betaFeedback

WorksMathematics

What is... An Exceptional Set Estimate?

Abstract

Exceptional set estimates are a fundamental topic of interest in projection theory. We introduce the topic through discrete geometry, and use this perspective to motivate some key results in the realm of geometric measure theory. To appear in the "What is?" series in the Notices of the AMS.

Shadows are Typically Large

Suppose you are standing outside on a sunny day from dawn til dusk, observing your shadow as it changes. Most of the time, you will see a large portion of your shadow; but at noon, when the sun is right above you, your shadow will fall right beneath your feet. In this way, at almost every time of day, your shadow is relatively large, and, more quantitatively, it is very exceptional for your shadow to be small. In projection theory, a field at the intersection of discrete geometry and geometric measure theory, one seeks to understand how general this phenomenon is in a mathematically precise way.

The concept of shadow here is made precise by orthogonal projections. Given θ∈S1⊂R2\theta\in\mathbb{S}^{1} \subset\mathbb{R}^{2}, we let Pθ:R2→RP_{\theta}:\mathbb{R}^{2} \to\mathbb{R} denote the orthogonal projection onto the line through the origin in direction θ\theta. In other words, Pθ(x)=θ⋅xP_{\theta}(x)=\theta\cdot x where ⋅\cdot is the Euclidean dot product. Given X⊂R2X\subset\mathbb{R}^{2} Borel, how large could the projection, Pθ(X)P_{\theta}(X), be? Answers to this question depend on how large the underlying set XX is itself. By Fubini’s theorem, if XX has positive planar Lebesgue measure, then Pθ(X)P_{\theta}(X) must have positive linear (1-dimensional) Lebesgue measure for all θ\theta. As such, we consider sets which have null Lebesgue measure but that are large in another quantitative sense, namely Hausdorff dimension.

Before continuing, we briefly note that the Hausdorff dimension of a set XX in a metric space, denoted dim⁡X\dim X throughout, can take on any real nonnegative value bounded by the dimension of the ambient space. Hausdorff dimension is appropriate for quantifying the size of Lebesgue measure zero sets. For instance, while the standard middle thirds Cantor set has Lebesgue measure zero, it has Hausdorff dimension log⁡2log⁡3\frac{\log2}{\log3}. Heuristically, dim⁡X=α\dim X=\alpha if the minimum number of δ\delta-balls needed to cover XX, ∣X∣δ|X|_{\delta}, is ∼δ−α\sim\delta^{-\alpha} for sufficiently small δ\delta. This heuristic would be made precise if we were instead using box dimension, but this heuristic is sufficient for our purposes. For more background on Hausdorff dimension, see e.g. [6].

Using the fact that orthogonal projection is a Lipschitz map onto a 1-dimensional subspace, it is immediate to see that

dim⁡Pθ(X)≤min⁡{dim⁡X,1} for all θ∈S1.\dim P_{\theta}(X)\leq\min\{\dim X,1\}\text{ for all }\theta\in\mathbb{S}^{1}.

Intuitively, all this is saying is that the size of the shadow (dim⁡Pθ(X)\dim P_{\theta}(X)) is at most the size of the object casting the shadow (dim⁡X\dim X), and at most the size of what the shadow is being cast upon (dim⁡R=1\dim\mathbb{R}=1). In 1954, Marstrand [5] showed that in fact,

dim⁡Pθ(X)=min⁡{dim⁡X,1} for a.e. θ∈S1.\dim P_{\theta}(X)=\min\{\dim X,1\}\text{ for a.e. }\theta\in\mathbb{S}^{1}.

(Here, almost every is with respect to the surface measure on the circle.) In other words, almost all of the time, the orthogonal projection of XX is as large as it could be, matching our intuition from our nice sunny day outside.

However, projection theorists sought to probe this question even further by estimating the size of the set of θ\theta where the above equality fails. These are known as exceptional set estimates.

