Riemannian 3-spheres that are hard to sweep out by short curves

The approach to bounding $\operatorname{scg}(M)$ by min-max methods is essentially a quantitative version of the proof that $M$ contains a closed geodesic. The existence of closed geodesics in $M$ was proven by applying a “geometric calculus of variations” to the free loop space $\Lambda M$, which is the space of continuous maps $S^{1}\to M$ given an appropriate topology [41]. Let $\Lambda^{0}M$ denote the space of constant loops on $M$. Since closed geodesics are critical points in $\Lambda M$ with respect to the length functional, we can, in some sense, find those critical points by applying the calculus of variations to continuous maps $f:(D^{d},\partial D^{d})\to(\Lambda M,\Lambda^{0}M)$, where $f$ represents a nontrivial element in $\pi_{d}(\Lambda M,\Lambda^{0}M)$, and $d\geq 1$ is the minimal degree such that $\pi_{d}(\Lambda M,\Lambda^{0}M)\neq 0$. We call such a map $f$ a sweepout of $M$ by free loops. The notion of a sweepout allows us to define the min-max value

It can be proven that $\lambda(M)$ is the length of some nonconstant closed geodesic in $M$. Indeed, a similar argument was used by A. I. Fet and L. A. Lyusternik to prove that $M$ must have at least one nonconstant closed geodesic [40].¹¹1Surveys of their proof are available in [8, 49]. Thus $\operatorname{scg}(M)\leq\lambda(M)$, and a natural strategy to bound $\operatorname{scg}(M)$ is to bound $\lambda(M)$ by constructing a single sweepout of $M$ that is “efficient” in the sense that every free loop in the sweepout has bounded length.

Such a bound on $\lambda(M)$ may be possible for some combinations of geometric parameters of $M$ but not others. For example, as a consequence of work by Y. Liokumovich, Nabutovsky, and Rotman, every Riemannian 2-sphere $M=(S^{2},g)$ satisfies $\lambda(M)\leq 664\sqrt{\operatorname{area}(S^{2},g)}+2\operatorname{diam}(S^{2},g)$ [37, Theorem 1.3]. Yet $\lambda(S^{2},g)$ cannot be bounded in terms of diameter alone or area alone: Sabourau presented a family of Riemannian 2-spheres $(S^{2},g)$ for which, after scaling, $\operatorname{area}(S^{2},g)\leq 1$ but $\lambda(S^{2},g)$ grows without bound [55, Remark 4.10]. In a similar vein, Liokumovich constructed a family of Riemannian 2-spheres $(S^{2},g)$ for which $\operatorname{diam}(S^{2},g)\leq 1$ but $\lambda(S^{2},g)$ grows without bound [32]. Evidently, the geometry of closed geodesics in $M$ can be probed by either constructing efficient sweepouts, or by constructing “pathological” Riemannian manifolds on which any sweepout must be inefficient.

Theorem 1.1 gives a family of Riemannian 3-spheres $M$ for which $\operatorname{diam}(M),\operatorname{vol}(M)\leq 1$ but $\lambda(M)$ grows without bound. Such results proving that $\lambda(M)$ is “large” should be understood with the caveat that even if $\lambda(M)$ is large, $M$ may still contain short closed geodesics. This is illustrated by the aforementioned results on $\operatorname{scg}(S^{2},g)$ and $\lambda(S^{2},g)$. Furthermore, min-max values are in some sense the lengths of closed geodesics of “positive Morse index,” so they do not lead to estimates of the lengths of closed geodesics that are local minima of the length functional.

Our result also does not imply that $\operatorname{scg}(M)$ cannot be bounded by a fixed function of $\operatorname{diam}(M)$ and $\operatorname{vol}(M)$. What we can conclude is that if such a bound can be proven using min-max methods, then it is not enough to use sweepouts that represent nontrivial elements of $\pi_{2}(\Lambda M,\Lambda^{0}M)$. More complicated types of sweepouts would be necessary, such as maps $f:(X,X_{0})\to(\Lambda M,\Lambda^{0}M)$ where $X$ is not a disk, or sweepouts by objects like 1-cycles that generalize free loops.

Our constructions are inspired by some Riemannian $n$-disks whose boundaries are “hard to contract” through surfaces of controlled $(n-1)$-volume. These Riemannian disks were constructed to answer questions of Gromov [23] and P. Papasoglu [50] which were special cases of the following general question: Given a Riemannian $n$-disk $D$, can $\partial D$ be homotoped to a point through $D$ while passing through only surfaces whose $(n-1)$-volumes are bounded by a given combination of the geometric parameters of $D$ and $\partial D$?

