Let S^2 be the unit sphere in R^3, let sigma be surface measure normalized
by sigma(S^2) = 4 pi, and for bounded measurable complex g on S^2 put
Eg(x) = int_(S^2) g(omega) exp(2 pi i <x, omega>) dsigma(omega),
x in R^3. (Ext)
The content of the result is that the extension operator maps bounded
data on the sphere into every L^p(R^3) with p > 3, which is the whole
conjectured open range:
|| Eg ||_(L^p(R^3)) <= C_p || g ||_(L^inf(S^2, sigma)), p > 3. (*)
Sharpness at p = 3 is an elementary computation. Take g = 1 and use
spherical coordinates about the axis x, writing r = |x|. The azimuthal
integral contributes 2 pi and the remaining integral is elementary:
E1(x) = 2 pi int_(-1)^1 exp(2 pi i r u) du = 2 sin(2 pi r) / r.
For p > 0,
int_(|x| > 1) |E1|^p dx = 4 pi int_1^inf r^(2-p) |2 sin(2 pi r)|^p dr,
whose integrand has positive mean comparable to r^(2-p) over each period
of sin, so the integral diverges exactly when p <= 3. The endpoint is
therefore excluded by a test function and (*) is the exact open range:
the estimate is asserted for p > 3 and for no smaller exponent. The
input class matters. A diagonal estimate L^p(S^2) -> L^p(R^3) controls
all L^p sphere data and hence also all bounded data; a bounded-data
estimate controls only || g ||_inf. Both forms appear in what follows and
they are distinguished where used.
The routes to (*) run through four bodies of work. Classical oscillatory
integral theory begins with Carleson and Sjolin, who reduced the disc
multiplier to a curvature condition on families of phase functions, and
with Fefferman's ball multiplier theorem, which showed that the
curvature is essential. Cordoba tied the radial multiplier problem to a
Kakeya maximal function, and Hormander isolated the general class of
oscillatory integral operators with Carleson-Sjolin curvature. Stein's
treatment of interpolation and the Tomas-Stein argument give the L^2
theory, which reaches p >= 2(d+1)/(d+3) but never p = 3 in three
dimensions. Every later gain on the sphere came from adding geometric
information about how tubes concentrate, rather than from the L^2
method.
pdf/src-1512.03238.pdf
Bourgain's work on Besicovitch type maximal functions and on
factorization of the extension operator is the source of the idea that
restriction bounds follow from incidence bounds for tubes, with a
loss that depends on what fraction of a tube is occupied. Wolff's
Kakeya maximal bounds, the bilinear cone estimate of Tao, Vargas and
Vega, and Wolff's sharp bilinear cone restriction estimate introduced
the bilinear and multilinear reductions of the problem; in the critical
regime the multilinear theory loses nothing. Bourgain and Guth combined
multilinear restriction with a transverse interaction analysis and
reached p > 33/10 on the sphere. Guth's polynomial partitioning reached
p > 13/4, and the weighted and two-ends refinements of Wang and of Wang
and Wu reached p > 42/13 and then p > 22/7.
pdf/src-1012.3760.pdf
pdf/src-1407.1916.pdf
pdf/src-1802.10186.pdf
Multilinear restriction in its sharp form is due to Bennett, Carbery and
Tao, whose abstract states the mechanism: "[d]-linear analogues of the
classical restriction and Kakeya conjectures in R^d [are proved]. [The]
approach involves obtaining monotonicity formulae pertaining to a
certain evolution of families of gaussians, closely related to heat
flow."
pdf/src-math-0509262.pdf
Polynomial partitioning entered through Guth, whose abstract records both
the setting and the loss: "If S is a smooth compact surface in R^3 with
strictly positive second fundamental form, and ES is the corresponding
extension operator, then for all p > 3.25, ||ES f||_(L^p(R^3)) <=
C(p,S) ||f||_(L^inf(S)). The proof uses polynomial partitioning arguments
from incidence geometry."
pdf/src-1407.1916.pdf
The oscillatory half of the argument below is a propagation of packet
arrays, and its analytic input is l^2 decoupling for compact
hypersurfaces with definite second fundamental form, proved by Bourgain
and Demeter: "[The] l2 Decoupling Conjecture [is proved] for compact
hypersurfaces with positive definite second fundamental form and also
for the cone." Refined decoupling in the form used for tube counts, and
the exact packet frame identities, follow the treatment of Guo, Oh,
Wang, Wu and Zhang, which builds the frame from Fourier series and
controls every change of frame with polynomial losses.
pdf/src-1403.5335.pdf
The geometric half needs quantitative incidence for tubes and for pins.
Guth, Iosevich, Ou and Wang proved the planar distance-set theorem by a
weighted restriction argument, Orponen, Shmerkin and Wang proved Kaufman
and Falconer type estimates for radial projections, Ren and Wang settled
the planar Furstenberg-set conjecture, and Orponen and Shmerkin supplied
the ABC-type sum-product input for Furstenberg sets. These enter as an
incidence count for pins at angular resolution beta, stated in the
projection part below. The Kakeya set conjecture in three dimensions was
settled by Wang and Zahl, whose abstract fixes the input precisely: "We
study sets of delta tubes in R^3, with the property that not too many
tubes can be contained inside a common convex set V." The estimate here
uses the same toolkit and does not reprove that theorem.
pdf/src-1808.09346.pdf
pdf/src-2209.00348.pdf
pdf/src-2308.08819.pdf
pdf/src-2301.10199.pdf
pdf/src-2502.17655.pdf
Two technical debts are discharged in the body: the entropy bookkeeping
of the calibrated profile, which follows the local entropy averaging
method of Hochman and Shmerkin and the inverse entropy theorems of
Hochman, and the almost-everywhere differentiability used to realize
tangents, which is the Rademacher theorem for Lipschitz functions on
Euclidean domains.
Two mechanisms meet in one weight. For a centered ellipse E with
semiaxes L >= a > 0 put
W(E) = L^(1+kappa) a^(1-kappa), 0 < kappa < 1/10. (W)
This overweights narrow ellipses, and it is the exact quantity that both
the incidence count and the packet count produce. The geometric
mechanism propagates capacity through entropy profiles and excludes
every aspect velocity of a calibrated path by a backward comparison,
producing a ball exponent q_* >= 1 + 2s - 10 kappa. The packet mechanism
propagates finite packet arrays through refined decoupling and a
fourth-moment estimate, and pins the amplification by two extrema and a
cancellation, producing gamma <= 10 kappa. Each leaves the margin 10
kappa, and kappa is arbitrary in (0, 1/10). The global assembly converts
the ball exponent into (*) by a layer cake and an L^4 bound.
Rotation and a partition of the sphere reduce (*) to finitely many
graph patches. On one patch, after a smooth cutoff and after absorbing
the area density into the data,
F_0(x, t) = int f(xi) chi(xi) exp(2 pi i (x . xi + t phi(xi))) dxi,
(Graph)
(x, t) in R^2 x R,
with chi a smooth nonnegative window equal to one on a slightly smaller
patch and f bounded and measurable, with no differentiability assumed.
The phase phi is smooth on a fixed compact frequency patch, its Hessian
is positive definite with bounds above and below, and the map xi |->
-grad phi(xi) is therefore a bi-Lipschitz parametrization of the
velocity region. A partition into finitely many such patches, the
triangle inequality in L^p, and the invariance of Lebesgue measure under
rotations leave only || F_0 ||_p <= C_p || f ||_inf to be proved, with a
constant depending on the finitely many fixed choices.
Lines, capacities and the input hypothesis. Write a line of R^2 in
parametric form X_v = V + vU with U the velocity and V the initial
position, and index the pair as P = (U, V). For a centered ellipse E
with semiaxes L >= a > 0, with W(E) as in (W), a capacity test at a
time v compares two containments with the same shape and independent
centers: the velocity falls in a translate of E and the position X_v
falls in a translate of E. The input hypothesis is
Pr(U in u + E, V in v + E) <= K W(E) (H)
simultaneously for all centers u, v, for all ellipses E with both
semiaxes in a fixed width range, and with K independent of E and of the
centers. The same inequality with the time index dropped is imposed on
the source positions. No independence between U and V is assumed; (H)
is a quantitative bound on joint concentration, and it is scale free
apart from the width range.
A sublaw of a probability measure mu is a finite measure nu <= mu. Along
a sequence of bases M -> infinity, a sublaw has large mass when its
total mass is M^(-o(1)); normalizing such a sublaw changes every upper
bound by at most M^(o(1)). For an exponent q, the propagation assertion
is the existence, for each instance satisfying (H), of a large sublaw on
which, simultaneously for all times v in [0,1), all centers and all
ellipses in an admissible width range,
Pr(T in I, U in u + E, X_v in w + M^(-1)E)
<= M^(-q + o(1)) K W(E). (Ball)
The time index I is the dyadic interval of the random time variable at
depth determined by v, and the factor M^(-1) in the position test is the
same eccentricity normalization that appears in the sharpness
computation: a single ball at scale M^(-1) in the position variable
carries capacity comparable to W(E) M^(-1). The common ellipse
orientation in (Ball) imposes nothing on the law of the two spatial
variables, and the two centers u, w are independent.
The exponent is the infimum over universal admissible values. Let
q(eps, D) be the supremum of the exponents q for which (Ball) holds
universally along sequences satisfying (H) and the width restriction
M^(-D) <= a <= L <= 1 with a time error at most eps, and put
q_* = lim over eps decreasing to 0 of inf over finite D of q(eps, D).
The limit exists by monotonicity in eps, and bounded changes in the
coordinate ranges are absorbed by covering and isotropic normalization.
The whole geometric half of the argument is the following statement.
For 0 < s <= 1 and 0 < kappa < 1/10,
q_* >= 1 + 2s - 10 kappa. (Geo)
The parameter s measures the sparsity in time: conditioning on the
whole velocity information and on a dyadic prefix of the time variable
of depth j, at most M^(eps) 2^(s j) prefixes carry mass, and no single
prefix carries more than M^(eps) 2^(-s j). The extremal families
constructed from packets in the second half have this property with
base M = m^2, and (Geo) is what their ball exponent must satisfy. The
proof of (Geo) occupies the geometric part; the packet part proves that
such families cannot beat 10 kappa.
Two elementary consequences of (Ball) fix the normalization. First, a
terminal test with time index dropped implies V = X_v - vU in a
translate of O(E), since 0 <= v < 1, so the mass is at most a constant
times K W(E); hence q_* >= 0 together with the trivial bound. Second,
isotropic shrinking is free: if both spatial variables are shrunk by a
fixed power B^(-1), then the transformed constant is K B^2 while
W(E/B) = B^(-2) W(E), and the product, hence the failing comparison, is
unchanged. This is the mechanism that moves a test inside the interior
of the admissible width range without changing tightness, and it is used
whenever a band endpoint would otherwise be unavailable.
Mesh epochs, assigned boxes, heavy boxes. Fix a base M = 2^n with n
large and a small positive eta. A mesh epoch is a dyadic time depth y at
which the time variable is truncated to intervals of length
M^(-y) = 2^(-floor(yn)). A box of shape (L, a) at an epoch is a
rectangle of side lengths L and a whose long axis is an oriented line in
the velocity plane; the associated ellipse has the same semiaxes, and
all bounds below are unchanged under replacing one by the other at the
cost of an absolute constant, since a rectangle is covered by boundedly
many translates of a concentric ellipse and conversely.
Fix a time bin I. A box C is heavy at tolerance e when the sublaw
carries mass at least
mass(C) >= H W(E) M^(-e), (Floor)
where E is the ellipse with the semiaxes of C and H is a fixed positive
power of M that serves as the baseline. To each particle z one assigns
one heavy box C(z) containing it, chosen by a measurable selection; the
assignment is part of the data. The first geometric statement compares
the assigned direction with a fixed box.
Slant bound. Let R be a rectangle of sides L and a with L >= a > 0, and
let R' be a copy rotated by an angle phi in [a/L, 1]. Then
area(R intersect R') <= 4 L a,
and R intersect R' lies in a rectangle with semiaxes
L/2 and a^2/(L phi). (Slab)
The first assertion is clear. For the second, the short direction of R'
cuts R in a strip of width 2 a/sin(phi) measured along that direction,
and the long direction of R' keeps the long coordinate bounded by L;
the intersection of the two slabs therefore has short extent at most
2 a^2/(L phi) after intersecting with R, using sin(phi) >= phi/2 and
a <= L phi. The displayed semiaxes follow after recentering.
Consequence for the weight. Suppose a <= L phi. Then
W(L, a^2/(L phi)) = L^(1+kappa) (a^2/(L phi))^(1-kappa)
= W(L, a) (a/(L phi))^(1-kappa) <= W(L, a), (Slant)
because a/(L phi) <= 1. A slant test is thus strictly stronger than the
axis test by the power in (Slant), and the eccentricity bias in (W) is
exactly what makes the two weights comparable up to that power.
The slant inequality. Let theta in [a/L, 1] and let S be a set of
particles in a fixed box C_0 of shape (L, a) such that every z in S has
an assigned heavy box C(z) whose long axis makes an angle at least theta
with that of C_0. Then
mu(S) <= H W(E) (a/(L theta))^(2(1-kappa)) M^(2e+o(1)). (Brush)
Since a/(L theta) <= 1, this implies the weaker bound with the exponent
2 kappa, which is the form consumed later; the display (Brush) is what
the two-slab computation produces.
Proof. Replace every ellipse by its rectangle. Consider pairs (z, z')
of particles of S. The velocity difference dU = U - U' lies in
C_0 - C_0, a rectangle of sides 2L and 2a. The same difference lies in
C(z) - C(z') up to a translation, and since the assigned axis is at
angle at least theta from that of C_0, the component of dU along the
short direction of C(z) is at most 2a. Apply (Slab) with phi = theta:
the set of admissible velocity differences is contained in a rectangle
with semiaxes L and a^2/(L theta), and (Slant) bounds its capacity by
W(E) (a/(L theta))^(1-kappa).
The same argument is then run in the position variable, at the lag y
chosen so that M^y a = L theta, that is with M^y comparable to
L theta / a. At that lag the short semiaxis of the assigned box in the
lagged position variable equals L theta, so the slant angle theta has
the same size in the lagged variable as in the velocity variable, and
the intersection computation (Slab) applies verbatim. The lagged
position differences dX = M^y (X_v - X'_v) therefore satisfy the same
bound, with the constant absorbed into M^(o(1)) since M^y is a power of
M fixed by L, a, theta.
Apply (H) to the two coordinate tests separately. Conditioning on the
velocity test and using the chain rule for entropy, the joint mass of
the pair set is at most the product of the two upper bounds, and the
offending configurations, namely those already accounted for by the
coarser axis tests at angle less than theta, are counted first at the
smaller angles. Summing over the dyadic shells phi in [theta, 2 theta]
and using
sum over dyadic phi >= theta of (a/(L phi))^(1-kappa)
<= C_kappa (a/(L theta))^(1-kappa),
the two contributions combine into the square displayed in (Brush). The
factor M^(2e) records the two mass floors of the assigned heavy boxes,
and M^(o(1)) the angular and dyadic rounding. When a/(L theta) is
bounded below by a fixed positive constant, both sides are comparable
to H W(E) and the trivial bound mu(S) <= mu(C_0) is no worse, so
(Brush) never contradicts the trivial bound in that regime.
Heavy boxes are few. With the same notation, an orientation grid of
spacing comparable to rho = a/L, and the family H of heavy boxes at
tolerance e over all time bins,
card H <= (H W(E))^(-1) M^(O_kappa(e) + o(1)). (List)
Proof. Put A = H W(E), c_kappa = 4/kappa, and choose tau_M -> 0 slowly
enough that every bounded and logarithmic factor is M^(o(tau_M)) while
tau_M log M -> infinity. Set Theta_0 = rho M^(c_kappa (e + tau_M)).
Remove from each C in H the particles belonging to another heavy box
whose axis differs from that of C by more than Theta_0. Summing
(Brush) over angular shells bounds the deleted mass by
A M^(2e - 2 kappa c_kappa (e + tau_M) + o(tau_M)) = o(A M^(-e)),
because 2 kappa c_kappa = 8 exceeds 3. Every heavy box therefore
retains a set G_C of mass at least A M^(-e)/2. A particle of G_C
belongs only to heavy boxes whose long axis lies within Theta_0 of the
axis of C, hence to at most
1 + Theta_0/rho = M^(c_kappa (e + tau_M) + o(tau_M))
grid boxes. Summing the masses of the G_C against the total mass bound
one gives
card H <= (A M^(-e)/2)^(-1) M^(c_kappa (e+tau_M) + o(tau_M)),
which is (List) after absorbing tau_M into o(1).
Inheritance across dyadic dilations. Let Q = M^d range over a fixed
finite set of dyadic dilations. If a particle z lies in M^d D for a
heavy box D, then the axis assigned to z at the base scale and the axis
of D agree up to the enlarged resolution
angle(C(z), D) <= (a/L) M^(C_kappa (e + d) + o(1)),
C_kappa = (c_kappa + 10)/(2 kappa). (Axis)
Otherwise (Brush) at the enlarged scale, with baseline A_Q = Q^2 A
since W dilates quadratically, mass floor A_Q M^(-(e+2d)), and angular
cutoff Theta_Q = Q rho M^(c_kappa (e+d+tau_M)), would be violated on a
set of deleted mass at most
M^(-7e - (c_kappa+4) d - 10 tau_M + o(tau_M)),
which is power small. The factor Q^2 is the homogeneity of W under
dilation by Q, and the angular grid of spacing rho produces the leading
a/L. Inequality (Axis) is what allows a test at a coarse scale to
identify the axis attached to a particle at a fine scale, at the price
of the exponent C_kappa (e + d).
Backward width band. Suppose that at a mesh epoch y the current large
sublaw has assigned boxes of widths L, a and satisfies the obstruction
just recorded. Let y' < y be the previous mesh epoch. After a further
large restriction the particles have assigned high prelaw boxes at y',
with a common pair of dyadic widths both belonging to
[a M^(-(y-y')), L], (Band)
up to bounded endpoint rounding, and the tolerance e_{y'} may be chosen
to tend to zero sufficiently slowly relative to e_y and to the
propagation error.
Proof. Call a prelaw box at epoch y', with both widths in (Band), high
when its mass is at least
K W(F) M^(-q_* y' - e_{y'}). (Floor')
If a large sublaw consisted of particles belonging to no such high box,
then every test in the band would have the improved cap with factor
M^(-e_{y'}), because a test above that threshold is itself a high prelaw
box containing those particles. Finite grids and bounded coverings make
the statement simultaneous over centers, orientations and band widths.
Partition that contrary sublaw by the prefix v' = T_{y'} and by
isotropic parent cells of width L for the velocity variable and h' L for
the position variable X_{v'}, where h' = 2^(-floor(y'n)). Let a parent
have centers u_Q, x_Q and mass m_Q, and normalize by
U' = (U - u_Q)/L, V' = (X_{v'} - x_Q)/(h' L), S = (T - v')/h'.
The residual time S lies in [0,1), and its terminal partition has base
M_res = 2^(floor(yn) - floor(y'n)) = M^(y - y' + o(1)). (Res)
Let s_0 be the left endpoint of a tested residual interval of length
M_res^(-1). On that event v = v' + h' s_0 = T_y, and the position
relation becomes
V' + s_0 U' = (X_v - x_Q - (v - v') u_Q)/(h' L). (Nor)
Both translations in (Nor) are fixed by the parent and by this residual
prefix, so the normalized line family satisfies the same two-coordinate
container test as the original, with the same constant and, by (Res),
with a position scale that agrees with the original one up to the
declared dyadic rounding:
h'/M_res = 2^(-floor(yn)) = M^(-y + o(1)). (Scl)
Apply the propagation assertion in this normalized parent, with base
M_res and with the residual time variable taking the role of the time
variable. The residual interval under test has length M_res^(-1), and
its propagation requires an input width that is a fixed multiple of
M_res^(-1) a/L. By (Band) and (Scl) this coincides with the lower
endpoint of the band up to a bounded factor, so the band cap supplies
the normalized input bound with constant
K L^2 M^(-q_* y' - e_{y'})/m_Q,
up to subpower factors, the factor L^2 being W(L F) = L^2 W(F). The
parent time profile is supplied by the time-sparsity hypothesis at the
appropriate level, and its complexity is bounded for the fixed mesh and
fixed y', so the hypotheses of the propagation assertion hold with
parameters fixed in advance.
Recombination. Multiplication by the original parent mass cancels
m_Q^(-1), and multiplication by W(E/L) cancels L^2, so the resulting
sublaw of the parent satisfies the original terminal test with bound
K W(E) M^(-q_* y - e_y + o(1)), (Recomb)
where the exponent e_y absorbs e_{y'} together with the propagation
error. Recombination costs only a bounded factor. A terminal bin
determines its prefix, and X_{v'} = X_v + (v' - v) U with
|v' - v| <= M^(-y'). As U varies in a terminal ellipse of long radius L,
the rebased position varies by O(M^(-y') L), so the terminal cell meets
boundedly many isotropic parent cells in both variables. The surviving
mass is thus a fixed fraction of the parent mass, which is what makes the
contrary sublaw fail and proves the band statement with the improved
cap.