A Discrete Model and Sharpness

It is natural to first approach these estimates by studying a discrete model where X⊂R2X\subset\mathbb{R}^{2} is finite. To begin, it is immediate to show that for all but finitely many θ\theta,

∣Pθ(X)∣=∣X∣.(1)|P_{\theta}(X)|=|X|. \tag*{(1)}

Indeed, in order for Pθ(X)P_{\theta}(X) to be smaller, there must be two distinct points x,y∈Xx,y\in X such that θ⋅(x−y)=0\theta\cdot(x-y)=0, and XX can only determine at most (∣X∣2)\binom{|X|}{2} such directions. Thus, let us bound the number of θ\theta for which (1) fails. In particular, given s≤∣X∣s \le|X|, we seek to bound the cardinality of

Es(X):={θ∈S+1:∣Pθ(X)∣<s}.E_s(X) := \{\theta\in\mathbb{S}^{1}_{+} : |P_\theta(X)| < s\}.

We consider S+1\mathbb{S}^{1}_{+}, i.e. angles between [0,π)[0,\pi), so as to avoid double-counting. As a starting point, let’s consider two model examples of sets X⊂R2X \subset\mathbb{R}^{2}. If XX is a collection of NN points on the yy-axis, then Es(X)E_s(X) is simply the direction θ=(1,0)\theta=(1,0) for all s∈(1,N]s \in(1,N]. This matches our sunny day picture: The only time in which XX can cast an exceptionally small shadow is at “noon.” If XX is a generic set of NN points, which is to say no line contains more than 2 points of XX, then

Es(X)=∅for s≤N/2.E_s(X)=\varnothing\quad\text{for }s\le N/2.

To see this, notice that for all θ∈Es(X)\theta\in E_s(X), by the pigeonhole principle, one of the fibers of PθP_\theta must contain at least three points of XX which is a contradiction. For the interested reader, we leave as an exercise to consider E⌈N/2⌉+1(X)E_{\lceil N/2\rceil+1}(X) when XX is a regular NN-gon, in comparison to the set of directions between pairs of points in XX.

Let us now prove bounds on Es(X)E_s(X) for general X⊂R2X \subset\mathbb{R}^{2}. One estimate we can note is that

∣E∣X∣1/2(X)∣≤1.(2)|E_{|X|^{1/2}}(X)| \le1. \tag*{(2)}

Indeed, if ∣E∣X∣1/2(X)∣≥2|E_{|X|^{1/2}}(X)| \ge2, there exist two nonantipodal directions θ,θ′\theta,\theta' such that we can cover XX in <∣X∣1/2<|X|^{1/2} many parallel lines in directions θ\theta and (θ′)⊥(\theta')^\perp respectively. We denote these lines/fibers Lθ\mathcal{L}_\theta and Lθ′\mathcal{L}_{\theta'}. Since both sets of lines cover XX, it follows that

∣X∣≤∣Lθ∣∣Lθ′∣<∣X∣1/2⋅∣X∣1/2=∣X∣,|X| \le|\mathcal{L}_\theta||\mathcal{L}_{\theta'}| < |X|^{1/2}\cdot|X|^{1/2}=|X|,

which is a contradiction.

To obtain stronger estimates, we consider an incidence problem between points and lines. To do so, first note that by definition, for all θ∈Es(X)\theta\in E_s(X), it takes at most ss-many lines in direction θ⊥\theta^\perp to cover XX. Call this set of fibers Lθ\mathcal{L}_\theta (see Figure 1), and let

L:=⋃θ∈Es(X)Lθ.\mathcal{L}:=\bigcup_{\theta\in E_s(X)}\mathcal{L}_\theta.

Then, we consider the set of incidences between XX and L\mathcal{L}; that is,

I(X,L):={(x,ℓ)∈X×L:x lies in ℓ}.\mathcal{I}(X,\mathcal{L}):=\{(x,\ell)\in X\times\mathcal{L}:x\text{ lies in }\ell\}.

Figure 1. Covering XX by fibers Lθ\mathcal{L}_\theta.

Notice that since Lθ\mathcal{L}_\theta covers XX and Lθ∩Lθ′=∅\mathcal{L}_\theta\cap\mathcal{L}_{\theta'}=\varnothing for all θ≠θ′\theta\ne\theta', it follows that

∣X∣∣Es(X)∣=∑x∈X∣I(x,L)∣=∣I(X,L)∣.|X||E_s(X)|=\sum_{x\in X}|\mathcal{I}(x,\mathcal{L})|=|\mathcal{I}(X,\mathcal{L})|.