Riemannian 2-disks whose boundaries are “hard to contract” through curves of length bounded in terms of the disk’s diameter were constructed by S. Frankel and M. Katz. More precisely, for each each $C>0$ they constructed a Riemannian 2-disk $D^{2}_{C}$ such that $\operatorname{diam}(D^{2}_{C}),\operatorname{length}(\partial D^{2}_{C})\leq 1$, but every nullhomotopy of $\partial D^{2}_{C}$ must pass through a curve longer than $C$ [18].²²2Nevertheless, the boundary of a Riemannian 2-disk $D$ can always be contracted through curves of length at most $\operatorname{length}(\partial D)+200\operatorname{diam}(D)\max\Big{\{}1,\ln\frac{\sqrt{\operatorname{area}(D)}}{\operatorname{diam}(D)}\Big{\}}$, as proven by Liokumovich, Nabutovsky, and Rotman [37]. Similarly in dimension 3, P. Glynn-Adey and Z. Zhu constructed a family of Riemannian 3-disks $D^{3}_{C}$, ranging over each $C>0$, such that $\operatorname{diam}(D^{3}_{C}),\operatorname{vol}(D^{3}_{C})\leq 10$ and $\operatorname{area}(\partial D^{3}_{C})=4\pi$, but every null-homotopy of $\partial D^{3}_{C}$ must pass through some surface of area greater than $C$ [21]. Their construction of $D^{3}_{C}$ was inspired by a previous construction by D. Burago and S. Ivanov of Riemannian 3-tori with small asymptotic isoperimetric constants [9].

Our constructions in Theorem 1.1 are essentially a combination of $D^{3}_{C}$ and a modification of $D^{2}_{C}$ by Liokumovich [34]. Let us compare our constructions with $D^{3}_{C}$ and their predecessor constructions by Burago and Ivanov. Burago and Ivanov constructed Riemannian 3-tori that contained 3 intertwining solid tori, each homeomorphic to $D^{2}\times S^{1}$. The metric on each solid torus is a product metric of a short metric on the $S^{1}$ factor with a Euclidean metric on the $D^{2}$ factor scaled by a large positive number. In contrast, $D^{3}_{C}$ is a Riemannian 3-disk containing two disjoint and linked solid tori $T_{1}\cup T_{2}$, each $T_{i}$ being homeomorphic to $D^{2}\times S^{1}$. The metric on each $T_{i}$ is the product of a short metric on the $S^{1}$ factor and a hyperbolic metric $g$ on the $D^{2}$ factor with negative curvature of a large magnitude, so that $(D^{2},g)$ has small diameter but large area. The metric $M$ we constructed is similar to $D^{3}_{C}$, except that in each $T_{i}$ we replace the hyperbolic metric on the $D^{2}$ factor with another metric with small diameter and large area.

Our choice of the metric on the $D^{2}$ factor in each $T_{i}$ can be motivated from the fact that we are considering maps $F:S^{3}\to M$ of nonzero degree and studying the lengths of the loops $F(S^{1}_{st})$. Each solid torus $T_{i}$ only “sees” part of each loop, namely $F(S^{1}_{st})\cap T_{i}$, which, assuming that the intersection is transverse, is either a closed curve or a union of arcs with endpoints on $\partial T_{i}$. In other words, $F(S^{1}_{st})\cap T_{i}$ is a relative 1-cycle in $(T_{i},\partial T_{i})$. These relative 1-cycles form a 2-parameter family ranging over $s$ and $t$; we will demonstrate that some 1-parameter subfamily of those relative 1-cycles will project onto $D^{2}$ to produce a “sweepout of $D^{2}$ by relative 1-cycles,” a notion that will be defined in Section 1.3. Proving this involves certain technical complications that we will resolve.

In light of the above, we will construct $M$ so that $D^{2}$ factor in each $T_{i}$ has a Riemannian metric that is difficult to sweep out by short relative 1-cycles. This would imply that one of the relative 1-cycles in some $T_{i}$, and therefore one of the $F(S^{1}_{st})$’s, has to be long. This metric on $D^{2}$ will be adapted from Riemannian 2-spheres that are hard to sweep out by short 1-cycles, which were constructed by Liokumovich [34] following inspiration from [18, 32].

Roughly speaking, if we take the min-max theory for sweepouts by free loops and replace free loops by relative cycles, we obtain a new min-max theory called Almgren-Pitts min-max theory that has been central to the study of minimal submanifolds.

