=============================================================================
AN EXACT HAUSDORFF GAUGE FOR SLE
a complete, friendly guide for curious teenagers
=============================================================================
rewritten from
"An exact Hausdorff gauge for SLE: a moment-integral
and finite-batch construction" (OpenAI, Sept 2026)
This version is shorter and slower than the original. It does not
skip any of the ideas in the original; it merely explains them in
plain language, with small running examples, before using them.
Everything is self-contained. The only prerequisites are:
- you know what a function and a limit are,
- you have met the idea of a probability (a number between 0
and 1 measuring how often an event happens).
-----------------------------------------------------------------------------
CONTENTS
-----------------------------------------------------------------------------
Part A: Background
A.1 How big is a jagged set? (Hausdorff measure)
A.2 Dimension as a critical exponent
A.3 Gauges: better rulers than r^d
A.4 What is SLE_kappa?
A.5 The exact question this paper answers
A.6 A quick map of the whole argument
Part B: The main theorem
Part C: The proof machine -- ingredients
C.1 Conformal radius and the inner-time clock
C.2 Centered laws: biasing toward one point
C.3 The multipoint estimate
C.4 A canonical mass measure on the trace
C.5 The disk model and its mass moments
Part D: Turning moments into a gauge
D.1 Why moment batches work
D.2 The finite lists and the gauge formula
Part E: Two covering arguments
E.1 The lower bound (positivity)
E.2 The upper bound (finiteness)
Part F: Full proofs and where things live
-----------------------------------------------------------------------------
PART A -- BACKGROUND
-----------------------------------------------------------------------------
A.1 HOW BIG IS A JAGGED SET? (HAUSDORFF MEASURE)
Imagine a wiggly path P in the plane. It is a fractal: if you zoom
in, it still looks just as wiggly. Ordinary length is not useful:
the path has infinite length, because any polygonal approximation
that keeps up with the wiggles uses more and more segments.
Hausdorff had a famous trick. Pick a positive number s. For any
covering of P by small planar blobs (disks or squares) with
diameters delta_1, delta_2, ..., define
cost_s(covering) = delta_1^s + delta_2^s + ...
Take the smallest possible cost over all countable coverings whose
blobs have diameter at most some threshold delta. Call that number
H^s_delta(P). Finally let delta shrink to 0:
H^s(P) = lim_{delta -> 0} H^s_delta(P).
This is the s-dimensional Hausdorff measure of P.
Sanity checks:
- If P is a smooth line segment and s = 1, H^s(P) is its length.
- If P is a smooth surface and s = 2, H^s(P) is its area.
- If s is too small, H^s(P) = infinity.
- If s is too large, H^s(P) = 0.
A.2 DIMENSION AS A CRITICAL EXPONENT
Cover the interval [0,1] by N segments of length 1/N. Each term
in cost_s is (1/N)^s and there are N of them, so the total is
N^{1-s}. As N grows, this goes to 0 for s > 1 and to infinity for
s < 1: the critical exponent is 1.
Cover the unit square by N^2 squares of side 1/N. Cost is
N^2 * (1/N)^s = N^{2-s}. Critical exponent 2.
For any set, exactly one s separates "measure infinite" from
"measure zero". That s is the Hausdorff dimension, dim_H(P).
Beffara (2008) proved for SLE_kappa with 0 < kappa < 8:
dim_H(curve) = d = 1 + kappa / 8 .
For kappa in (0,8) this d lies strictly between 1 and 2: bigger
than a line, smaller than a region.
Useful shorthand we will keep returning to:
a = 2 / kappa
d = 1 + 1/(4a) = 1 + kappa/8
alpha = 2 - d = 1 - 1/(4a)
p = 4a - 1
Then 0 < alpha < 1. The algebra below keeps using the same identity:
2 * a * alpha = p / 2 .
The "complementary exponent" alpha controls how fast mass piles up
as you zoom to the tip of the curve. p controls an angle weight
sin^p(theta) that will appear in the change of law.
A.3 GAUGES: BETTER RULERS THAN r^d
Hausdorff's "ruler" r^s is not the only option. Replace r^s by any
continuous increasing function h(r) with h(0) = 0 and h(r) > 0 for
r > 0. Such an h is called a "Hausdorff gauge" or "dimension function".
Redo the construction with h in place of r^s:
cost_h(covering) = h(delta_1) + h(delta_2) + ...
Then take the same two limits (smallest cost with cap delta, then
delta -> 0). The result is H^h(P), the h-Hausdorff measure of P.
Any choice of gauge h yields a legitimate measure. The question is
whether it gives a useful number for our particular curve.
Power gauges cannot help beyond d. If h(r) = r^s with s < d, then
H^h(P) = infinity. If s > d, H^h(P) = 0. At the critical power
h(r) = r^d, Rezaei (2018, Theorem 1.1) proved that for chordal
SLE_kappa in the standard normalization the answer is zero:
H^{r^d}(trace) = 0 almost surely.
So the "dimension" power gauge is not enough. But gauges are more
flexible. For instance
h(r) = r^d * ( log log(1/r) )^{beta}
is strictly bigger than r^d near r = 0 whenever beta > 0 (because
log log(1/r) grows as r shrinks). So H^h could be positive even
when H^{r^d} is zero.
Schramm (2006, Problem 7.1) asked: does some gauge h give a
sigma-finite Hausdorff measure on the SLE_kappa trace? He suggested
r^d log log(1/r) as a candidate. This paper proves a positive
answer (existence) along a different route: instead of guessing a
closed-form h, it BUILDS one from computable deterministic numbers.
The gauge built this way may, or may not, agree asymptotically with
the log-log candidate; the paper leaves that comparison open.
A.4 WHAT IS SLE_KAPPA? (BRIEF INTRO)
Schramm-Loewner Evolution (SLE_kappa) with parameter kappa > 0 is
a family of random plane curves in the upper half-plane
HH = { z = x + iy : y > 0 }.
Each curve starts at the origin and grows off to infinity. SLE
curves are the canonical objects of two-dimensional critical statmech:
they appear as scaling limits of critical Ising clusters, percolation
interfaces, loop-erased walks, and many others.
The construction is best described by a differential equation.
Let B(t) = B_t be a standard Brownian motion: B_0 = 0, continuous
random path, and B_t - B_s is normally distributed with mean 0 and
variance t - s. Fix kappa, put
a = 2 / kappa,
and for each starting point z in HH solve the ordinary differential
equation
dg_t(z)/dt = a / ( g_t(z) - B_t ), g_0(z) = z.
This is the Loewner equation. For each time t >= 0 the map z |-> g_t(z)
is conformal (angle-preserving), and it takes some remainder domain
H_t = HH minus the curve up to time t
onto HH itself. The curve itself is recovered as the tip:
gamma(t) = limit of g_t^{-1}( U_t + i*epsilon )
as epsilon -> 0,
where U_t is a suitable boundary point on the real axis. The
important fact (Rohde-Schramm 2005, Theorem 5.1) is that for the
Brownian driver B_t these limits exist and produce a continuous
curve that grows out of 0 and goes to infinity.
The full trace is Gamma = gamma([0, infinity)). Some facts:
- For 0 < kappa <= 4 the curve is a simple arc (no self-touching).
- For 4 < kappa < 8 the curve is a self-touching arc: it touches
itself but does not cross.
- In every case the curve is "simple enough" to be a fractal of
dimension 1 + kappa/8.
- The trace is never revisits itself for a positive amount of
capacity time.
There are two natural notions of "how much time has passed" on the
curve. In the paper's main theorem the parametrization uses half-
plane capacity 2t; inside the Loewner equation we adopt an internal
time ("a*t"), which is what the identity dg/dt = a/(g - B_t) is
normalized to give. The two differ only by a constant rescaling,
which does not affect "positive finite" statements about Hausdorff
measure on positive-time subintervals. From now on we use the
internal time, and simply write B_t.
A.5 THE EXACT QUESTION THIS PAPER ANSWERS
Let h be any fixed Hausdorff gauge. Consider chordal SLE_kappa
(0 < kappa < 8) with trace Gamma. We ask:
Does there exist a deterministic gauge h = h_kappa,
depending only on kappa, such that
0 < H^{h_kappa}( gamma([s,t]) ) < infinity
for every real 0 < s < t < infinity,
simultaneously, on a single event of probability 1?
"Deterministic" means h_kappa does not depend on the sample curve.
"Simultaneously" means we do not get a different event for each
interval -- one exceptional null set works for all of them at once.
Because H^{r^d} = 0 (Rezaei), we are not allowed to take h(r) = r^d.
We need something strictly bigger near zero. We cannot take
h(r) = r^s for s < d because that would make the measure infinite.
So h must be a non-power gauge: bigger than r^d by a slowly-varying
factor, but not so big that the measure blows up to infinity.
The paper gives an affirmative answer, and in fact produces an
explicit h_kappa, specified by deterministic integrals of a Gaussian
process. That is the content of the Main Theorem (Part B below).
A.6 A QUICK MAP OF THE WHOLE ARGUMENT
We are going to prove the Main Theorem by the following steps.
Each one is a self-contained object that we introduce, use, and
cite back to later.
Step 1. Define a mass measure mu on the trace. Think of mu as
"how much of the curve has passed through here". It is
an adapted, increasing measure on HH that starts empty
at t = 0 and grows as the curve grows. Key property:
its conditional future mean is an explicit Green-type
expression G_S(x) * (area).
Step 2. Bias the curve toward a marked point x. Freeze at the
time when x's conformal radius reaches a prescribed
smaller value. The bias ("centered law" P*_x) is a
probability measure on stopped paths that makes x look
like the origin of a unit disk. The bias has mean one
and exists at all finite inner levels.
Step 3. Measure local mass in a fixed subdisk of radius 1/4
centered at x. Call the accumulated mass up to inner
time ell the variable L_ell, and its limit L. Prove
that every integer moment m_n = E[ L^n ] is finite
and strictly positive, using the multipoint estimate.
(The multipoint estimate is a uniform bound on the
probability that many points simultaneously reach
their own prescribed radii.)
Step 4. Show that the numbers m_n(ell), m_n are computable from
kappa alone via Wiener integrals, angle averages, and
area integrals, with no reference to a specific random
sample.
Step 5. Build the gauge from these numbers alone. Fix an order
n. Keep a level ell_n where 3/4 of m_n is already there.
Repeat dyadic thresholds A_k with specific multiplicity
N_{nk} (chosen from m_n and m_{n+1}). These form one
"batch". Each batch has a cleanup entry that catches
points missed by all tests. Convert every entry into a
radius r_{ni} and a threshold b_{ni}, and take the
gauge
h_kappa(r) = inf over all entries of
b_{ni} * r_{ni}^d * max{1, (r/r_{ni})^2}.
Step 6. Lower bound: H^h is positive almost surely on every
nontrivial interval. This uses a "past-mass identity"
that pays off overfull squares by mass that has already
gone by, plus the summable tail estimate from Step 5.
Step 7. Upper bound: H^h is finite almost surely on every
bounded box. This uses successful finite-scale tests
along the curve: greedy selection makes their disks
disjoint, one actual terminal mass pays for them, and
the cleanup entry pays for the few cells the tests
missed.
Step 8. Assemble. Intersect the probability-one events of the
lower and upper bounds. Because each finite-time image
gamma([s,t]) is compact and hence lies in some box B_m,
the finite-sigma-finite claim follows.
The rest of the document expands each step in this order, first
carefully (Part C) as standalone results, then in the way that
they combine (Parts D and E), with full proofs in Part F.
-----------------------------------------------------------------------------
PART B -- THE MAIN THEOREM
-----------------------------------------------------------------------------
Fix 0 < kappa < 8. Let gamma be chordal SLE_kappa, starting at 0
and going to infinity in the upper half-plane HH, parametrized by
half-plane capacity 2t (equivalently, in internal time as given by
the Loewner equation dg_t/dt = a/(g_t - B_t) with a = 2/kappa).
Let Gamma = gamma([0, infinity)) be the entire trace.
THEOREM (Main). There exists a deterministic function
h_kappa : [0, infinity) --> [0, infinity),
continuous, nondecreasing, with h_kappa(0) = 0 and h_kappa(r) > 0
for every r > 0, depending only on kappa, with the following
properties.
(S1) POSITIVITY ON INTERVALS. On a single event of probability 1,
0 < H^{h_kappa}( gamma([s,t]) ) < infinity
holds simultaneously for every real 0 < s < t < infinity.
(S2) EXPLICIT FORMULA. The gauge is the infimum
h_kappa(r)
= inf over all pairs (n, i) with n >= 1 and 1 <= i <= I_n
of the quantity
b_{ni} * r_{ni}^d * max{1, (r/r_{ni})^2},
with h_kappa(0) = 0. The numbers b_{ni} > 0 and r_{ni} > 0
are explicit deterministic sequences built in Part D. Each
batch index n contributes finitely many entries (i=1..I_n);
all batches together give the family over which the infimum
is taken.
(S3) DETERMINISM. The numbers b_{ni} and r_{ni} are specified by
deterministic integrals involving only kappa: a Wiener integral
of the Loewner equation, an average over the angle law
pi_a(theta) proportional to sin^{4a}(theta), and an integral
of a fixed indicator over the unit disk. In particular,
h_kappa depends only on kappa. It does not depend on the
sample curve, and it does not depend on the interval [s,t]
appearing in S1.
(S4) FINITENESS IN BOXES. For every integer m >= 1, put
B_m = [-m,m] + i[0,m]. Then
E[ H^{h_kappa}( Gamma cap B_m ) ] < infinity.
Consequently, on a single event of probability 1,
H^{h_kappa}( Gamma cap B_m ) is finite for every m
simultaneously.
Several clarifications:
* SIGMA-FINITENESS. The trace outside the origin has nonzero
sigma-finite h_kappa-measure. Because H^h is countably additive
and Gamma sits inside the increasing union of boxes B_m, and
because each box has finite measure, the sigma-finiteness follows
directly from S4.
* POSITIVITY IS ASSERTED ONLY ON POSITIVE-TIME INTERVALS. S1
excludes intervals that touch s = 0. SLE has zero local mass at
the origin (in any time normalization), so this restriction is
natural: the interesting claim is about positive-time subarcs.
* ONE EVENT FOR ALL INTERVALS. The exceptional null set where
S1 fails does depend on kappa but not on the sample nor on the
particular interval. So "for every 0 < s < t" is a single
probability-1 event, not a separate almost-sure claim for each
choice of [s,t].