Indeed, notice that for every x∈Xx\in X and every θ∈Es(X)\theta\in E_s(X), there is a distinct fiber in Lθ\mathcal{L}_\theta containing xx. Hence, if we have an upper bound on the right hand side, we can hope to conclude estimates on ∣Es(X)∣|E_s(X)|. For instance, using that two points determine one line through both, one can use Cauchy–Schwarz to verify that

∣I(X,L)∣≤∣X∣∣L∣1/2+∣L∣.|\mathcal{I}(X,\mathcal{L})|\le|X||\mathcal{L}|^{1/2}+|\mathcal{L}|.

Since ∣Lθ∣≤s|\mathcal{L}_\theta|\le s, one has ∣L∣≤s∣Es(X)∣|\mathcal{L}|\le s|E_s(X)|. Hence, if s<∣X∣/2s<|X|/2 (and thus the first term dominates) one can conclude

∣Es(X)∣≲s.(3)|E_s(X)|\lesssim s. \tag*{(3)}

While (3) is not sharp, it does mark a significant improvement on the trivial estimate

∣Es(X)∣≤(∣X∣2)∼∣X∣2.|E_s(X)|\le\binom{|X|}{2}\sim|X|^2.

The sharp estimate was given by Szemerédi and Trotter [11], who proved

∣I(X,L)∣≤C(∣X∣2/3∣L∣2/3+∣X∣+∣L∣)|\mathcal{I}(X,\mathcal{L})|\le C\left(|X|^{2/3}|\mathcal{L}|^{2/3}+|X|+|\mathcal{L}|\right)

for some universal constant CC. Hence, assuming s<∣X∣/3Cs<|X|/3C, it follows that

∣Es(X)∣≲max⁡{s2∣X∣−1,1},(4)|E_s(X)|\lesssim\max\{s^2|X|^{-1},1\}, \tag*{(4)}

which is sharp. We now turn to the continuous analogues of these problems in the framework of geometric measure theory.

The Problem and Its History

Let us begin by stating what the continuum analogue of Es(X)E_s(X) is: Given X⊂R2X \subset\mathbb{R}^2 and some 0<s<min⁡{dim⁡X,1}0 < s < \min\{\dim X,1\}, let

Es(X)={θ∈S+1:dim⁡Pθ(X)≤s}.E_s(X)=\{\theta\in\mathbb{S}_+^1:\dim P_\theta(X)\leq s\}.

The aforementioned result of Marstrand shows Es(X)E_s(X) has null measure for all ss. Hence, it is natural to ask for upper bounds on the dimension of Es(X)E_s(X).

Each of the discrete bounds we derived has an interesting continuum counterpart, with the analogue of (4) being conjectured by D. Oberlin in 2008 [7]. To go from the discrete to the continuum, let us go back to our heuristic that ∣X∣δ∼δ−dim⁡X|X|_\delta\sim\delta^{-\dim X} for all small δ\delta. Based on the discrete setting, we might expect that δ\delta-balls should be no different from points for small enough δ\delta. As such, by (4) we might conjecture

∣{θ:∣Pθ(X)∣δ≲δ−s}∣δ≲max⁡{(δ−s)2∣X∣δ−1,δ0}∼δ−max⁡{2s−dim⁡X,0}.\left|\{\theta:|P_\theta(X)|_\delta\lesssim\delta^{-s}\}\right|_\delta\lesssim\max\{(\delta^{-s})^2|X|_\delta^{-1},\delta^0\} \sim\delta^{-\max\{2s-\dim X,0\}}.

