Yufei Zhao's blog
https://yufeizhao.com/
Wed, 24 Mar 2021 00:08:04 -0400Wed, 24 Mar 2021 00:08:04 -0400The cylindrical width of transitive sets<p>The high dimensional sphere comes many surprising and counterintuitive features. Here is one of them.</p>
<p>We say that a subset $X$ of the unit sphere in $\mathbb{R}^d$ is <em>transitive</em> if, informally, one cannot distinguish one point of $X$ from another if we are allowed to rotate/reflect.
Formally, we say that $X$ is transitive if the group of isometries of $X$ acts transitively on it. This is equivalent to saying that $X$ is the orbit of some point under the action of some subgroup of the orthogonal group.</p>
<p>Perhaps you, like me, can only visualize in three-dimensions. Here are some examples:</p>
<ul>
<li>Every regular polytope (platonic solid) is transitive.</li>
<li>Start with an arbitrary vector, and consider all points obtained by permuting its coordinates, as well as switching the signs of some of the coordinates. The resulting set is transitive (this is a generalization of the <a href="https://en.wikipedia.org/wiki/Permutohedron">permutahedron</a>.) A specific example is given below.</li>
<li>The whole sphere itself is transitive</li>
</ul>
<p>The finite examples look like they can be fairly spread out on the sphere.
However, this intuition turns out to be completely wrong in high dimensions.
All <em>finite</em> transitive sets in high dimensions have small width, meaning that they lie close to some hyperplane (in sharp contrast to the whole unit sphere, which has width 2).</p>
<p>I had conjectured the following statement, which was proved by Ben Green.</p>
<p><strong>Theorem</strong> (<a href="https://dev.arxiv.org/abs/1802.01904">Green</a>). Every finite transitive set of unit vectors in $\mathbb{R}^d$ lies within distance $O\bigl(\frac{1}{\sqrt{\log d}}\bigr)$ of some hyperplane.</p>
<figure>
<img src="/blog/images/transitive-set-width.png" width="500" />
<figcaption>A finite transitive set lies close to a hyperplane</figcaption>
</figure>
<p>This bound is actually tight for the example obtained by taking all signings and permutations of the unit vectors</p>
\[\frac{1}{\sqrt{H_d}}\Biggl(\pm 1, \frac{\pm 1}{\sqrt{2}}, \frac{\pm 1}{\sqrt{3}}, \ldots, \frac{\pm 1}{\sqrt{d}}\Biggr), \quad \text{where } H_d = 1 + \frac{1}{2} + \cdots + \frac{1}{d} \sim \log d\]
<p>Green’s proof, among other things, uses the classification of finite simple groups.</p>
<h3 id="our-new-results-finite-transitive-sets-have-small-cylindrical-width">Our new results: finite transitive sets have small cylindrical width</h3>
<p>In our new paper, <a href="https://arxiv.org/abs/2101.11207">The cylindrical width of transitive sets</a>, coauthored with Ashwin Sah and Mehtaab Sawhney, we generalize Green’s result to show that not only do transitive sets lie close to a hyperplane, but they also lie close to some codimension-$k$ subspace for every $k$ that is not too large.</p>
<p><strong>Theorem</strong> (<a href="https://arxiv.org/abs/2101.11207">Sah–Sawhney–Zhao</a>). There exists some constant $C$ so that for every $1 \le k \le d/(\log d)^C$, every finite transitive set of unit vectors in $\mathbb{R}^d$ lies within distance $O\bigl(\frac{1}{\sqrt{\log (d/k)}}\bigr)$ of codimension-$k$ subspace.</p>
<p>The bound $O\bigl(\frac{1}{\sqrt{\log (d/k)}}\bigr)$ is tight up to a constant factor. We conjecture that the hypothesis $1 \le k \le d/(\log d)^C$ can be removed:</p>
<p><strong>Conjecture.</strong> For every $1 \le k \le d$, every finite transitive set of unit vectors in $\mathbb{R}^d$ lies within distance $O\bigl(\frac{1}{\sqrt{\log (2d/k)}}\bigr)$ of codimension-$k$ subspace.</p>
<p>Finally, we conjecture that not only do finite transitive sets lie close to subspaces, but they can be completely contained in a small cube. All examples known to us are consistent with this conjecture, even though I still find it very counterintuitive.</p>
<p><strong>Conjecture</strong> (Finite transitive sets lie in a small cube). Every finite transitive set of unit vectors in $\mathbb{R}^d$ is contained in a cube of width $O\bigl(\frac{1}{\sqrt{\log d}}\bigr)$.</p>
<p>Even proving a bound of $o(1)$ (as $d\to \infty$) would be quite interesting.</p>
<p>Both conjectures are generalizations of Green’s result, but the relationship between the two conjectures is unclear.</p>
Thu, 28 Jan 2021 00:00:00 -0500
https://yufeizhao.com/blog/2021/01/28/cylindrical-width-transitive-set/
https://yufeizhao.com/blog/2021/01/28/cylindrical-width-transitive-set/blogAshwin Sah and Mehtaab Sawhney win the Morgan Prize<figure>
<img src="/blog/images/ashwin-mehtaab-2020.jpg" height="200" />
<figcaption>Ashwin Sah and Mehtaab Sawhney</figcaption>
</figure>
<p>Congratulations to <a href="http://www.mit.edu/~asah/">Ashwin Sah</a> and <a href="http://www.mit.edu/~msawhney/">Mehtaab Sawhney</a> for winning the <a href="http://www.ams.org/news?news_id=6435">2021 AMS-MAA-SIAM Frank and Brennie Morgan Prize</a> for Outstanding Research in Mathematics by an Undergraduate Student!
This is the most prestigious prize in the US for undergraduate mathematics research (though I must say that their work far surpasses that of “undergraduate research”).
It is also the first time in the <a href="http://www.ams.org/prizes-awards/pabrowse.cgi?parent_id=19">history of the Morgan Prize</a> for it to be jointly awarded to more than one recipient.</p>
<p>From the prize citation:</p>
<blockquote>
<p>The award recognizes the duo’s innovative results across a broad range of topics in combinatorics, discrete geometry, and probability.</p>
</blockquote>
<blockquote>
<p>The prize committee notes, “Working alongside one another, Sah and Sawhney settled longstanding conjectures and improved results by established mathematicians. The research of Sah and Sawhney is both deep and broad, tackling questions at the very forefront of current research, yet extending across the entire gamut of modern combinatorics, with significant contributions to extremal graph theory, graph limits, additive combinatorics, Ramsey theory, algebraic combinatorics, combinatorial geometry, random graphs and random matrix theory. They have demonstrated a significant amount of ingenuity, originality and technical ability, resulting in a research record, which is extremely rare for undergraduate students.”</p>
</blockquote>
<blockquote>
<p>Combined, Sah and Sawhney have authored 30 papers (11 together), and published in top journals, including <em>Inventiones Mathematicae</em>, <em>Advances in Mathematics</em>, <em>Mathematical Proceedings of the Cambridge Philosophical Society</em>, the <em>Journal of Combinatorial Theory Series B</em>, and <em>Combinatorica</em>.</p>
</blockquote>
<p>(The paper count in the award citation was taken from earlier this year and is now already quite a bit out of date.)</p>
<p>I got to know both Ashwin and Mehtaab as students in my <a href="/a34/">Putnam Seminar</a> and <a href="/gtac/">graduate combinatorics class</a> during their very first undergraduate semester at MIT. Since then, they have led an intensively productive and ongoing collaboration, producing an incredible amount of fantastic research results across a wide spectrum of topics in combinatorics, probability, number theory, and algorithms. It was a great pleasure for me to be working with Ashwin and Mehtaab on many of these projects.