* IT IS NOT r^d log log(1/r). The gauge h_kappa is not claimed to
match any closed-form expression at small scales. A companion
paper (OpenAI 2026, an "explicit exact Hausdorff gauge") proves
that the regular closed-form gauge
r^d * (log log(1/r))^{(2-d)/2}
is also exact on chordal SLE, by a different method (quantitative
tails and dense visits). The present paper builds its own gauge
from moments and leaves the asymptotic comparison between the
two open.
* NORMALIZATION. The main theorem states its result in the
"capacity 2t" parametrization. The rest of the document works
in internal time (the parametrization of the Loewner equation
dg/dt = a/(g - B_t) with a = 2/kappa), which differs from
capacity 2t by a constant rescaling. Constant rescaling does not
change "positive finite" statements about Hausdorff measure on
positive-time subintervals, because images of [s,t] under either
parametrization are the same set of geometric points; only the
time labels differ.
Here is what remains. Part C introduces the ingredients: conformal
radius and inner time (C.1), centered laws (C.2), the multipoint
estimate (C.3), the mass measure mu on the trace (C.4), and the
centered disk moments (C.5). Part D turns the moments into batches
and states the gauge formula D.2 with its entries explicitly. Part
E proves the two covering bounds: positivity (E.1) and finiteness
(E.2). Part F assembles the pieces and states every intermediate
lemma and proposition used, with its proof.
-----------------------------------------------------------------------------
PART C -- THE PROOF MACHINE: INGREDIENTS
-----------------------------------------------------------------------------
C.1 CONFORMAL RADIUS AND THE INNER-TIME CLOCK
A conformal map is an angle-preserving, locally invertible complex
map. The Loewner maps g_t: H_t -> HH are the prototypes.
The conformal radius of a domain U at a point x in U is defined as
R_U(x) = |F'(0)|,
where F: DD -> U is a conformal map from the unit disk DD onto U
with F(0) = x, normalized so that F'(0) is a positive real number.
(Riemann's mapping theorem says such an F exists and is unique up
to rotation at 0; the choice of normalization makes R_U(x) a single
positive real number.) Intuitively, R_U(x) is "how big U looks
around x" measured in units of the unit disk.
Useful concrete cases:
R_HH(z) = 2 Im(z) for z in the upper half-plane,
R_DD(w) = 1 - |w|^2 for w in the unit disk.
Conformal covariance:
if J: U -> V is conformal and y = J(x), then
R_V(y) = |J'(x)| * R_U(x).
And the Koebe quarter theorem:
(1/4) R_U(x) <= dist(x, boundary U) <= R_U(x).
Now return to the Loewner evolution. Take a fixed starting point
x inside HH (or inside a fixed conformal image D = J(HH)). Let
R_t(x) = R_{H_t}(x)
be the conformal radius of the current domain H_t at x. As t grows,
the curve eats away at H_t, so R_t(x) shrinks.
The key computation from the Loewner equation: writing Z_t = g_t(z)-B_t
where z = J^{-1}(x), and Y_t = Im(Z_t), and putting
V_t = Y_t / |Z_t|^2 > 0,
we get
d log R_t(x) = - 2 a V_t^2 dt.
(Read `d` as the small change in t; the equality holds with
probability one, simultaneously for all t.) So R_t(x) shrinks
exponentially at a rate controlled by V_t^2, which depends on how
close the tip is to x. Define the inner-time clock
s_x(t) = integral_0^t V_u^2 du.
Then
R_t(x) = R_0(x) * exp( - 2 a s_x(t) ).
So inner time is the natural local clock at x, and reducing the
conformal radius from R_0(x) to e^{-2a * ell} * R_0(x) takes inner
time exactly ell.
The level times are, for a chosen level ell > 0,
tau_ell(x) = inf { u : s_x(u) = ell while x is alive },
and tau_ell(x) = +infinity if x is never alive up to that
level.
"Alive" means x is inside the current domain H_t. A point that has
been swallowed has been "frozen" out, and after being swallowed it
can no longer be reached by the shrinking domain. On the event
tau_ell(x) < infinity, the current conformal radius is exactly
R_0(x) * e^{-2a * ell}.
Two more facts on the same theme, stated without proof but as
direct consequences of the definitions:
* The physical clock and the inner clock are related by
d t / d s = Y_s^2 / sin^4(theta_s),
where theta_s = arg(Z_s). In words, the physical time advances
slowly when the tip is near a boundary angle (sin theta small),
and quickly when the tip is well away from the boundary.
* The angle theta_t itself evolves in inner time s according to
the SDE
d theta_s = V_s dB_s + (1 - 2a) cot(theta_s) V_s^2 ds,
whose volatility V_s dB_s is that of a Brownian motion when
reparametrized by inner time. This SDE is central in C.2.
The whole purpose of the inner-time reparametrization is to isolate
the following picture: for a marked point x, once we "measure
time" with s_x, the current conformal radius around x shrinks at a
fixed exponential rate. Everything else -- the angle of x relative
to the tip -- is encoded by the diffusion theta_s.
C.2 CENTERED LAWS: BIASING TOWARD ONE POINT
Fix a starting point x inside HH (or inside a fixed conformal image
D = J(HH)). The idea of the centered law is to "redraw" the SLE
curve with a bias in favor of passing close to x. The bias is not
ad hoc: it is exactly what the Green function of the remaining
trace -- a measure of "how much future mass remains near x" --
would predict.
The Green weight. Define
G_t(x) = R_t(x)^{-alpha} * sin^p( theta_t(x) ),
with theta_t(x) = arg(Z_t) as in C.1, and set G_t(x) = 0 when x is
dead (not in the current domain H_t). The exponents are the ones
from Part A: alpha = 2-d and p = 4a-1.
Fact A (local martingale). While x is alive,
d G_t(x) = G_t(x) * p * cot( theta_t ) * V_t * dB_t.
(Derived directly from C.1 by Ito's Formula.) So G_t(x) is a
nonnegative local martingale with volatility p * cot(theta) * V
times its own value. It is NOT a true martingale: its mean decays
with t.
Fact B (mean formula). For every finite deterministic t,
E[ G_t(x) ] = G_0(x) * P_x^*( t < T^*(x) ),
where P_x^* is a certain probability on stopped paths (defined
below) and T^*(x) is an a.s. finite lifetime (see "Finite
lifetime" below). Because T^*(x) < infinity a.s., we get
E[G_t(x)] -> 0 as t -> infinity.
Change of law (finite-tilt). For each fixed level ell > 0, put
M_ell(x) = ( G_{tau_ell(x)}(x) / G_0(x) )
* 1{ tau_ell(x) < infinity }.
This is a nonnegative random variable measurable with respect to
the stopped path up to time tau_ell(x). Its expectation is
E[ M_ell(x) ] = 1.
So M_ell(x) is a probability density on the stopped sigma-field,
and we may define a new probability on stopped paths by
P_x^*[ A ] = E[ M_ell(x) * 1(A) ],
for every event A determined by the path up to time tau_ell(x).
We call P_x^* the centered law (or "tilted law") at x.
Because the weights M_ell are consistent as ell increases
(M_{ell_1} = E[M_{ell_2} | F_{tau_{ell_1}}] when ell_1 < ell_2),
we get one probability on stopped paths up to T^*(x) by taking
ell -> infinity.
PROOF THAT E[ M_ell(x) ] = 1.
The weight M_ell is the ratio of the Green weight at the level
time tau_ell(x) to the initial Green weight, times the indicator
that the level was reached. We localize, apply Girsanov, and let
the cutoff go to infinity.
Localization. For integer N >= 1 set
zeta_N = inf { t : theta_t not in (1/N, pi - 1/N),
Y_t < 1/N, or t > N }.
Here theta_t = arg(Z_t(x)) and Y_t = Im(Z_t(x)). The stopped
weight M_ell^{(N)} = 1{tau_ell < zeta_N} G_{zeta_N}(x)/G_0(x) is
bounded and adapted: on {tau_ell < zeta_N} the path has
theta_t in [1/N, pi - 1/N] and Y_t >= 1/N up to zeta_N, so both
factors R_t^{-alpha} and sin^p(theta_t) are bounded.
Girsanov. From the local-martingale identity (proved below using
only Ito and p = 4a - 1, 2 a alpha = p/2),
d G_t(x) = G_t(x) * p cot(theta_t) V_t dB_t,
the density G_{zeta_N}(x)/G_0(x) on F_{zeta_N} has the required
form. Under the localized tilted law P^{*,N}_x, B tilts to a
new Brownian by drift -p cot(theta) V*dt, so the angle SDE
becomes
d theta_s = d W_s + 2a cot(theta_s) ds
in inner time s = s_x(t), after using (1-2a) + p = 2a. The
tilted diffusion is global in inner time s: the Lyapunov
estimate below shows sin(theta_s) stays strictly positive on
every finite s, and dt/ds = Y_s^2 sin^{-4}(theta_s) stays
finite on every finite inner interval.
Lyapunov bound. Fix 0 < q < p and F_q(theta) = sin^{-q}(theta).
Under the tilted generator L^* = 1/2 d^2 + 2a cot(theta) d,
L^* F_q = (q/2) F_q - (q(p-q)/2) cot^2(theta) F_q
<= (q/2) F_q.
So e^{-qs/2} F_q(theta_s) is a nonnegative supermartingale.
For the first exit sigma_delta of theta_s from (delta, pi - delta),
P^{*,N}_x( sigma_delta <= s )
<= e^{qs/2} F_q(theta_0) sin^q(delta) -> 0.
So neither 0 nor pi is hit in finite inner time. The vertical
coordinate obeys d log Y_s = -a sin^{-2}(theta_s) ds, so Y_s
never reaches 0 on finite inner intervals either.
Conclusion. On {tau_ell < zeta_N}, M_ell^{(N)} = M_ell. Since
the tilted law is global, P^{*,N}_x(tau_ell < zeta_N) -> 1 as
N -> infinity (the spatial caps become eventually inactive, and
physical time of the level tau_ell(x) is almost surely finite).
Monotone convergence gives E[M_ell(x)] = 1.
Consistency M_{ell_1} = E[M_{ell_2} | F_{tau_{ell_1}}] for
ell_1 < ell_2 follows by the same argument applied to the
stopped ratio G_{tau_{ell_2}} / G_{tau_{ell_1}}: conditional on
F_{tau_{ell_1}}, the future is a fresh SLE in H_{tau_{ell_1}},
and G is again a local martingale with the same exponent.
Derivation of the local-martingale identity. Ito gives
d(theta_t) = V_t dB_t + (1 - 2a) cot(theta_t) V_t^2 dt.
With d log R_t = -2 a V_t^2 dt (which is pure drift), and
d(sin^p(theta_t))/sin^p(theta_t)
= p cot(theta_t) V_t dB_t
+ V_t^2 [ (1-2a) p cot^2(theta_t)
+ (p(p-1)/2) cot^2(theta_t)
- p/2 ] dt,
the drift of dG_t/G_t is V_t^2 dt times
[ 2 a alpha + (p(1-2a) + p(p-1)/2) cot^2(theta_t) - p/2 ].
The cot^0 term vanishes because 2 a alpha = p/2 (equivalent to
d = 1 + 1/(4a)). The cot^2 term vanishes because
p(1-2a) + p(p-1)/2 = (p/2)(p - 4a + 1) = 0,
which is exactly the definition p = 4a - 1. QED.
Tilted angle dynamics. Under P_x^*, the tip's angle theta
(parametrized by inner time ell) solves the SDE
d theta_s = dW_s + 2 a * cot( theta_s ) * ds,
whose drift has been shifted from the original (1-2a) cot theta
by exactly +p cot theta (recall p = 4a-1). The generator is
L^* = (1/2) d^2/d theta^2 + 2a * cot(theta) * d/d theta.
Stationary law. The tilted diffusion has a stationary density
pi_a( d theta ) = c_a * sin^{4a}( theta ) * d theta,
where c_a > 0 is chosen so pi_a has total mass 1 on (0, pi).
Check by the standard formula for a one-dimensional diffusion:
for dX = dW + b(X) ds, the density satisfies rho'/rho = 2 b;
with b(theta) = 2a cot(theta), we get log rho = 4a log sin(theta),
i.e. rho proportional to sin^{4a}(theta).
Neither endpoint theta = 0 nor theta = pi is reached in finite
inner time under P_x^*. The vertical drift 2a cot(theta) pushes
away from the boundaries, and the diffusion stays in the interior
of (0, pi) forever.
Uniform mixing. There is a deterministic s_a < infinity
(depending only on a) such that, for every s >= s_a and every
starting angle theta_0 in (0, pi),
(1/2) pi_a(.) <= P_x^*[ theta_s in . ] <= 2 pi_a(.).
In particular, any nonnegative bounded angle function phi(theta)
has conditional expectation under P_x^* at inner time s >= s_a
between one half and twice its pi_a-average. This "mixing
between 1/2 and 2" bound is what makes later arguments valid
regardless of the initial angle.
PROOF OF UNIFORM MIXING (spectral gap).
Substitute v = cos(theta). Then d/dtheta = -sin(theta) d/dv =
-(1-v^2)^{1/2} d/dv, and the tilted generator L^* = (1/2)
d^2/dtheta^2 + 2a cot(theta) d/dtheta becomes
A = (1/2)(1 - v^2) dv^2 - (1/2 + 2a) v dv.
Integration by parts against the density
varpi_a(dv) = c' (1 - v^2)^{2a - 1/2} dv
makes A symmetric on polynomials: the boundary term is order
(1 - v^2)^{2a + 1/2}, which vanishes at v = +-1 because
2a + 1/2 > 0. So A is negative semidefinite on L^2(varpi_a).
Let P_m be the orthonormal polynomials for varpi_a with P_0 = 1.
Since A preserves polynomial degree and vanishes on constants,
A P_m = -lambda_m P_m with eigenvalue lambda_m = m(m + 4a)/2
(read off from the leading-degree coefficient).
The orthonormal polynomials grow subexponentially: for m+1
pairwise-separated intervals I_1, ..., I_{m+1} in (-1/2, 1/2) of
length comparable to 1/(m+1), each I_k contains a point v_k with
|P_m(v_k)|^2 <= C_a (m+1) (because varpi_a is bounded below
there and int P_m^2 d varpi_a = 1). Lagrange interpolation at
these points (separated by >= c/(m+1)) gives
|P_m(v)| <= exp(C_a (m+1) log(m+2)) on [-1/2, 1/2],
and the same holds on [-1, 1] with C_a increased.
Consequently the heat-kernel series
k_s(v, w) = 1 + sum_{m >= 1} exp(-s lambda_m) P_m(v) P_m(w)
converges absolutely and uniformly on [-1, 1]^2 for every s > 0.
Ito's formula applied to P_m under the tilted generator A gives,
for the actual transition law,
E_v [ P_m(v_s) ] = exp(-s lambda_m) P_m(v).
(The local martingale A-integral is a true martingale: P_m is
bounded on compact subsets and A P_m P_m is dominated by an
integrable bound.) So the polynomial moments of the true
transition law and of the signed kernel k_s(v, w) varpi_a(dw)
agree. Polynomial density in C([-1, 1]) identifies them, so
k_s is a nonnegative reproducing kernel for the transition.
Since lambda_m -> infinity (in fact lambda_m = m(m + 4a)/2)
and the L^inf growth of P_m is only exp(C_a m log m), the sum
sum_{m >= 1} exp(-s lambda_m) P_m(v) P_m(w) goes to 0 uniformly
in (v, w) as s -> infinity. So there is s_a < infinity with
(1/2) <= k_s(v, w) <= 2 for all v, w and all s >= s_a.
Translating back to theta = arccos(v) and using
(1-v^2)^{2a-1/2} = sin^{4a-1}(theta) up to normalisation gives
(1/2) pi_a(.) <= P_x^*[ theta_s in . ] <= 2 pi_a(.)
for all s >= s_a and all theta_0 in (0, pi). QED.
Inverse-angle integrability. The choice of p is exactly such that:
integral_0^pi sin^{-p}(theta) pi_a(d theta)
= c_a * integral_0^pi sin^{4a-p}(theta) d theta
= c_a * integral_0^pi sin(theta) d theta
= 2 * c_a < infinity.
The middle equality uses 4a - p = 1, which follows directly from
p = 4a - 1. This integrability makes the change-of-law density
well behaved near the extreme angles theta = 0 and theta = pi.
Finite lifetime. Under P_x^*, the stopping time
T^*(x) = lim_{ell -> infinity} tau_ell(x)
is finite almost surely. In other words, x ALWAYS dies in finite
inner time under the centered law. There is no path where x
survives to infinity.
PROOF THAT T^*(x) < INFINITY ALMOST SURELY.
Fix 0 < q < p and 0 < epsilon < a/2. The uniform mixing bound
gives a deterministic B < infinity with
E_x^*[ sin^{-q}(theta_s) ] <= B for all s >= s_a
and every starting angle theta_0 in (0, pi).
Let F_q(theta) = sin^{-q}(theta) and recall d log Y_s =
-a sin^{-2}(theta_s) ds. The Lyapunov estimate from the
finite-tilt proof says M_s = e^{-q s / 2} F_q(theta_s) is a
nonnegative supermartingale under P_x^*.
Fix integer m >= s_a. Doob's maximal inequality on [m, infinity)
with initial value M_m gives
P_x^*( sup_{s >= m} M_s > e^{-q m / 2 + q epsilon m} )
<= E_x^*[M_m] / e^{-q m / 2 + q epsilon m}
<= e^{-q m / 2} B / e^{-q m / 2 + q epsilon m}
<= B e^{-q epsilon m}.
Summable in m. Borel-Cantelli: for all sufficiently large m,
sup_{s >= m} M_s <= e^{-q m / 2 + q epsilon m}.
Taking m = floor(s) for large s gives
F_q(theta_s) <= e^{q epsilon s + O(1)}.
So sin^{-4}(theta_s) <= C e^{4 epsilon s} for all sufficiently
large s.
Combining with Y_s <= Y_0 e^{-a s} (from d log Y_s <= -a ds):
dt/ds = Y_s^2 sin^{-4}(theta_s)
<= Y_0^2 e^{-2 a s} * C e^{4 epsilon s}
= C Y_0^2 e^{-(2 a - 4 epsilon) s}.
Since epsilon < a / 2, this integral over [0, infinity) is
finite, so
T^*(x) = int_0^infty Y_s^2 sin^{-4}(theta_s) ds < infinity.
QED.
What is really going on. The weight G_t(x) = R_t(x)^{-alpha}
sin^p(theta) is the "future mean mass density" at x from the
mass construction (C.4). Using it as a change-of-law density
converts the SLE tip into a "biased" trajectory that is drawn
toward x. The tilt is chosen so that the surviving angular degree
of freedom becomes stationary with the sin^{4a} density and
uniform mixing kicks in after O(1) inner time. All the later
probability bounds are conditioned on this well-behaved angle
process.
A technical remark. Each M_ell is a density only on the
sigma-field of the stopped path up to time tau_ell(x); it is not
a density on the whole infinite future. The "limit as ell ->
infinity" is a projective limit over these finite stopped laws.
This lets us evaluate functionals of the stopped path without
ever committing to absolute continuity on an infinite horizon.
C.3 THE MULTIPOINT ESTIMATE
The multipoint estimate is the workhorse behind the moments m_n in
C.5. It is a uniform bound on the probability that many points in
a domain SIMULTANEOUSLY reach their own prescribed conformal
radii before being swallowed.
Setup. Fix x_1, ..., x_m distinct points inside D = J(HH).
Choose positive constants b_1, ..., b_m (the "caps") such that
b_i <= R_D(x_i) and b_i <= |x_i - x_j|
for every j != i. So the b_i are smaller than the initial conformal
radius at x_i and smaller than the distance to any other marked
point. Fix further target radii 0 < e_i <= b_i.
Event. Let E_m be the event that, for every i in {1,...,m}, the
point x_i reaches conformal radius e_i while still alive (i.e. at
some instant, its current R_t(x_i) is exactly e_i and it has never
been swallowed up to that instant). The points may finish at
different times.
THEOREM (Multipoint estimate). There is a constant C_{m,a} <
infinity depending only on m and a = 2/kappa (in particular
independent of the domain D, the caps b_i, the targets e_i, and
the initial angles) such that
P( E_m ) <= C_{m,a} * prod_{i=1}^m (e_i / b_i)^alpha.
For target radii that may exceed caps, we use the version
P( E_m ) <= C_{m,a} * prod_{i=1}^m min{ 1, (e_i / b_i)^alpha },
which follows by dropping the constraints whose target exceeds its
cap and readjusting the constant for the remaining number of
constraints.
Why the exponent is alpha = 2 - d. The conformal radius e_i is
the natural scale of x_i. A point at radius e_i is (roughly)
e_i^alpha "fractional-dimensional mass" of the curve. So the
hitting probability is a power alpha of the ratio e_i / b_i. The
product over i is what "independence" would predict; the constancy
of C_{m,a} says this is essentially correct even though the
points DO interact.
TWO THINGS TO NOTICE:
(1) The constant C_{m,a} does NOT grow like m! (factorial).
It is finite and depends only on m and a, through explicit
constants (see below). For the upper layers of the argument to
close, we need the moments of the mass to be finite without
huge growth in m. That is what this bound delivers.
(2) The bound is completely UNIFORM in D, in the caps b_i, in
the targets e_i, in the initial angles, and in the ORDER and
TIMES at which the hits happen. This uniformity is crucial
later when we apply the bound inside a disk model whose
boundary angle depends on the drivers.
Proof strategy. Standard "one-point-at-a-time" martingale
arguments fail because the points interact: when two points are
still alive and both active, their angular evolutions are
correlated, and their mutual distance drifts. We handle this
via a PRODUCT SUPERMARTINGALE with a regularized factor for each
point.
Define for each point i a "clock" u_i(t) = integral_{sigma_i}^t V_i^2
ds where sigma_i is the first time R_t(x_i) reaches its cap b_i,
and V_i is V-t of point i. Define U_i = (1/(2a)) log(b_i/e_i), the
amount of clock needed to shrink conformal radius from b_i to e_i.
Then point i finishes at its target at exactly the moment u_i(t) =
U_i, provided it is still alive.
Freeze-and-decay factor. Define
delta_i(u) = exp( - beta * min( u, U_i - u ) ),
f_i(u, theta) = sin^p(theta) + delta_i(u) * ( 1 + sin^q(theta) ),
where the constants q, p, M, beta are all explicit functions of a
alone (derived in detail inside the next paragraph). The role of delta_i is to
be 1 at the boundaries of the activity interval [0, U_i] and small
in the middle. The exponent q < p is chosen so that SOME angular
positivity is available even when sin theta is small. The q and
beta choices land us in a regime where all the sums converge.
Product supermartingale. Consider
Q_i(t) = (1/3) * e^{p u_i(t)/2} * f_i( u_i(t), theta_i(t) )
during activity, and Q_i(t) = 1 before sigma_i (when R_t(x_i) is
still above b_i) and Q_i(t) = (1/3) e^{p U_i / 2} after T_i (when
the point has finished). Set Q(t) = Q_1(t) * ... * Q_m(t).
The claim is: Q is a supermartingale after an explicit
stochastic discount. Precisely, define a positive drift k_t by
k_t = C_1 sum_{i active} delta_i(t) V_i^2
+ C_3 sum_{i 0,
C_2 = (p+q) 3^eta,
r = 1 - 2 eta.
For m > 1 choose epsilon = c_1 / (2(m-1)); for m = 1 there are
no cross terms, so any epsilon is fine. Set
C_3 = C_epsilon = r C_2^{1/r} (eta/epsilon)^{2 eta / r}.
Both C_1 and C_3 depend only on m and a, as claimed. They are
the coefficients appearing in k_t.
Where the drift bound comes from. Recall s = sin(theta_i) and
sigma = delta_i' / delta_i, with |sigma| <= beta <= 1. The exact
drift of f_i is
(d/du + L + p/2) f_i / f_i
= (delta_i / f_i) [ p/2 + ((p-q)(1+q))/2 s^q
- k s^{q-2} + sigma (1 + s^q) ].
Because |sigma(1 + s^q)| <= 1 + s^q <= 2 + s^q (absorbing the
boundary case s ~ 1 into A by an extra +2 there), the bracket
is at most A - k s^{q-2}. Now add the c_1 D_i term with
D_i = (delta_i/f_i) s^{q-2}:
- For s < h_0, we have s^{q-2} > h_0^{q-2} = 2A/k, so the
bracket is at most A - 2A = -A. The added c_1 D_i term is at
least (k/2)(delta_i/f_i) * (2A/k) = A delta_i/f_i, so the
sum is at most 0 <= C_1 delta_i.
- For s >= h_0, use f_i >= s^p >= h_0^p. The bracket is at
most (delta_i/f_i) A <= A h_0^{-p} delta_i = C_1 delta_i.
Both cases yield
(d/du + L + p/2) f_i / f_i <= C_1 delta_i - c_1 D_i.
Where the Young inequality comes from. Angular differentiation
gives |d_theta f_i| / f_i <= (p+q) delta_i^{-1} s^{q-1}. Using
delta_i <= f_i and eta(q-2) = q-1,
|d_theta f_i| / f_i
<= (p+q) 3^eta delta_i^{-1-eta} D_i^eta
= C_2 delta_i^{-1-eta} D_i^eta.
Product formula cross-term. Ito's product formula contributes,
for each active unordered pair (i, j), a term bounded by
C_2 (delta_i delta_j)^{-1-eta}
(D_i V_i^2)^eta (D_j V_j^2)^eta
(V_i V_j)^{1-2eta}.
Weighted Young with exponents (eta, eta, 1-2eta) yields
|cross| <= epsilon (D_i V_i^2 + D_j V_j^2)
+ C_3 (delta_i delta_j)^{-M} V_i V_j.
Choose epsilon = c_1 / (2(m-1)); summing over all pairs, the
first term is absorbed into -c_1 sum_i D_i V_i^2 (each i meets
at most m-1 partners, so the total absorbed is at most c_1/2).
The second term is exactly the C_3 pair contribution in k_t.
With these constants, the discounted product
M_t = Q_t * exp( - integral_0^t k_v dv )
is a nonnegative local supermartingale,
and the discount integral is deterministically bounded:
integral_0^infinity k_t dt <= K_{m,a} < infinity,
with
K_{m,a} = 2m C_1 / beta + C_3 * binom(m,2) * I(a, M beta).
Here I(a, M beta) = e^{2 M beta} / (2a (1 - e^{2 M beta - a})) <
infinity, because 2M beta <= a/2 < a by the choice of beta.
Pair interaction. The key estimate behind the k_t bound is the
following geometric inequality. For two points i, j both
surviving and both active, define
L_{ij}(t) = log | (Z_i(t) - conj Z_j(t)) / (Z_i(t) - Z_j(t)) |,
where Z_i(t) = g_t(z_i) - B_t (as in C.1). A direct computation
from the Loewner equation gives
d L_{ij} / dt = - 2 a V_i V_j.
So L_{ij}(t) is a NONNEGATIVE process (it equals -log of a ratio
in the unit disk) that is strictly decreasing at rate 2a V_i V_j.
Meanwhile, the analogous inequality for the actual geometric
distance gives
L_{ij}(t) <= e^{ - a max( u_i(t), u_j(t) ) } when at least
one of the two points has been active for a while.
That is to say: two points that survive into each other's activity
interval must necessarily be nearly "conformally adjacent". The
distance between them decays exponentially in the leading clock.
Summing over all pairs, this bounds the accumulated cross terms in
k_t. The interaction integral inequality is
integral_{joint activity} e^{ c (u_i + u_j) } V_i V_j dt
<= I(a, c) for 2c < a.
Both the diagonal k_t = C_1 sum delta_i V_i^2 term and the off-
diagonal pair term are absorbed by the negative drift from Q_i's
own dynamics (which is -c_1 * D_i V_i^2, with D_i = (delta_i/f_i)
sin^{q-2} inside the drift bound, plus the negative contribution
of the sin^p term when angles are near the boundary). The
optimizer chooses the constants c_1 and C_3 so that the sum of all
positive terms is <= 0 away from what the discount k_t accounts
for.
Conclusion of the proof. At the stopping time T_all = max_i T_i
(the first time all points have simultaneously finished), the
product Q takes the deterministic value
Q_{T_all} = 3^{-m} * prod_{i=1}^m (b_i / e_i)^alpha.
Since M_t = Q_t e^{-int_0^t k} is a supermartingale with M_0 = 1
(or slightly less, depending on whether sigma_i = 0 for all i),
we have, using a standard localization argument to handle the
bounded cutoff on angles/heights/time and passing to the limit,
E[ M_{T_all} ; T_all < infinity ] <= 1.
Rearranging:
P( T_all < infinity ) <= e^{ K_{m,a} } * 3^m *
prod_{i=1}^m (e_i / b_i)^alpha.
Setting C_{m,a} = 3^m e^{K_{m,a}} proves the proposition.