This thus naturally leads to the following conjecture of Oberlin: For X⊂R2X\subset\mathbb{R}^2 Borel,

dim⁡Es(X)≤max⁡{2s−dim⁡X,0}.(5)\dim E_s(X)\leq\max\{2s-\dim X,0\}. \tag*{(5)}

He obtained the statement for dim⁡X≤1\dim X\leq1 and the special case s=dim⁡X2−ϵs=\frac{\dim X}{2}-\epsilon for all ϵ>0\epsilon>0 using restricted Radon transforms, which is what motivated this conjecture. The case s=dim⁡X2s=\frac{\dim X}{2} for the entire range of dim⁡X∈[0,2]\dim X\in[0,2] was obtained by Bourgain in 2010 [1] and can be thought of as an analogue of (2).

The analogue of (3) was proven by Kaufman in 1968 [4] who proved

dim⁡Es(X)≤s.(6)\dim E_s(X)\leq s. \tag*{(6)}

Notice however, heuristically, we should expect that the larger dim⁡X\dim X is, the smaller dim⁡Es(X)\dim E_s(X) should be (i.e. the larger the object, the fewer directions with small shadow). To this end, though it does not have a matching discrete analogue above, it is worth noting Falconer’s exceptional set estimate from 1982 [3]:

dim⁡Es(X)≤max⁡{1+s−dim⁡X,0}.\dim E_s(X)\leq\max\{1+s-\dim X,0\}.

Both Kaufman and Falconer’s projection theorems were quite difficult to improve upon for general ss. The fundamental problem in improving upon Kaufman’s exceptional set estimate is demonstrated by a hypothetical subring example. If A⊂RA\subset\mathbb{R} were an ss-dimensional subring, then for a∈Aa\in A and θ\theta in direction (1,a)(1,a) we have

Pθ(A×A)=ca(A+aA)=caAP_\theta(A\times A)=c_a(A+aA)=c_aA

where ca=(1+a2)−1/2c_a=(1+a^2)^{-1/2}. Thus, each such θ\theta is in Es(A×A)E_s(A\times A). In other words, s≤dim⁡Es(A×A)s\leq\dim E_s(A\times A), which shows (6) is sharp for this hypothetical example.

However, Bourgain’s discretized sum-product theorem [1] gives that there does not exist an ss-dimensional Borel subring of R\mathbb{R} for any s∈(0,1)s\in(0,1). Using this, Orponen and Shmerkin [8] proved a groundbreaking ϵ\epsilon-improvement on Kaufman’s theorem when s<dim⁡X≤1s<\dim X\leq1. They proved that for such ranges of ss and t=dim⁡Xt=\dim X, there exists an ϵ:=ϵ(s,t)\epsilon:=\epsilon(s,t) such that

dim⁡Es(X)≤s−ϵ.\dim E_s(X)\leq s-\epsilon.

This small improvement opened the floodgates on a number of problems in geometric measure theory, ultimately leading to the resolution of Oberlin’s conjecture (5) due to Orponen–Shmerkin [9] and Ren–Wang [10] in 2023.

A Peek Beyond the Floodgates

Let us conclude by noting a few directions one can go from here. Firstly, it is worth observing that at its core, exceptional set estimates grapple with the question: How can points lie on lines? To this end, exceptional sets are closely related to Furstenberg sets. The discrete analogue of the Furstenberg set problem asks: Given integers s≥2s\geq2 and t≥1t\geq1, how small can a finite set X⊂R2X\subset\mathbb{R}^2 be if there exists a family L\mathcal{L} of at least tt distinct lines such that ∣X∩ℓ∣≥s|X\cap\ell|\geq s for every ℓ∈L\ell\in\mathcal{L}. Using the Szemerédi–Trotter theorem, one can prove

∣X∣≳min⁡{st,s3/2t1/2}.|X|\gtrsim\min\{st,s^{3/2}t^{1/2}\}.

The Furstenberg set problem correspondingly asks: For $0<s\leq 1$ and 0<t≤20<t\leq2, given X⊂R2X\subset\mathbb{R}^2 such that there exists a set of lines $\mathcal{L} with dim⁡L=t\dim\mathcal{L}=t and dim⁡(X∩ℓ)≥s\dim(X\cap\ell)\ge s for all ℓ∈L\ell\in\mathcal{L}, how small can dim⁡X\dim X be? Building on work of Orponen–Shmerkin [9], Ren–Wang [10] proved the sharp bound for this problem as well:

dim⁡X≥min⁡{s+t,3s+t2,s+1}.\dim X \ge\min\left\{s+t,\frac{3s+t}{2},s+1\right\}.