More precisely, for any Riemannian $n$-manifold $M$ with boundary, consider the space of relative flat $k$-cycles in a $M$ with coefficients in $G$, denoted by $\mathcal{Z}_{k}(M,\partial M;G)$. Intuitively one can think of it as the group of relative singular $k$-cycles in $M$ endowed with a topology where two relative cycles are “close” when their difference can be filled by a $(k+1)$-chain of small volume in $M$; formal definitions are available in [16, 17]. One can define a sweepout of $M$ by relative $k$-cycles to be a continuous map $f:N\to\mathcal{Z}_{k}(M,\partial M;G)$ from a simplicial complex $N$ so that $f^{*}(\iota)\neq 0$, where $\iota\in H^{n-k}(\mathcal{Z}_{k}(M,\partial M;G);G)$ is the fundamental cohomology class of $\mathcal{Z}_{k}(M,\partial M;G)$. More generally, for any integer $p\geq 1$, $f$ is called a $p$-sweepout by cycles if $f^{*}(\iota^{p})\neq 0$, where $\iota^{p}$ denotes the $p^{\text{th}}$ cup power. A min-max value called the $p$-width can be defined as:

Henceforth we will write $\operatorname{width}_{p}^{k}(M)$ to denote $\operatorname{width}_{p}^{k}(M;\mathbb{Z})$ when $M$ is orientable, and $\operatorname{width}_{p}^{k}(M;\mathbb{Z}_{2})$ otherwise. The $p$-widths form a non-decreasing sequence: $\operatorname{width}_{1}^{k}(M)\leq\operatorname{width}_{2}^{k}(M)\leq\operatorname{width}_{3}^{k}(M)\leq\dotsb$

The $p$-widths of a closed Riemannian manifold $M$ are realized as the volumes of minimal submanifolds in $M$, which may contain a “small” singular set. The study of $p$-widths have led to existence proofs for minimal submanifolds and the solutions of several conjectures about minimal submanifolds. Almgren-Pitts Min-max theory and some its applications to these conjectures are surveyed in [12, 48].

Since every free loop in $M$ is an integral 1-cycle, we have $\operatorname{width}_{1}^{1}(M)\leq\lambda(M)$. Like $\lambda(M)$, curvature-free bounds on the $p$-widths of $M$ in terms of geometric parameters of $M$ are much better understood when $\dim M=2$. For every closed Riemannian surface $S$ and $p\geq 1$ we have $\operatorname{scg}(S)\leq\operatorname{width}_{p}^{1}(S)$ due to the recent work of O. Chodosh and C. Mantoulidis [11]. F. Balacheff and Sabourau proved that $\operatorname{width}_{1}^{1}(S)\leq 10^{8}\sqrt{\operatorname{genus}(S)+1}\sqrt{\operatorname{area}(S)}$ [2]. However, Liokumovich proved that $\operatorname{width}_{1}^{1}(S)$ (and, by extension, $\operatorname{width}_{p}^{1}(S)$ for all $p$) cannot be bounded solely in terms of $\operatorname{diam}(S)$, answering a question of Sabourau [55]: the counterexamples from [32] can be adapted into a family of Riemannian surfaces $S$ for which $\operatorname{diam}(S)=1$ but whose values of $\operatorname{width}_{1}^{1}(S)$ grow without bound [34]. This exemplifies the potential for bounds on $\lambda$, and counterexamples to conjectured bounds, to shed light on the $p$-widths, and by extension the geometry of minimal submanifolds.

We adapted Liokumovich’s Riemannian surfaces of small diameter but large 1-width into Riemannian 2-disks with small diameter but large 1-width. These Riemannian 2-disks are hard to sweep out by short relative 1-cycles, and they serve as ingredients in our construction, as explained at the end of Section 1.2.

For manifolds of dimension at least 3, bounds on some of their $p$-widths have been established under certain curvature assumptions [22, 24, 13, 20, 56, 36, 38]. However, there are no known curvature-free bounds on the $p$-widths of manifolds of dimension 3 and above in terms of geometric parameters such as diameter and volume. In fact, for closed 3-manifolds $M$ and any $p$, $\operatorname{width}_{p}^{2}(M;\mathbb{Z}_{2})$ cannot be bounded in terms of $\operatorname{diam}(M)$ and $\operatorname{vol}(M)$. This can be shown in two steps: roughly speaking, $M$ can be “cut” into two regions of equal volume by a hypersurface of area $\operatorname{width}_{1}^{2}(M;\mathbb{Z}_{2})$.³³3This follows due to an argument adapted from [34, p. 396]. On the other hand, Papasoglu and E. Swenson constructed, for each $C>0$, a Riemannian 3-sphere of diameter and volume at most 1 for which any such “cutting hypersurface” must have area greater than $C$ [51].⁴⁴4Riemannian 3-disks that are “hard to cut” in this manner were independently constructed by Glynn-Adey and Zhu [21], except that they only studied cutting hypersurfaces that were embedded disks.

The preceding argument would not apply to widths $\operatorname{width}_{p}^{k}(M)$ where $k\leq\dim M-2$. Nevertheless, our Theorem 1.1 may serve as a first step towards proving that the $p$-widths $\operatorname{width}_{p}^{1}$ of Riemannian 3-spheres cannot be bounded in terms of diameter and volume.

Our result may also offer insights in the study of waists, which are another class of min-max values that were defined by Gromov in [22]. They arise from sweepouts of Riemannian $n$-manifolds with boundary $M$ by relative $k$-cycles that are obtained as “slices” of $M$. More precisely, given a continuous map $M\to\mathbb{R}^{n-k}$, its fibers can be considered as “slices” of $M$. We can define the $k$-waist of orientable $n$-manifolds $M$ to be

where the infimum is taken over maps $\sigma$ whose fibers are Lipschitz $k$-cycles, such that the map $\mathbb{R}^{n-k}\to\mathcal{Z}_{k}(M,\partial M;\mathbb{Z})$ given by $t\mapsto\sigma^{-1}(t)$ is continuous.

The known bounds on waists in terms of the geometric parameters of $M$ follow a pattern similar to the other min-max values that we introduced earlier. By the definitions we have $\operatorname{width}_{1}^{k}(M)\leq\operatorname{waist}_{k}(M)$. For a Riemannian 2-sphere $(S^{2},g)$, Liokumovich proved that $\operatorname{waist}_{1}(S^{2},g)\leq 52\sqrt{\operatorname{area}(S^{2},g)}$ [33].⁵⁵5A similar bound on $\operatorname{waist}_{1}$ for other closed Riemannian surfaces is implicit via the monotone sweepouts involved in the proof of [35, Theorem 1.1]. For manifolds of dimension at least 3, some bounds on waists have been obtained under curvature assumptions [38, 36, 56].

There have been more results on $\operatorname{waist}_{k}(M)$ when $k=\dim M-1$; the aforementioned Riemannian 3-spheres that are “hard to cut” by hypersurfaces demonstrate that $\operatorname{waist}_{2}(M)$, which is bounded from below by $\operatorname{width}_{1}^{2}(M)$, cannot be bounded in terms of $\operatorname{diam}(M)$ and $\operatorname{vol}(M)$ when $\dim M=3$. On the flipside, it is currently an open question whether $\operatorname{waist}_{k}(M)$ can be bounded in terms of the geometric parameters of $M$ when $\dim M\geq 3$ and $k\leq\dim M-2$. Within this context, L. Guth conjectured that when $M$ is a Riemannian 3-torus, then $\operatorname{waist}_{1}(M)\leq C\sqrt[3]{\operatorname{vol}(M)}$ for some constant $C$ [25, Section 7]. Our Theorem 1.1 could serve as a first step towards disproving this conjecture.

Our result can be contrasted with a recent result of Nabutovsky, Rotman, and Sabourau which proves bounds on another min-max value related to waists. To define this min-max value, for each closed Riemannian $n$-manifold $M$ and an integer $k\geq 0$, consider a continuous map $F:N\to M$ of nonzero degree from a closed $n$-dimensional pseudomanifold $N$. “Slice” $N$ similar to before using a continuous map $\sigma:N\to K$ to a finite simplicial complex $K$ whose fibers $\sigma^{-1}(t)$ are $k$-dimensional simplicial complexes. Then define

where the infimum range over all $\sigma$ and $F$ that satisfy the stipulated criteria.

These min-max values are related to waists via $W_{k}(M)\leq\operatorname{waist}_{k}(M)$. The bounds $W_{k}(M)\leq c_{n}\sqrt[n]{\operatorname{vol}(M)}$ and $W_{k}(M)\leq c_{n}^{\prime}\operatorname{diam}(M)$ were proven by Nabutovsky, Rotman, and Sabourau for some dimensional constants $c_{n}$ and $c_{n}^{\prime}$ [47]. However, it remains to be seen whether $W_{k}(M)$ is attained as the volume of some minimal object. Since each $F(\sigma^{-1}(t))$ may not be a loop or a cycle (see [47, Example 1.2]), the min-max theory for minimal submanifolds does not fit here. For the same reason, there is no direct comparison between $W_{1}(M)$ and $\lambda(M)$ or $\operatorname{width}_{p}^{1}(M)$.