Note on the choice of stopping-time cutoffs. To make M_t
well-behaved, one stops when the angle of any unfinished point
leaves (1/N, pi - 1/N), when the height Y_i of any such point
falls to 1/N, or at physical time N. Call this cutoff zeta_N. On
the event zeta_N = infinity AND T_all < infinity, all targets
have been reached, and the supermartingale argument applies. P(
T_all < zeta_N ) is bounded by the displayed inequality. As N
grows, the events { T_all < zeta_N } increase up to { T_all <
infinity } (by continuity of the paths, and because a completed
point does not need any further cutoff). So the tail gives the
proposition.
COMPARISON WITH THE LITERATURE. The same proposition also
follows from Rezaei-Zhan (2017, Theorems 1.1 and 1.2) together
with the Koebe quarter theorem (Section 3.2 of Beliaev 2015,
specifically Theorem 3.2.5 and Corollary 3.2.6). The direct
proof above is self-contained and is the one used throughout.
COROLLARY (Integrated level-moment bound, "level moments").
For every nonnegative bounded f supported in a compact K that is
compactly contained in D, and for every integer m >= 1, define
A_ell( f ) = integral_K f(x) * G_{tau_ell(x)}(x) *
1{ tau_ell(x) < infinity } dx dy.
Then
sup_{ell > 0} E[ A_ell(f)^m ] < infinity.
Moreover, if x_0 in D is an additional fixed point with target
radius 0 < e_0 <= R_D(x_0), and E_0 is the event that x_0 reaches
conformal radius e_0 while alive, then
sup_{ell > 0} E[ 1_{E_0} * A_ell(f)^m ] <= C * e_0^alpha,
with C depending only on m, a, the size and location of f, and a
positive lower bound for R_D on K union {x_0}.
Proof. Apply the multipoint estimate of C.3 (Proposition
mp:estimate, with the explicit C_{m,a} derived there) to the
m points where A_ell(f) has mass, together with the cap-and-
target choices e_i = R_D(x_i) e^{-2a ell} <= b_i and
b_i <= min_{j != i} |x_i - x_j|. Integrating the collision
kernel |x_i - x_j|^{-2 alpha} over the support K^m gives the
m-th moment bound (this uses only 2 alpha < 2, the same fact
that gave atomlessness of mu_infinity in C.4). For the
center-moment estimate with E_0, choose x_0 as an additional
marked point with cap b_0 <= R_D(x_0) and target e_0, and note
that the C.3 proposition contains (e_0/b_0)^alpha <= 1 as a
multiplicative factor. QED.
C.4 A CANONICAL MASS MEASURE ON THE TRACE
Goal. Build an adapted increasing measure mu_t on the open upper
half-plane HH (or on a conformal image D = J(HH)) with the
following meaning: mu_t(A) is "how much of the curve up to time t
has passed through A". It starts empty at t = 0 and grows on the
trace.
THEOREM (Construction and conditional kernel). There is an
adapted increasing family ( mu_t ) for 0 <= t <= infinity of
locally finite measures on HH, with the properties below.
LEVEL APPROXIMANTS. Definition first. For an integer j >= 1
define
M_t^{(j)}( dx dy ) = G_{tau_j(x)}(x)
* 1{ tau_j(x) <= t and tau_j(x) < inf }
dx dy,
where tau_j(x) is the level time from C.1 (the first inner time s
at which x's conformal radius reaches R_0(x) * e^{-2a j}), and
G_s(x) = R_s(x)^{-alpha} sin^p(theta_s) is the Green weight from
C.2. So M^{(j)} is supported on the set of points whose conformal
radius around them is exactly R_0(x) e^{-2a j}, weighted by the
Green function.
CONVEX EXTRACTION (full argument).
(i) Polish space. The space M_{lf}(HH) of locally finite positive
Radon measures with the vague topology is Polish. Its Borel sigma-
field is generated by evaluations phi |-> mu(phi) for phi in
C_c(HH). Vague convergence sets and limit maps are therefore
Borel in the product space of sequences.
(ii) Countable determining family. Fix a countable dense-in-C_c
family { phi_k } (e.g. rational-coefficient bumps on a rational
ball basis), countably many rational times q, and a compact
exhaustion { K_m }. A locally finite positive measure is
determined by its values on { phi_k } alone. So agreement on
all (phi_k, q) pairs implies the measure-valued equality mu_q =
mu_q' everywhere.
(iii) Tightness and compactness. By the integrated level-moment
bound (Corollary mp:level-moments) applied at each fixed level j
on a compact K cc HH, the test-masses M_q^{(j)}(phi_k) have
bounded first moments, uniformly in j, and are therefore tight
in the sigma(M_{lf}, C_c)-topology along any subsequence of j.
The double sequence of the test-masses (indexed by j on the one
hand and (phi_k, q) on the other) is bounded in L^1 and admits a
diagonal subsequence j_l along which
M_q^{(j_l)}(phi_k) converges for every (Phi_k, q).
Standard Cantor-diagonal over the countable family.
(iv) Cesaro smoothing. The subsequence limit may depend on
measurable null-set choices. Replace with a tail Cesaro average
over a sparsely chosen sequence j_L -> infinity (each j_{L+1}
chosen so large that for every (phi_k, q) with k, q <= L, the
increment |M_q^{(j_{L+1})}(phi_k) - M_q^{(j_L)}(phi_k)| <= 2^{-L}).
Then
mu_t := weak-lim_{L -> infinity} ( 1/L ) sum_{l <= L}
M_t^{(j_l)}
is a well-defined positive measure-valued functional, measurable
in the driving path, vaguely continuous at rational times by
construction, and measurable at all times by right-continuity
of t |-> mu_t(phi_k) (each M_t^{(j)} is right-continuous).
In particular the convex-combination choice is deterministic
in kappa alone: it depends on moment bounds, not on the sample.
(v) Consistency and kernel. Because the SAME sequence (j_l) is
used for ALL (phi_k, q), the resulting mu_q is one fixed object.
Apply M5 as an identity of conditional kernels to each M_q^{(j_l)}
against the finite determining family, average over l, and pass
to the L^1 limit: the kernel identity M5 holds for mu. Then
Monotone Class lifts from countable tests to all nonnegative
F_t x B(HH)-measurable tests.
(vi) Uniqueness removes the choice. If a second sequence of
convex weights gave a different mu~_t, both mu and mu~ would be
continuous increasing adapted finite-variation test-mass processes
starting at zero with the same remaining conditional potential
G_S dx dy (from M5). By the UNIQUENESS paragraph below,
mu = mu~. So the specific convex coefficients do not matter.
QED.
Now the properties.
M1. mu_0 = 0, and mu_t -> mu_infinity weakly as t -> infinity.
M2. Almost surely, simultaneously for every 0 <= v < t < inf,
the increment ( mu_t - mu_v ) is carried by
gamma([v,t]) cap HH.
M3. The terminal measure mu_infinity is nonatomic and has
moments of every integer order on compact sets. (The
exponent alpha < 1 is what makes atomlessness true; see
F.4.)
M4. The map t |-> mu_t is locally continuous in total
variation: for any compact K and any time t, the total
variation of mu_{t+delta} - mu_t restricted to K tends
to 0 as delta -> 0.
M5. Almost surely, for every stopping time S (possibly
infinite) and every nonnegative test function f that is
measurable with respect to F_S,
E[ mu_infinity( f ) - mu_S( f ) | F_S ]
= integral f(x) G_S(x) dx dy.
Here G_S(x) = R_S(x)^{-alpha} sin^p(theta_S) on the event
{ x in H_S } and G_S(x) = 0 outside H_S. The identity
holds as a conditional measure kernel, so it includes
random nonnegative tests known at S.
M6. Taking S = 0 recovers the unconditional mean
E[ mu_infinity ] = G_0(x) dx dy.
M7. Almost surely, for every 0 <= v < t < infinity,
( mu_t - mu_v )( HH ) > 0.
The positive mass can always be detected on an interior
compact.
M8. First-visit identity. Put v(x) = inf{ t : gamma(t) = x }. Almost
surely v(x) < infinity for mu_infinity-almost every x,
and simultaneously for every finite t,
mu_t( dx dy ) = 1{ v(x) <= t } * mu_infinity( dx dy ).
In words: a later geometric return to an already visited
point brings no new mass. The measure is a "first visiter
counter".
M9. Conformal covariance. For a fixed conformal map J : HH -> D
define the pushed-forward measure
mu_t^D( dx dy ) = J_* ( |J'|^d mu_t )
= the pushforward of the measure
|J'(z)|^d mu_t(dz)
under z |-> J(z).
Then the kernel M5 holds with G_t^D in place of G_t,
the support property M2 holds with J(gamma([v,t])) in
place of gamma([v,t]), and everything else carries over.
Reason: conformal radius transforms as R_D(J(z)) =
|J'(z)| * R_HH(z), area transforms as |J'|^2, and the
combination "radius^{-alpha} times area" becomes
|J'|^d times the original.
M10. Pathwise restart. Fix a stopping time S < infinity. Put
J_S = (g_S - B_S)^{-1} : HH -> H_S. Let mu~ be the
canonical mass of a fresh Brownian driver. Then almost
surely, as measures on H_S, simultaneously for all
0 <= t <= infinity,
( mu_{S+t} - mu_S )|_H_S
= ( J_S )_* ( |J_S'|^d mu~_t ).
Equivalently: after stopping at S, the future of mu is
exactly the canonical mass of a fresh SLE, carried through
the (random) conformal map J_S. Once this measure equality
holds, any random nonnegative Borel weight on H_S can be
integrated without picking an additional version.
PROOF OF M10 (pathwise restart).
(a) Terminal integrability of both sides. Write mu_t^{(S)} :=
mu_{S+t} - mu_S restricted to H_S, and nu_t := (J_S)_*(|J_S'|^d
mu~_t) for the pushforward of the fresh-SLE mass. Take phi in
C_c(H_S). By strong Markov for the SLE driver, conditioning on
F_S gives E[mu_infinity^{(S)}(phi) | F_S] = int phi(x)
G_S^{restart}(x) dx dy where G_S^{restart} is the restarted
Green weight at the current time S (with angle theta_S).
Because E[G_S^{restart}|F_0] <= G_0 under the change of
variables J_S, this has finite expectation <= int phi(x) G_0(x)
dx dy after integrating. The same computation gives the
terminal mean of nu_∞. The two conditional means agree.
(b) Kernel at S + t. Fix t >= 0. The kernel identity M5 gives,
for the actual process, on F_{S+t}:
E[mu_infinity^{(S)}(dx) - mu_t^{(S)}(dx) | F_{S+t}]
= G_{S+t}^{actual}(x) dx dy.
For the fresh process (J_S)_*(|J_S'|^d mu~_t) under strong
Markov + op:tested-potential, the same identity applied in
the restarted domain H_S at inner time t yields
E[nu_infinity(phi) - nu_t(phi) | F_{S+t}^{fresh}]
= int phi(x) G_t^{fresh}(x) dx dy
with G_t^{fresh}(x) dx dy the fresh-SLE kernel at level t.
Because J_S maps concrete points but conformal covariance of
G is exact (|J'|^{-alpha} x |J'|^2 = |J'|^d), the two
conditional potentials are the same positive form once both
are expressed in fresh coordinates.
(c) Uniqueness. Now mu^{S + ·} - mu_S and nu_· are two
continuous increasing adapted R-valued processes starting from
zero on H_S, of locally finite total variation, and with the
same remaining conditional potential at every fixed t. Their
difference is a continuous martingale of finite variation and
starts at zero, hence vanishes (see UNIQUENESS below). This
identifies the two measure-valued processes as one, on a common
probability-one event.
(d) Simultaneity. Choose the countable determining family of
tests and rational times. The comparison in (c) is made once
per (test, rational t) on a null event; the union over a
countable family is a null event. Intersect with the null
event of M5 for mu and the null event of M5 for the fresh
mu~. On the remaining probability-one event M10 holds
simultaneously for all t >= 0 and all test functions phi in
the countable family, hence (by continuity in t and weak
convergence on tests) for all phi in C_c(H_S) and all real
t >= 0. QED.
UNIQUENESS. If A_t and A~_t are two continuous increasing
adapted finite-variation test-mass processes starting at zero
with the same remaining conditional potential G_S dx dy (a
conditional-kernel identity), then A_t - A~_t is a continuous
martingale of finite variation and starts at 0, so it vanishes.
Doing this for a countable determining family of compact tests
proves uniqueness of mu. In particular the extracted mu is
independent of the particular convex weights used.
SCALING. Let B^{(c)}_t = c^{-1} B_{c^2 t} be the SLE driver of the
scaled curve. Then almost surely, for all t and all Borel sets A,
mu_t( B^{(c)} ) ( A ) = c^{-d} * mu_{c^2 t}( B )( c A ).
This is a direct consequence of uniqueness (the scaled process has
the same kernel, continuity provenance and terminal integrability
as the canonical process).
HOW TO READ THE PROPERTIES.
- M5 is the crucial "the future looks like G_S times area"
identity. It is why mu can be compared to the canonical
Green potential and why a change-of-law argument can be
used later.
- M7 is why mu detects mass on every nontrivial interval. A
nontrivial interval contains a nontrivial rational interval,
and start-of-interval brownian germ sigma-field plus restart
gives positivity.
- M8 is the "past mass" property used in the lower Hausdorff
bound. Mass at a point is only counted the first time the
curve visits it. Subsequent geometric returns do not add
mass.
- M9 and M10 are what allow us to "start afresh" at a stopping
time and compare to a fixed disk model. This is the key point
where the proof passes from the actual curve to an
arbitrary-angle reference model.
What remains. C.5 defines the centered disk model and the mass
moments L, L_ell, m_n, m_n(ell) used to build the gauge.
C.5 THE DISK MODEL AND ITS MASS MOMENTS
We now freeze everything into a reference model where one marked
point has been placed at the origin of the unit disk. All of the
infinite-dimensional mass mu from C.4 is compressed into a single
scalar quantity: the accumulated mass in a fixed quarter-radius
subdisk.
Conformal identification. For a fixed angle theta in (0, pi),
define the disk map
J_theta(zeta) = ( zeta - e^{i theta} ) / ( zeta - e^{-i theta} ),
which is a conformal (Mobius) map from the upper half-plane HH onto
the unit disk DD, taking the boundary point e^{i theta} to the
marked center 0. In this view the marked boundary point of the
drivers becomes the interior center of the disk.
The chosen test region inside the disk is
K = { x in DD : |x| <= 1/4 },
a fixed closed disk of radius 1/4 around 0.
Centered laws at the disk model. For a fixed theta, let
P_theta^* be the centered law (from C.2) which biases the curve
toward theta's image in the disk. Stack these up according to the
invariant angle law pi_a:
P_mix^*( d theta ) = integral_{0}^{pi} P_theta^* pi_a( d theta ).
In words: pick the driver angle at random from pi_a, then run the
centered SLE with that start. P_mix^* is a probability on stopped
paths, and expectation under it is denoted E_mix^*.
The centered mass functional. Let mu^DD_{tau_ell(0)} be the
disk-image mass (property M9 from C.4) evaluated at the level
time tau_ell(0) of the marked center 0. Let
L_ell = mu^DD_{tau_ell(0)}( K ),
the accumulated mass inside the fixed subdisk K by inner time ell.
Let
L = lim_{ell -> infinity} L_ell
be the eventual limit; the limit exists almost surely because
mu is increasing.
Note: this is one scalar per path, and it depends only on the total
mass pushed into a fixed quarter-disk by inner time ell. All the
infinite-dimensional structure has been summed away.
THEOREM (Finite positive disk moments). For every integer n >= 1,
the deterministic numbers
m_n( ell ) = E_mix^*[ L_ell^n ],
m_n = E_mix^*[ L^n ]
= lim_{ell -> inf} m_n( ell )
satisfy
0 < m_n < infinity.
Moreover L > 0 almost surely, and for each fixed starting angle
theta in (0, pi), the centered law satisfies the uniform bound
sup_{ell > 0} E_theta^*[ L_ell^n ]
<= C_{n,a} * sin^{-p}(theta).
The important content:
(i) m_n is positive AND finite for EVERY integer n. The whole
gauge will need all n simultaneously, so the "positive every
order" is not a decoration.
(ii) The bound sin^{-p}( theta ) is integrable against pi_a,
because int sin^{-p} sin^{4a} = int sin = 2 by the identity
4a - p = 1 from C.2. So the exceptional values theta near 0
and theta near pi are covered by the angle law pi_a.
Proof ideas (full version in F.5):
* The moment bound at a fixed theta is essentially a two-step
application of the multipoint estimate (C.3) at the marked
point 0 and at one more interior point.
* Positivity L > 0 uses the mass positivity M7 of C.4 combined
with the fact that finitely much mass has been pushed into K
by inner time ell. As ell grows, the mass keeps increasing.
* The "sin^{-p}(theta)" prefactor comes from the density of the
finite-tilt change of law at inner time ell evaluated at the
moment the level-time disk closes.
EXPLICIT DETERMINISTIC FORMULA (the coefficient prescription).
The moments m_n and m_n(ell) are deterministic numbers computable
from kappa alone. Here is the explicit integral definition of
these numbers.
For a fixed theta, define the following deterministic integrands
on the path space of a standard Wiener process b (Brownian motion
started at 0):
P_ell( theta, b ) =
1{ tau_ell(0; theta, b) < infinity }
* e^{ 2 a alpha ell }
* sin^p( theta_{tau_ell(0)}(0) ) / sin^p(theta).
Here tau_ell(0) = tau_ell(0; theta, b) depends on the Loewner
driver b, and theta_{tau_ell(0)}(0) is the angle of the marked
center just before it dies.
For the discrete compensator, choose dyadic mesh sizes
delta_h = 2^{-h}, t_i = i * delta_h, N_h = h * 2^h.
For a path b and an independent Wiener path c, let b (+)_t c be
the concatenation: agrees with b up to t, then continues as
b(t) + c( . - t ) after t. Define
q_{h,i}( theta, b )
= Psi_{t_i}( theta, b )
- integral Psi_{t_{i+1}}( theta, b (+)_{t_i} c )
W( d c ),
where Psi_t( theta, b ) = integral_K G_t( x; theta, b ) dx dy is the
Green weight mass inside K.
Then the discrete compensator is
D_h( theta, b; ell )
= sum_{ i = 0 }^{ N_h - 1 }
1{ t_{i+1} <= tau_ell(0; theta, b) }
q_{h,i}( theta, b ).
THEOREM (Explicit coefficients). The moments from the previous
proposition are given by
m_n( ell )
= lim_{u -> infinity} lim_{h -> infinity}
integral_{0}^{pi}
integral
P_ell( theta, b )
( [ D_h( theta, b; ell ) ]_0^u )^n
W( d b )
pi_a( d theta ),
m_n
= lim_{ell -> infinity} m_n( ell ).
Here [ x ]_0^u = min{ max{ x, 0 }, u } is the interval truncation.
All limits can be taken along integer subsequences, so the
coefficients are purely deterministic numbers determined by kappa.
WHY THIS IS THE RIGHT KIND OF FORMULA. The inner integral over b
is an average over the Loewner driver, sampled from Wiener measure.
The outer integral over theta averages the driver start angle by
pi_a. The quantities P_ell and D_h are the "change of law" weight
and its discrete-compensator approximation to the actual mass,
respectively. The double limit recovers m_n(ell) first in the
mesh h -> infinity, then truncated to mass level u -> infinity.
FURTHER COMMENTS.
(a) Sampling construction. The Wiener integral above can be
realized as a countable product of independent standard
Gaussians: choose each dyadic increment of b as an independent
normal variable, and connect midpoints by the standard bridge
construction. The polygonal interpolants converge uniformly
on compacts with probability one. So there is no
"continuous-time randomness" needed: a countable product of
Gaussians suffices.
(b) Loewner quantities. For fixed theta the Loewner maps g^{(j)}
can be built by explicit Euler iteration with a truncated
vector field
F_j( t, z ) = a * conj( z - b(t) ) /
max( |z - b(t)|^2, j^{-4} ).
At every finite time the truncated solution is identical to
the true Loewner solution, because the truncation only kicks
in when the point is not yet the tip.
(c) Clock and radius. In the disk model the level time is
tau_ell(0) = inf{ q rational positive : q alive and
s_0(q) > ell }.
On this event the clock amounts to exactly ell and the
conformal radius around 0 is R_0(0) e^{-2a ell}. Using
rationals makes tau_ell(0) both deterministic-measurable
and available path-by-path.
(d) No further normalization. Because the coefficients are
defined by these explicit integrals, m_n carries no hidden
choice of convex extraction or mass normalization: everything
has been fixed once and for all.
Compressed statement. The moment sequence m_1, m_2, m_3, ... is
a deterministic sequence of positive real numbers, computable
from kappa alone, with the property that m_n(ell) -> m_n as ell
-> infinity for every n. This is the input to Part D.
-----------------------------------------------------------------------------
PART D -- TURNING MOMENTS INTO A GAUGE
-----------------------------------------------------------------------------
D.1 WHY MOMENT BATCHES WORK
We have all we need from Part C: a deterministic moment sequence m_1,
m_2, ..., m_n positive finite, with m_n(ell) -> m_n as ell -> inf.
And we have mu_t on every sample (up to null event) with the properties
M1-M10.
The question of Part D is:
How do we convert the moments into a Hausdorff gauge h,
and why should that gauge work?
Two opposing forces.
On one hand, to prove the trace has POSITIVE Hausdorff measure with
h, we need not-too-many bad squares. Informally, a small square of
side r that intersects the trace should carry mass roughly proportional
to h(r). If we measure too small a ruler (h too small), too many
squares look "overfull" and no cover exists.
On the other hand, to prove the trace has FINITE Hausdorff measure
with h, we need to actually CONSTRUCT a small cover of the trace at
each scale. If we measure too large a ruler (h too large), the cost
of any cover blows up.
The sweet spot. The critical power d is both too small (measure
zero) and too big (measure too large) at once, in the exponent sense.
The trick is to build h(r) from a mixture of scale-r data at
DIFFERENT test scales. Each test scale is a dyadic threshold
A_k = 2^k that reads "is the mass in a fixed subdisk at least A_k".
These events are more and more unlikely as A_k grows, with a
power-law tail controlled by the moments m_n and m_{n+1}. Repeat
each dyadic threshold MANY TIMES (multiplicity N_{nk}) and the sum
of "cost per hit" over all repetitions becomes a deterministic
power of r that the actual trace can afford.
The refinement scale. For each order n we also choose a level ell_n
that captures at least 3/4 of the total moment m_n:
ell_n = 1 + sum_{j=1}^{infinity} 1{ m_n(j) <= 3 m_n / 4 }.
Since m_n(j) is increasing to m_n, this is finite: it stops the
sum the moment m_n(j) exceeds 3 m_n / 4. The summands count how
many integer levels are "too small".
The batch threshold. Also define
T_n = 4 m_{n+1} / m_n,
a deterministic ratio of successive moments. T_n is the mass level
at which "large enough to be useful but small enough to control
by the next moment" begins.
The list of tests. For fixed n, form a finite list. In increasing
order of k in Z, include A_k = 2^k with multiplicity
N_{nk} = 1{ 2^k <= T_n } * floor( 4^n 2^{kn} / m_n ).
Then append the number n exactly ONCE as the "cleanup entry". The
resulting entries are
b_{n1}, b_{n2}, ..., b_{nI_n},
where I_n = I_n' + 1 is the total number of entries in batch n
(I_n' being the sum of the multiplicities N_{nk}).
Finiteness and lower bound. One can check that:
- The list is finite, because N_{nk} > 0 requires
m_n^{1/n} / 4 <= A_k <= T_n.
So the "active" dyadic levels A_k lie in a fixed finite
interval. When A_k is outside the interval, N_{nk} = 0.
- Every entry satisfies
b_{ni} >= c_0 := min{ 1, m_1 / 4 } > 0.
(In particular, the batch thresholds are bounded below by
a positive number depending only on kappa.)
So the entries are finite in number, they lie in [c_0, infinity),
and they are explicit functions of (n, kappa, m_n, m_{n+1}).
The threshold interpretation. Why does the multiplicity look like
4^n 2^{kn} / m_n? Roughly:
- The expectation of L_ell^n at ell = ell_n is at least 3/4 m_n.
So there are lots of paths with L^n near m_n.
- Dyadic mass level A_k = 2^k is being tested many times.
The relative "cost" of getting above A_k is charged as
(A_k)^n / m_n, which measures how "rare" that mass level
is with respect to order n.
- Repeating each level with multiplicity proportional to
(A_k)^n / m_n makes the total expected "rewards" from
successful tests roughly 2^n (many chances), while the
total expected "losses" from misses is bounded by
(1 / T_n)^alpha times the (n+1)st moment -- i.e. summable.
Two-sided probability bounds. With the multiplicities above, one
proves:
UPPER TAIL. For every n >= 1,
sum_{k in Z} N_{nk} * P_mix^*{ L >= 16 A_k }
<= 4^{-n} / ( 1 - 2^{-n} ).
So the total "extreme mass" cost across all dyadic levels is
small and summable in n.
LOWER SUCCESS. For n >= 3,
sum_{k in Z} N_{nk} * P_mix^*{ L_{ell_n} >= A_k }
>= 2^{n-2} - 1 / (2 (1 - 2^{-n}))
>= 2^n / 8.
So the test battery supplies about 2^n independent chances
of success, at the level ell_n that captures 3/4 of the
moment.
Together these two bounds are exactly what the lower and upper
covering arguments need. The upper tail makes the probability that
the curve has "overfull" squares summable. The lower success makes
the probability that a given scale has no "successful" ball
vanishingly small.
Cleanup. On top of the tests, one cleanup entry per batch of
threshold b = n: this pays for any cell that the tests miss, in
both cost and probability. Summability of the cleanup over n is
direct: sum_{n >= 1} P_mix^*{ L >= 16 n } <= E_mix^* [ floor(L/16) ]
<= m_1 / 16 < infinity.
D.2 THE FINITE LISTS AND THE GAUGE FORMULA
D.1 defined the thresholds b_{ni} per batch. Now assign to each
entry its radius r_{ni}, and take the infimum of the corresponding
"quadratic envelopes" to define the gauge itself.
Radii. For batch n and slot i = 1, ..., I_n:
r_{ni} = exp{ -n^2 - 2 a (i-1) (ell_n + n) }.
So radii decrease in i within a batch: r_{n1} is the largest,
r_{nI_n} the smallest. Across batches they interleave: r_{n+1,1}
is (roughly) e^{-(n+1)^2} which is < e^{-n^2} = r_{n1}, but the
last radii of batch n may be much smaller than r_{n+1,1}.
Everything is interleaved and there is no ordering between
batches.
Gap structure inside a batch. The factor 2a(ell_n+n) between
consecutive radii is deliberate: it makes each successive test
occur after a fixed inner-time gap of (ell_n + n) relative to
its predecessor, which is more than enough for the angle mixing
in C.2 to kick in. And the n^2 leading factor makes the radius
of batch n+1 strictly smaller than the starting radius of batch
n, so the "cleanup" of the previous batch (which pays for points
missed by its tests) is comfortably absorbed by the next batch's
testing at a much finer scale.
THE GAUGE ITSELF. Define
H_{ni}(r) = b_{ni} * r_{ni}^d * max{ 1, (r / r_{ni})^2 },
for r > 0, and set
h_kappa(r)
= inf over all n >= 1 and 1 <= i <= I_n
of H_{ni}(r)
for r > 0, and h_kappa(0) = 0.
So the gauge at radius r is the minimum over every entry (n,i) of
"cost of the entry" times "the quadratic envelope at r". The
quadratic envelope switches from linear-in-r (when r >= r_{ni})
to quadratic-in-r (when r <= r_{ni}). The switch is the fact that
below the entry's own radius, we're counting sub-squares, and a
planar cover requires a factor (r / r_{ni})^2 to convert from "cost
at radius r_{ni}" to "cost at radius r".
Two key facts.
FACT 1 (valid gauge). h_kappa is a Hausdorff gauge:
- h_kappa is continuous on [0, infinity);
- h_kappa is nondecreasing;
- h_kappa(0) = 0;
- h_kappa(r) > 0 for every r > 0;
- the quadratic envelope satisfies the doubling property
h_kappa(C r) <= C^2 h_kappa(r) for every C >= 1 and r >= 0.
In particular, h_kappa is one of the "legitimate" gauges described
in Part A.
FACT 2 (finite-exclusion lemma). Let H_{ni}(r) be as above.
Suppose F: [0, infinity) -> [0, infinity) satisfies
(i) F(r) <= M < infinity for all r;
(ii) F(r) <= C * H_{ni}(r) for every n >= n_0, 1 <= i <= I_n,
and every r > 0 (with C a fixed finite constant).
Then for some finite C' depending only on n_0, C, and M,
F(r) <= C' * h_kappa(r) for every r > 0.
In other words, "forgetting about the finitely many initial
batches" loses only a bounded factor. This is the finite-exclusion
statement used later to pass from "the density bound holds for
large batches" to "the density bound holds for all batches, with a
bigger constant".
WHY THESE FORMULAS WORK.
(a) Positivity at radius r.
- The full infimum is over infinitely many entries. For
a fixed r > 0, entries with r_{ni} < r have
H_{ni}(r) >= c_0 r^2 r_{ni}^{d-2} -> infinity
as r_{ni} -> 0.
So the infimum is attained on the finitely many entries
with r_{ni} >= r, hence is positive and finite.
- This uses d < 2 (i.e. d - 2 = -alpha < 0), which is
exactly the condition kappa < 8.
(b) Continuity at zero.
- At the cleanup radius rho_n = r_{nI_n}, the formula
gives h(rho_n) <= n rho_n^d.
- rho_n <= e^{-dn^2} -> 0.
- So h(rho_n) -> 0, and monotonicity + continuity on (0,inf)
give h(0) = 0.
(c) Doubling.
- Each H_{ni} satisfies H_{ni}(Cr) <= C^2 H_{ni}(r).
- The infimum of a family of functions each satisfying the
property also satisfies it, because C is applied pointwise.
(d) The envelope matches Peres-Solomyak.
- For a step profile phi(s) = inf_{r_{ni} >= s} b_{ni} r_{ni}^d,
the identity
h_kappa(s) = inf_{0 < r <= s} (s/r)^2 phi(r)
is the quadratic regularization used by Peres-Solomyak
(2005, Lemma 1.2 and Section 5) in covering arguments for
projections of fractals.
- The batch data and the lower bound c_0 > 0 are what make
this profile nondegenerate.
(e) The finite-exclusion lemma.
- For a fixed r > 0, entries with r_{ni} >= r are finite
in number (because r_{n1} <= e^{-n^2} -> 0 and each batch
is finite).
- Entries with r_{ni} < r become arbitrarily large as r
shrinks (from positivity (a)), so they can be discarded
in a lower bound.
- For the finitely many "small-batch" entries (n < n_0)
and for all r, the pointwise bound
H_{ni}(r) >= b_{ni} r_{ni}^d > 0
gives a fixed positive lower bound. So F(r) <= M
implies F(r) <= (M / b_{ni} r_{ni}^d) * H_{ni}(r) for
each such entry.
- Taking C' = max{ C, max_{n 0. Part D gave us h_kappa. Part E.1
proves that this number is POSITIVE almost surely, for every
positive-time subinterval.
The strategy.
(S1) Do a SPATIAL COMPARISON at a stopped restart. At every
conformal-radius-restart point z, mass in a square of side
r is comparable to the disk-model mass L_ell in a fixed
quarter-disk, up to explicit constants.
(S2) Do a PAST-MASS identity. If a cell in the grid is overfull
(mu_inf(Q) >> 2D b r^d), then at least half of its terminal
mass arrives AFTER the first instant at which the
accumulated mass in Q exceeds D b r^d. This past-mass
"pays off" the overfull cell by mass that is already on
the trace.
(S3) Sum the resulting expected loss over all bad cells and all
entries. Summability of the upper tail from D.1 makes these
losses finite. So almost surely, only finitely many bad
cells survive at any point.
(S4) On the surviving "good set" S_{n_0} (after a batch cutoff
n_0), derive a density bound: any set U of diameter s meets
at most 16 max{1,(s/r_{ni})^2} grid cells of side r_{ni},
so its outer measure is bounded by 32D b_{ni} r_{ni}^d
times that count.
(S5) Apply the finite-exclusion lemma of D.2 to pass from
"the density bound holds for late batches" to "the density
bound holds for ALL batches", with a bigger constant C_nu.
(S6) Conclude: for any countable cover {U_j} of gamma([u,v]),
the mass nu(HH) is bounded by C_nu * sum_j h(diam U_j),
so the outer measure of gamma([u,v]) is positive with
respect to h.
Let's unpack the constants and each step.
SPATIAL COMPARISON AT A STOPPED RESTART.
Fix z in HH and r > 0 with 100r < R_0(z) = 2 Im(z). Let S be the
first time R_S(z) = 100r while z alive. On {S < infinity}:
(a) There is a conformal disk map F: DD -> H_S with F(0) = z,
normalized so |F'(0)| = 100r. This is the standard "restart
at conformal radius 100r" disk identification. The disk
has exactly the markings of the disk model C.5.
(b) The Koebe growth and distortion estimates give the explicit
bounds:
overline B(z, 16r) subset F(K) subset
overline B(z, 400r/9) subset
overline B(z, C_* r),
where K = {|w| <= 1/4}, and
c_* r <= |F'(w)| <= C_*' r for w in K,
with the explicit constants
C_* = 200, c_* = 192/5, C_*' = 8000/27.
(c) Because F(K) is compactly contained in H_S, its accumulated
mass at time S is zero. Proposition M10 (restart) identifies
the actual subsequent measure on F(K), simultaneously at
ALL later times, with the pushforward of the disk measure
weighted by |F'|^d.
(d) Hence if T denotes the time at which the inner clock reaches
one more level ell after S, then
(c_* r)^d * L_ell <= (mu_T - mu_S)(F(K))
<= (C_*' r)^d * L_ell.
Here L_ell is the disk-model mass from C.5 at inner time ell
in this restarted disk. This is the "restarted mass is
equivalent to disk mass" statement.
PAST-MASS IDENTITY.
Fix a deterministic closed square Q compactly contained in HH
and a mass threshold H > 0. Let M_q = mu_q(Q) be the accumulated
mass in Q by time q. Let v(z) be the first visit time of z by
the trace.
The identity says:
E[ integral_Q
1{ there exists rational q >= 0 with q < v(z) and
M_q > H }
mu_infinity( dz ) ]
= integral_Q G_0(z) *
P_z^*[ there exists rational q >= 0 with
q < T^*(z) and M_q > H ] dA(z).
The strict inequality q < v(z) (or q < T^*(z)) is part of the
statement. Without it, mass arriving exactly at the terminal
time would be mishandled.
Intuition. The left side is "how much terminal mass sits on
cells that were already overfull before z was first visited by
the past trace". The right side says this equals the Green
potential G_0(z) times the probability, under the CENTERED law
at z, that mass in Q exceeds H before z's lifetime runs out.
This is a "look backwards from z, but only count mass that
arrived before z was first visited" identity.
PROOF.
Step 1 (finite rational times). Fix a finite, increasing list
of rational times 0 <= q_1 < ... < q_N. Define the disjoint
events
A_j = { M_{q_j} > H and M_{q_k} <= H for all k < j },
the "q_j is the first successful time so far" event. Each
A_j belongs to F_{q_j} and depends only on the driving path,
not on the integration point z.
Step 2 (first-visit identity). By the first-visit identity
mu_t(dz) = 1{v(z) <= t} mu_infinity(dz), the remaining mass at
time q_j is
(mu_infinity - mu_{q_j})(dz)
= 1{q_j < v(z)} mu_infinity(dz).
The deterministic-time a.s. equality uses nonatomicity of
mu_infinity at the singleton {v(z) = q_j}, which has zero
terminal mass (property M3 of C.4).
Step 3 (first success reduces the integral). For fixed z, the
event {exists j <= N with q_j < v(z) and M_{q_j} > H} is
equivalent to {exists j <= N with A_j and q_j < v(z)}, and
since A_j are disjoint, the indicator is
sum_{j=1}^N 1{A_j} 1{q_j < v(z)}.
Combining Steps 2 and 3,
int_Q 1{exists j <= N, q_j < v(z), M_{q_j} > H}
mu_infinity(dz)
= sum_{j=1}^N int_Q 1{A_j} (mu_infinity - mu_{q_j})(dz).
Step 4 (kernel identity). Apply the conditional kernel M5 to
the bounded nonnegative F_{q_j}-measurable indicator
1{A_j} 1_Q:
E[ 1{A_j} (mu_infinity - mu_{q_j})(Q) ]
= E[ 1{A_j} int_Q G_{q_j}(z) dA(z) ]
= int_Q E[ 1{A_j} G_{q_j}(z) ] dA(z),
using that A_j is F_{q_j}-measurable and Fubini on the compact
Q.
Step 5 (finite-tilt evaluated at q_j). By op:tested-potential
applied with H = 1{A_j} (bounded, F_{q_j}-measurable) and
t = q_j,
E[ 1{A_j} G_{q_j}(z) ] = G_0(z) P_z^*(A_j, q_j < T^*).
Because the A_j are disjoint and T^* is fixed-path,
sum_{j=1}^N P_z^*(A_j, q_j < T^*)
= P_z^*(exists j <= N with q_j < T^* and M_{q_j} > H).
Step 6 (assemble). Summing Step 4 over j and using Step 5,
E[ int_Q 1{exists j <= N, q_j < v(z), M_{q_j} > H}
mu_infinity(dz) ]
= int_Q G_0(z)
P_z^*(exists j <= N, q_j < T^*, M_{q_j} > H) dA(z).
Step 7 (pass to all rational times). Exhaust Q_{>=0} by an
increasing sequence of finite subsets, sorting each finite
subset before forming the A_j (the "first success" partition
depends only on the sorted list). The existence events on both
sides increase with the finite subset. Monotone convergence
gives the full identity. The measurable choice of
P_z^*(A, q < T^*) in z was established in C.2 (the
parameter-measurability note inside the finite-tilt proof).
QED.
BAD CELLS.
Fix a compact E compactly contained in HH. Set
D = 16 * (C_*')^d.
For batch entry (n,i), write r = r_{ni} and b = b_{ni}. Use the
square grid of side r in the plane. Call a grid cell Q "bad for
entry (n,i)" when
mu_infinity(Q) > 2 D b r^d.
CLAIM (Summability of bad cells). For every sufficiently large
n, uniformly over all its entries and all cells meeting E,
E[ mu_infinity(Q) ; Q is bad for entry (n,i) ]
<= 4 ( integral_Q G_0(z) dA(z) ) *
P_mix^*[ L >= 16 b ].
Interpretation: The expected TERMINAL mass sitting on a bad cell
is at most the GREEN potential of that cell times the small
probability that the disk-model mass exceeds the threshold b.
Because sum over all (n,i) of P_mix^*[L >= 16 b_{ni}] is
summable in n (Lemma D.1 "tail summability"), the sum over
all bad cells of the expected mass they carry is finite.
Consequently, almost surely, mu_infinity-almost every point of E
lies in NO bad cell for any entry of any sufficiently late batch.
PROOF.
Step 1 (split the mass). Put H = D b r^d so "Q bad" means
mu_infinity(Q) > 2H. On that event, mu_infinity(Q) - H >
mu_infinity(Q) / 2, hence mu_infinity(Q) < 2 (mu_infinity(Q)
- H). So
E[ mu_infinity(Q); Q bad ]
<= 2 E[ 1{Q bad} (mu_infinity(Q) - H) ].
It remains to bound the right side.
Step 2 (past-mass identity). Let q* = inf { q >= 0 : M_q > H }
with M_q = mu_q(Q). Continuity and monotonicity of q |-> M_q
(M4) give M_{q*} = H on the event {exists q: M_q > H}; this
event holds a.s. on {Q bad}. On {q* < v(z)}, the first-success
event in the past-mass identity fires for some rational q in
(q*, v(z)); on {q* >= v(z)} it cannot fire. So the left side
of the past-mass identity at this H equals
E[ int_Q 1{q* < v(z)} mu_infinity(dz) ].
Step 3 (relate to Q bad). On {Q bad},
(mu_infinity - mu_{q*})(Q) = mu_infinity(Q) - H
> mu_infinity(Q) / 2,
and off {Q bad} this quantity is <= mu_infinity(Q)/2.
Combine Step 2 with the explicit form of the past-mass
identity:
E[ 1{Q bad} (mu_infinity(Q) - H) ]
<= E[ (mu_infinity - mu_{q*})(Q); Q bad ]
<= E[ int_Q 1{q* < v(z)} mu_infinity(dz) ]
= int_Q G_0(z)
P_z^*(exists q < T^*, M_q > H) dA(z).
Step 4 (restart comparison). Fix z in Q and n large enough
that r = r_{ni} <= e^{-n^2} and (2a)^{-1} log(R_0(z)/(100r))
>= s_a (this inner gap is uniform on the compact E for
sufficiently large n). Let S be the first time R_S(z) = 100r
while z alive. On {S < T^*}, the disk map F: DD -> H_S with
|F'(0)| = 100r exists, and the Koebe growth and distortion
estimates of C.4 give
B(z, 16r) subset F(K) subset B(z, C_*r),
c_* r <= |F'(w)| <= C_*' r (w in K).
Since Q is a square of side r centred at z, Q subset F(K).
By the spatial comparison of C.4, for every q < T^*,
mu_q(Q) <= mu_q(F(K)) <= (C_*'r)^d L,
where L is the disk-model mass at the corresponding inner
level. (Past mass mu_S(Q) = 0: Q lies in interior H_S.) So
{exists q < T^*: M_q > H} is contained in
{L > H/(C_*'r)^d} = {L > 16 b}, using H = 16 (C_*')^d b r^d.
Step 5 (mixing at restart). In inner time, S is reached after
at least (2a)^{-1} log(R_0(z)/(100r)) >= s_a. So the uniform
mixing bound of C.2 applies to the restart angle theta_S:
its law has density at most 2 relative to pi_a. Conditional
on F_S, the future driver and hence L is independent of the
past; more precisely, the conditional law of L given
theta_S agrees with the disk law P_{theta_S}^*. Hence
P_z^*(exists q < T^*, M_q > H)
<= E_z^*[ P_{theta_S}^*(L > 16 b) ]
<= 2 P_mix^*[L > 16 b].
Combining with Step 3 and Step 1,
E[ mu_infinity(Q); Q bad ]
<= 4 (int_Q G_0(z) dA(z)) P_mix^*[L > 16 b].
Step 6 (Tonelli and cleanup). For the countable family of
cells of every grid of every batch entry, apply Tonelli to
the nonnegative integer-valued sum over all cells of Q:
- the number of cells meeting E in the grid of side r_{ni}
is at most O(r_{ni}^{-2}), and the total expected
sum of the per-cell bounds is
sum_n 4 (int_E G_0) sum_i P_mix^*[L > 16 b_{ni}],
- which is finite by Lemma D.1 "tail summability" and
the cleanup-tail bound sum_n P_mix^*[L > 16 n]
<= m_1 / 16 < infinity.
Hence, almost surely, mu_infinity-almost every point of E
lies in only finitely many bad cells among all entries of
all batches. Grid lines carry zero terminal mass by
E[mu_infinity(dz)] = G_0(z) dA(z) and because lines have zero
two-dimensional Lebesgue measure. So there is a batch cutoff
n_0 such that no cell at a batch n >= n_0 is bad on a set of
full mu_infinity|_E-measure. QED.
DENSITY BOUND.
Fix a nontrivial rational interval [u,v] with 0 < u < v. The
increment mu_v - mu_u has positive total mass (by M7) on some
compact E compactly contained in HH. Restrict to the "surviving
good set" S_{n_0}: E minus the union of every bad cell for
entries (n,i) with n >= n_0. Then S_{n_0} is Borel, S_{n_0}
increases with n_0, and the union over all S_{n_0} carries the
full increment on E. Choose n_0 so that (mu_v - mu_u)(S_{n_0})
is strictly positive, and set
nu = (mu_v - mu_u) restricted to S_{n_0}.
Then nu is finite, nonatomic, carried by gamma([u,v]), and
nu(HH) > 0.
For every entry (n,i) with n >= n_0, every grid cell has
nu(Q) <= 2 D b_{ni} r_{ni}^d.
(Bad cells have zero nu-mass by construction of S_{n_0}; good
cells have mass bounded by 2D b_{ni} r_{ni}^d; cells disjoint
from E have zero nu-mass.)
Now take any subset U of the plane with diameter s > 0. How
many grid cells of side r_{ni} can U meet? Each coordinate
projection has length at most s, so it meets at most (s/r_{ni})
+ 2 grid intervals, hence U meets at most
(s/r_{ni} + 2)^2 <= 16 * max{ 1, (s/r_{ni})^2 }
cells if s/r_{ni} >= 1 (the "+2" absorbs constants) and trivially
bounded otherwise. Then
nu^*(U) <= sum over cells met of nu(cell)
<= 16 max{1,(s/r_{ni})^2} * 2 D b_{ni} r_{ni}^d
= 32D b_{ni} r_{ni}^d max{1,(s/r_{ni})^2}.
This is true for every n >= n_0 and every entry (n,i).
PASSING TO THE FULL GAUGE.
Define F(s) = sup_{diam U <= s} nu^*(U). This is bounded by
nu(HH) < infinity, and satisfies F(s) <= 32D H_{ni}(s) for
every n >= n_0 and every entry (n,i) of batch n, where H_{ni}
is the function from D.2. So F(s) satisfies the finite-exclusion
lemma of D.2 with C = 32D:
F(s) <= C_nu * h_kappa(s) for every s > 0,
where
C_nu = max{ 32D, max_{n < n_0, i}
nu(HH) / (b_{ni} r_{ni}^d) }.
Since the set {n < n_0, i} is finite, C_nu is finite.
For a set of diameter zero, nonatomicity of nu gives
nu^*(empty-ish U) = 0.
CONSEQUENCE.
For any countable covering {U_j} of gamma([u,v]) by blobs of
small diameter, outer-measure subadditivity gives
nu(HH) = nu^*(gamma([u,v])) <= sum_j nu^*(U_j)
<= C_nu * sum_j h_kappa(diam U_j).
Taking the infimum over covers with diameter cutoff -> 0,
H^h_kappa(gamma([u,v])) >= nu(HH) / C_nu > 0.
There are only countably many nontrivial rational intervals.
Every real 0 < s < t contains a nontrivial rational closed
subinterval. By monotonicity of Hausdorff measure, positivity
of each rational subinterval implies positivity of every real
positive-time subinterval. All this on the single
probability-one event from the summability of bad cells.
E.2 THE UPPER BOUND (FINITENESS)
Part E.1 showed H^h(gamma([s,t])) is positive. Part E.2 shows
H^h(Gamma cap B_m) is FINITE (in expectation, and then almost
surely), where B_m = [-m, m] + i[0, m]. Together they prove the
Main Theorem.
The strategy.
(T1) A single-cell visit lemma: if a closed square Q of side r
centered at z = x + iy with y > 4r meets the trace, then z
reaches conformal radius 8r while alive. This gives an
initial probability bound on Q being visited.
(T2) Batched finite tests. At the final mesh of batch n (the
smallest radius in the batch), run the J = I_n - 1
non-cleanup tests at fixed inner-time gaps. Success at
test i forces the ball of the entry to be a "candidate",
i.e. its terminal mass is at least c_*^d times the disk
mass threshold.
(T3) Failure probability. With the batch multiplicities, the
probability that a cell is visited and has NO candidate
is at most C_a r_f^alpha y^{-alpha} exp(-2^n / 16).
(T4) Greedy covering of candidate cells: substitute a disjoint
subcollection, whose triples cover every cell that had a
candidate. Disjointness is what lets one actual terminal
mass pay for the entire chosen subcover.
(T5) Cleanup for missed cells: every remaining visited cell is
paid for individually by the cleanup entry of the batch.
Its expected cost is O(r_f^d) per cell, and there are
O(r_f^{-2}) cells in the box, so the total cleanup cost
is bounded.
(T6) Assemble: each batch n gives a finite cover of Gamma cap B_m
with total expected gauge cost bounded by a uniform constant
C_m independent of n. Take n -> infinity and use Fatou.
The single-cell visit lemma.
Let Q be a closed square of side r with center z = x + iy (y > 0).
Suppose y > 4r and Q intersects Gamma. Then z reaches conformal
radius 8r while alive.
Proof. If z is swallowed before R reaches 8r, the current domain
H_t contains a ball B(z, 2r) for every t < T_z (the lifetime of
z), by the Koebe quarter theorem. The past trace avoids this open
ball. But the ball B(z, 2r)
lies, at the moment z is swallowed, in the bounded complementary
component of HH minus gamma([0, T_z]) -- since z itself is in
that component and the ball is connected and disjoint from the
trace. The future trace lies in the closure of the SURVIVING
domain H_{T_z}, which is disjoint from that bounded component, so
the future trace cannot hit the ball (or Q inside it). So Q
being visited forces R > 8r at some point before z is swallowed.
Quantitatively, this gives
P( Q intersects Gamma ) <= C_a r^alpha y^{-alpha},
for every r > 0 and y > 0. The proof uses the finite-tilt change
of law at conformal-radius level 8r, plus inverse-angle
integrability to absorb the boundary angle. For y/r bounded (so
the crossing time is short), the trivial bound P <= 1 absorbs
the constant.
Setup for batch n.
Fix m (the box) and take a sufficiently large batch n. Write
J = I_n - 1, (number of non-cleanup tests)
r_f = r_{n, I_n}, (final mesh radius)
r_i = r_{ni}, b_i = b_{ni} for 1 <= i <= J.
By Lemma D.1, J is at least 1 for all sufficiently large n
(there are always at least the batch-multiplicity entries).
Cover B_m by finitely many closed squares of side r_f, with
lower-left corners at x = -m + k r_f (k = -m/r_f, ...) and
heights y_j = (j + 1/2) r_f starting from 0. Retain the whole
squares in the outermost rows and columns. Centers are at
z_{j,k} = (-m + (k + 1/2) r_f) + i( j + 1/2 ) r_f.
Candidate definition. For a batch entry (b_i, r_i) at a center
z, define a "candidate":
mu_infinity( overline B(z, C_* r_i) cap HH )
>= c_*^d b_i r_i^d.
Equivalently, the terminal mass inside the ball of radius
C_* r_i around z is at least c_*^d b_i r_i^d.
Every candidate ball CONTAINS the final cell centered at z.
There are finitely many candidate checks in a batch, they
involve only terminal mass (so they can be evaluated AFTER the
sample is drawn), and candidate selection itself is done after
the sample.
Estimating the probability that a cell is VISITED but has NO
candidate.
Fix a center z with height y >= e^{-n}. A visit to the final
cell centered at z requires z to reach conformal radius 8r_f by
Lemma E.2 above.
For each test i in the batch:
- S_i = first time R_t(z) = 100 r_i while z alive.
- T_i = first time R_t(z) = 100 r_i e^{-2a ell_n} while alive.
- On {tau_f < infinity}, both finite, and the inner-clock
gap from S_i to T_i is exactly ell_n.
- Let F_i: DD -> H_{S_i} be the disk map (as in C.5), and
define the stopped test mass
X_i = integral_{F_i(K)} |F_i'(F_i^{-1}(x))|^{-d}
( mu_{T_i} - mu_{S_i} )( dx ).
On a reached endpoint, X_i equals the disk mass L_{ell_n}
for the restart angle. Set X_i = 0 otherwise.
KEY. Measurable at T_i, and only paths reaching tau_f enter
the relevant event below.
Success. Success X_i >= b_i implies
( mu_{T_i} - mu_{S_i} )( F_i(K) )
>= c_*^d b_i r_i^d,
so by the Koebe containment F_i(K) subset B(z, C_* r_i) and
monotonicity of mass the ball qualifies as a candidate.
Gap structure. The inner-clock gaps are deterministic:
sigma_1
= (1/(2a)) log( 2y / (100 e^{-n^2}) )
>= (n^2 - n - log 50) / (2a),
sigma_{i+1} - (sigma_i + ell_n) = n (1 <= i < J),
sigma_f - (sigma_J + ell_n)
= n + (1/(2a)) log(100/8),
where sigma_i = s_z(S_i), sigma_f = s_z(tau_f).
For large n each gap permits the mixing lemma of C.2.
Iterated conditional expectations: Let q_i = P_mix^*{ L_{ell_n}
>= b_i }. Conditional on the whole stopped history through
T_{i-1}, the angle at S_i has density at least 1/2 relative
to pi_a (mixing after the deterministic gap n). The disk law of
X_i is determined by that angle (via restart M10). So
conditional probability of success is at least q_i / 2.
Applying the lemma D.1's "lower success" inequality, the
probability that X_i < b_i for ALL i <= J is at most
prod_{i=1}^J (1 - q_i / 2) <= exp( - (1/2) sum q_i )
<= exp(- 2^n / 16 ).
Changing law for the FULL failure event.
The event "cell visited and has no candidate" is contained in
A = { tau_f < infinity } cap (cap_{i=1}^J { X_i < b_i } ).
All test endpoints precede tau_f, so A is measurable at that
stopping time. Only this enlarged event is used when changing
law. In particular, no terminal candidate information is put
into a finite Radon-Nikodym formula.
Let that~ = theta_{tau_f}(z) be the final angle of z. Conditional
on the history through T_J, the final deterministic gap gives
E_z^*[ sin^{-p}( that~ ) | F_{T_J} ]
<= 2 integral sin^{-p}(theta) pi_a(d theta) < infinity.
Since the failure-of-all-tests event F_fail is known at T_J, we
have
P( cell at z is visited and has no candidate )
<= P( A )
= (8 r_f)^alpha G_0(z) E_z^*[ sin^{-p}(that~) ; F_fail ]
<= C_a r_f^alpha y^{-alpha} exp(-2^n / 16)
for y >= e^{-n}.
For centers below e^{-n}, use the unconditional estimate
P( cell at z is visited ) <= C_a r_f^alpha y^{-alpha}
obtained from the single-cell visit lemma. This also covers the
bottom row (center height r_f/2), where the trivial bound
absorbs the constant.
Greedy covering.
Order all candidate balls of the batch by decreasing radius
r_i, with a fixed deterministic tie-break. Choose the first
ball, delete every ball meeting it, and continue. The selected
closed balls are pairwise disjoint.
For a deleted ball of radius r_i <= r_{i'} (the selected ball
that deleted it), the concentric triple of the selected ball
contains the deleted ball. Therefore the triples of the
selected balls cover EVERY final cell that had a candidate.
Cost. For a selected ball with entry (b_i, r_i), the triple
has radius 3C_* r_i, diameter 6C_* r_i. The gauge formula
(D.2) gives
h_kappa( 6 C_* r_i )
<= ( 6 C_* )^2 b_i r_i^d
<= ( 6 C_* )^2 / c_*^d *
mu_infinity( overline B(z, C_* r_i) cap HH ).
For all sufficiently large n, these balls lie (after
intersection with HH) inside the fixed bounded set
Btilde_m = [-m-2, m+2] + i(0, m+2].
Disjointness: the total gauge cost of the selected triples is
bounded by
( 6 C_* )^2 / c_*^d * mu_infinity( Btilde_m ).
This is the actual terminal mass of a fixed region on the same
sample. It has finite expectation:
E mu_infinity( Btilde_m )
= integral_{Btilde_m} G_0(z) dA(z)
<= (2^{-alpha} (2m+4)(m+2)^{1-alpha}) / (1-alpha)
< infinity.
This works even at the real boundary.
Cleanup for missed cells.
Add every final cell that meets Gamma and has no candidate.
The cleanup entry of the batch gives the individual cost
h_kappa( sqrt(2) r_f ) <= 2 n r_f^d.
Number of cells. There are O_m(r_f^{-1}) columns and
O_m(r_f^{-1}) rows in the box. Since 0 < alpha < 1, the sum
over half-integer row heights satisfies
sum_{all cells} y^{-alpha}
<= C_{a,m} r_f^{-2},
sum_{y < e^{-n}} y^{-alpha}
<= C_{a,m} r_f^{-2} (e^{-n} + r_f)^{1-alpha}.
(E.g. sum_{j=0}^N (j + 1/2)^{-alpha}
<= C_alpha (N+1)^{1-alpha} proves both.)
Expected added cost. Multiplying the per-cell probability
bounds by the cleanup per-cell cost h(sqrt(2) r_f) <= 2n r_f^d
and using d + alpha = 2 (which is alpha = 2 - d from Part A):
C_{a,m} [ n exp(-2^n / 16)
+ n (e^{-n} + r_f)^{1-alpha} ].
Both terms are bounded uniformly in n and tend to 0 as
n -> infinity. In particular, the bottom-row contribution is
at most C_m n r_f^{d-1}, which is o(r_f^{1-d}) as required.
Measurability.
The candidate family is finite with measurable terminal-mass
indicators. The greedy order is deterministic. For a fixed
closed cell, the event of meeting the trace is measurable on
every compact time interval (by continuity); a countable union
over integer time horizons gives the full trace event. All
upper-bound centers, entries, and boxes together constitute a
countable family, so the stopped measure identities used in
constructing their tests can be imposed on one event.
Assembly.
