would not satisfy the cap. So it drops
linearly.
The same linearity appears in the horizon: the horizon is
approximately k * d_* full-horizon, and the time rate along the
nested charts is preserved at rate k. This is a similar extremal
argument applied to the chart time lengths.
Why nested? Because we need a rigid chain: two charts at depths
r_1 < r_2, with r_2 - r_1 a fixed positive amount, have
predictable time-length ratio and predictable density drop. This
"connectivity across scales" is what lets us build a locally-
determined contradiction later.
4.3 Narrowness: the last degree of freedom
The critical path is nearly determined. Its density profile is
linear (rate d_*), its time rate is k, and its charts are nested
at fixed depth spacings. The only undetermined parameter is the
aspect ratio of the chart boxes: how much thinner is the smaller
spatial direction than the time scale?
Define the "narrowness" of a chart as
ell_chart = (1/s) * log_{N_0}( q^2 / J_c ),
where q is the chart's time length, J_c is the spatial Jacobian
(determinant of the spatial box matrix D_c), and s is the current
depth. The exponent q^2 / J_c is the "second-axis thinness":
in R^4 (two spatial dimensions) a chart's box is q by q by (q *
2^{-ell s}) up to subpower, and the Jacobian is q^2 times the
thinner side, so q^2 / J_c captures precisely how thin the
thinner side is.
Under composition (Lemma 2.5), narrowness adds:
ell_(composite) = ell_(parent) + ell_(child).
So narrowness of a chart system is the sum of the parent and child
narrownesses. This is what makes it a good invariant: we can
track it up and down nested chains.
For a critical path, take the infimum of ell over maximizing
sequences of true terminal systems. Call it ell_opt. This ell_opt
is what we will classify. It lies in [0, infinity].
ell_opt = 0. Charts are nearly isotropic: both spatial
directions have length comparable to q.
This is the "isotropic case."
0 < ell_opt < inf. There is a definite aspect ratio between
the two spatial directions. The chart is
"narrow" but not infinitesimally narrow.
ell_opt = infinity. Every chart is arbitrarily skinny in one
spatial direction. This is the "unbounded
narrowness" case.
Why exactly these three cases? Because ell is a real number in
[0, infinity], and "isotropic" and "infinite" are the two ends.
The intermediate regime handles finite positive aspect ratios.
Every chart falls into exactly one of these cases, so if we can
produce contradictions in all three, we have excluded all charts
and hence all counterexamples.
Each case gives a different geometric obstruction and requires a
different analytic engine:
ell = 0 Multilinear Kakeya and polynomial fits
(Section 6).
0 < ell < infinity Planar rigidity and scalar reduction
(Sections 7-8).
ell = infinity Full-time localization and scalar
projection (Section 9).
The three sections below operate independently; each produces the
same type of final object: a "better chart." Section 4.4 defines
that object.
4.4 The candidate-upgrade criterion
Everything in Sections 5-9 is working toward a single finite-
sample contradiction: produce, from the critical problem, a
"known atlas" at a horizon exponent clearly below the critical
k. That is literally incompatible with criticality (which
demands the horizon tend to k), so no counterexample survives.
The bridge from "local construction" to "global contradiction"
is this criterion.
Definition 4.2 (Time-improvement candidate). A time-improvement
candidate at tangent depth r, of "reference depth" s (r <= s),
consists of:
-- a true packet Q of the original problem, with thickness
a_r and time length q_s (the terminal time scale at
depth s);
-- a degree-2 test P_new on Q's trajectories, with
|P_new| <= K a_fit * (q_s), for a_fit < 1,
throughout the packet time block;
-- a selected incidence subset (of Q's incidences) with
mass at least K^{-1} q_s nu_Q * (a_fit/a_r)^d, i.e. a
definite fraction of what the d-cap allows, weighted by
the improvement factor;
-- the hidden-derivative bound |d_w P_new| >= K^{-1} on
that subset, so P_new is a legitimate chart test.
The candidate is "better than Q" if a_fit < a_r, i.e. the new
test fits the trajectories to a smaller thickness than the
original, on a definite fraction of Q's incidences.
The following proposition is the bridge we use, and whose proof
is the whole of Sections 5-9 at the local level.
Proposition 4.3 (Candidate-upgrade criterion). Consider a
critical path with parameters d_* < 1, k > 0, and an optimized
narrowness ell_opt as in Section 4.3. Suppose that in every
substantial residual of the critical path, and at some tangent
depth r, there is a time-improvement candidate that fits Q at
thickness a_fit = a_r * N_0^{-e} for some fixed e > 0. Then:
(i) The candidate can be upgraded to a true atlas at depth
r + delta_r whose time horizon is strictly below
k * (r + delta_r) by an amount proportional to e, on
some subsequence;
(ii) That atlas, when lifted back to the original problem,
contradicts the criticality of k.
The mechanism of the upgrade is: (a) candidate -> known atlas
via greedy coverage and the sheet-matching lemma; (b) known atlas
at shallower horizon -> forbidden by Proposition 4.1 (critical
path) restricted to the deeper sub-problem.
So from Section 5 onwards, our job is exactly this: in every
substantial residual, and for the appropriate narrowness regime,
produce a time-improvement candidate with a_fit strictly smaller
than a_r. Section 6 does it for ell = 0, Sections 7-8 for finite
positive ell, Section 9 for infinite ell.
SECTION 5. SCALAR CONCENTRATION AND POLYNOMIAL FITTING
5.0 What this section does
Sections 2-4 set up the extremal problem. Sections 5-9 must deliver
a "time-improvement candidate" in each narrowness regime. That
candidate is a lower-degree polynomial fit with a strictly smaller
thickness than the current chart, on a definite fraction of the
chart's incidences.
The engine behind every candidate is the SAME. It is the
following technical fact:
If a large family of incidence points has SCALAR SUPPORT
(meaning: at each time, the values of the scalar signal X
occupy only p^{-d} bins of size p, with d < 1), then the
family must CLUSTER in its low-degree coefficients. That
clustering, combined with a planar dot-product/rigidity
estimate imported from sticky Kakeya, produces a polynomial
fit.
The polynomial fit is a compact representation of a large chunk
of the data. In the isotropic case (Section 6), the fit leads to
a conservative field and a chart improvement. In the finite-
narrowness case (Sections 7-8), it leads to a planar two-strip
rigidity and then to a wide-interval improvement. In the
unbounded-narrowness case (Section 9), it leads to a scalar
projection and then to a full-time chart improvement.
So the two central results of this section are what every later
section leans on, and if you understand them you understand the
whole "improvement" mechanism.
Read this section as:
5.1 -- the scalar model, the setup, the definitions.
5.2 -- "support forces clustering": one lemma.
5.3 -- the planar dot-product alternative: one theorem,
imported.
5.4 -- "clustering gives a polynomial fit": the first-jet
proposition.
5.5 -- "higher jets and higher-degree fits": the final
polynomial-fitting proposition.
5.1 The scalar-normal data
We work with a trajectory family of the following shape:
y(t) = y_0 + t V, |y(t)|, |V| <= K,
x(t) = x_0 + t x_1 + t^2 x_2, |x_0|, |x_1|, |x_2| <= K.
So the "base" y is affine in time (in R^j for j = 1 or 2), and the
"scalar" x is quadratic. This is exactly the trajectory model of
Section 2, with the hidden w rescaled to be the scalar x. The
native signal is X(t) = P_*(t, y(t), w(t)), which is quadratic
along trajectories.
The prior on trajectories is nu, total mass at most K. A time
label set E_l is a subset of [0, 1]; the labeled incidence set
{(l, t) : t in E_l} has mass delta under nu (x) dt.
A "scalar-normal row" is a time-dependent linear functional
F(t) in (R^j)^* applied to y(t). The scalar-normal data is the
quadruple
(nu, y, x, F, E)
with the following property.
Definition 5.2 (Scalar-normal data). Fix j in {1, 2}, d < 1,
and a floor zeta. Given a trajectory prior nu, affine y and
quadratic x with coefficients bounded by K, a label set E_l
of incidence mass at least inverse-subpower, and a
measurable, time-only row F(t) with |F(t)| <= K, write
S_l(t) = x_l(t) - F(t) y_l(t)
for the "residual". The data is scalar-normal if:
(a) For every interval J of length r and mesh q in
[zeta, r], the set of residuals {(S_l(t), y_l(t)) :
t in E_l} has at most K (r/q)^{j+d} q-bins.
(b) The row F is slowly varying: |F(t) - F(u)| <= K(|t - u|
+ zeta).
(c) The residual slope R_l(t) = x_l'(t) - F(t) y_l'(t) is
also slowly varying in the sense that nearby spacetime
points have nearby R values.
(d) Given any coefficient bin B of side rho (all polynomial
coefficients of x and y to resolution rho), the
conditional spatial cap holds: for every spatial cube
Q of side q in [zeta, rho],
nu(B intersect {t in E_l, (x_l, y_l)(t) in Q})
<= K (q/rho)^{j+d} nu(B).
The point of the definition is that scalar support (bound a) and
the conditional cap (d) are qualitative conditions on the
incidence distribution, not specific numbers. The data is
scalar-normal if it satisfies them uniformly in a fixed range of
scales.
The purpose of the row F: it is the "leave-out" that makes the
scalar residual S small. Without F, the residual S would just be
x, which is too big. With F chosen correctly, S is the deviation
of x from a prediction that depends only on the base y. This is
exactly the shape of a chart: base y varies in R^j, scalar x is
predicted by a linear-in-y row F(t).
The lemma at the heart of Section 5 is:
Lemma 5.1 (Scalar coefficient clustering). Let x, y be as
above, with the trajectory prior satisfying the coefficient
cap (nu of a p-ball in coefficient space is at most K p^d for
p in [p_*, 1]). Suppose the labeled incidence measure, of
mass at least inverse-subpower, occupies at most K p^{-1-d}
squares of side p in the (t, x)-plane, for p = N^{-c+o(1)}
fixed. Then some coefficient bin of side p has trajectory
mass at least p^d / K.
In plain words: if the scalar signal X has small support in the
(t, X)-plane (size at most p^{-1-d} cells of side p, so d < 1 cuts
it below the naive bound p^{-2}), and the coefficient distribution
is not more concentrated than p^d anywhere, then some p-bin of
coefficients must be heavy, of mass at least p^d.
The intuition is a pigeonhole over the 2D plane. Total mass is at
most the number of occupied (t, X)-cells times the coefficient
mass per cell. If the peak coefficient mass were below p^d, the
total would be below p^{-1-d} * p^d = p^{-1}, but total mass is at
least inverse-subpower. So a heavy coefficient bin must exist.
5.2 How the clustering argument works
The proof of Lemma 5.1 is a neat trick: convert a 1D scalar
control problem into a 2D planar-motion problem, and apply the
weighted plank estimate from Section 2.7 there.
Step 1: lift. Adjoin an independent uniform coordinate D in [0,
1] to each trajectory, and consider the planar motion
z_(l, D)(t) = ( x_(0,l) + t(x_(1,l) - D), D + t x_(2,l) ).
It is affine in t: its two components are polynomials of degree
at most 1 in t. The key identity:
(1, t) . z_(l, D)(t) = x_l(t).
So the scalar value x_l(t) is a linear projection of the lifted
planar motion.
Step 2: count occupied cells. The support hypothesis for X
bounds the (t, x)-cells; each cell lifts to at most 1/p
choices of D-coordinate, and then at most K choices of the
first planar coordinate. So the lifted planar motion occupies
at most K p^{-2-d} spacetime cells in (t, z_1, z_2).
Step 3: weighted plank estimate. Apply Lemma 2.7 (the weighted
full-time plank estimate) to the lifted family. The traces
satisfy a plank cap (from the coefficient cap of the
trajectory prior combined with the flexibility of D). The
total number of marks is at least p^{-1} / K. The number of
visited cells is at most K p^{-2-d}. Their ratio is at least
p^{1+d} / K. So there is a heavy full-time plank: an affine
planar strip of widths u, v with
p / K <= u <= v <= K
whose lifted trajectory mass is at least K^{-1} p^{d-1} u v.
Step 4: bound the D-acceptance. For fixed scalar coefficients
(x_0, x_1, x_2), varying D changes the lifted motion by
D(-t, 1). The narrow width v bounds the D's that fit the
plank: at most C u of them. So the coefficient mass in the
plank is at least K^{-1} p^{d-1} v.
Step 5: compare with the coefficient cap. All quadratics that
fit this plank lie in one coefficient ball of radius K v
(differences of two plank-lifted motions are O(v) everywhere
in t, and a quadratic is determined by three samples). The
coefficient cap says such a ball has mass at most K v^d. So
K^{-1} p^{d-1} v <= K v^d,
hence v <= K p. Substituting back: the coefficient mass in the
plank is at least K^{-1} p^{d-1} * p = p^d / K, so within the
plank, one coefficient p-bin has mass at least p^d / K. That is
the lemma.
The essential point: the scalar support deficit d < 1 is
compensated by a "collapse" onto a thin planar strip. The
weighted plank estimate is what turns the deficit into
clustering.
5.3 The planar dot-product alternative
Clustering (Lemma 5.1) is only useful if it can be combined with
something that says: "you cannot have that much clustering unless
the trajectories have a specific structure." The next imported
black box does exactly that.
It comes from Hong Wang and Joshua Zahl's work on sticky Kakeya,
Theorem 5.2 of arXiv:2210.09581v2 at the start of Section 8, with
Definition 4.2. We use it as-is.
Theorem 5.4 (Planar dot-product alternative). For every eps
> 0 there are eta, delta_0 > 0 such that the following holds
for 0 < p <= delta_0. Let A, B be subsets of [0,1]^2 with
covering numbers satisfying
N_s(E intersect B(z, w)) <= p^{-eta} w/s
for E in {A, B}, p <= s <= w <= 1.
Then at least one of two alternatives holds:
(1) Expansion. There are affine lines l and l^perp (a line
and its perpendicular) such that both
N_p(A intersect {z : dist(z, l) <= p})
N_p(B intersect {z : dist(z, l^perp) <= p})
are at least p^{eps - 1}.
(2) Projection collapse. Every restricted triple set H
subset A x B x B with N_p(H) >= p^{eta - 3} admits a
scale s >= p and an interval J of length at least
p^{-eta} s, for which
N_s(J intersect {a.(b - b') : (a, b, b') in H})
>= (|J|/s)^{1 - eps}.
In plain words: two planar sets with well-behaved covering
numbers are either (1) both close to a specific pair of
perpendicular lines, or (2) for any restricted triple family,
the dot products a.(b - b') fail to spread out. "Fails to
spread out" means: they collapse into a small interval at some
scale, contradicting the naive bound (|J|/s) that a
non-collapsing distribution would give.
Read this as an alternative. You either get the STRUCTURE
(directions nearly aligning with a perpendicular pair), or you
get a COLLAPSE (dot products pile up into a small interval).
The argument in the rest of Section 5 either exploits the
alignment (if (1) holds) or drains mass due to the collapse
(if (2) holds). One way or another we extract something usable.
5.4 Beyond support: a polynomial fit
We are ready to combine the pieces. Support bounds give
clustering (Lemma 5.1). Clustering plus the dot-product
alternative (Theorem 5.4) gives a ROW estimate. That row plus a
final entropy argument gives a POLYNOMIAL FIT.
Two propositions. The first is the first-jet version, and the
second extends it to arbitrary finite degree.
Proposition 5.5 (Scalar-normal first jet). Assume
scalar-normal data. Fix a finite nested hierarchy of dyadic
time lengths with finest length H_* and vertex floor zeta
< H_*^3 / K. Then a dense restriction of the single-time
labels admits a time-only row A(t), bounded by K, such that
|F(v) - F(u) - (v - u) A(u)| <= K H^{11/10}
whenever retained active times u, v lie in the same hierarchy
block of length H. Define
J_1(l, t) = x_l'(t) - F(t) y_l'(t) - A(t) y_l(t).
Then for any finite presampled time tuple and any dense tested
event on those tuples, the entropy of J_1(l, t) at width r,
given the tested time and the scalar residual S_l(t) at width
r, is o_* provided r >= H_*^{(1-d)/64}.
Read this as: the row F, which was our "linear-in-y" predictor,
has a FIRST-ORDER Taylor correction A(t) y. After adding A, the
residual J_1 has essentially no entropy given (time, scalar).
That is a strong statement: the residual J_1 is nearly determined
by (t, S).
Idea of proof. Two stages.
(i) Fit each row component of F by an affine function on
selected times in each block. The clustering lemma and the
planar dot-product alternative give this fit on half the
block, by arguing that the only obstruction is that the
base y is confined to a short strip near one specific line,
which can be absorbed into F.
(ii) Compare the affine fits across hierarchy lengths. The
block-to-block slope differences A_I - A_{I'} are
controlled by comparing values at two shared times; that
gives A(u) via a limit. Then check that J_1 (the residual
after subtracting A y) has small entropy using the
triangle inequality for entropy.
The role of the "prepared labels" is technical: before fitting
each component, we restrict to times where the current labels
actually live and every block has a definite mass. This is the
hypothesis the rest of the argument relies on: mass must be
there, not just theoretically. The densities and the caps are
what let us enforce it.
Proposition 5.6 (Higher-jet polynomial fitting). Fix d < 1,
M >= 2, and a dyadic block length h = N^{-c+o(1)} with c > 0
fixed. Take the jet hierarchy H_1 >> H_2 >> ... >> H_* fixed
in advance (ratios chosen so that each scale is one full
power of the next). Take the floor zeta <= H_*^3/K. Then given
scalar-normal data, a dense set of labels (l, t) has the
following property: there is a row-valued polynomial P of
degree M, with coefficients bounded by K, such that on labeled
times v in E_l of length at least h/K including v = t,
|F(v) - P(v)| <= K h^{M+1/4}.
The polynomial may depend on the incidence.
The proof is induction using the first-jet result as a base case.
At each level of the jet hierarchy, divide the jet difference
between two nearby times by the time gap. That gives a
higher-order jet estimate at a smaller scale. Recurse to build up
polynomial coefficients Taylor-step by Taylor-step. The entropy
control from Proposition 5.5 ensures the recursion stays tight.
For high school readers, the essential content is: "you get a
degree-M polynomial fit to the row on a definite chunk of time,
from a purely combinatorial support hypothesis." That is the
engine that will run in every narrowness regime.
That closes Section 5. The improvement mechanism is now in hand:
cluster scalar support, import a planar dot-product
alternative, fit a row, extend to a polynomial.
Everything left is applying this to specific narrowness
geometries.
SECTION 6. THE ISOTROPIC IMPROVEMENT
6.0 What this section does
We now aim at the ell = 0 case: charts that are nearly isotropic
in the spatial plane, with the two spatial widths both comparable
to the time scale (up to subpower factors). This is the regime
where the geometric structure is closest to classical Kakeya-type
analysis, and it is where the multilinear Kakeya theorem does its
main work.
The goal: build a chart at terminal depth r + delta that is
strictly better than the critical path, contradicting
Proposition 4.1 (critical path). By Proposition 4.3
(candidate-upgrade criterion), it suffices to produce a
"time-improvement candidate": a polynomial test with strictly
smaller thickness a_fit < a_r on a positive fraction of the
trajectories, in the same normalized chart.
The argument has four stages:
6.1 Determinant multilinear Kakeya. An imported black box: an
analytic bound controlling how tubes in independent
directions can overlap.
6.2 The velocity graph. Using the imported estimate, build a
vector row G at every fine physical cell that nearly
matches the actual velocities.
6.3 Interpolation to a polynomial field. From the discrete
velocity graph, extract a polynomial T(u) matching G on a
dense set.
6.4 An approximately conservative field. Show that T is
close to being a gradient, extract a potential P.
6.5 Degree restoration. If P is not quadratic, either encode
the obstruction or find a hidden coordinate that forces
it to be quadratic.
The output of all this is a degree-2 test with thickness strictly
below a_r, contradicting the isotropic critical path.
This in turn gives a stronger statement: no scalar-model version 1
problem in C(d) has positive horizon. That statement feeds into
Section 9.
6.1 Determinant multilinear Kakeya
We import one analytic black box. It is the equal-dimension,
wedge-weighted form of the determinant multilinear Kakeya theorem,
due to Bennett-Carbery-Tao, with the endpoint case due to Guth
and a wedge-weighted version due to Bourgain-Guth.
Theorem 6.1 (Determinant multilinear Kakeya). For 2 <= n <= 4
there is a constant C_n < infinity such that the following
holds. Let T_i (i = 1, ..., n) be weighted families of radius-r
neighborhoods of full lines in R^n, with unit directions e_T
and nonnegative weights w_T. Write M_i = sum_T w_T. Then
int_{R^n} ( sum_{T_1, ..., T_n}
|det(e_T_1, ..., e_T_n)|
(prod_i w_T_i) 1_{T_1 int ... int T_n}(x)
)^{1/(n-1)} dx
<= C_n r^n prod_i M_i^{1/(n-1)}.
In words: if you take n families of thin tubes in R^n, with one
tube from each family, and their directions are independent (
determinant bounded away from zero ), then those tubes cannot
overlap "too much" across all cells they visit. The integral on
the left is a measure of how much total tube-overlap-weight is
present, discounted by (1/(n-1)) of the direction determinant.
Concretely: supposing you place M lines in each of n independent
families, the multilinear Kakeya theorem says that across all
space points, the total overlap is bounded by C_n r^n (prod M_i)^
{1/(n-1)}. If the lines were just random, you would get this
"trivially"; the theorem says the bound holds even for the
worst-case arrangement, provided only that the directions were
independent.
We use the theorem with n = m + 1, where m is the ambient
dimension of the base (m = 1 or 2). The four-dimensional ambient
case (n = 4) is where the multilinear Kakeya bound is exactly at
the endpoint; that is why a 4D theorem needs a 5th-direction
surcharge.
The round line families are usually constructed from "affine
approximations" of trajectory families: take each trajectory,
replace it in a small time window by its affine tangent line, then
enlarge into a radius-r tube. The determinant of tangent
directions captures the "independence" of the family.
6.2 The velocity graph
The first step is to turn the "many incidences" hypothesis into a
"velocity field" object.
Proposition 6.2 (Physical velocity graph). For a sufficiently
fine fixed power b, there is a row G = (G_0, G_y), constant on
each fine physical cell W_b and bounded by K, such that on a
substantial reference family,
|X'(t) - G(W_b) . v| <= b^kappa,
where v = (1, V) is the unit direction, and kappa > 0 depends
only on the base dimension m and on 1-d. Moreover, for every
required dyadic s_0 >= b^kappa, the values of G to width s_0
have at most K possibilities in a fixed coarse physical cell
W_{s_0}.
Concretely: at each spatial cell W_b (which has side b in
physical coordinates), define a vector G(W_b). The proposition
says that the trajectory velocity X'(t) is nearly in the direction
of G(W_b).v at every incidence in that cell.
The "graph" terminology comes from the graph-of-a-function form:
G lives in "velocity space", and its value is (nearly) the
derivative along the trajectory, as a function of the base point
and the direction. It is a graph of "velocity as a function of
position and direction."
The proof uses determinant multilinear Kakeya: if no such G
existed, then successive independent samples from a physical cell
would give velocity directions spanning R^{m+1}, and multilinear
Kakeya would force the visited cells to exceed the support bound.
The number of visited cells is controlled by the scalar caps; the
number forced by multilinear Kakeya would be too large. Hence a
good G must exist.
The "at most K possibilities" part is a regularity statement: at
each coarse cell, G is confined to a short list of small balls.
That is what makes G usable as a "field" for interpolation: you
only need a finite list of values to describe it in each cell.
6.3 Interpolating the velocity graph to a polynomial field
Now we turn the discrete graph G into a continuous field. The
central result is:
Proposition 6.3 (Polynomial field interpolation). There is a
row-valued polynomial T(u) of degree at most M in each of the
base and hidden coordinates (with coefficients bounded by K)
such that on the reference family,
|G(W_b) - T(u)| <= K eps_T,
for some eps_T with b^{kappa} << eps_T << 1. Here u is the
normalized physical coordinate (u = (z - z_c)/h). In particular,
for every long reference line in the family (weighted duration
at least h/K in the block),
|X'(t) - T(u_l(t)) . v_l| <= K eps_T
throughout the whole Q-block of that line.
Translation: after interpolating, the field G is a global
degree-M polynomial in the base coordinates, and the trajectory
velocities are within eps_T of the polynomial predicted values.
This is a "propagation" statement: G is a discrete object known
only in a list-per-cell, but the lists are short, and combining
them via Lagrange interpolation over a product grid gives a
global object that reproduces G on a substantial fraction of
cells and hence on all reference lines (by Remez, since
polynomial values are bounded above and the errors are small).
Proof strategy. The base coordinates u = (u_1, ..., u_m) live in
an m-dimensional grid of step xi (with m = 1 or 2 for our
applications). In each fiber (fixed u_1, ..., u_k, varying u_k+1
to u_m), G(W_b) is a list-bounded set of affine or low-degree
functions. Take a "product grid" of m directions; on each axis,
choose a fiber polynomial uniformly from its list. A pigeonhole
gives one combination of fiber choices that retains a positive
fraction of cells. Then use Lagrange interpolation to combine
the fiber polynomials into a single tensor polynomial T(u). The
error telescopes because each fiber choice has small error, and
the Lagrange weights are bounded by K (which is why we take a
grid with fixed steps).
The passage from cells to whole reference lines uses the
comparison argument from the graph Prop 6.2: the "at most K
possibilities" property means the polynomial fit at a coarse cell
predicts the polynomial fit at fine sub-cells, and this propagates
all the way to the whole Q-block.
This is one of the places where the "<=" inequalities have to be
done carefully. The choices of fiber polynomials, the coarse-to-
fine propagation, and the Remez step all lose a constant factor
subpower in N_0. Those losses add up. The final choice of eps_T
must be larger than the sum of all these subpower losses, but
strictly smaller than the target accuracy b^{kappa}. Since
b^{kappa} is a fixed power of b and the losses are subpower in N_0,
b^{kappa} always eventually dominates.
The row T therefore acts as a "polynomial predictor" for the
velocity X' on the entire reference family. Trajectory by
trajectory: X'(t) is within K eps_T of T(u(t)).v.
That is the point at which the trajectory family becomes a
POLYNOMIAL OBJECT. Everything after uses T as a global
polynomial field.
6.4 The field is approximately a gradient
Here is the pivotal step. Recall that the scalar signal is
X(t) = P_*(t, y(t), w(t)).
The chain rule gives, for a trajectory (t, y, w) satisfying the
native velocity relation:
X'(t) = partial_t P_*(t, y, w)
+ partial_y P_*(t, y, w) . y'(t)
+ partial_w P_*(t, y, w) . w'(t).
Now here is a crucial simplification. In the normalized physical
coordinates on a chart, the hidden coordinate w is an affine
function of t (in Version 2), and the y coordinate is affine in t.
So X'(t) is a linear functional in (1, V) applied to something.
In the "vertical-compatible" form of P_* (which is what we can
reduce to by the model-reduction of Section 2), the hidden
derivative contribution to X' is constant along trajectories, and
gets absorbed into the row G_0. Then
X'(t) = G_*(W_b) . v + O(error)
for v = (1, V), where G_* is a specific combination of the partials
of P_* evaluated at the current trajectory point. From Section 6.3,
T(u) reproduces G_*, so in fact
X'(t) ~ T(u(t)) . v
on reference lines. This is the starting point.
Conservativity. The function X(t) is a SCALAR. The gradient-like
combination T.v is its derivative. For a scalar function, the
"cross partial" structure of T must be symmetric. Precisely: if
T = (T_0, T_y) is a row-valued vector field, and X'(t) = T_0 + T_y
. V for every trajectory velocity V that appears (V ranges over
a spanning set of velocities in the chart), then T_y must be
symmetric -- because it is the Jacobian of a scalar function y ->
partial_y X restricted to those slopes.
More carefully: consider the matrix
M(u) = D T(u) as a function of u,
i.e. the Jacobian matrix. The relation X'(t) = T(u(t)).v, being
consistent for all trajectories passing through u with different
velocities, forces M to be approximately symmetric:
|M - M^T| <= K eps_T / (some natural scale)
where the error is small compared to the chart's own scale. The
antisymmetric part of M is bounded by a fixed power of eps_T.
Proposition 6.4 (Conservativity). Fix eps_0 > 0 (a power
depending on the argument). On the retained reference family,
after further restriction to a dense submeasure,
sup = max over pairs of trajectories through u of
|skew(DT(u))| <= eps_0
up to subpower factors, where "skew" means the antisymmetric
part of the Jacobian matrix of T at u.
Given this, there is a potential (a scalar function) P such that
grad P(u) ~ h T(u)
with error bounded by K a h^{c_0} for some fixed c_0 > 0. The
h factor comes from a change of variables: the reference family
has a natural "time" scale h (the Q-block length), and the chain
rule transcribes the field T into a gradient with a factor of h.
Theorem 6.5 (Potential existence). Under the conservativity
hypothesis, the polynomial
P(u) = h * integral_0^1 T(tu) . u dt
(defined by differentiating under the integral sign) satisfies
|grad P(u) - h T(u)| <= K a h^{c_0}
on the required enlarged box, and all fixed positive
derivatives of P are bounded by K h.
Wait, but "K a h^{c_0}" is the error at the terminal thickness
scale. If P is a genuine degree-2 polynomial, this is already a
candidate at thickness a h^{c_0}.
The catch: P need not be degree 2. Usually P will be degree > 2,
and then the "chart" constraint (which requires a degree-2 test at
thickness a) is violated. Restoring degree 2 is the content of
Section 6.5.
6.5 Degree restoration
The potential P from Theorem 6.5 is a genuine polynomial, but its
degree is usually above 2. The chart requires a degree-2 test.
We now show: if P has higher degree, then its higher-degree
terms have a very specific structure -- they are multiples of a
single quadratic form q -- and that structure lets us encode them
into the hidden coordinate, producing a new degree-2 test at a
strictly smaller thickness.
The structural fact comes from linear algebra on tensor
contractions over a nonsingular conic in direction space.
Lemma 6.6 (Tensor kernels over a nonsingular conic). Fix 7
distinct real projective points p_1, ..., p_7 lying on a
nonsingular conic q(x) = 0 in the direction space. Then:
(a) The cubic two-copy test T -> (T(p_i, p_i, .))_{i=1..7},
on totally symmetric 3-tensors, has kernel exactly
{ q time L : L linear }.
(b) The 1+2 copy quartic test has kernel exactly the
multiples of q^2.
(c) The 1+1+1 copy quintic test has trivial kernel on the
exact-conic family.
This is an elementary fact of projective geometry. A cubic
vanishing at all 7 points of a nonsingular conic is divisible by
its equation (over C: the binary sextic parametrization has 7
zeros, so it vanishes identically). Euler's identity then
iterates: cubic = q L, quartic = lambda q^2, quintic = q^2 L'.
The "least positive singular value" version, used quantitatively,
follows from compactness of the coefficient family plus
semialgebraic bounds.
Applying to our potential P: on a subsequence where the cubic,
quartic, and quintic Taylor coefficients are all bounded by K
(they are, by the algebraic norm control from Section 3), the
higher-order terms must be
cubic term of P = q(u) . L_1(u),
quartic term of P = lambda_0 . q(u)^2,
quintic term of P = q(u)^2 . L'_1(u),
with q a specific quadratic form, L_1 and L'_1 linear, and
lambda_0 a scalar. All are bounded by K.
The quadratic form q is called the obstruction conic. It is
determined (up to scale) by any five points of the 7-point
configuration, and in the isotropic regime the trajectory
family forces q to be a particular concrete conic.
Now encode q in the hidden coordinate. Define
H'(u, w) = (w - L'(u)) / S_0,
where L' is the affine line fitted to the trajectory (this is
the "affine defect" from Section 5.5) and
S_0 = inf over affine L of ||f - L||_{C^0(U)}. Then
q(u - u_c) ~ H'(u, w) up to error ~ K a_2,
because q vanishes identically near the fitted line and the
line deviates from f by S_0.
Proposition 6.7 (Restoration to a quadratic test). Define
P_new(z, w) = P_s(z, w)
- a ( R_2(u) + H'(u,w) L_1(u) + lambda_0 H'(u,w)^2 )
where P_s is the original chart test. Then P_new is a
degree-2 polynomial in (z, w), and on retained incidences
in a fixed positive fraction of the reference family,
|P_new| <= K a_r N_0^{-e} and |partial_w P_new| >= B/K.
for some fixed e > 0.
The value bound is small compared with a_r = a (the terminal
thickness). So P_new is a TIME-IMPROVEMENT CANDIDATE: a
quadratic test at strictly smaller effective thickness a_r
N_0^{-e} on a positive fraction of trajectories. Degree
restoration is done.
6.6 Concluding the isotropic case
The pieces now fit together. From Sections 6.1-6.5, in every
substantial residual of an isotropic critical path, we have
constructed a time-improvement candidate. By Proposition 4.3
(candidate-upgrade criterion), it upgrades to a better true
atlas, whose existence contradicts Proposition 4.1 (critical
path). The isotropic case cannot carry a hypothetical Kakeya
counterexample.
In fact the argument is stronger. It works for a purely scalar
version-1 problem (single spatial coordinate, quadratic trajectory
hidden, no hidden coordinate w at all). Applying the same
conservativity-potential-degree-restoration chain, but with the
0-dimensional hidden coordinate (so the "H'" trick is not
available, and we must have q identically zero), we obtain:
Corollary 6.8 (Zero horizon in scalar model). For any d < 1
and any admissible S, eta > 0, every version-1 exact problem
in C(d; S, eta) has horizon H(E) = 0.
This corollary is used in Section 9 as the "scalar projection"
input for the unbounded-narrowness case.
So Section 6 finishes:
- the isotropic narrowness regime (ell = 0), via the
multilinear Kakeya / polynomial potential / degree
restoration chain;
- a corner-case result (Corollary 6.8) that will feed forward.
Sections 7, 8, 9 handle finite positive ell and infinite ell.
SECTION 7. RETAINED-EDGE PLANAR TOOLS
7.0 What this section does
Section 6 handled the isotropic case completely: ell = 0. Now we
need the finite positive narrowness case: 0 < ell < infinity. In
this regime one spatial axis is much shorter than the time scale,
but not infinitesimally so. The chart is a very thin slab, and the
geometry within the slab must be understood in two dimensions.
The result this section proves is one of the most surprising in
the paper. It says:
Theorem 7.4 (Row-fit rigidity). Let J be a bounded row from a
planar domain to R. Suppose J evaluates to a nearly constant
value along a large fraction of many lines. Then, on many
incidences, J agrees with a constant row plus a multiple of
the "rotated" row Jz = (-Z, t).
The 90-degree rotation Jz = (-Z, t) is the key structure. It
comes from a very specific geometric configuration: most lines
must lie in two parallel strips whose width equals the mesh (up
to subpower) and whose common slope is the perpendicular
direction to the strips.
The proof has three technical steps:
7.1 Graph BSG. A graph version of the Balog-Szemeredi-Gowers
theorem, imported from additive combinatorics.
7.2 Common endpoint sets. Combining the graph output over
many slopes to extract a common pair of endpoint sets.
7.3 Rigidity of restricted sumsets. From the endpoint sets,
extract the two parallel strips.
Then Section 7.4 assembles the row-fit conclusion.
The name "retained-edge" refers to the crucial point: the graph
BSG argument is applied in a form that retains ORIGINAL edges,
not just a compact-model count. Without that, we could not lift
the conclusion back to "incidence mass in the original problem."
The retained edges are what make the argument compatible with
the rest of the paper's density bookkeeping.
7.1 The graph version of Balog-Szemeredi-Gowers
The Balog-Szemeredi-Gowers theorem is one of the central facts of
additive combinatorics. In its classic form: if a finite set A of
integers has many representations of elements of A + A as a + a',
then some large subset A' of A has a sumset of size comparable to
|A'|. "Many representations" means the graph of the sum relation
is dense; the conclusion is that the arithmetic structure is not
far from an arithmetic progression.
The version we use is due to Sudakov-Szemeredi-Vu and is
recast in a graph form by Tao and Vu. We use it as an imported
black box.
Theorem 7.1 (Graph BSG). There is an absolute constant C <
infinity with the following property. Let A, B be subsets of
a metric mesh at scale h in R^2, with |A|, |B| in [L/K, K L]
where L = h^{-1}/K (up to subpower). Let G be a bipartite
graph on A x B with |G| >= L^2/K. Suppose the sumset
Sigma_G = { (x + kappa y) : (x, y) in G, kappa a slope } has
at most K L cells at mesh h, after rounding.
Then there are subsets A_* subset A, B_* subset B with
|A_*|, |B_*| >= L / K, |G cap (A_* x B_*)| >= L^2 / K,
and N_h(A_* + kappa B_*) <= K L for every slope kappa in the
(fixed finite) slope list.
In plain words: a dense sumset in the mesh gives a large
regular subgraph, whose endpoint sets still have near-minimal
sumsets. The word "retained" (in the section name) refers to the
fact that (x, y) still come from the ORIGINAL graph G, not from
any approximation. This is what makes the argument compatible
with incidence mass: every original edge stays original.
The proof is a three-path argument. Given a dense subgraph, one
finds many length-3 paths (x, y', x', y) with y - y' in a fixed
sumset. Those paths are used to build sub-sets A_*, B_* by
"averaging" over the paths. The averaging is what produces the
concentration, and the three-path structure is what keeps
original edges. For our applications, we only ever invoke the
conclusion, never the internal path argument.
7.2 Common endpoint sets
A single application of Theorem 7.1 gives one pair (A_*, B_*) for
one graph. We have MANY slopes kappa in play simultaneously (the
row-fit problem comes with a slope field). We need to combine the
BSG outputs across slopes to extract a SINGLE pair (A_*, B_*)
that works for all of them at once, up to subpower loss.
Lemma 7.2 (Common endpoint sets). Under the graph hypotheses of
a row-fit problem, after restricting the index law to a
subpower fraction of its mass, there is a specific slope
kappa_0 in the family, and fixed subsets A_* subset A, B_*
subset B, with the following property. For any fixed m >= 1
and H >= 1, the scalar combination set
c_1 E + ... + c_m E + d_1 F + ... + d_m F,
where E = A_*, F = -kappa_0 B_*, and each c_i, d_i is a sum of
at most H signed products of at most H elements of the
restricted slope family R = { kappa / kappa_0 : kappa in
family }, has at most K_{m,H} L cells at mesh h.
The proof is a two-step averaging. First, apply Theorem 7.1
independently across all slopes. Each slope kappa gives a BSG
output (A_kappa, B_kappa). Second, take the average intersection
of the (A_kappa x B_kappa) rectangles over all kappa (weighted
by their subsists). The intersection of two "large" rectangles
in a common universe is "large" by set theory:
|R_1 cap R_2| >= |R_1| |R_2| / |universe|,
and since each rectangle has mass at least L^2/K, some pair has
intersection at least L^2/K^6. Keep only the pair with maximal
intersection, reweight by another subpower factor, and iterate.
The conclusion says: after the retraction, all the scalar
combinations of "directions" and "positions" (with fixed small
complexity coefficients) have sumset of size O(L). That's the
"scalar closure" hypothesis we need for the next step.
7.3 Rows with small restricted sumsets have planar rigidity
Here is the actual geometric content.
Theorem 7.3 (Two-strip rigidity). Suppose rows with small
restricted sumsets. Let A_*, B_*, kappa_0 be as in Lemma 7.2. Then, on
the retained incidences, there is a fixed slope s_0 (bounded
subpower) and, for each cell x in the (Z, t)-plane, a pair of
parallel strips
S_1(x) = { z : |orth_s(z - x)| <= h^{1 - o(1)} },
S_2(x) = { z : |orth_s(z - x - c)| <= h^{1 - o(1)} },
where c is a fixed vector, such that, restricted to the
retained incidences, most edges (a, b) in the graph have
their endpoints in S_1 and S_2 respectively, with s_0 as
their slope and half-width roughly h.
The strips have width equal to the mesh h, up to subpower loss.
"Perpendicular" (the direction orth_s) is a rotation by 90 degrees.
The slope s_0 is a fixed, subpower-bounded value, and both strips
have the same slope.
Why "two strips"? Because the graph provides BOTH endpoints a
and b = a + (something), and the sumset rigidity says the pair
must collapse onto a low-dimensional structure. Low-dimensional
in the (Z, t)-plane with both endpoints constrained means each
endpoint lies in a specific strip. The slope of that strip is
determined by kappa_0 and the "rotated" Jz structure.
Why "perpendicular"? Because the slope kappa_0 from Lemma 7.2
is one of the "slopes" that parametrizes the SUM, i.e. Sigma_G,
and typical sums a + kappa b live near a specific scalar value
(the row's constant). The linear combination of that value with
the row-fit condition produces a rotated structure, and the 90-
degree rotation (-Z, t) appears from the specific relation between
Z and t that forces "constant row + rotated row" structure.
For high-school readers, the essence is: you cannot have a
row that evaluates to the same value on lots of very different
lines. The lines must all be squeezed into two parallel strips
specific to the row.
7.4 The row-fit conclusion
The two-strip rigidity is what we use, but the statement we
actually want is a bit abstract. We record it as a single
proposition.
Proposition 7.4 (Row-fit). Fix a bounded smooth row F on a
planar chart. Suppose that, on a substantial set S of line
incidences (each incidence is a pair (z, v) with z in the
chart and v in a compact set of unit vectors), the directional
evaluation
partial_t F(z) . 1 + partial_Z F(z) . v
is nearly constant: it lies in a fixed interval of length
bounded by a small subpower.
Then on a positive-measure subset of S, there exist constants
C in R and c in R^2 such that
|F(z) - C - c . Jz| <= h^{1 - o(1)},
where Jz = (-Z, t) is the 90-degree rotation in the (t, Z)
plane.
Read this as: "if a row F, evaluated along many lines with many
different directions, gives nearly the same directional derivative
everywhere, then F is very close to C + c . Jz on many incidences."
The constant C is a global scalar; c is a global vector. Together
they describe a (basis-dependent) linear-plus-constant pattern,
with the linear part always having the specific "rotated" form
c . Jz.
Why must c appear with Jz, not with z itself? Because directional
derivative is Jz . (1, v) = -Z + t v. If the directional
derivative were constant, it would say: -Z + t v = const, i.e.
Z(t) = -const + t v, which is a line through (-const) with velocity
v. That is precisely the geometry of "the row F(z) is being
evaluated on a family of lines that all pass through a common
point."
The general case, where the derivative is only NEARLY constant,
already allows a lot; but the two-strip rigidity from Theorem 7.3
constrains the deviations. The two strips trap the endpoint
positions, and the common slope s_0 (which is subpower-bounded)
turns into the constant c.
That is the row-fit conclusion. It is used in Sections 8 and 9 to:
- Section 8 (finite narrowness): reduce the row-fit problem to
a scalar projection along c . Jz.
- Section 9 (unbounded narrowness): reduce to a scalar model
and then apply Corollary 6.8 (zero horizon in scalar model).
Together with Proposition 4.3 (candidate-upgrade criterion), this
closes the finite positive ell and infinite ell cases.
Section 7 is done.
SECTION 8. THE FINITE-NARROWNESS REDUCTION TO A HORIZONTAL MODEL
8.0 What this section does
We now handle the case 0 < ell < infinity: charts whose smaller
spatial axis is much shorter than the time scale, but the ratio
is not arbitrary. Concretely, if the time scale is q and the two
spatial widths are (with q_1 <= q_2), we have
q_1 / q ~ N_0^{-ell}, q_2 / q ~ N_0^{o(1)}.
So the smaller spatial direction "collapses" at rate ell, and the
larger direction stays at the time scale.
Geometrically: the chart is a very thin slab. Its "wasted" thinner
spatial direction does not contribute to base geometry, so the
problem essentially lives in the (t, Z) plane, where Z is a
coordinate along the longer spatial axis.
The rest of the paper treats this case by reducing it to a scalar
model in one spatial coordinate (the "horizontal model"), and
then -- once it is a scalar problem -- by pushing through the
scalar machinery from Section 5.
The section proves two structural propositions:
8.1-8.2 The horizontal model. In any finite-positive-narrowness
residual, we can switch to a chart where one spatial
coordinate is "wasted" and the effective base is
(t, Z) in the plane. On this chart, the trajectory family
becomes a version-1 (scalar) problem in disguise, with the
scalar signal X = the zero-width-spatial coordinate.
8.3 Wide intervals. Once in the horizontal model, we can find
"wide intervals" -- nontrivial time intervals on which the
scalar signal is not just locally consistent but has
definite slope. Wide intervals are what turn the local
polynomial fits from Section 5 into a global improvement.
Section 9 then case-analyses by the critical rate: how much time
the wide intervals actually cover. The three sub-cases (below,
at, above critical rate) each get a separate analysis.
8.1 Preparing the horizontal model
We start from a critical path in the finite-positive-ell regime.
The row-fit conclusion (Proposition 7.4) says: on many incidences,
the effective row F agrees with C + c . Jz where J is the 90-degree
rotation. That means the chart's "row-like" predictor is
structurally very rigid: it is a constant plus a rotation.
Concretely, let Z be the longer spatial axis (with q_2 ~ q), and
let Y be the shorter one (with q_1 ~ q N_0^{-ell}). After
paragraph 8.0's rescaling, the chart is a box of side ~1 in
(t, Z) times a box of side ~N_0^{-ell} in Y.
The row-fit conclusion, applied to the chart's row-restriction,
says the row is (approximately) of the form
F(t, Y, Z) ~ C + c . (-Z, t).
We record this structural fact:
Proposition 8.1 (Rigid row structure). On a substantial
residual of a finite-positive-narrowness critical path, after
possibly shrinking the mass by a subpower factor, the chart's
row F is nearly constant plus a linear functional in
J(t, Z) = (-Z, t) with fixed coefficients C, c bounded by K.
That is the practical content of Section 7 in this context. Now
we can proceed to convert the chart into a horizontal model. The
key is: the Y-direction (the narrow one) is "invisible" to the
row structure. Whatever happens in Y does not affect F's
row-like behavior.
So: integrate out Y, and the chart becomes a "horizontal"
(planar) chart with the base reduced to (t, Z). Define the
horizontal model by keeping:
horizontal base: (t, Z) in a compact chart in R^2;
horizontal mean: the Y-averaged hidden coordinate w;
horizontal signal: X ~ the actual scalar signal.
The averaged version is a version-1 (scalar) problem in disguise.
8.2 The horizontal model
We can now state the reduction properly.
Proposition 8.2 (Horizontal model). Let E be a
finite-positive-narrowness exact problem in C(d) with
horizon H(E) > 0. Then there is a further dense restriction of
E and a version-1 (scalar) exact problem E_hor in C(d_hor)
with the same horizon and d_hor < 1, such that E_hor:
- has base (t, Z) in a compact planar chart,
- has trajectory prior the (t, Z)-projection of the original
paths, and
- has scalar signal X equal to a fixed scalar linear
functional of (Z(t), t) plus small error.
The point: a finite-positive narrowness problem in 4 dimensions
reduces to a scalar problem in 1+1 dimensions. The Y-direction
does not contribute geometry. What remains is a classic "1D
Kakeya problem": a scalar signal over time, with many paths
producing the same scalar value at many sampled times.
Once we are in E_hor, all the machinery of Section 5 applies
directly. In particular:
- scalar-normal data hypotheses in (t, Z);
- the clustering lemma (5.1);
- the planar dot-product alternative (5.4);
- the higher-jet polynomial fit (5.6).
The result is a degree-M polynomial P(t, Z) fitting the scalar
signal along many incidences on long time intervals.
What "long" means is the next question, and it is the whole
content of Section 9.
8.3 Wide intervals
The scalar machinery gives a polynomial fit on a chunk of time
of length at least h/K. That "h/K" is what we called a "wide
interval". But wide intervals alone are not immediately a
contradiction; we need them to be "many" and "spread across the
chart" enough that the critical-path horizon bound fires.
Proposition 8.3 (Wide intervals). Under the hypotheses of
Proposition 8.2, there is a fixed s_0 > 0 and, on a positive
fraction of trajectories of E_hor, a collection of disjoint
wide intervals where the polynomial fit holds, whose total
labeled time is at least N_0^{-2 k s_0}. Moreover, the
interval lengths are at least N_0^{-k s_0} (up to subpower
factors).
The intervals are "wide": not of order h (the chart time scale),
but of a definite shorter scale. The parametrization uses the
critical rate k: how badly does the horizon of E_hor degrade
relative to the parent problem? The intervals are as long as
"wide" means.
The exponent ratio between 2ks_0 (total time covered) and ks_0
(individual interval length) is what Section 9 exploits: three
different comparisons, depending on how 2k compares to 1.
Section 8 is complete: the finite-narrowness problem has been
reduced to a scalar problem in the plane, and the scalar problem
produces wide intervals with definite length. Section 9
case-analyses by the critical rate.
SECTION 9. THE CASE ANALYSIS BY CRITICAL RATE
9.0 What this section does
We have reduced the finite-positive-narrowness case to a scalar
problem E_hor in the plane, with a definite parameter k > 0 -- the
"critical rate". Wide intervals exist, of length at least
N_0^{-k s_0}, total time at least N_0^{-2 k s_0}.
Whether these intervals are useful depends on how 2k compares to 1:
Case 2k > 1 (off-critical, wide). Wide intervals are short
relative to the total labeled time. The polynomial fits
exist and are short, but there are MANY of them, and their
total length is large. This is the "wide" regime, and the
off-critical wide proposition applies.
Case 2k = 1 (critical). The critical rate sits exactly at the
boundary. Wide intervals have length comparable to total
labeled time, but the individual intervals are exactly
N_0^{-1/2 s_0} which is short in absolute terms. This is
where two-color matching has to be used.
Case 2k < 1 (below critical). Wide intervals are long. Their
length can be improved by a parameter bootstrap: adding a
finer parameter to the chart gradually raises the
resolution at which intervals can be fit.
Each case emits a "time-improvement candidate" at thickness
strictly smaller than a_r. By Proposition 4.3, this upgrades to
a better true atlas and contradicts the critical path. So the
finite-positive-narrowness case is closed by Section 9.
The section proves three propositions, one for each sub-case:
9.1 (k > 1/2) Off-critical wide.
9.2 (k < 1/2) Parameter bootstrap.
9.3 (k = 1/2) Critical rate.
9.1 The setup: what "improvement candidate" needs
Recall Proposition 4.3 (candidate-upgrade criterion). To upgrade a
critical path to a contradiction, it is enough to produce a
time-improvement candidate on a substantial residual: a
polynomial test Q of degree 2, evaluated at thickness
a_fit = a_r N_0^{-e} for some fixed e > 0, that holds to positive
accuracy on a positive fraction of trajectories.
So the goal in every sub-case is: build a degree-2 polynomial Q
with |Q| <= a_r N_0^{-e} on many incidences, and derivative floor
|partial_w Q| >= B/K.
The wide-interval fits from Proposition 8.3 give us degree-M
polynomial fits on many incidences of total time
N_0^{-2 k s_0}. The question is whether this is enough to
construct a degree-2 candidate at the required thickness.
The answer depends on the value M and on how many independent
directions the fit must match across. Roughly:
-- M = 1: the fit is a scalar linear-in-(Z, t). Crossing
trajectories at two velocities already determines it; not
enough to force a "time-improvement" without an extra
structural input. This is the critical-rate case.
-- M = 2: the fit is a quadratic. If it holds on enough
distinct trajectories, the quadratic coefficients are
overdetermined and can be improved. This is the off-critical
and below-critical regimes.
The three sub-cases therefore differ in how they use the fits:
9.1 (k > 1/2): fits exist on many short intervals. The
intervals at scale N_0^{-k s_0} are just long enough to
distinguish two different quadratic fits. Overdetermined
least-squares upgrade fires.
9.2 (k < 1/2): fits exist on a few long intervals. The long
intervals are longer than the critical scale, so a
fine-parameter "bootstrap" simulates a critical case with
a finer base resolution.
9.3 (k = 1/2): fits exist on intervals of exactly the critical
length. This is the hardest case; two-color matching is
required to force any improvement.
9.4 Critical rate: k = 1/2
The critical case. When 2k = 1, every wide interval has length
exactly N_0^{-s_0/2}: long enough that a degree-M fit exists, but
not so long that the coefficient is determined by many independent
intervals. Pigeonhole and least-squares do not fire; the mass is
concentrated on a single scale.
The contradiction here comes from the two-color matching
machinery of Section 3: specifically Corollary 3.4 (local
matching of scalar graphs).
Proposition 9.3 (Critical rate). Suppose 2k = 1. Then a
substantial residual of the horizontal model carries a
time-improvement candidate with e >= c_0 s_0, for some fixed
c_0 > 0.
Setup. Fix a substantial residual in the horizontal model. Pick
a single trajectory and one of its wide intervals I, of length
|I| = N_0^{-s_0/2}. Split I into two colors: even-indexed
fine-time sub-bins are color 1, odd-indexed ones color 2. Both
colors sample the SAME underlying signal X but at disjoint
sample times.
On each color, Proposition 8.3 (wide-interval polynomial fit)
supplies a degree-M polynomial that fits X at the color's
sample times, at thickness a_r N_0^{-e_0}.
Two facts about the two polynomial fits.
(a) Value agreement. Both fits are close to X at interleaved
fine-time samples, hence close to each other (within a_r
N_0^{-e_0}).
(b) Norm difference. If the two fits agreed NORM-wise on the
whole interval -- i.e. their coefficients matched to
within the chart scale -- then merging the two colors
would produce a single fit on all of I with half the
color-local loss. That would contradict the optimality
of the wide-interval construction from Proposition 8.3.
So on any successful I, the fits must differ in norm by
at least a_r as polynomials.
Machine. Apply Corollary 3.4 with
M = a_r (the norm gap),
Delta = a_r N_0^{-e_0} (the value gap),
R = M / Delta = N_0^{e_0},
eps <= eps_* small, d <= d_0 < 1.
Both color lists of sheets (the polynomial fits, viewed as
algebraic sheets) have size at most R^{d+eps} -- subpower.
The list-index-base marginal condition: each fit must support at
least R^{-eps} mass on the color. This holds because the color
itself is a definite-mass restriction (Proposition 8.3 gives
positive mass to wide intervals).
The value-distance bound: (a) above gives |f_1 - f_2| <= Delta
R^eps on interleaved samples.
The norm-distance bound: (b) above gives ||f_1 - f_2|| >= M
R^{-eps} on the whole interval.
Corollary 3.4 declares these hypotheses jointly impossible.
Contradiction.
Conclusion: the critical rate cannot occur. A wide interval at
exactly the critical scale -- long enough to fit a polynomial
per color, but not long enough to overdetermine it across colors
-- forces an impossible list configuration through the value/
norm gap mechanism.
This closes Section 9. Combined with Sections 6-8, the
finite-positive-narrowness case 0 < ell < infinity is
completely excluded:
2k < 1: bootstrap reaches 2k >= 1 (Prop 9.2)
2k = 1: critical rate (Prop 9.3)
2k > 1: off-critical wide (Prop 9.1)
Section 10 handles ell = infinity and assembles the final proof.
SECTION 10. UNBOUNDED NARROWNESS AND THE FINAL ASSEMBLY
10.0 What this section does
The last narrowness case: ell = infinity. The chart at depth r
is arbitrarily skinny in one spatial direction; the aspect ratio
q_1/q_2 of the two spatial axes goes to zero faster than any
fixed power of N_0. No fixed "shorter axis" survives.
Unlike the finite-narrowness case (Section 8), there is no
coordinate system we can reduce the chart to using row-fit
alone. Any projection loses information; the chart's internal
geometry is now essentially concentrated in the direction of
the ONE longer spatial axis, which the chart rotates with
depth.
The argument: project to a scalar coordinate perpendicular to
the current longer axis. The projection is a version-1 scalar
problem. Apply Corollary 6.8 (zero horizon in the scalar model)
to it. That contradicts the assumed horizon of the parent
problem, and hence unbounded narrowness cannot carry a
counterexample.
10.1 Preparation
Take a critical path in the ell = infinity regime. From
Proposition 4.1:
- Nested true systems at depths 0 = r_0 < r_1 < ... with
density profile f(r) = d_* r + o(1) and horizon rate
h(r) = k r + o(1).
- The time scale is h_r = N_0^{-k r + o(1)}.
- The two spatial widths are q_1(r) << q_2(r), with
q_2(r) = N_0^{-r + o(1)} and q_1(r) trending to zero
faster than N_0^{-ell' r} for any fixed ell' > 0 along the
path.
Following the retained-edge (Section 7) and horizontal-model
(Section 8) constructions adapted to the ell = infinity limit,
we prepare:
- A scalar functional Z(.) selecting a coordinate along the
longer spatial axis.
- The projected trajectory (t, Z(t)) is affine in time:
Z(t) = Z_0 + V_z t for bounded |Z_0|, |V_z|.
- The projected measure is a submeasure of the trajectory
distribution restricted to (t, Z), with cell-count and
marginal-density bounds (the "prepared bounds").
- Trajectories of the projected problem are disjointly labeled
at each depth; label mass at depth r_i is at least K^{-1}
times the projected reference measure at that depth.
The projected problem is thus a version-1 (scalar) exact
problem in the plane.
10.2 The scalar projection
The key step: the projected problem inherits the horizon of the
parent, in the following precise sense.
Lemma 10.1 (Horizon inheritance under projection). If the
parent critical path has horizon rate k > 0, then the
projected scalar problem has horizon rate at least k/2 > 0
on a subpower-mass restriction.
The factor 1/2 is fixed but slack; the argument only needs "some
positive fraction". The claim follows from the parent's
definition of horizon plus a careful accounting of what is lost
in projection: at most a factor-2 loss from the two-sided
density concentration, and a subpower loss from sampling
restriction.
We are now in the situation: a scalar problem with POSITIVE
horizon. But Corollary 6.8 (from Section 6) says the scalar
model cannot have positive horizon. Contradiction.
So unbounded narrowness cannot carry a hypothetical Kakeya
counterexample.
10.3 Assembly of the main theorem
We now prove Theorem 1.1.
Theorem (Restatement). Every Kakeya set K subset of R^4 has
Hausdorff dimension exactly 4.
Proof. The ambient bound dim_H K <= 4 is trivial. We prove the
reverse.
Suppose for contradiction that K is a Kakeya set with
dim_H K < 4. By Theorem 2.1 (reduction), we get a fixed
sigma > 0 and an exact problem E in C(d; S, eta) with
d < 1 and horizon H(E) >= sigma > 0.
Apply Section 4.1 (extremal argument) to get the critical
problem: a problem E with minimal d_*, and then, within that
d_*, maximal normalized horizon k > 0. The problem is
extremal in both coordinates.
Apply Section 4.2 (critical path). Get nested true systems
with linear density profile f(r) = d_* r + o(1), horizon rate
h(r) = k r + o(1), and narrowness invariant ell_opt.
Classify by ell_opt.
Case 1: ell_opt = 0 (isotropic). Section 6 constructs a
time-improvement candidate in every substantial residual.
By Proposition 4.3 (candidate-upgrade criterion), this upgrades
to a better true atlas. That contradicts Proposition 4.1
(critical path) restricted to the deeper sub-problem. So
ell_opt = 0 is impossible.
Case 2: 0 < ell_opt < infinity (finite positive narrowness).
Section 8 reduces to a scalar problem in the plane.
Section 9 shows that this scalar problem produces a
time-improvement candidate for any critical rate k > 0,
via one of the three sub-cases:
- k < 1/2: bootstrap (Prop 9.2) reaches off-critical regime.
- k > 1/2: off-critical wide intervals (Prop 9.1).
- k = 1/2: two-color matching at critical scale (Prop 9.3).
Proposition 4.3 upgrades to a better true atlas. Contradicts
Proposition 4.1. So 0 < ell_opt < infinity is impossible.
Case 3: ell_opt = infinity (unbounded narrowness).
Section 10.2 projects to a scalar problem with positive
horizon. Corollary 6.8 (zero horizon in scalar model)
forbids this. Contradiction. So ell_opt = infinity is
impossible.
All three cases are impossible. Therefore no Kakeya set K
with dim_H K < 4 exists. That is, dim_H K = 4. QED.
10.4 What this structure buys us
The proof above is exactly what the original paper does.
Two structural features made it necessary.
First, no measurability was ever assumed. The reduction in
Section 2 uses outer measure in the analytic step ("what
fraction of the unit segment lands in the covering ball?").
This is well-defined for arbitrary sets. All density
bookkeeping is done via selections from outer measure, never
requiring measurability of K.
Second, no compactness or stickiness of the witnessing lines
was imposed. The proof handles dense families just as well
as sparse. The existential quantifier "a_e exists for each
e" does not need to select a_e measurably or boundedly.
These are exactly the "worst-case" Kakeya sets one expects to
be hardest. The argument treats them without any structural
assumption. Theorem 1.1 is now proved.
SECTION 11. GEOMETRIC CONSEQUENCES
11.0 What this section does
Theorem 1.1 is the specifically 4-dimensional Kakeya statement.
But several other geometric results follow directly by combining
Theorem 1.1 with known transfer theorems from the literature.
These transfer theorems convert a Kakeya-set lower bound into a
lower bound for other incidence problems: Nikodym sets, curved
Kakeya sets, projections. The transfers are due to other authors;
this paper supplies the input.
This section records the concrete consequences. It is the
shortest mathematical section in the paper.
11.1 Packing and Minkowski dimensions
The Hausdorff dimension is one way to measure a set; there are
others. Two standard notions are packing dimension (a "dual"
of Hausdorff dimension) and the Minkowski (or box-counting)
dimensions.
Every Kakeya set K in R^4 also has packing dimension 4, since
Hausdorff dimension is bounded above by packing dimension, which
is in turn bounded above by the ambient dimension 4.
If K is also bounded, then its lower and upper Minkowski
dimensions are both exactly 4. Recall that these are defined by
dim_M,lower K = liminf_{delta -> 0} log N_delta(K) / log(1/delta),
dim_M,upper K = limsup_{delta -> 0} log N_delta(K) / log(1/delta),
where N_delta(K) is the least number of balls of radius delta
covering K. The Minkowski dimension is NOT a topological
invariant, unlike Hausdorff dimension, so it is a strictly
stronger geometric conclusion.
Corollary 11.1 (Minkowski dimension). Every bounded Kakeya
set K in R^4 has lower and upper Minkowski dimension 4.
The proof is: dim_H <= lower Minkowski <= upper Minkowski <= 4,
and dim_H = 4 by Theorem 1.1.
11.2 Nikodym sets
A Nikodym set is a measurable set Omega in R^d with the
following property: for almost every point p, there is a line
through p such that Omega is "dense near p" on that line,
meaning Omega has positive density at p along the line, but
Omega has measure zero globally.
Nikodym conjectured in the 1930s that no such set exists in
R^d for d >= 2. This was disproved: Nikodym sets exist. But how
large can they be? That is a dimension question.
Gao, Liu, and Xi (2025) proved a general transfer: in any
constant-sectional-curvature Riemannian manifold of dimension
d, the Hausdorff dimension of every Nikodym set is at least
the Kakeya-set dimension in R^d. Theorem 1.1 gives the Kakeya-
dimension-4 lower bound in R^4. Combining:
Corollary 11.2 (Four-dimensional Nikodym). Let (M^4, g) be
a Riemannian manifold of constant sectional curvature. Then
every Nikodym set Omega subset of M has dim_H Omega = 4.
This resolves Conjecture 1.4 of Gao-Liu-Xi in dimension four.
The reverse inequality dim_H Omega <= 4 holds trivially for any
subset of a 4-manifold.
11.3 Fixed-chart curved Kakeya sets
Replace "straight lines" with a family of curves. A "curved
Kakeya set" is a set E such that for every apex y, there is a
curve gamma_y that starts at y and stays in E. The curves are
supplied by a smooth map phi via an implicit equation.
For a specific case (fixed-chart phase; see Gao-Liu-Xi), the
core curves can be straightened by a single smooth change of
coordinates. Specifically: there is an open time interval, an
open direction ball, and a smooth coordinate change kappa such
that
phi(kappa(u, s); y) = u . y + s h(y) + q(y) + f(s),
with det Hessian of h nonzero. In these coordinates, the core
curves are straight lines with directions nu(y) = (-grad h(y), 1)
/ sqrt(1 + |grad h(y)|^2).
Corollary 11.3 (Fixed-chart curved Kakeya). Under the
fixed-chart hypotheses above, any curved Kakeya set E subset
of R^4 has dim_H E = 4.
The proof: apply the straight-line Theorem 1.1 in the kappa
coordinates. Hausdorff dimension of a set is preserved by any
smooth local diffeomorphism. Then pull back the conclusion.
11.4 Three-dimensional quadratic curved Kakeya
A different curved Kakeya problem: curves parameterized by two
variables (y, r) with a specific quadratic structure. For a real
2x2 matrix B of rank 1, Nadjimzadah (2025) shows that the
corresponding curved segments, after a quadratic lift to R^4,
become straight-line segments in every direction.
Lift: define pi(x, z, t) = (x + t U z, t) where B = U V is a
rank-1 factorization. Then the lifted trajectories have
projection back to R^3 as precise quadratic curves.
Corollary 11.4 (Quadratic curved Kakeya). Let E be a curved
Kakeya set in R^3 whose curves are the projection of
straight-line segments under a rank-1 quadratic lift. Then
dim_H E = 3.
The proof: apply Theorem 1.1 in the 4-dimensional lift,
then use the linear-projection dimension transfer.
11.5 What is not claimed
The transfers in this section are due to other authors. This
paper does not improve them. What it does is supply the
dimension-4 Kakeya input they need.
In particular, we do NOT address:
- curved Kakeya sets whose curves cannot be straightened by a
single smooth coordinate change (the "variable-chart" case);
- Nikodym sets on manifolds with non-constant curvature;
- curved Kakeya sets for a phase phi with mixed rank
vanishing.
These require their own transfer theorems, which the present
paper does not supply. The fixed-chart and constant-curvature
cases are the ones that Theorem 1.1 answers.
That closes the mathematical content of the paper.
APPENDIX A. A MATH TOOLKIT FOR HIGH SCHOOL READERS
A.0 Purpose
The main text assumes calculus, basic linear algebra, and (in
Section 3) complex numbers. If you have those, you can read
most of the paper. This appendix fills a few specific gaps that
come up in the presentation.
A.1 The N_0 business: exponents and logarithms
Throughout, N_0 -> infinity is a large parameter. "At
resolution N_0" means "after discretizing to cell size 1/N_0."
Most quantities in the paper are powers of N_0.
We write A ~ N_0^{alpha} to mean "A is N_0^{alpha}, up to
subpower factors" (next subsection). When you see such an
expression, read it as "the exponent alpha is what matters."
Rules of exponent arithmetic apply as usual:
N_0^{alpha} * N_0^{beta} = N_0^{alpha + beta}
N_0^{alpha} / N_0^{beta} = N_0^{alpha - beta}
(N_0^{alpha})^k = N_0^{alpha k} for any fixed k.
Logarithms grow slower than any fixed power: for any fixed C
and any eps > 0,
(log N_0)^C << N_0^eps for N_0 large enough.
That is why a "polynomial bound in log N_0" never interferes
with a "subpower factor."
A.2 Subpower factors
A quantity K_0 >= 1 is subpower if
log K_0 / log N_0 -> 0 as N_0 -> infinity.
Equivalently: for every eps > 0, once N_0 is large enough,
K_0 <= N_0^eps. In words: a subpower factor is smaller than any
fixed power of N_0.
Examples of subpower quantities: (log N_0)^C for fixed C,
2^{sqrt(log N_0)}, and any constant like 100 or 10^{10}. (A
constant is subpower because log(constant) / log N_0 -> 0.)
Throughout the paper we prove things of the form:
error <= K_0 * N_0^{-alpha}, K_0 subpower, alpha > 0.
Such an error is asymptotically negligible compared with
N_0^{-beta} for any fixed beta < alpha. That is why we
"cheerfully drop" subpower factors: they never interfere with
a strictly positive power gap. If a bound says
"<= K_0 N_0^{-alpha}", it means "= o(N_0^{-beta}) for any beta
< alpha."
A.3 Vectors and linear algebra basics
R^n is n-tuples of real numbers. A vector v = (v_1, ..., v_n)
can be added to another vector and multiplied by a scalar. The
dot product v . w = v_1 w_1 + ... + v_n w_n produces a
scalar.
An m x n matrix A maps R^n to R^m. If f: R^n -> R^m is
differentiable, its Jacobian Df(u) is an m x n matrix whose
(i, j) entry is partial f_i / partial x_j evaluated at u.
Symmetric and skew parts. Any square matrix A splits uniquely
as S + K, where S = (A + A^T)/2 is symmetric (S^T = S) and
K = (A - A^T)/2 is skew-symmetric (K^T = -K). A gradient field
grad P has symmetric Jacobian, because cross partials commute:
partial_i partial_j P = partial_j partial_i P. Conversely, a
vector field with symmetric Jacobian is a gradient, up to
constant. This is the reason a "nearly symmetric Jacobian" is
enough in Section 6.4 to get a potential.
Tensors. The k-th derivative of a nice function is a symmetric
k-tensor: an array indexed by k coordinates, invariant under
permuting the coordinates. The "cubic Taylor coefficient" of a
function is a symmetric 3-tensor; the "quartic Taylor
coefficient" is a symmetric 4-tensor. In Lemma 6.6 we test
these by slotting in 7 specific points and taking all
contractions.
Norms and determinants. For a vector v, |v| = sqrt(v . v). For
a matrix A, ||A|| = max_{|v|=1} |A v| (operator norm). The
determinant det A scales volumes: det A is the signed volume
of the image under A of the unit cube. In 2 dimensions,
det [[a, b], [c, d]] = ad - bc.
The 90-degree rotation. In the (t, Z)-plane, the matrix
J = [[0, -1], [1, 0]] rotates by 90 degrees: J(t, Z) = (-Z, t).
Note J^2 = -I. This J appears in Section 7 as the "rotated
row" and is the geometric heart of the row-fit statement.
A.4 Complex numbers and polynomial roots
Complex numbers C = { a + bi : a, b in R }, with i^2 = -1.
They form a field: addition, subtraction, multiplication,
division all work with the usual rules. Any real number is a
complex number (b = 0). The modulus |a + bi| = sqrt(a^2 + b^2).
The fundamental theorem of algebra. Every non-constant
polynomial with complex coefficients has at least one complex
root. Equivalently: a polynomial of degree d has exactly d
roots in C, counted with multiplicity. Equivalently again:
P(z) = c (z - r_1)(z - r_2) ... (z - r_d),
where c is the leading coefficient and r_1, ..., r_d are the
roots (possibly repeated).
Multiplicity. A root r has multiplicity k if (z - r)^k divides
P but (z - r)^{k+1} does not. The total count, with
multiplicity, is always d.
Simple roots. A root is simple if its multiplicity is 1. If P
has a simple root at r and holomorphic coefficients nearby,
there is a holomorphic branch of the root r(z) defined in a
small neighborhood of r. This is the implicit function theorem
for polynomial roots: if partial_w P(r) != 0, then w -> P(w)
has a nonzero derivative at r, so a nearby zero moves
smoothly in r.
Algebraic sheets. A holomorphic function f(u) is an "algebraic
sheet of degree at most D" if there is a nonzero polynomial
Q(u, f(u)) = 0 of total degree at most D. The prototypical
example is f(u) = the branch of the equation w^2 - S(u) = 0,
i.e. f(u) = sqrt(S(u)), where S is a nice function. This is an
algebraic sheet of degree 2.
Why the Section 3 lemmas are what they are. In
Lemma 3.1 (Bernstein-Remez norm comparison), the statement is:
"bounded sheet on a positive-measure subset must be bounded on
the whole ball." For a genuinely algebraic sheet, this follows
from the fundamental theorem of algebra applied to the relation
Q(u, w) = 0: roots cannot "escape" without going through a
branch point, and the clearance from branch points is what the
norm bound controls.
In Lemma 3.2 (continuation), the statement is: "extending an
algebraic sheet along a path costs at most a polynomial in the
inverse clearance." This is because each "step" is either a
small Taylor-extension of a holomorphic piece (costs A per
unit clearance) or a change-of-branch at a branch point (costs
A^fixed total). Fixed-degree bounds make these cost-dependencies
polynomial.
A.5 Hausdorff dimension, briefly
We already introduced Hausdorff dimension in Section 1.3 by the
cover-sum formula:
H^s(K) = lim_{delta -> 0} inf sum_i (diam U_i)^s,
over countable covers of K by sets U_i of diameter at most
delta. The Hausdorff s-dimensional measure H^s is either 0 or
infinity for most K (with a single critical exponent), and that
critical exponent is the Hausdorff dimension.
Two facts we use repeatedly:
1. Lipschitz invariance. Smooth (or Lipschitz) transformations
of coordinates preserve Hausdorff dimension. This is why
"straightening coordinates" in Sections 10 and 11 is fine:
dimension is a coordinate-independent notion under nice maps.
2. Lower bounds via measure. To prove dim_H K >= s, it
suffices to construct a probability measure mu on K such
that mu(B_r) <= C r^s for every ball B_r of radius r
(Frostman's lemma). This is the standard conduit from
"many incidence" to "dimension."
A.6 Outer measure and arbitrary sets
The Hausdorff measure above is defined using "inf over covers."
That inf exists for ANY set K, measurable or not. So the
definition of Hausdorff dimension is well-defined for every
subset of R^n, without any measurability assumption.
The rest of the proof respects this. When Section 2 says "take
a ball cover of K at scale delta," the cover exists for
arbitrary K. When it counts "volume of the ball times the
fraction of the unit segment in the ball," that's an integral
against a measure defined via inf-cover construction. Nothing
requires K to be measurable.
Why this matters. There exist Kakeya sets that are not Lebesgue
measurable. If the proof used measurability, it would only
cover "nice" Kakeya sets, not the worst-case ones. Section
2's reduction is by outer measure precisely to avoid this.
A.7 What to skim on a first read
If you want to understand the shape of the proof without
absorbing every detail, skim on the first pass:
- Section 3: the proof of Theorem 3.3 (two-color matching).
You can accept it as a black box on first read.
- Section 5: the entropy-chain details behind Propositions
5.5 and 5.6. The statements are geometric; the proofs are
long entropy bookkeeping.
- Section 6: the technical details of velocity-graph
interpolation and degree restoration. Read the statement
of Proposition 6.7 (restoration to a quadratic test) and
accept the rest.
- Section 9: the internals of Proposition 9.2 (bootstrap).
Read the statement; the iteration is bookkeeping.
The other sections should be readable at their own pace. The
overall skeleton -- Sections 1, 2, 4, and the final assembly in
Section 10.3 -- is the shortest coherent summary.
FURTHER READING
The proof in this paper rests on two kinds of prior work:
the Kakeya problem itself, and the analytic / combinatorial
tools imported at specific steps. Here is a guide.
On the Kakeya problem.
Besicovitch (1928). Constructed Kakeya sets with Lebesgue
measure zero in any dimension. The starting point of the
whole subject.
Davies (1971). Proved dim_H = 2 for Kakeya sets in R^2.
Wolff (1995). Introduced the hairbrush method. Proved
dim_H K >= (n+2)/2 in R^n, which gives dim >= 3 in R^4.
Bourgain (1999). Connected Kakeya to additive combinatorics
via sumsets. Katz-Tao (2002) and Katz-Laba-Tao (2000)
developed the arithmetic method. Laba-Tao (2001) improved
the upper Minkowski bound in R^n for n >= 4.
Guth-Zahl (2018). Polynomial method; proved the "planebrush"
bound 3 + 1/40 in R^4. The start of the recent wave of
4D lower bounds.
Katz-Rogers (2019). Improved to 3.059 using planebrush
techniques.
Rai-Choudhuri (2026). Proved the "sticky" 13/4 = 3.25 bound
for Kakeya sets sharing their lines with a common family.
Wang-Zahl (2025). Proved the full three-dimensional Kakeya
theorem (Hausdorff dimension 3). Simplification by
Guth-Wang-Zahl (2026).
On the analytic input.
Bennett-Carbery-Tao (2006). Multilinear Kakeya and
multilinear restriction. Guth (2010) proved the endpoint
multilinear Kakeya theorem.
Bourgain-Guth (2011). Wedge-weighted versions of
multilinear Kakeya, as used in Section 6.1.
Carbery-Valdimarsson (2009). Determinant multilinear
Kakeya, as used in Section 6.1 with n = m + 1.
Wang-Zahl (2025), start of Section 8. The planar
dot-product alternative (Theorem 5.4 in this paper),
imported from sticky Kakeya work.
On the combinatorial input.
Sudakov-Szemeredi-Vu (2005). Graph version of the
Balog-Szemeredi-Gowers theorem, with retention of original
edges.
Tao-Vu (2006). Textbook formulation of the graph BSG.
On the transfer theorems.
Gao-Liu-Xi (2025). Transfers Kakeya-set lower bounds to
Nikodym sets on constant-sectional-curvature manifolds,
and to curved Kakeya sets for fixed-chart phases.
Nadjimzadah (2025). Quadratic lifts from R^3 curved
Kakeya problems to R^4 straight-line Kakeya.
Keleti-Mathe. General-direction Kakeya transfers used in
Section 11.4 for curved Kakeya consequences.
Each of these is discussed in the corresponding place in the
main text. The present paper's contribution is the
4-dimensional Kakeya dimension theorem itself, and its
integration into the specific machinery developed in
Sections 2-10.
THE END.