The techniques we used to prove Theorem 1.1 can be adapted for yet another min-max theory that arises from using sweepouts by paths instead of free loops or 1-cycles. This has led to a result about the geometry of orthogonal geodesic chords, which were defined earlier as geodesics in a Riemannian manifold $M$ that meet a fixed submanifold $N$ orthogonally at its endpoints. Orthogonal geodesic chords are 1-dimensional analogues of free boundary minimal submanifolds, which are submanifolds $P\subset M$ with vanishing mean curvature such that $\partial P\subset N$ and $\partial P$ meets $N$ orthogonally.⁶⁶6Some results about the existence and regularity of free boundary minimal submanifolds are surveyed in [31]. Orthogonal geodesic chords are also related to brake orbits, special types of periodic orbits in certain Hamiltonian systems [19].

When $\dim N=0$, an orthogonal geodesic chord is simply a geodesic with specified endpoints. The existence of such geodesics and bounds on their length were studied in [57, 59, 46, 45, 10, 4]. Lyusternik and L. Schnirelmann proved that every convex domain $M\subset\mathbb{R}^{n}$ with boundary $N$ contains $n$ orthogonal geodesic chords [39]. W. Bos extended this existence result to Riemannian $n$-disks $M$ with convex boundary $N$ [7]. For Riemannian 2-disks $M$ with strictly convex boundary $N$, J. Hass and P. Scott [27] and D. Ko [29] showed that one can even arrange the orthogonal geodesic chords to be simple, that is, avoid self-intersection.

The existence and geometry of orthogonal geodesic chords may be probed using min-max techniques as follows. Let $\Omega_{N}M$ denote the space of piecewise smooth paths in $M$ that start and end on $N$, topologized as in [41, p. 88]. Let $d\geq 1$ be the smallest degree for which $\pi_{d}(\Omega_{N}M,\Lambda^{0}N)\neq 0$. Then define

where $E$ is the energy of a path, $E(\alpha)=\int_{0}^{1}\lVert\alpha^{\prime}(t)\rVert^{2}\,dt$. (Using length instead of energy would not significantly affect the resulting min-max theory, because the Hölder inequality relates length to energy.) X. Zhou proved that when $M$ is a complete and homogeneously regular Riemannian manifold and $N$ is a closed submanifold such that $\pi_{1}(M,N)=0$ and $\pi_{2}(M,N)\neq 0$, then $\lambda_{\mathrm{rel}}(M,N)>0$ is the energy of an orthogonal geodesic chord [61].

Recent results have estimated the lengths of orthogonal geodesics chords in various spaces $M$ and $N$, where $N$ may or may not be $\partial M$. When $M$ is a Riemannian 2-disk with convex boundary $N$, I. Beach proved that $M$ contains at least two distinct simple orthogonal geodesic chords whose lengths are bounded by $f(\operatorname{diam}(M),\operatorname{area}(M),\operatorname{length}(\partial M))$ for some function $f$ [5]. In addition, recent work by Beach, H. C. Peruyero, E. Griffin, M. Kerr, Rotman, and C. Searle implies that when $M$ is a closed Riemannian manifold and $N$ is an analytic 2-sphere embedded in $M$, then $M$ contains an orthogonal geodesic chord whose length is bounded by $f^{\prime}(\dim M,\operatorname{diam}(M),\operatorname{diam}(N),\operatorname{area}(N))$ for some function $f^{\prime}$ [6]. Both of these results were obtained by proving that either $\lambda_{\mathrm{rel}}(M,N)$ is bounded by the relevant function ($f$ or $f^{\prime}$) due to the existence of a sweepout of $M$ by curves of energy (equivalently, length) bounded by that function, or else the obstruction to the existence of such a sweepout is an orthogonal geodesic chord of length bounded by that function.

When $M$ is a Riemannian 3-sphere and $N$ is an embedded sphere, we proved that $\lambda_{\mathrm{rel}}(M,N)$ cannot be bounded by any function of $\operatorname{vol}(M)$, $\operatorname{diam}(M)$, and $\operatorname{diam}(N)$.