Mixed-norm Brasamp-Lieb inequalities

Abstract: Christ [Chr01] looked at a certain boundedness problem for trilinear operators. In this paper we set up a framework of mixed-norm Brascamp-Lieb inequalities and identify a new interesting regime not covered by the study of classical Brascamp-Lieb. Positive results in [Chr01] can be viewed as the first nontrivial progress in this new regime. We will also prove another mixed-norm Brascamp-Lieb inequality, showcasing how one can use a tensor product trick to slightly sharpen Christ's argument and also obtain the endpoint case. We then discuss the connections between mixed-norm Brascamp-Lieb and the Kakeya Conjectures, as well as unique new difficulties for these mixed-norm Brascamp-Lieb inequalities compared to the classical setting.

• The paper develops a tensor-product transference principle that converts uniform discrete restricted weak-type estimates into sharp continuous endpoint mixed-norm inequalities.
• It proves a quadrilinear inequality at exponents $(p_1,p_2,p_3,p_4)=(2/7,2/7,2/7,1/7)$ with mixed norm $L_x^{7/4}L_y^{7/3}$, extending known results into the genuinely new $q<2$ regime.
• The results show that mixed-norm inequalities are sensitive to rational projection directions, lack perturbative stability, and could imply Hausdorff Kakeya dimension bounds approaching the conjectured optimal value.

This paper by Ruixiang Zhang develops a framework of mixed-norm Brascamp–Lieb inequalities in $\mathbb{R}^2$, identifies a regime genuinely outside the scope of the classical theory, proves a new endpoint quadrilinear inequality, and establishes a formal bridge from these inequalities to the Kakeya Conjectures. The central objects are inequalities of the form

$$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$

where $B_1,\dots,B_m$ are orthogonal projections onto one-dimensional subspaces $H_1,\dots,H_m$ of $\mathbb{R}^2$. After normalization $\frac{1}{q}+\frac{1}{r}=1$ and $\sum_j p_j = 1$, the case $q \ge 2$ follows from classical Brascamp–Lieb via mixed-norm Hölder; the case $q < 2$ is where genuinely new phenomena appear.

The $q<2$ regime and its benchmarks

The paper first records structural facts that distinguish the mixed-norm setting. A necessary condition (Lemma 2 in the paper's numbering) shows that if \eqref{BLmix} holds for some $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$0, then necessarily $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$1: testing with sums of $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$2 characteristic functions forces $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$3, which fails for $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$4 when $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$5. This contrasts sharply with the classical inequality, where in dimension two with all fibers one-dimensional, nothing is gained by taking $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$6—the $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$7 case plus Hölder suffices. In the mixed-norm setting, even in dimension two, one must go beyond the bilinear case to find any nontrivial inequality.

For $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$8 with equal exponents $$\Big\| \prod_{j=1}^m f_j(B_j(x,y))^{p_j} \Big\|_{L_x^q L_y^r} \le C \prod_{j=1}^m \|f_j\|_{L^1}^{p_j},$$9, Christ [christ2001certain] established a dichotomy: if $B_1,\dots,B_m$0 are disjoint rational subspaces not equal to the $B_1,\dots,B_m$1-axis, the inequality holds for some $B_1,\dots,B_m$2; if two subspaces are rational and one is irrational, it fails for every $B_1,\dots,B_m$3. This dependence on rationality of the projections has no analogue in the classical characterization of Bennett–Carbery–Christ–Tao [bennett2008brascamp], which depends only on scaling and dimension conditions.

A transference principle via tensor products

The paper's main technical contribution is an upgrade theorem: restricted weak type estimates on discrete groups imply honest endpoint mixed-norm inequalities. Specifically, suppose all projections are rational, of the form $B_1,\dots,B_m$4 with coprime integers. If, for every torsion-free abelian group $B_1,\dots,B_m$5, finite sets $B_1,\dots,B_m$6 and a set $B_1,\dots,B_m$7 whose elements satisfy $B_1,\dots,B_m$8 and whose fibers over the first coordinate have size at most $B_1,\dots,B_m$9 obey

$H_1,\dots,H_m$0

with uniform constant, then the full continuous mixed-norm inequality holds.

The proof proceeds in two steps. Dyadic pigeonholing of each $H_1,\dots,H_m$1 and of the inner sum reduces to the restricted weak type estimate, but incurs a logarithmic loss $H_1,\dots,H_m$2. The loss is removed by a tensor product trick: applying the logarithmic estimate to the $H_1,\dots,H_m$3-fold product functions on $H_1,\dots,H_m$4 yields an $H_1,\dots,H_m$5-th power of the desired left-hand side against only a linear-in-$H_1,\dots,H_m$6 logarithmic factor, and letting $H_1,\dots,H_m$7 eliminates the loss entirely. Transference from $H_1,\dots,H_m$8 to $H_1,\dots,H_m$9 then follows by approximation and Fatou's lemma. The consequence is that one may work entirely with combinatorial, restricted weak type estimates—a substantially easier class of problems—and still obtain sharp endpoint inequalities.

An endpoint quadrilinear inequality

As an application, the paper proves the quadrilinear estimate

$\mathbb{R}^2$0

with exponents $\mathbb{R}^2$1 matching the norm pair $\mathbb{R}^2$2, so that $\mathbb{R}^2$3. The key input is a sharpened version of a Katz–Tao arithmetic projection bound [katz1999bounds]: for an abelian group without order-$\mathbb{R}^2$4 elements, if $\mathbb{R}^2$5 satisfies $\mathbb{R}^2$6, $\mathbb{R}^2$7, $\mathbb{R}^2$8, $\mathbb{R}^2$9, and fibers over $\frac{1}{q}+\frac{1}{r}=1$0 have size at most $\frac{1}{q}+\frac{1}{r}=1$1, then

$\frac{1}{q}+\frac{1}{r}=1$2

The proof combines counting of coincidences among quadruples of points with two algebraic claims showing that $\frac{1}{q}+\frac{1}{r}=1$3 is determined by three other projections, bounding tuples by $\frac{1}{q}+\frac{1}{r}=1$4. The paper notes this lemma already appears implicitly in Tao's entropy estimates [Taoentropy] and in related work of Pohoata–Zakharov [pohoata2024generalized]; its contribution here is recognizing that the transference principle upgrades it to the endpoint continuous inequality, whereas Christ's interpolation-based method covers only an open range of exponents and misses endpoints.

Connection to Kakeya

The paper proves that progress on \eqref{BLmix} toward $\frac{1}{q}+\frac{1}{r}=1$5 would resolve the Hausdorff Kakeya Conjecture in all dimensions. Precisely: if \eqref{BLmix} holds for rational projections with some $\frac{1}{q}+\frac{1}{r}=1$6, then every Kakeya set in $\frac{1}{q}+\frac{1}{r}=1$7 has Hausdorff dimension at least $\frac{1}{q}+\frac{1}{r}=1$8. The mechanism is the "sum-difference" property $\frac{1}{q}+\frac{1}{r}=1$9 of Katz–Tao [katz1999bounds, katz2002new]: the inequality implies $\sum_j p_j = 1$0 for $\sum_j p_j = 1$1, which transfers to $\sum_j p_j = 1$2 via embeddings preserving finitely many linear relations, and $\sum_j p_j = 1$3 for $\sum_j p_j = 1$4 bounds Kakeya dimensions. While the Minkowski-dimension version of this implication was known, the Hausdorff version required new work: the paper shows in an appendix that Bourgain's argument [bourgain1999dimension], using recent quantitative Szemerédi bounds of Leng–Sah–Sawhney [leng2024improved] in place of Heath-Brown–Szemerédi, yields the Hausdorff conclusion. This motivates the paper's conjecture that for every $\sum_j p_j = 1$5 there exist $\sum_j p_j = 1$6, $\sum_j p_j = 1$7, and rational projections with $\sum_j p_j = 1$8 satisfying \eqref{BLmix}; by the theorem above, this conjecture implies the Hausdorff Kakeya Conjecture.

Two caveats temper this program. First, Katz [katz2006elementary] showed that elementary graph-theoretic methods cannot prove $\sum_j p_j = 1$9 for $q \ge 2$0, ruling out the Katz–Tao approach used here for the strongest targets. Second, the paper leaves open whether \eqref{BLmix} for $q \ge 2$1 has Kakeya implications when the projections cannot be simultaneously made rational under affine transformations—the irrational case behaves differently, as Christ's dichotomy already indicates.

Failure of perturbed versions

A further structural obstruction is established: no perturbed version of \eqref{BLmix} holds when $q \ge 2$2. In the classical setting, the inequality is stable under small perturbations of projection directions, both at the level of constants [bennett2018stability] and in local forms underlying the multilinear Kakeya inequality [bennett2006multilinear, guth2010endpoint]. The paper constructs counterexamples showing that the analogous statement for the mixed-norm inequality fails unconditionally.

The construction uses Dirichlet approximation: choosing an irrational $q \ge 2$3 and coprime $q \ge 2$4 with $q \ge 2$5, one builds $q \ge 2$6 tubes in each of $q \ge 2$7 families, each family covering a grid of $q \ge 2$8 unit balls, with tube directions within $q \ge 2$9 of prescribed directions $q < 2$0. The left-hand side of the would-be perturbed inequality is then $q < 2$1 while the right-hand side is $q < 2$2, contradicting $q < 2$3 as $q < 2$4. Consequently, the three principal techniques for proving perturbed classical inequalities—heat flow [bennett2006multilinear], polynomial partitioning [guth2010endpoint], and multiscale analysis [guth2015short]—cannot be imported wholesale into the mixed-norm setting. The author notes that adapting the heat flow extrapolation seems partially viable but encounters an obstructing term, while no adaptation is apparent for the other two methods.

Limitations and open questions

Several limitations are explicit. The transference principle requires rational projections; the behavior for genuinely irrational projections remains open, including whether such inequalities have any Kakeya content. The main theorem is proved for one specific exponent configuration chosen for technical simplicity, and the author states that further results from [katz1999bounds, katz2002new] could likely be upgraded similarly, though this is not carried out. The conjecture driving the Kakeya connection—for arbitrary $q < 2$5, existence of \eqref{BLmix} with $q < 2$6—remains open, and the Katz barrier at $q < 2$7 constrains available methods below that threshold. Finally, the absence of perturbed versions means the powerful stability machinery of the classical theory is unavailable, and the paper does not identify a replacement technique.

Conclusion

The paper reframes Christ's trilinear analysis as part of a broader theory of mixed-norm Brascamp–Lieb inequalities, isolates $q < 2$8 as a regime with genuinely different structure—requiring $q < 2$9, sensitive to rationality of projections, and admitting no perturbed stability—and supplies a transference principle converting uniform discrete restricted weak type estimates into endpoint continuous inequalities. Its concrete output is the endpoint quadrilinear estimate with exponents $q<2$0 at $q<2$1, and its conceptual output is the equivalence, mediated by sum-difference properties and modern quantitative Szemerédi bounds, between approaching $q<2$2 in these inequalities and proving the Hausdorff Kakeya Conjecture.