Let C_n be the total gauge cost of the cover of Gamma cap B_m
produced by the batch-n construction. The bounds above show
that sup_n E C_n < infinity (after omitting finitely many
initial batches). Every cover diameter is at most
delta_n = 6 C_* e^{-n^2} -> 0.
Since H^h_delta(Gamma cap B_m) <= C_n for every late n:
Fatou's Lemma gives
E[ liminf_n C_n ] <= liminf_n E C_n < inf,
so liminf C_n is finite almost surely. Since covers with
delta_n <= delta are all admissible for H^h_delta, and taking
delta -> 0:
H^h_kappa( Gamma cap B_m ) <= liminf_n C_n < inf a.s.
Measurability of the left side.
Fix integer T; then K_T = gamma([0,T]) cap B_m is a measurable
compact random set. Indeed, for a fixed closed F,
{K_T cap F != empty} equals the event
min_{0 <= t <= T} dist(gamma(t), B_m cap F) = 0,
and the minimum can be taken over rational times.
For compact-set measurability of the map K |-> H^h(K),
approximate H^h(K) by countable infima over finite unions of
rational balls, and note that a fixed-cutoff countable
infimum is measurable in K. This identifies H^h
Gamma cap B_m = lim_T H^h(K_T) as measurable.
Then
H^h( Gamma cap B_m ) = lim_T H^h(K_T)
by continuity from below (Hausdorff outer measure is a Borel
measure). Integrating over the Fatou bound proves S4.
A countable intersection now gives almost-sure finiteness in
every integer box B_m simultaneously.
Theorem assembly.
Intersect the probability-one events of E.1 (lower bound) and
E.2 (upper bound).
On that event:
- The lower bound holds for every real 0 < s < t < infinity.
- The image gamma([s,t]) is compact (continuity of gamma),
so it lies in some integer box B_m.
- H^h(gamma([s,t])) is finite on that box.
- H^h(gamma([s,t]) >= positivity from E.1.
Therefore
0 < H^h_kappa( gamma([s,t]) ) < infinity
simultaneously for every real positive-time subinterval.
The constant rescaling between capacity time and internal time
preserves the statement: the sets gamma([s,t]) are the same
set of curve points, only the time labels differ.
This proves the Main Theorem (S1)-(S4).
-----------------------------------------------------------------------------
PART F -- ASSEMBLY, PROOF MAP, REFERENCES
-----------------------------------------------------------------------------
F.1 THE WHOLE PROOF IN SIX LINES
Despite all the machinery, the proof is a sequence of six concrete
statements. Read this even if you skipped Part E.
Line 1. There is a canonical mass measure mu on the trace
(C.4) with conditional kernel mu_infinity - mu_S given
by G_S(x) dx dy (M5), a pathwise restart identity M10
at stopping times, and first-visit accounting M8.
Line 2. The centered law P_x^* at a marked point x is well
defined (C.2): it makes the angle theta stationary
with density pi_a(theta) proportional to sin^{4a}(theta)
and mixes uniformly after a fixed inner time s_a. The
lifetime T^*(x) is finite almost surely.
Line 3. The disk-model mass L_ell = mu_{tau_ell(0)}(K) has all
finite POSITIVE moments 0 < m_n < infinity (C.5),
and each moment is a deterministic number computable
from kappa via explicit Wiener / angle / area integrals.
Line 4. Convert the moments into a family of dyadic-scale tests
A_k = 2^k with multiplicities N_{nk} (D.1-D.2). The
upper tail is summable in n (extreme mass cost), the
lower success is at least 2^n / 8 (many useful
chances). One cleanup entry per batch covers the
leftover.
Line 5. LOWER BOUND (E.1): in each grid of side r_{ni}, cells
whose terminal mass exceeds 2 D b_{ni} r_{ni}^d are
paid off by PAST mass (mass that arrived before the
cell's center was first visited). Summability of the
upper tail means only finitely many overfull cells
survive. The surviving set carries mass nu > 0, which
respects the density bound nu(U) <= C_nu h_kappa(U).
Any countable cover costs at least nu(HH) / C_nu, so
H^h_kappa(gamma([s,t])) > 0.
Line 6. UPPER BOUND (E.2): at every large batch n, run finite
tests. If a test succeeds at radius r_i, the ball of
radius C_* r_i around the center is a "candidate" for
the cover. Greedy selection over a batch (in decreasing
radius order) leaves DISJOINT selected balls whose
triples cover every non-empty candidate cell. The mass
in one bounded region Btilde_m pays for all of them at
once. Empty cells are paid for by the cleanup entry,
giving h(sqrt2 r_f) <= 2 n r_f^d per cell. The total
expected cost is bounded by a constant C_m uniform in
n. Letting n -> infinity, Fatou gives
H^h_kappa(Gamma cap B_m) < infinity almost surely.
That is S1-S4 of the Main Theorem. Everything else is paperwork.
F.2 WHY THE CONSTRUCTION IS "THE RIGHT ONE"
A reader might wonder: why THESE moments, why THESE batches, why
THIS gauge formula? Here is the honest answer.
* The mass measure mu is not arbitrary. It is FORCED by the
"reasonable" requirement that the future conditional mean is
a canonical Green kernel G_S dx dy with G_S = R_S^{-alpha}
sin^p(theta_S). Uniqueness (C.4) says there is exactly one such
adapted continuous finite-variation measure process. So the
existence of mu and its kernel is not a choice; it is a
consequence.
* The centered laws P_x^* are FORCED by the same mass construction
with the finite-tilt weight M_ell = G_{tau_ell(x)} / G_0(x). The
change of law has expectation one and gives the tilted angle
diffusion. Mixing follows from the explicit spectral
decomposition of that diffusion.
* The disk moments m_n are FORCED once you decide to look at
mass in a fixed subdisk of a canonical disk model. The values
are deterministic numbers (a Wiener integral, an angle
average, an area integral) because the centered law is a
deterministic average. They are not a "choice".
* The batch construction is FORCED, modulo minor tweaks, by the
following demand. We want a gauge h such that:
(a) for every h-bounded cover of a subset of the trace, the
total mass sitting on the trace is at most C * h(cover);
(b) for every batch, at every cell, either the cell contains
a positive mass witness (candidate), or the mass is
bounded by h(cell).
Requirement (a) pulls one direction (h must be "large enough"
so that the density bound is achievable). Requirement (b)
pulls the other (h must be "small enough" so that the mass
bound is finite). The "quadratic envelope"
max{1, (r/r_{ni})^2}
is the minimal way to bridge these two pulls: it gives a cost
at scale r_{ni} and quadratic growth for subdividing subgrids.
Any gauge satisfying (a) and (b) must dominate this envelope
on each entry; the infimum over entries is the minimal such
gauge at every scale.
* The specific numbers (4^n, 2^{kn}, m_n, T_n) are chosen so that
the two probability bounds (upper tail, lower success) are in
balance. Larger multiplicity would blow up the upper-tail cost;
smaller would make the failure probability non-summmable. The
factor 4 is a free parameter that could be adjusted within a
small range; nothing in the proof is tight.
F.3 WHAT IS PROVED HERE AND WHAT IS EXTERNAL INPUT
Fully proved in this document (every lemma, proposition, and
computation is derived explicitly in Parts C-E, in the order used):
- A.1-A.3: Hausdorff measure, dimension, gauges, the Schramm
question.
- A.4: SLE via Loewner equation (standard background).
- C.1: conformal radius, inner time, level times.
- C.2: Green weight, local martingale property, finite-tilt
change of law (with the explicit Optional-Stopping /
Girsanov argument for E[M_ell] = 1), stationary angle,
orthogonal-polynomial proof of uniform mixing, Lyapunov
and Borel-Cantelli proof that T^*(x) < infinity.
- C.3: the multipoint estimate at every fixed m, including
the explicit A, h_0, C_1, c_1, C_2, C_3 from the multipoint
supermartingale, and the Young + cross-term absorption that
gives the drift k_t.
- C.4: existence of mu via the explicit Polish-space +
Cantor-diagonal + Cesaro-average argument; properties M1-M10,
including the strong-Markov + uniqueness proof of the
pathwise restart identity M10, and measurability of the mass
functional in the marked point.
- C.5: the disk model, its moments, the explicit Wiener / angle
integral formulas.
- D.1: batches, upper tail, lower success, cleanup.
- D.2: gauge definition, admissibility, finite-exclusion.
- E.1: spatial comparison, past-mass identity via the
first-success partition + conditional kernel M5 + monotone
convergence, bad-cell summability via the restart comparison
and mixing, density bound, positive lower Hausdorff measure.
- E.2: single-cell visit, batched tests, greedy covering,
cleanup, measurable cover, finite expected upper measure,
Borel measurability of H^h on compacts via rational-ball
infima and continuity from below, assembly of S1-S4.
External inputs used, cited in A.4 and C.1-C.2:
- continuous trace generation for chordal SLE (Rohde-Schramm
2005, Theorem 5.1): existence and continuity of gamma;
- classical Koebe estimates for conformal maps (Beliaev 2015,
Section 3.2: Theorems 3.2.5, 3.2.9, 3.2.11 and Corollary
3.2.6): the quarter theorem, growth, and distortion bounds.
Every other argument -- centered laws, multipoint bounds, mass
construction and restart, disk moments, batches, gauge, and both
covering proofs -- is proved self-contained below. The former
"sketched in the source paper" list has been fully absorbed into
the body; no claim in the document depends on external unproved
material beyond the two inputs named above.
F.4 REFERENCES
The full paper cites these works. Here is a plain list for
further reading.
Schramm--Loewner basics:
- O. Schramm, "Conformally invariant scaling limits: an
overview and a collection of problems," ICM 2006. Contains
the original Problem 7.1 on Hausdorff measures of SLE.
- S. Rohde, O. Schramm, "Basic properties of SLE," Annals of
Mathematics 161 (2005), 883-924. Existence and continuity
of the SLE trace; the Loewner equation we use here.
Dimension and Hausdorff measure of SLE:
- V. Beffara, "The dimension of the SLE curves," Annals of
Probability 36 (2008), 1421-1452. The formula d = 1 + kappa/8.
- M. A. Rezaei, "Hausdorff measure of SLE curves," Stochastic
Processes and their Applications 128 (2018), 884-896. Proves
H^{r^d} = 0 for chordal SLE, which is why we need a gauge
strictly bigger than r^d.
Natural parametrization and Minkowski content:
- G. F. Lawler, S. Sheffield, "A natural parametrization for
the Schramm-Loewner evolution," Annals of Probability 39
(2011), 1896-1937.
- G. F. Lawler, W. Zhou, "SLE curves and natural
parametrization," Annals of Probability 41 (2013),
1556-1584.
- G. F. Lawler, M. A. Rezaei, "Minkowski content and natural
parameterization for the Schramm-Loewner evolution," Annals
of Probability 43 (2015), 1082-1120.
- G. F. Lawler, M. A. Rezaei, "Basic properties of the natural
parametrization for the Schramm-Loewner evolution,"
arXiv:1203.3259 (2012).
Multipoint estimates and Green functions:
- G. F. Lawler, B. M. Werness, "Multi-point Green's functions
for SLE and an estimate of Beffara," Annals of Probability 41
(2013), 1513-1555.
- M. A. Rezaei, D. Zhan, "Higher moments of the natural
parameterization for SLE curves," Ann. IHP Prob. Stat. 53
(2017), 182-199.
- M. A. Rezaei, D. Zhan, "Green's functions for chordal SLE
curves," Prob. Theory Rel. Fields 171 (2018), 1093-1155.
Conditional decompositions along the trace:
- L. S. Field, "Two-sided radial SLE and length-biased chordal
SLE," arXiv:1601.03374 (2016).
- D. Zhan, "Decomposition of Schramm-Loewner evolution along
its curve," Stochastic Processes and their Applications 129
(2019), 129-152.
- D. Zhan, "Optimal Holder continuity and dimension properties
for SLE with Minkowski content parametrization," Prob.
Theory Rel. Fields 175 (2019), 447-466.
Variation and modulus of continuity:
- N. Holden, Y. Yuan, "Regularity of the Schramm-Loewner
evolution: up-to-constant variation and modulus of
continuity," Annals of Probability 54 (2026), 367-420.
Conformal mapping toolbox:
- D. Beliaev, "Conformal Maps and Geometry," Oxford lecture
notes (2015). Contains the Koebe estimates used here.
Quadratic regularization:
- Y. Peres, B. Solomyak, "The sharp Hausdorff measure condition
for length of projections," Proc. AMS 133 (2005), 3371-3379.
The "quadratic envelope" regularization of gauge functions
goes back to this paper.
The companion paper:
- OpenAI, "An explicit exact Hausdorff gauge for SLE: a regular
formula from quantitative tails and dense visits," OpenAI
Math Release (2026). Proves that the CLOSED-FORM gauge
r^d (log log(1/r))^{(2-d)/2}
is also exact on chordal SLE by a different method. This
paper's moment-defined gauge is a priori different; the
asymptotic comparison between it and the closed-form gauge
is left open here.
F.5 WHAT TO READ NEXT
If this has caught your interest and you want to go further:
- The companion paper (OpenAI 2026) proves that the closed-form
log-log gauge is also exact on chordal SLE. The natural open
question is whether its gauge and the moment-gauge built here
agree asymptotically at small radii.
- The multipoint Green's function machinery (Lawler-Werness,
Rezaei-Zhan) has many other applications in SLE theory.
- Natural parametrization and its Minkowski content are the
standard "fair clock" on the SLE trace; the canonical mass mu
built here is the object that natural parametrization measures,
up to normalization.
- The moment-batch covering trick generalizes: whenever all
moments of a localized random measure are bounded, one can
often construct a gauge that measures the underlying set.
==================================================================
ACKNOWLEDGMENTS
==================================================================
This document is a pedagogical re-derivation of the original
technical paper "An exact Hausdorff gauge for SLE". Every
lemma and proposition in Parts C, D, E is proved from first
principles in this text. The external inputs used are:
- continuous trace generation for chordal SLE
(Rohde-Schramm 2005, Theorem 5.1), and
- the classical Koebe estimates for conformal maps
(Beliaev 2015, Section 3.2).
Everything else -- Green weight as local martingale, finite
centered laws, uniform mixing, multipoint radius bound,
canonical mass construction with conditional kernel, explicit
disk moments, finite batches, gauge construction with admissibility,
past-mass identity, bad-cell summability, single-cell visit lemma,
greedy covering, cleanup, and both Fatou applications -- is
derived below (Parts C, D, E) without deferring to the paper.
===== END OF DOCUMENT =====
# ^C interrupted by user!