Some of their work were previously featured on this blog:</p>
<ul>
<li><a href="/blog/2018/05/12/the-number-of-independent-sets-in-an-irregular-graph/">The number of independents in an irregular graph</a></li>
<li><a href="/blog/2018/09/26/a-reverse-sidorenko-inequality/">A reverse Sidorenko inequality</a></li>
<li><a href="/blog/2020/05/20/ashwin-sahs-new-diagonal-ramsey-number-upper-bound/">New upper bounds on diagonal Ramsey numbers</a></li>
<li><a href="/blog/2020/10/13/jain-sah-sawhney-singularity/">Singularity of discrete random matrices</a></li>
</ul>
<p>Ashwin and Mehtaab are now both first-year PhD students at MIT. I’m sure that we will see a lot more coming from them.</p>
Thu, 29 Oct 2020 00:00:00 -0400
https://yufeizhao.com/blog/2020/10/29/sah-sawhney-morgan-prize/
https://yufeizhao.com/blog/2020/10/29/sah-sawhney-morgan-prize/blogJain–Sah–Sawhney: Singularity of discrete random matrices<p><a href="https://math.mit.edu/~visheshj/">Vishesh Jain</a>, <a href="http://www.mit.edu/~asah/">Ashwin Sah</a>, and <a href="http://www.mit.edu/~msawhney/">Mehtaab Sawhney</a> just posted two amazing papers: <em>Singularity of discrete random matrices</em> <a href="https://arxiv.org/abs/2010.06553">Part I</a> and <a href="https://arxiv.org/abs/2010.06554">Part II</a>.</p>
<figure>
<img src="/blog/images/vishesh-jain-2020.jpg" height="200" />
<img src="/blog/images/ashwin-sah-2020.jpg" height="200" />
<img src="/blog/images/mehtaab-sawhney-2020.jpg" height="200" />
<figcaption>Vishesh Jain, Ashwin Sah, and Mehtaab Sawhney</figcaption>
</figure>
<p><strong>The singularity problem for random matrices:</strong> what is the probability that a random $n\times n$ matrix with independent 0/1 entries is singular?</p>
<p>The first non-trivial bound on this problem was by Komlós in 1967. After a long series of works by many mathematicians including Kahn, Komlós, Szemerédi, Tao, Vu, Bourgain, Wood, Rudelson, and Vershynin, a breakthrough was recently achieved by <a href="https://annals.math.princeton.edu/2020/191-2/p06">Tikhomirov</a>, who showed that, when the matrix entries are iid uniform from \(\{0,1\}\), the probability of singularity if $(1/2 + o(1))^n$, which is tight since with probability $2^{-n}$ the first row is zero.</p>
<p>The new papers of Jain–Sah–Sawhney consider a natural extension of the singularity problem where each entry of the $n\times n$ matrix is iid: 1 with probability $p$ and 0 with probability $1-p$, for some value of $p$ other than 1/2.</p>
<p>A folklore conjecture in this area states that the main reason for the singularity of such random matrices is the presence of two equal rows/columns or some zero row/column:</p>
<p><strong>Conjecture.</strong> For a fixed $0 < p < 1$, the probability that a random $n\times n$ matrix with iid Bernoulli(<em>p</em>) entries is singular is $1+o(1)$ times the probability that the matrix has one of the following: (a) two equal rows, (b) two equal columns, (c) a zero row, (d) a zero column.</p>
<p>Note that even Tikhomirov’s breakthrough result does not give precise enough results to answer this conjecture. However, we now have the following incredible result.</p>
<p><strong>Theorem.</strong> (Jain–Sah–Sawhney) The above conjecture is true for all $p \ne 1/2$.</p>
<p>(The case $p = 1/2$ remains open.)</p>
<p>In fact, they prove something much more general: namely that the above result remains true if the random matrix entries are iid discrete random variables that are <em>non-uniform on its support</em>. Furthermore, they prove nearly tight bounds on the probability of these matrices having small least singular values. Very impressive!</p>
Tue, 13 Oct 2020 00:00:00 -0400
https://yufeizhao.com/blog/2020/10/13/jain-sah-sawhney-singularity/
https://yufeizhao.com/blog/2020/10/13/jain-sah-sawhney-singularity/blogGunby: Upper tails for random regular graphs<p><a href="https://sites.google.com/view/benjamingunby/home">Ben Gunby</a>, a final-year PhD student at Harvard whom I advise, just posted his paper <a href="https://arxiv.org/abs/2010.00658"><em>Upper tails of subgraph counts in sparse regular graphs</em></a>. This is a substantial paper that unveils surprising new phenomena for the problem of large deviations in random graphs.</p>
<figure>
<img src="/blog/images/ben-gunby-2020.jpg" width="300" />
<figcaption>Ben Gunby</figcaption>
</figure>
<p><strong>Problem.</strong> What is the probability that a random graph has far more copies of some given subgraph than expected?</p>
<p>This is called the “upper tail” problem for random graphs.
It is a problem of central interest in probabilistic combinatorics.
There was a lot of work on this problem for the classic Erdős–Rényi random graph $G(n,p)$ (with edges appearing independently at random), with a lot of exciting recent developments.
Ben considers the upper tail problem for <em>random regular graphs</em>, another natural random graph model.
Random regular graphs are usually harder to analyze since the random edge appearance are no longer independent, and this increased difficulty also shows up for the upper tail problem.</p>
<p>Let me highlight one specific result in Ben’s paper (Theorem 2.9).</p>
<p><strong>Theorem.</strong> (<a href="https://arxiv.org/abs/2010.00658">Gunby</a>) <em>Let \(G_n^d\) be a random $n$-vertex $d$-regular graph, where $d = d(n)$ depends on $n$. Assume that $dn$ is even and and \(n^{14/15}\log n \le d = o(n)\). Let $X$ denote the number of copies of</em>
<img src="/blog/images/k24plusedge.svg" width="80" />
<em>in \(G_n^d\). Then for every fixed \(\delta > 0\) (here \(\sim\) means up to $1+o(1)$ factor),</em></p>
\[-\log \mathrm{Pr}(X \ge (1+\delta)\mathbb{E} X) \sim \frac{(18\delta)^{1/3}}{2} \frac{d^3}{n} \left(\log \frac{n}{d}\right)^{2/3} \left(\log\log \frac{n}{d}\right)^{1/3}\]
<p><br /></p>
<p>This result is surprising.
The appearance of the fractional power of the log factor has not been observed in previous works on the upper tail problem.
It is even more impressive that Ben was able to prove such a sharp estimate.</p>
<p>Understanding these large deviation probabilities roughly corresponds to understanding the “dominant reason” for rare events. Roughly speaking, the fractional power of log appears because the dominant reason <strong>is not the presence of some fixed subgraph (e.g., a large clique or hub), unlike all earlier results,</strong> which have asymptotics of the form \(n^2 p^d \log(1/p)\).
Something else more subtle is happening.
While we still do not understand the complete picture, Ben’s result provides strong evidence why the upper tail problem for random regular graphs may be substantially more intricate than that of $G(n,p)$.</p>
<p>In this blog post, I will tell you some history and background of the upper tail problem in random graphs.
I will try to explain why Ben’s result is surprising.
Lastly, I will also mention a recent paper I have with Yang Liu on the <a href="https://arxiv.org/abs/1910.02916">upper tail problem in random hypergraphs</a>, where a different new phenomenon is observed.</p>
<h3 id="chasing-the-infamous-upper-tail">Chasing the infamous upper tail</h3>
<p>Roughly, and intuitively, to gain a precise understanding of the upper tail problem, we are really asking:</p>
<p><em>What are the dominant reasons for the rare event that a random graph $G(n,p)$ has way too many triangles?</em></p>
<p>We can now say confidently the following:</p>
<p><em>It is primarily because the such a rare random graph contains roughly a large clique or a large hub.</em></p>
<p>Here a “hub” is a set of vertices adjacent to all other vertices.
It is not hard to see that planting a large clique or hub in a $G(n,p)$ would automatically generate many additional triangles (these triangles may use also the edges of the random graph).</p>
<p>For example, the following theorem
is the culmination of a long line of work by many researchers.</p>
<p>(We use the following asymptotic notation, which is standard in this line of work: \(f \ll g\) means $f = o(g)$.)</p>
<p><strong>Theorem.</strong> <em>Let $X$ denote the number of triangles in $G(n,p)$. Then, for $n^{-1/2} \ll p \ll 1$ and fixed $\delta > 0$, one has</em></p>
\[-\log \mathrm{Pr}(X \ge (1+\delta) \mathbb{E} X)
\sim \min \left\{\frac{\delta}{3}, \frac{\delta^{2/3}}{2}\right\} n^2p^2\log\frac{1}{p}\]
<p>The problem of estimating the upper tail $\mathrm{Pr}(X \ge (1+\delta) \mathbb{E})$ has a long history.
Janson and Rucinski called this problem the <a href="http://www2.math.uu.se/~svante/papers/sj137.pdf">infamous upper tail</a>, which had resisted a wide range of techniques that researchers have used to tackle this problem.
This upper tail problem was an excellent litmus test in the development of concentration inequalities.</p>
<figure>
<img src="/blog/images/infamous-upper-tail.svg" width="400" />
</figure>
<p>Several significant breakthroughs then occurred on this problem over the past decade.</p>
<p><a href="https://arxiv.org/abs/1005.4471">DeMarco–Kahn</a> and <a href="https://arxiv.org/abs/1003.3498">Chatterjee</a> independently determined the order of magnitude of the log-probability in the above upper tail problem, for triangle counts in a random graph.
Their work closed a log-factor gap left by previous methods developed by <a href="https://doi.org/10.1002/rsa.10113">Kim–Vu</a>.</p>
<p>The next goal was then to pinpoint the tail probability. This led to further developments of powerful tools.</p>
<p>The seminal work of <a href="https://arxiv.org/abs/1008.1946">Chatterjee and Varadhan</a> developed a <strong>large deviation principle</strong> for the upper tail problem in random graphs, thereby reducing the problem to a <strong>variational problem</strong> in the space of graphons, i.e., graph limits. Their work paved the way for much further development, which has two complementary strands:</p>
<p><strong>1. Extending the large deviation framework to sparser graphs.</strong> Chatterjee and Varadhan’s proof uses Szemerédi’s graph regularity lemma, which only allows us to handle dense graphs (here for random graphs $G(n,p)$ <em>dense</em> means constant $p$, whereas <em>sparse</em> means $p = o(1)$). Further work, starting with <a href="https://arxiv.org/abs/1401.3495">Chatterjee–Dembo</a>’s “non-linear large deviations”, as well as further improvements by <a href="https://arxiv.org/abs/1612.04346">Eldan</a>, <a href="https://arxiv.org/abs/1809.11148">Cook–Dembo</a>, <a href="https://arxiv.org/abs/1810.01558">Augeri</a>, <a href="https://arxiv.org/abs/1904.08212">Harel–Mousset–Samotij</a>, and <a href="https://arxiv.org/abs/1912.11410">Basak–Basu</a>, allow us to handle much sparser densities of graphs.</p>
<p>For certain subgraphs, including cliques, we now have a very precise understanding thanks to <a href="https://arxiv.org/abs/1904.08212">Harel–Mousset–Samotij</a>, who gave a short and clever <a href="https://arxiv.org/abs/1801.04584">container</a>-inspired combinatorial argument in the case of triangles.</p>
<p><strong>2. Solving the variational problem.</strong> The above large deviation frameworks reduces the problem of estimating the tail probability to another problem—a variational problem in the space of graphons.</p>
<p><em>Variational problem for upper tails:</em> Among edge-weighted graphs with some prescribed subgraph density, determine the minimum possible relative entropy (a certain non-linear function of the edge-weights)</p>
<p>These variational problems are more difficult than classical problems in extremal graph theory. In fact, despite <a href="https://arxiv.org/abs/1210.7013">some progress</a>, we still do not understand how to solve such problems exactly in the “dense” regime.</p>
<p>For sparse graphs $G(n,p)$ with $p \to 0$, the variational problem was solved asymptotically in the case of clique subgraphs by <a href="https://arxiv.org/abs/1402.6011">Lubetzky–Zhao</a>, and for more general subgraph counts by <a href="https://arxiv.org/abs/1507.04074">Bhattacharya–Ganguly–Lubetzky–Zhao</a>.</p>
<p>The two complementary strands together pinpoint upper tail probabilities. Similar to the case of triangles, the main message is, roughly speaking,</p>
<p><em>The dominant reason for a random graph to have too many copies of some fixed $H$ is that it contains a large clique or a large hub.</em></p>
<p><strong>Theorem.</strong> <em>Let $H$ be a graph with maximum degree \(\Delta\). Let $\delta > 0$.
Let $X$ be the number of copies of $H$ in $G(n,p)$. Provided that $n^{-\alpha_H} \le p = o(1)$ (where $\alpha_H$ is some explicit constant that has improved over time),
one has</em></p>
\[-\log \mathrm{Pr} (X \ge (1+\delta)\mathbb{E} X) \sim c(H, \delta) p^{\Delta} n^2 \log \frac{1}{p},\]
<p><em>where \(c(H,\delta) > 0\) is an explicit constant determined in <a href="https://arxiv.org/abs/1507.04074">Bhattacharya–Ganguly–Lubetzky–Zhao</a>.</em></p>
<p>While this is a pretty satisfactory formula, our understanding is still not complete.</p>
<ol>
<li>
<p>Is the result valid for the “full” range of $p$?</p>
</li>
<li>
<p>What precise statements can you make about the conditioned random graph? (see <a href="https://arxiv.org/abs/1904.08212">Harel–Mousset–Samotij</a>)</p>
</li>
</ol>
<h3 id="gunbys-work-on-random-regular-graphs">Gunby’s work on random regular graphs</h3>
<p>Ben Gunby’s work studies the upper tail problem for random regular graphs:</p>
<p><strong>Problem.</strong> Let $X$ be the number of copies of some given subgraph $H$ in a random $d$-regular graph $G_n^d$. Estimate $-\log \mathrm{Pr}(X \ge (1+\delta)\mathbb{E} X)$, when \(n^{1-c} \le d = o(n)\).</p>
<p>There is a similar dichotomy of steps as above:</p>
<ol>
<li>
<p>Proving a large deviations framework. The upper bound to the probability can be deduced from the corresponding theorems for $G(n,p)$, while the lower bound requires new and lengthy arguments.</p>
</li>
<li>
<p>Solving the variational problem – this is where new difficulties and innovations lie.</p>
</li>
</ol>
<p>A recent paper of <a href="https://arxiv.org/abs/1909.03045">Bhattacharya–Dembo</a> (this is by Sohom Bhattacharya and not my coauthor <a href="http://www-stat.wharton.upenn.edu/~bhaswar/">Bhaswar Bhattacharya</a> mentioned earlier) essentially solved the variational problem when $H$ is a regular graph (Ben had also independently worked out the case when $H$ is a cycle).</p>
<p>Ben Gunby’s new paper solves the variational problem for “most” graphs $H$ (those that satisfy some technical condition), which includes all graph $H$ with average degree greater than 4, as well as pretty much all small graphs.</p>
<p>For most graphs $H$, the upper tail has “expected” behavior, meaning that the dominant reason for seeing many copies of $H$ is, very roughly speaking, the presence of some large planted subgraph. The resulting tail probability formulas also resemble what we have seen earlier:</p>
\[-\log \mathrm{Pr}(X \ge (1+\delta) \mathbb{E} X) \sim \rho(H, \delta) n^2 (d/n)^{2+\gamma(H)} \log(n/d).\]
<p>for some graph parameter \(\gamma(H)\), whose role here is far from obvious.</p>
<p>For every graph $H$, Ben manages to bound the log-probablity within a log-factor gap:</p>
\[n^2 (d/n)^{2+\gamma(H)} \lesssim -\log \mathrm{Pr}(X \ge (1+\delta) \mathbb{E} X) \lesssim n^2 (d/n)^{2+\gamma(H)} \log(n/d)\]
<p>The most surprising part of Ben’s paper is that for some $H$, the upper tail is not “expected” as above, i.e., the asymptotic formula two display equations earlier is incorrect for this $H$.</p>
<p>For the graph <img src="/blog/images/k24plusedge.svg" width="60" />, Ben solves the variational problem, which led to the formula at the beginning of this post. The solution turns out to come not from “planting a subgraph” but rather from some other more delicate construction.</p>
<p>It remains to be seen what happens to other graphs $H$.
Will we need even more intricate constructions?
What does this mean for techniques towards solving the upper tail problem for the full range of degree parameters? There is still lot that we do not understand.</p>
<h3 id="large-deviations-for-random-hypergraphs">Large deviations for random hypergraphs</h3>
<p>In a recent paper with <a href="https://arxiv.org/abs/1910.02916">Yang Liu</a> (to appear in <em>Random Structures & Algorithms</em>), we considered the extension of the upper tail problem to random hypergraphs.
Yang Liu is a PhD student at Stanford. We started this work while Yang was an undergraduate student at MIT.</p>
<figure>
<img src="/blog/images/yang-liu-2020.jpg" width="200" />
<figcaption>Yang Liu</figcaption>
</figure>
<p>Consider the random $k$-uniform hypergraph \(G^{(k)}(n,p)\), where every unordered $k$-tuple of vertices appears as an edge independently with probability $p$.</p>
<p><strong>Problem.</strong> Given a $k$-uniform hypergraph $H$. Let $X$ denote the number of copies of $X$ in \(G^{(k)}(n,p)\). Estimate $\mathrm{Pr}(X \ge (1+\delta)\mathbb{E} X)$.</p>
<p>As before, there are two complementary strands (1) developing large deviation framework and (2) solving the variational problem.
Some of the work towards (1) for random graphs can be transferred over to the hypergraph setting. Towards (2), we solve the variational problem asymptotically for some $H$, and make a conjecture for what we believe should happen for general $H$.</p>
<p>We show that one cannot naively extend the upper tail results from random graphs to random hypergraphs.
In our paper, we take the readers on a long discussion through several iterations of naive conjectures and counterexamples, eventually building up to our final conjecture, which takes some effort to state, but very roughly, says that</p>
<p><em>The dominant reason for upper tails in random hypergraphs is the presence of a collection of “stable mixed hubs.”</em></p>
<p>For both random regular graphs and random hypergraphs, these recent results highlight unexpected phenomena.
They illustrate the richness of the problem and lead us to more questions than answers.</p>
Mon, 05 Oct 2020 00:00:00 -0400
https://yufeizhao.com/blog/2020/10/05/random-regular-tails/
https://yufeizhao.com/blog/2020/10/05/random-regular-tails/blogJoints of varieties<p>The summer has come to an end.
The temperature in Boston seems to have dropped all the sudden.
Our fall semester has just started, though without all the new and returning students roaming around campus this time (most classes are online this term).
I quite miss the view from my MIT office, and all the wonderful people that I would run into the department corridors.</p>
<figure>
<img src="/blog/images/office-figurines.jpg" width="300" />
<figcaption>The view from my (currently unoccupied) MIT office</figcaption>
</figure>
<p>I’ve been working from home for the past several months and connecting with my students via Zoom.
In this and the upcoming several posts, I would like to tell you about some of the things that we have been working on recently.</p>
<p><br /></p>
<p>With PhD student Jonathan Tidor and undergraduate student Hung-Hsun Hans Yu, we recently solved the <a href="https://arxiv.org/abs/2008.01610">joints problem for varieties</a> by a developing a new extension of the polynomial method. (Also see <a href="https://www.youtube.com/watch?v=dsnU8iPL-WI">my talk video</a> and <a href="/research/slides/joints_varieties.pdf">slides</a> on this work.)</p>
<figure>
<img src="/blog/images/jonathan-tidor-2020.jpg" height="200" />
<img src="/blog/images/hans-yu-2020.jpg" height="200" />
<figcaption>Jonathan Tidor and Hung-Hsun Hans Yu</figcaption>
</figure>
<p>I discussed the joints problem in an <a href="/blog/2019/11/21/joints-tightened/">earlier blog post</a>. Recall that the joints problem asks:</p>
<p><strong>Joints problem:</strong> At most how many joints can $N$ lines in $\mathbb{R}^3$ can make?</p>
<p>Here a <em>joint</em> is a point contained in three non-coplanar lines.</p>
<figure>
<img src="/blog/images/joints-3-planes.png" width="200" />
<figcaption>A joint</figcaption>
</figure>
<p>Here is an example of a configuration of lines making many joints. Take <em>k</em> generic planes in space, pairwise intersect them to form \(N = \binom{k}{2}\) lines, and triplewise intersect to form \(J = \binom{k}{3} \sim \frac{\sqrt{2}}{3} N^{3/2}\) joints.</p>
<figure>
<img src="/blog/images/joints-optimal.png" width="300" />
<figcaption>Planes, lines, and joints</figcaption>
</figure>
<p>The joints problem was raised in the <a href="https://doi.org/10.1016/0925-7721(92)90009-H">early 1990’s</a>.
It is a problem in <a href="https://en.wikipedia.org/wiki/Incidence_geometry"><em>incidence geometry</em></a>, which concerns possible incidence patterns of objects like points and lines.
The joints problem has also received attention from harmonic analysts due to its connections to the <a href="https://en.wikipedia.org/wiki/Kakeya_set">Kakeya problem</a>.</p>
<p>The big breakthrough on the joints problem was due to Guth and Katz, who showed that the above example is optimal up to a constant factor:</p>
<p><strong>Joints theorem</strong> (<a href="https://arxiv.org/abs/0812.1043">Guth–Katz</a>). $N$ lines in \(\mathbb{R}^3\) form $O(N^{3/2})$ joints.</p>
<p>Guth and Katz solved the problem using a clever technique known as the <strong>polynomial method</strong>, which was further developed in their subsequent breakthrough on the <a href="https://arxiv.org/abs/1011.4105">Erdős distinct distances problem</a>.
The polynomial method has since then grown into a powerful and indispensible tool in incidence geometry, and it also led to important recent developments in harmonic analysis.</p>
<p>I learned much of what I know about the polynomial method from taking a course by Larry Guth while I was a graduate student.
Guth later developed the course notes into <a href="https://bookstore.ams.org/ulect-64/">a beautifully written textbook</a>, which I highly recommend.
A feature of the expository style of this book, which I wish more authors would adopt, is that it frequently motivates ideas by first explaining false leads and wrong paths. As I was reading this book, I felt like that I was being taken on a journey as a travel companion and given a lens towards the thinking process, rather than somehow being magically teleported to the destination.</p>
<figure>
<img src="/blog/images/guth-book.jpg" />
</figure>
<p>In my previous paper on the joints problem with Hans Yu, which I discussed in <a href="/blog/2019/11/21/joints-tightened/">earlier blog post</a>, we sharpened the Guth–Katz theorem to the optimal constant.</p>
<p><strong>Theorem</strong> (<a href="https://arxiv.org/abs/1911.08605">Yu–Zhao</a>). $N$ lines in \(\mathbb{R}^3\) have \(\le \frac{\sqrt{2}}{3} N^{3/2}\) joints.</p>
<p>Everything I have said so far are also true in arbitrary dimensions and over arbitrary fields, but I am only stating them in \(\mathbb{R}^3\) here to keep things simple and concrete.</p>
<h3 id="joints-of-flats">Joints of flats</h3>
<p>Our <a href="https://arxiv.org/abs/2008.01610">new work</a> solves the following extension of the joints problem.</p>
<p><strong>Joints problem for planes.</strong> At most how many joints can $N$ planes in \(\mathbb{R}^6\) make?</p>
<p>Here a <em>joint</em> is a point contained in a triple of 2-dimensional <em>planes</em> not all contained in some hyperplane.</p>
<p>The example given earlier for joints of lines extends easily to joints of planes.</p>
<p><strong>Example.</strong> Take $k$ generic 4-flats in \(\mathbb{R}^6\), pairwise intersect them to form $N = \binom{k}{2}$ planes, and triplewise intersect them to form \(\binom{k}{3} = \Theta(N^{3/2})\) joints.</p>
<p>Joints of planes concerns incidences between high dimensional objects, which in general tend to be more intricate compared to incidence problems about points and lines.</p>
<p>I first learned about this problem from <a href="https://arxiv.org/abs/1606.08525">a paper of Ben Yang</a>, who did this work as a PhD student of Larry Guth.
Yang proved that $N$ planes in \(\mathbb{R}^6\) form at most $N^{3/2+o(1)}$ joints.
His proof uses several important recent developments in incidence geometry, and in particular uses the <a href="https://arxiv.org/abs/1102.5391">polynomial partitioning technique</a> of Guth and Katz (as well as subsequent extensions due to <a href="https://arxiv.org/abs/1103.2926">Solymosi–Tao</a> and <a href="https://arxiv.org/abs/1410.8871">Guth</a>).
This is a fantastic result, though with two important limitations due to the nature of the method:</p>
<ol>
<li>There is an error term in the exponent $N^{3/2+o(1)}$ that is conjectured not to be there. One would like a bound of the form $O(N^{3/2})$</li>
<li>The result is restricted to planes in \(\mathbb{R}^6\) and does not work if we replace \(\mathbb{R}\) by an arbitrary field. This is because the polynomial partitioning method uses topological tools, in particular the ham sandwich theorem, which crucially requires the topology of the reals.</li>
</ol>
<p>Our work overcomes both difficulties. As a special case, we prove (here \(\mathbb{F}\) stands for an arbitrary field, and the hidden constants do not depend on \(\mathbb{F}\)):</p>
<p><strong>Joints of planes</strong> (<a href="https://arxiv.org/abs/2008.01610">Tidor–Yu–Zhao</a>). $N$ planes in \(\mathbb{F}^6\) have $O(N^{3/2})$ joints.</p>
<p>And more generally, our results hold for varieties instead of flats.
Here a <em>joint</em> is a point contained in a triple of varieties, such that it is a regular point on each of the three varieties and the three tangent plane do not lie inside a common hyperplane.</p>
<p><strong>Joints of varieties</strong> (<a href="https://arxiv.org/abs/2008.01610">Tidor–Yu–Zhao</a>). A set of 2-dimensional varieties in \(\mathbb{F}^6\) of total degree $N$ has $O(N^{3/2})$ joints.</p>
<p>The above statements are special cases of our results. The joints problem for lines is known to hold under various extensions (arbitrary dimension and fields, several sets/multijoints, counting joints with multiplicities), and our theorem holds in these generalities as well.</p>
<h3 id="proof-sketch-of-the-joints-theorem-for-lines">Proof sketch of the joints theorem for lines</h3>
<p>Here is a sketch of the proof that $N$ lines in \(\mathbb{R}^3\) make $O(N^{3/2})$ joints (following the simplifications due to <a href="https://arxiv.org/abs/0906.0558">Kaplan–Sharir–Shustin</a> and <a href="https://arxiv.org/abs/0906.0555">Quilodrán</a>; also see <a href="http://math.mit.edu/~lguth/PolyMethod/lect4.pdf">Guth’s notes</a> for details of the proof).</p>
<p>The first two steps below are both hallmarks of the polynomial method, and they are present in nearly every application of the method.
The third step is a more joints-specific argument.</p>
<ol>
<li>
<p><strong>Parameter counting.</strong> The dimension of the vector space of polynomials in \(\mathbb{R}[x,y,z]\) up to degree $n$ is \(\binom{n+3}{3}\).
We deduce that if there are $J$ joints and \(J < \binom{n+3}{3}\), then one can find a nonzero polynomial $g$ of degree $\le C J^{1/3}$ (for some constant $C$) that vanishes on all the joints (as there are more coefficients to choose from than there are linear constraints). Take such a polynomial $g$ with the minimum possible degree.</p>
</li>
<li>
<p><strong>Vanishing lemma.</strong> A nonzero single variable polynomial cannot vanish more times than its degree.</p>
</li>
<li>
<p>If all lines have $> CJ^{1/3}$ joints, then the vanishing lemma implies that $g$ vanishes on all the lines.
At each joint, $g$ vanishes identically along three independent lines, and hence its directional derivatives vanish in three independent directions, and thus the gradient \(\nabla g\) vanishes at the joint. Therefore, the three partial derivatives
\(\partial_x g\),
\(\partial_y g\),
\(\partial_z g\) all vanish at every joint, and at least one of them is a nonzero polynomial and has smaller degree than $g$, thereby contradicting the minimality of the degree of $g$. Therefore, the initial assumption must be false, namely that some line has \(\le C J^{1/3}\) joints. We can then remove this line, deleting \(\le CJ^{1/3}\) joints, and proceed by induction to prove that $J = O(N^{3/2})$.</p>
</li>
</ol>
<h3 id="some-wishful-thinking">Some wishful thinking</h3>
<p>It is natural to try to extend the above proof for the joints problem for planes.
One quickly runs into the following conundrum:</p>
<p><em>How does one generalize the <strong>vanishing lemma</strong> to 2-variable polynomials?</em></p>
<p>The vanishing lemma is such an elementary fact about how polynomial behaves on a line.
Yet it plays a crucial role in the polynomial method.
We would be able to adapt the above proof to joints of planes <em>if only something like the following would be true</em> (they are not):</p>
<p><em>Wishful thinking 1.</em> Show that every nonzero polynomial $g(x,y)$ of degree $n$ has $O(n^2)$ zeros.</p>
<p>Well, this is obviously nonsense. The zero set of $g(x,y)$ is typically a curve with infinitely many points.</p>
<p>How else might we try to extend the vanishing lemma to two dimensions?
<em>Bézout’s theorem</em> is one such extension.
It tells us that two polynomials $g(x,y)$ and $h(x,y)$ without common factors cannot have more than $(\deg g)(\deg h)$.
So, perhaps, instead of trying to find one polynomial vanishing at all joints, we try to find a pair of polynomials.</p>
<p><em>Wishing thinking 2.</em> Show that one can find two polynomials \(g, h \in \mathbb{R}[x_1, x_2, x_3, x_4, x_5, x_6]\) vanishing at all $J$ joints and $(\deg g)(\deg h) = O(J^{1/3})$, and such that $g$ and $h$ have no common factors when restricted to any of the given planes.</p>
<p>This approach is a natural one to try for the joints of planes problem, and is an approach that many people have thought about.
It also seems extremely hard to get it to work (likely impossible).
It seems closely related to the <a href="https://terrytao.wordpress.com/2012/09/25/a-partial-converse-to-bezouts-theorem/">inverse Bézout problem</a>,
which already illustrates some of the difficulties encountered.</p>
<p>Instead of asking a polynomial to vanish at every joint, we can demand more. Namely that we can ask the polynomial to vanish at each joint to some high order (equivalently, asking some derivatives of the polynomial to vanish at the joints).
This idea (sometimes called the “method of multiplicities”) has shown to be useful in a variety of contexts.
Perhaps we can modify Wishing thinking 1 to account for these multiplicities?</p>
<p><em>Wishful thinking 3.</em> Show that every nonzero polynomial $g(x,y)$ of degree $n$ can vanish to order $s$ at no more than $O(n^2/s^2)$ points.</p>
<p>The reason why one might hope for such a statement is that there are \(\binom{n+2}{2} \approx n^2/2\) degrees of freedom for choosing $g$, and this number is roughly the same as the number of constraints associated to forcing the polynomial to vanish to order $s$ at $n^2/s^2$ points.</p>
<p>However, Wishful thinking 3 is still too good to be true. For example, the polynomial $g(x,y) = y^s$ is a low-degree polynomial that vanishes to order $s$ at every point on the $x$-axis.</p>
<p>So what was wrong with the parameter counting intuition? It must be the case that some of the linear constraints associated to these high order vanishing constraints are linearly dependent.</p>
<h3 id="extending-the-polynomial-method">Extending the polynomial method</h3>
<p>So far all I have told you are what doesn’t work, and I haven’t said anything about what does work.
Where we left off above is a good starting point of what we actually do, which I will not describe here in this blog post but instead refer you to either <a href="https://www.youtube.com/watch?v=dsnU8iPL-WI">my talk</a> or <a href="https://arxiv.org/abs/2008.01610">our paper</a> for details. We have to account for the linear dependencies between these higher order vanishing conditions, and we come up with a process of assigning non-redundant vanishing conditions (each in the form of some higher order derivative operator evaluated at a joint).
We prove a new vanishing lemma tailored for this joints problem. Rather vaguely, it has the following form:</p>
<p><strong>A new vanishing lemma.</strong> Given a finite set of planes in \(\mathbb{R}^6\) forming some joints, and a positive integer $n$, and an integer (“handicap”) attached to each joint, we run a certain <em>process</em> that assigns, on each plane, a total of \(\binom{n+2}{2}\) derivative operators across the joints on this plane. Then for every nonzero polynomial \(g(x_1, \dots, x_6)\) of degree at most $n$, there is some joint $p$, three planes $F_1, F_2, F_3$ containing $p$, and a derivative operator $D_i$ assigned to $p$ on $F_i$ through our process for each $i =1,2,3$, so that $D_1 D_2 D_3 g(p) \ne 0$.</p>
<p>This new vanishing lemma overcomes the earlier difficulty and allows us to extend the polynomial method to solve the joints problem for planes.</p>
<p><br /></p>
<p>Let me finish with an open problem: while we can prove the optimal constant for the joints theorem for lines, we do not yet know how to prove the optimal constant for the joints theorem for planes.</p>
<p><strong>Conjecture.</strong> $N$ planes in \(\mathbb{F}^6\) have \(\le (\frac{\sqrt{2}}{3} + o(1)) N^{3/2}\) joints.</p>
Sat, 12 Sep 2020 00:00:00 -0400
https://yufeizhao.com/blog/2020/09/12/joints-varieties/
https://yufeizhao.com/blog/2020/09/12/joints-varieties/blogAshwin Sah's new diagonal Ramsey number upper bound<figure>
<img src="/blog/images/ashwin-sah-2020.jpg" width="200" />
<figcaption>Ashwin Sah</figcaption>
</figure>
<p><a href="http://www.mit.edu/~asah/">Ashwin Sah</a> just proved a new upper bound to diagonal Ramsey numbers. See <a href="https://arxiv.org/abs/2005.09251">his preprint on the arXiv</a>. This is the first improvement since Conlon’s upper bound published in <a href="https://arxiv.org/abs/math/0607788">Annals of Math in 2009</a>, which in turn built on earlier work of <a href="https://doi.org/10.1016/S0304-0208(08)73063-9">Thomason (1988)</a>.</p>
<p>Obtaining asymptotics of <a href="https://www.newyorker.com/magazine/2020/05/04/the-man-who-thought-too-fast">Ramsey</a> numbers is perhaps <em>the</em> central open problem of Ramsey theory in combinatorics. Research on this problem had led us to a lot of new mathematics that have proven to be useful elsewhere. Most notably, <a href="https://projecteuclid.org/euclid.bams/1183510596">Erdős’ proof of a lower bound to Ramsey numbers (1947)</a> initiated the probabilistic method, by now an indispensable method in combinatorics.</p>
<p>Ashwin’s improvement relies on new insights into graph quasirandomness properties.</p>
<p>Congratulations Ashwin! This is an impressive accomplishment.</p>
<p>Ashwin had just finished his undergraduate studies at MIT. I’m happy that he will be staying at MIT for his PhD. Ashwin already has an impressive <a href="http://www.mit.edu/~asah/research.html">list of papers</a>. I had previously mentioned Ashwin on this blog reporting on our joint work (together with Mehtaab Sawhney and David Stoner) on our work on <a href="/blog/2018/05/12/the-number-of-independent-sets-in-an-irregular-graph/">independent sets</a> and the <a href="/blog/2018/09/26/a-reverse-sidorenko-inequality/">reverse Sidorenko inequality</a>.</p>
Wed, 20 May 2020 00:00:00 -0400
https://yufeizhao.com/blog/2020/05/20/ashwin-sahs-new-diagonal-ramsey-number-upper-bound/
https://yufeizhao.com/blog/2020/05/20/ashwin-sahs-new-diagonal-ramsey-number-upper-bound/blogJoints tightened<p>I’m happy to announce a new paper titled <a href="https://arxiv.org/abs/1911.08605">Joints tightened</a> coauthored with Hung-Hsun Hans Yu, an undergraduate student at MIT. In this paper, we determine the tight constant in the joints problem.</p>
<figure>
<img src="/blog/images/hans-yu-2019.jpg" width="300" />
<figcaption>Hung-Hsun Hans Yu</figcaption>
</figure>
<p>The <strong>joints problem</strong> is a classic problem in <strong>incidence geometry</strong>. A typical question in incidence geometry concerns what kinds of configurations can be built using just points and lines.</p>
<p>Suppose you are allowed to draw <em>L</em> lines in space. What is the maximum possible number of points that lie on three lines (let’s call these points “triple intersections” for now)?</p>
<p>It turns out this is not a great question, since you can “cheat” by drawing a 2-dimensional picture from a grid as below. This configuration of <em>L</em> lines has \(\Theta(L^2)\) triple intersections, and you cannot do much better since <em>L</em> lines have at most $L^2$ intersections.</p>
<figure>
<img src="/blog/images/joints-grid-diagonal.png" width="200" />
<figcaption>Not a joints configurations</figcaption>
</figure>
<p>The issue with the above example is that it is really a 2-dimensional configuration, whereas we really want to ask a 3-dimensional question. One way to make the problem much more interesting is only count “truly 3-dimensional” intersection points.</p>
<p>Given a collection of lines in space, we define a <strong>joint</strong> to be a point that is contained in three of our lines and not all lying on the same plane.</p>
<figure>
<img src="/blog/images/joints-3-planes.png" width="200" />
<figcaption>A joint</figcaption>
</figure>
<p><strong>Joints problem.</strong> What is the maximum number of joints that can be formed by <em>L</em> lines in 3-dimensional space?</p>
<p>The joints problem was raised in <a href="https://doi.org/10.1016/0925-7721(92)90009-H">early 1990’s</a>, and it turned out to be a very interesting problem! Besides its intrinsic interest as a combinatorial geometry problem, it also caught the attention of analysts as the joints problem is connected to the <a href="https://en.wikipedia.org/wiki/Kakeya_set">Kakeya problem</a>, a central problem in analysis (this connection was popularized by <a href="http://amathe.web.elte.hu/modern/wolff_review.pdf">Wolff</a>).</p>
<p>Here are some examples of configurations of lines that have a lot of joints.</p>
<p><strong>Example 1.</strong> The easiest example to visualize is by considering a \(k \times k \times k\) grid of lines, which has \(L = 3k^2\) lines and \(J = k^3 = (L/3)^{3/2}\) joints:</p>
<figure>
<img src="/blog/images/joint-3d-grid.png" width="300" />
<figcaption>A grid of joints</figcaption>
</figure>
<p><strong>Example 2.</strong> Actually, there is another example that has even more joints (by a constant factor). Think about a tetrahedron, formed by 4 faces (planes), and whose edges form 6 lines, and the vertices form 4 joints. We can generalize the tetrahedron example by taking <em>k</em> generic planes in space, and have the planes pairwise intersect to make \(L = \binom{k}{2}\) lines, and triplewise intersect to make \(J = \binom{k}{3}\) joints. (This example appeared in the <a href="https://doi.org/10.1016/0925-7721(92)90009-H">original paper</a> on the joints problem.)</p>
<figure>
<img src="/blog/images/joints-optimal.png" width="300" />
<figcaption>Planes, lines, and joints</figcaption>
</figure>
<p>In both examples, the number of joints is \(\Theta(L^{3/2})\). A matching upper bound on the number of joints turned out to be much less obvious. A stunning breakthrough took place in 2008 when <a href="https://arxiv.org/abs/0812.1043">Guth and Katz</a> proved an \(O(L^{3/2})\) upper bound on the number of joints in using the polynomial method. Their solution was partly inspired by <a href="https://arxiv.org/abs/0803.2336">Dvir’s stunning solution</a> to the finite field Kakeya problem, and it was also a precursor to Guth and Katz’s subsequent celebrated <a href="https://arxiv.org/abs/1011.4105">solution of the Erdős distinct distances problem</a>.</p>
<p>Larry Guth conjectured in his book <a href="https://bookstore.ams.org/ulect-64/">The Polynomial Methods in Combinatorics</a> that the second example above is best possible. I first learned about this very nice conjecture as a graduate student in Guth’s class on the polynomial method.</p>
<p>The main result of our paper is a new improved upper bound on the joints problem, matching the constant in the second example above. More precisely, we prove that <em>L</em> lines in space make at most \(\frac{\sqrt{2}}{3}L^{3/2}\) joints, thereby confirming Guth’s conjecture up to a \(1+o(1)\) factor. Our results also hold in arbitrary dimensions and over arbitrary fields.</p>
<p>It was quite unexpected to us that such a tight bound can even be proved, especially given how long it took just to obtain an upper bound of the right order of magnitude on the joints problem. In pretty much all other classical problems in incidence geometry (i.e., the <a href="https://en.wikipedia.org/wiki/Szemer%C3%A9di%E2%80%93Trotter_theorem">Szemerédi–Trotter theorem</a>), the optimal constant factor is not known. In fact, our result might be the first one in incidence geometry that obtains a tight constant factor.</p>
Thu, 21 Nov 2019 00:00:00 -0500
https://yufeizhao.com/blog/2019/11/21/joints-tightened/
https://yufeizhao.com/blog/2019/11/21/joints-tightened/blogEquiangular lines with a fixed angle<p>I’m happy to announce our new paper <a href="https://arxiv.org/abs/1907.12466">Equiangular lines with a fixed angle</a> joint with four MIT coauthors: Zilin Jiang, Jonathan Tidor, Yuan Yao, and Shengtong Zhang.</p>
<figure>
<img src="/blog/images/equiangular_lines_team.jpg" />
<figcaption>(L->R) Shengtong Zhang, Yuan Yao, Jonathan Tidor, Zilin Jiang, me</figcaption>
</figure>
<p>Zilin is an Instructor (postdoc), Jonathan is my PhD student who just finished his second year, and Yuan and Shengtong are undergraduates who just finished their second and first respective years.</p>
<p>Our paper solves a longstanding problem in discrete geometry concerning <strong>equiangular lines</strong>, which are configurations of lines in space that are pairwise separated by the same angle. How many such lines can exist simultaneously in a given dimension?</p>
<p>We determine, for every fixed angle, the maximum number of equiangular lines with the given angle in high dimensions.</p>
<p>Our paper builds on a long sequence of earlier ideas. There were some important progress on this problem by Igor Balla, Felix Dräxler, Peter Keevash, and Benny Sudakov, as featured in this earlier <a href="https://www.quantamagazine.org/a-new-path-to-equal-angle-lines-20170411/">Quanta Magazine article</a> written for a lay audience.</p>
<p>In the end we finished off the problem in a clean and crisp manner, in a 10-page paper with a self-contained proof.</p>
<p>It has been known that the study of equiangular lines is closed linked to <strong>spectral graph theory</strong>, which seeks to understand graphs via their spectrum (i.e., eigenvalues). We prove our main result via a new insight in spectral graph theory, namely that a bounded degree graph must have sublinear second eigenvalue multiplicity.</p>
Tue, 30 Jul 2019 00:00:00 -0400
https://yufeizhao.com/blog/2019/07/30/equiangular-lines-with-a-fixed-angle/
https://yufeizhao.com/blog/2019/07/30/equiangular-lines-with-a-fixed-angle/blogImpartial digraphs<p>I just finished a new paper, <a href="https://arxiv.org/abs/1906.10482">Impartial digraphs</a>, coauthored with Yunkun Zhou, who just completed his undergraduate studies at MIT and will be moving to Stanford to start his PhD this fall.</p>
<figure>
<img src="/blog/images/yunkun-zhou-2019.jpg" />
<figcaption>with Yunkun Zhou (R) after his MIT commencement</figcaption>
</figure>
<p>The problem (along with the conjecture that we proved) was proposed by Jacob Fox, Hao Huang, and Choongbum Lee. Ron Graham had also told me about this problem with much enthusiasm during tea time one afternoon at the Berkeley Simons Institute a couple of years ago, after learning about it from Jacob.</p>
<p>I really enjoyed working on this paper with Yunkun. The problem is natural and simple to state, the answer appears somewhat unexpected (at least initially), and the solution ended up being quite tricky yet clean, employing a variety of tools in combinatorics, ranging from analytic methods (graph limits), algebraic considerations (polynomial identities and factorization), as well basic combinatorial techniques such as bijections.</p>
<p>We are going to talk about <strong>directed graphs</strong> (which graph theorists like to abbreviate as <strong>digraphs</strong>). These are graphs where every edge is given an orientation.</p>
<figure>
<img src="/blog/images/impartial-0.png" width="150" />
</figure>
<p>A <strong>tournament</strong> is a directed graph where there is a directed edge between every pair of vertices (think of it as recording the outcome of a round-robin tournament where every player plays against every other player).</p>
<figure>
<img src="/blog/images/impartial-1.png" width="150" />
</figure>
<p>Given a directed graph <em>H</em> (think of it as a pattern), we can ask: how many different copies of the pattern <em>H</em> appear in a given tournament?</p>
<figure>
<img src="/blog/images/impartial-2.png" width="500" />
</figure>
<p>Look at the following 4-vertex tournament <em>H</em>. It has the following curious property: all tournaments of a fixed size contains the same number of copies of H, no matter how the edges in the tournament are oriented.</p>
<figure>
<img src="/blog/images/impartial-2a.png" width="500" />
</figure>
<p>Hmm, why is this true? And are there other directed graphs <em>H</em> with the same property, namely that all tournaments of a given size contain the same number of copies of <em>H</em>?</p>
<p>In the good old mathematical tradition, we came up with a name for directed graphs <em>H</em> with this curious property: we called them <strong>impartial</strong> (as in: impartial judges of a tournament).</p>
<p>Okay. So why is the above directed graph impartial? Are there any other impartial directed graphs? What are all the impartial directed graphs?</p>
<p>Here is a simple example of an impartial digraph:</p>
<figure>
<img src="/blog/images/impartial-4.png" width="50" />
</figure>
<p>Here is another one:</p>
<figure>
<img src="/blog/images/impartial-5.png" width="80" />
</figure>
<p>Indeed, given any 10-vertex tournament, we can count the number of copies of the first edge (it doesn’t depend on the tournament), and then count the number of copies of the second edge among the remaining 8 vertices (it still doesn’t depend on the tournament).</p>
<p>Once these two edges are placed in a tournament, there are two ways to join their tips, but they result in identical patterns (one flipped from the other):</p>
<figure>
<img src="/blog/images/impartial-6.png" width="150" />
</figure>
<p>It must then follow that this directed graph is impartial as well:</p>
<figure>
<img src="/blog/images/impartial-8.png" width="60" />
</figure>
<p>We can follow the same logic and continue this procedure, iterate, and build up even larger impartial directed graphs. Each step, take two copies of the previous tree (keeping the same edge orientations), and then add a new edge between a twin pair of vertices.</p>
<figure>
<img src="/blog/images/impartial-9.png" width="150" />
</figure>
<figure>
<img src="/blog/images/impartial-10.png" width="150" />
</figure>
<figure>
<img src="/blog/images/impartial-11.png" width="300" />
</figure>
<figure>
<img src="/blog/images/impartial-12.png" width="300" />
</figure>
<p>We call any directed graph that can building up as above a <strong>recursively bridge-mirrored</strong> <strong>tree</strong>.</p>
<p>The same argument as earlier shows that every recursively bridge-mirrored tree is impartial.</p>
<p>Are there other impartial digraphs? Did we miss any?</p>
<p>It turns out, there are actually no others. We have found them all, and that is the main result of our paper:</p>
<p><strong>Theorem.</strong> A directed graph is impartial if and only if every component is recursively bridge-mirrored trees.</p>
<p>Just a word about how about this problem is related to a bigger picture. It is related to a number of other graph homomorphism inequalities that are central in extremal graph theory. It can be considered as the “equality case” for a directed analog of an open problem known as <strong>Sidorenko’s conjecture</strong>. See <a href="https://arxiv.org/abs/1906.10482">our paper</a> for some further discussions and open problems.</p>
Thu, 27 Jun 2019 00:00:00 -0400
https://yufeizhao.com/blog/2019/06/27/impartial-digraphs/
https://yufeizhao.com/blog/2019/06/27/impartial-digraphs/blogA reverse Sidorenko inequality<p>Together with undergraduates Ashwin Sah, Mehtaab Sawhney, and David Stoner (the same team that <a href="/blog/2018/05/12/the-number-of-independent-sets-in-an-irregular-graph/">proved Kahn’s conjecture on independent sets that I blogged about earlier</a>,
we are excited to announce our new paper <a href="https://arxiv.org/abs/1809.09462">A reverse Sidorenko inequality</a>. This paper solves several open problems concerning graph colorings and homomorphisms, including one of my favorite problems regarding maximizing the number of <em>q</em>-colorings in a <em>d</em>-regular graph. I had previously <a href="https://yufeizhao.wordpress.com/2016/10/29/extremal-regular-graphs/">blogged</a> about some of these problems.</p>
<p>It has been truly a pleasure working with Ashwin, Mehtaab, and David on this project. I have learned so much from them.</p>
<figure>
<img src="/blog/images/hom_coloring.png" width="400" />
</figure>
<p>Here are some highlights of our new results.</p>
<h3 id="1-graph-colorings">1. Graph colorings.</h3>
<p> Among <em>d</em>-regular graphs, which graph has the most number of proper <em>q</em>-colorings, exponentially normalized by the number of vertices of <em>G</em>?</p>
<p>We show that the extremal graph is the complete bipartite graph \(K_{d,d}\) (or disjoint unions thereof, as taking disjoint copies does not change the quantity due to the normalization).</p>
<h3 id="2-graph-homomorphisms">2. Graph homomorphisms.</h3>
<p>Among <em>d</em>-regular graphs, which triangle-free graph <em>G</em> has the most number of homomorphisms into a fixed <em>H</em>, again exponentially normalized by the number of vertices of <em>G</em>?</p>
<p>We show that the extremal graph <em>G</em> is the complete bipartite graph \(K_{d,d}\).</p>
<p>The triangle-free hypothesis on <em>G</em> is best possible in general. For certain specific <em>H,</em> such as \(K_q\), corresponding to proper <em>q</em>-colorings as earlier, the triangle-free hypothesis can be dropped. It remains a wide open problem to determine for which <em>H</em> can one drop the triangle-free hypothesis on <em>G</em>.</p>
<h3 id="3-partition-functions-of-ferromagnetic-spin-models">3. Partition functions of Ferromagnetic spin models.</h3>
<p>Among <em>d</em>-regular graphs, which graph maximizes the log-partition function per vertex of a given Ferromagnetic spin model (e.g., Ising model)?</p>
<p>We show that the extremal graph is the clique \(K_{d+1}\).</p>
<p>For each setting, we establish our results more generally for irregular graphs as well, similar to <a href="/blog/2018/05/12/the-number-of-independent-sets-in-an-irregular-graph/">our earlier work on independent sets</a>.</p>
<p>Our results can be interpreted as a reverse form of the inequality in <strong>Sidorenko’s conjecture</strong>, an important open problem in graph theory stating a certain positive correlation on two-variable functions.</p>
<p>One can also view our results as a graphical analog of the <strong>Brascamp–Lieb inequalities</strong>, a central result in analysis.</p>
<p>This paper resolves one of my favorite open problems on this topic (the number of graph colorings). It also points us to many other open problems. Let me conclude by highlighting one of them (mentioned in #2 earlier). I’ll state it in a simpler form for <em>d</em>-regular graphs, but it can be stated more generally as well.</p>
<p><strong>Open problem.</strong> Classify all <em>H</em> with the following property: among <em>d</em>-regular graphs <em>G</em>, the number of homomorphisms from <em>G</em> to <em>H</em>, exponentially normalized by the number of vertices of <em>G</em>, is maximized by \(G = K_{d,d}\).</p>
<p>Our results show that \(H = K_q\) works, even when some of its vertices are looped. Generalizing this case, we conjecture that all antiferromagnetic models <em>H</em> have the same property (though we know that this would not be the complete class, even if true).</p>
Wed, 26 Sep 2018 00:00:00 -0400
https://yufeizhao.com/blog/2018/09/26/a-reverse-sidorenko-inequality/
https://yufeizhao.com/blog/2018/09/26/a-reverse-sidorenko-inequality/blog