The Furstenberg set problem and its connection to exceptional set estimates is a rich story that we do not go into further here.

Secondly, it is worth noting that projection theorists study exceptional sets estimates for many types of projections. One important example to this end involves radial projections πx:R2∖{x}→S±1\pi_x:\mathbb{R}^{2}\setminus\{x\}\to\mathbb{S}^{1}_{\pm} given by

πx(y)=y−x∣y−x∣.\pi_x(y)=\frac{y-x}{\lvert y-x\rvert}.

Notice that the fibers of these projections are yet again lines, which might lead one to ask if our incidence theoretic approach carries over to this setting. While this line of questioning does give a number of interesting estimates, for more we refer the reader to [9].

Finally, in higher dimensions one can consider orthogonal projections onto subspaces of a fixed dimension. To this end, the reader can verify that (3) and (4) hold replacing projections onto lines in R2\mathbb{R}^{2} with projections onto hyperplanes in Rn\mathbb{R}^{n}, as the fibers of such projections are lines, and thus amenable to the incidence estimates above. In this same vein, Kaufman’s estimate (6) holds for hyperplane projections as well, and the sharp analogue of the Szemerédi–Trotter-type estimate was recently obtained by Cholak et al. [2] using [9, 10]. However, the search for sharp exceptional set estimates onto high codimensional planes is still underway, making this an exceptionally exciting time to study projection theory.

References

  1. [1]Jean Bourgain. “The discretized sum-product and projection theorems”. In: Journal d’Analyse Mathématique 112.1 (2010), pp. 193–236. ISSN: 1565-8538. DOI: 10.1007/s11854-010-0028-x.DOI
  2. [2]Peter Cholak et al. “Bounding the dimension of exceptional sets for orthogonal projections”. In: Proceedings of the American Mathematical Society 154.10 (2026), pp. 4257–4266. ISSN: 0002-9939. doi: 10.1090/proc/17615.DOI
  3. [3]K. J. Falconer. “Hausdorff dimension and the exceptional set of projections”. In: Mathematika 29.1 (1982), pp. 109–115. DOI: 10.1112/S0025579300012201.DOI
  4. [4]Robert Kaufman. “On Hausdorff dimension of projections”. In: Mathematika 15.2 (1968), pp. 153–155.DOI
  5. [5]J. M. Marstrand. “Some fundamental geometrical properties of plane sets of fractional dimensions”. In: Proc. Lond. Math. Soc. (3) s3-4.1 (1954), pp. 257–302.DOI
  6. [6]Pertti Mattila. Geometry of Sets and Measures in Euclidean Spaces: Fractals and Rectifiability. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 1995.
  7. [7]Daniel Oberlin. “Restricted Radon Transforms and Projections of Planar Sets”. In: Canadian Mathematical Bulletin 55.4 (2012), pp. 815–820. ISSN: 1496-4287. DOI: 10.4153/cmb-2011-064-6.DOI
  8. [8]Tuomas Orponen and Pablo Shmerkin. “On the Hausdorff dimension of Furstenberg sets and orthogonal projections in the plane”. In: Duke Mathematical Journal 172.18 (2023). ISSN: 0012-7094. DOI: 10.1215/00127094-2022-0103.DOI
  9. [9]Tuomas Orponen and Pablo Shmerkin. “Projections, Furstenberg sets, and the ABC sum-product problem”. In: arXiv preprint arXiv:2301.10199 (2023).arxiv.org/abs/2301.10199
  10. [10]Kevin Ren and Hong Wang. “Furstenberg sets estimate in the plane”. In: arXiv preprint arXiv:2308.08819 (2023).arxiv.org/abs/2308.08819
  11. [11]Endre Szemerédi and W. T. Trotter. “Extremal problems in discrete geometry”. In: Combinatorica 3.3–4 (1983), pp. 381–392. ISSN: 1439-6912. DOI: 10.1007/bf02579194.DOI

Paper details

vlnc:W_9djupszm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents