LOGARITHMIC IMPROVEMENTS IN THE WEYL LAW AND

EXPONENTIAL BOUNDS ON THE NUMBER OF CLOSED GEODESICS

ARE PREDOMINANT
arXiv:2204.11921v1 [math.DS] 25 Apr 2022

YAIZA CANZANI AND JEFFREY GALKOWSKI

Abstract. Let M be a smooth compact manifold of dimension d without boundary. We in-

troduce the concept of predominance for Riemannian metrics on M , a notion analogous to full
Lebesgue measure which, in particular, implies density. We show that there is Ω > 1 such that,
for a predominant C ν metric, the number, c(T, g), of closed geodesics of length ≤ T satisfies
log c(T, g) = O T Ω .

In addition, for g a Riemannian metric on M , let 0 = λ21 (g) < λ22 (g) ≤ λ23 (g) . . . be the
eigenvalues of −∆g . The Weyl law states that there is cd > 0 such that
#{j : λj (g) ≤ λ} = cd volg (M )λd + E(λ, g)
with E(λ, g) = O(λd−1 ) as λ → ∞. We show that for ν > 0 large enough there is Ω > 1 such
that for a predominant C ν metric
E(λ, g) = O(λd−1 /(log λ)1/Ω ).
After an application of recent results of the authors in the case of the Weyl law [CG20],
these estimates follow from a study of the non-degeneracy properties of nearly closed orbits for
predominant sets of metrics.

1. Introduction

We study properties of the geodesic flow and remainders in the Weyl law for ‘typical’ metrics on
a compact manifold without boundary. Since the space of Riemannian metrics, G , on a manifold
cannot be endowed with a non-trivial, translation invariant Borel measure, we introduce an analog
of full Lebesgue measure in infinite dimensions called predominance. We then study properties of
the geodesic flow and remainders in the Weyl law for predominant sets of metrics.
The notion of predominance has three important properties: 1) any predominant set is dense, 2)
a finite intersection of predominant sets is predominant, and 3) in finite dimensions, a predominant
set has full Lebesgue measure. Heuristically, a set G ⊂ G is predominant if there is a family of
submanifolds endowed with finite Borel measures {(Γg , µg )}g∈G such that g ∈ Γg , µg assigns a
positive measure to any neighborhood of g, the map g 7→ Γg is C 1 , and G ∩ Γg has full µg measure
for every g ∈ G . For the careful definition of this concept, see Definition 2.4 and Remark 2.5.

Remainders in the Weyl law. Let ν ≥ 0 and M be a compact C ν manifold without boundary,
of dimension d. Let G ν denote the space of C ν Riemannian metrics on M with the topology
induced from the C ν norm on symmetric tensors. For g ∈ G ν , let −∆g denote the (positive)
Laplace–Beltrami operator with eigenvalues 0 = λ21 (g) < λ22 (g) ≤ λ23 (g) ≤ . . . . Then, for λ > 0
1
2 YAIZA CANZANI AND JEFFREY GALKOWSKI

define the eigenvalue counting function

N (λ, g) := #{j : λj (g) ≤ λ}.
Our first theorem shows that the Weyl law has a logarithmic improvement for a predominant set
of metrics. In what follows, B1 denotes the ball of radius 1 in Rd , volRd (B1 ) denotes its volume,
and volg (M ) denotes the volume of M as measured by the metric g.
Theorem 1.1. Let d ≥ 2. There is ν0 > 0 and for all ν > 0 there is Ων > 0 such that the
following holds. If ν ≥ ν0 , M is a compact C ν -manifold of dimension d without boundary, and
Ω > Ων , then there is a predominant set GΩ ⊂ G ν (see Definition 2.4) such that for every g ∈ GΩ
 1

N (λ, g) = (2π)−d volg (M ) volRd (B1 )λd + Og λd−1 (log λ) Ω , λ → ∞.
In particular, GΩ is dense in G ν .

The constant Ων in Theorem 1.1 is explicit, and we can take

Ων := 1 + log2 (2(d − 1)(2d + 1)(max(ν, 6) + 3d − 1)) − log2 (2d − 1). (1.1)
Remark 1.1. The authors wish to stress that, although our original motivation was to study
typical properties of remainders in the Weyl law, the analysis in this article is dynamical in nature
and studies predominant properties of the geodesic flow. Once these dynamical properties are
established in Theorem 1.3 below, a direct application of the authors’ work [CG20, Theorem 2]
produces the Weyl remainder estimate of Theorem 1.1 as a corollary (See Section 3.3).

Defining E(λ, g) := N (λ, g) − (2π)−d volg (M ) volRd (B1 )λd , the Weyl law states that, for suffi-
ciently smooth metrics, E(λ, g) = Og (λd−1 ). This estimate is sharp on the round sphere and has
a long history dating back to the work of Weyl [Wey12], who proved (in a slightly different con-
text) that E(λ, g) = o(λd ). The estimate E(λ, g) = Og (λd−1 ) was proved by Levitan [Lev52] and
Avakumović [Ava56] after which Hörmander [Hör68] provided a general framework for studying
such remainders, reproving this estimate and making far reaching generalizations.
Using this framework, Bérard [Bér77] showed that E(λ, g) = Og (λd−1 / log λ) on both surfaces
without conjugate points and non-positively curved manifolds of any dimension. Duistermaat–
Guillemin [DG75] showed that E(λ, g) = o(λd−1 ) provided that the measure of the set of closed
geodesics in S ∗M is 0. Fifteen years later, Volovoy [Vol90a] provided estimates under dynamical
conditions guaranteeing that E(λ, g) = Og (λd−1 / log λ) and verified these conditions for certain
specific examples in [Vol90b]. The recent work of Bonthenneau [Bon17] improved a geometric
estimate in Bérard’s work, thus generalizing his result to manifolds without conjugate points of
any dimension. Finally, in [CG20], the authors provided estimates on E(λ, g) under assumptions
on the volume of nearly closed geodesics which improve the results of [DG75, Bér77, Vol90a].
For manifolds with boundary, the analog of [DG75] was proved by Ivrii [Ivr80]. (For a more
comprehensive history of the Weyl law, see [Ivr16].)
Manifolds where there are known polynomial improvements of the form E(λ, g) = Og (λd−1−ε )
are very rare. For instance, such estimates hold on the torus [Hux03, BW17], products of
spheres [IW21], and other special integrable systems [Vol90b]. Nevertheless, it has long been ex-
pected that, for a ‘typical’ metric g, there exists ε > 0 such that E(λ, g) = Og (λd−1−ε ). However,
until now, the best available result is that E(λ, g) = o(λd−1 ) for a Baire-generic set of g. This can
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 3

be recovered from the work of Sogge–Zelditch [SZ02] or can be seen by combining the remainder
estimates in [DG75] with the bumpy metric theorem of Anosov and Abraham [Ano82, Abr70].
Theorem 1.1 improves on these bounds in two important ways. First, Baire genericity is
replaced by the concept of predominance which is an analog of full Lebesgue measure in infinite
dimensions. Just as in finite dimensions a full Lebesgue measure set is much more ‘typical’ than a
Baire generic one (indeed, a Baire generic set may have measure 0), a predominant set in infinite
dimensions is much more ‘typical’ than a Baire generic set. Second, although the change from
o(λd−1 ) to O(λd−1 /(log λ)1/Ω ) may seem small, this improvement requires the development of new
ideas and requires subtle dynamical estimates. In addition, it is the only quantitative remainder
estimate available for typical metrics.
Remark 1.2. We have not attempted to make the value of ν0 from in Theorem 1.1 explicit.
However, it is likely that, following the arguments in [CG20] carefully, ν0 can be taken to be
ν0 = Cd for some C > 0.
Remark 1.3. Although we have kept careful track of the constant Ων in (1.1), we do not expect
it to be optimal. Indeed, we conjecture that Ων could be replaced by 1 + ε for any ε > 0. This
would also allow us to obtain the same estimate for predominant sets in G ∞ . At the moment,
when working in G ∞ , we obtain weaker remainder estimates for predominant sets of metrics (see
Remark 1.8).

Growth of the number of periodic geodesics. We next discuss the growth of the number
of periodic geodesics of a given length. We say that a geodesic γ ⊂ M is a primitive periodic
geodesic with length T > 0, if there is a diffeomorphism h : R/T Z → γ, such that |ḣ|g(h(t)) = 1,

(h(0), ḣ(0)) = (h(T ), ḣ(T )), (h(0), ḣ(0)) 6= (h(t), ḣ(t)) for t ∈ (0, T ).
That is, γ is a periodic geodesic and T is its minimal period. For T > 0 and g a Riemannian
metric on M let
c(T, g) := #{γ : γ is a primitive periodic geodesic for g with length ≤ T }.

We obtain the following theorem on the growth of c(T, g).

Theorem 1.2. Let ν ≥ 5, M be a C ν - compact manifold of dimension d without boundary, and
Ων as in (1.1). Then, for each Ω > Ων there is a predominant set GΩ ⊂ G ν such that for g ∈ GΩ
there exists C > 0 such that for all T > 0


c(T, g) ≤ exp CT Ω .
In particular, GΩ is dense in G ν .

Bounds on c(T, g) have a long history in the literature. For Baire generic metrics, a complete
picture of the non-quantitative behavior of c(T, g) is available. The bumpy metric theorem [Ano82,
Abr70] can be used to show that for Baire generic smooth metrics, g, and all T > 0, c(T, g) < ∞
(see also [KT72]). Furthermore, Hingston [Hin84] showed that limT →∞ c(T, g) = ∞ for a Baire-
generic set of metrics. Petkov and Stojanov have studied similar properties of closed billiards in
generic domains in Rd [Sto87, PS87a, PS87b].
4 YAIZA CANZANI AND JEFFREY GALKOWSKI

As with remainders in the Weyl law, quantitative estimates on c(T, g) are much more subtle.
Only recently, Contreras [Con10] showed that for g in an open dense subset of G 2 there is c > 0
such that
log c(T, g) ≥ cT − c. (1.2)
One can check that this lower bound is optimal for any dense set of metrics (so certainly for any
reasonable notion of a typical metric, including predominance). Indeed, in the case of manifolds
with negative curvature, the works of Margulis [Mar69] and Bowen [Bow72] show that, for such
a metric g, there is α > 0 such that
1
lim log c(T, g) = α.
T →∞ T

In particular, this shows that there are open sets of metrics such that c(T, g) grows exactly
exponentially and hence that [Con10] gives a complete picture for lower bounds on c(T, g) for
typical C 2 metrics. (See also [Kni98] for the case of compact rank 1 manifolds.)
On the other hand, as far as the authors are aware, Theorem 1.2 is the first quantitative
upper bound on c(T, g) for ‘typical’ metrics. One reason that lower bounds on c(T, g) are well
understood, but upper bounds are not is that one can find a structure, called a hyperbolic basic
set (see [Con10]), which is stable under perturbation and guarantees the lower bound (1.2).
Moreover, the existence of such a set can be guaranteed by studying the Poincaré maps associated
to genuinely periodic orbits.
Unfortunately, no such structure for upper bounds exists and one must understand not only
Poincaré maps of periodic orbits but also those of near periodic orbits. Indeed, although there is
no rigorous proof at present, there is strong evidence that there is no quantitative upper bound
on c(T, g) which is Baire generic. For instance, in the case of diffeomorphims, Kaloshin [Kal00]
showed that there is no growth rate for the number of periodic points that is Baire generic. (See
Section 2 for an additional heuristic discussion for lack of Baire genericity.)

Non-degeneracy of nearly periodic orbits. Our main theorem, which implies both Theo-
rems 1.1 and 1.2, controls how close two periodic orbits may be for a predominant set of metrics
and can be used, for instance, to control the volume of nearly periodic orbits (see Section 3.1).
The result will be stated in terms of how close dϕgt may be to the identity for this set of metrics,
where ϕgt is the geodesic flow for the metric g at time t acting on S ∗M .
To understand how this is connected to the distance between periodic orbits, let H|ξ|g denote
the Hamiltonian vector field associated to |ξ|g , and ϕgt := exp(tH|ξ|g ) : S ∗M → S ∗M . Observe
that if ρ is a t periodic point (i.e. ϕgt (ρ) = ρ for some t > 0), and v ∈ Tρ S ∗M/RH|ξ|g (ρ) is not in
the kernel of I − dϕgt , then any perturbation of the initial point ρ in the direction of v will not be
periodic with period near t. (See Figure 2 for a schematic of such an orbit.) In particular, if the
map
I − dϕgt : Tρ S ∗M/RH|ξ|g (ρ) → Tρ S ∗M/RH|ξ|g (ρ)
is invertible, then there are no vectors in this kernel and hence every small perturbation of ρ other
than those along H|ξ|g (ρ) will produce a point which is not periodic with period near t.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 5

Motivated by this, we say that a t periodic point, ρ, is non-degenerate if

I − dϕgt : Tρ S ∗M/RH|ξ|g (ρ) → Tρ S ∗M/RH|ξ|g (ρ)
is invertible. The famous bumpy metric theorem [Abr70, Ano82] states that, for a Baire generic
set of smooth metrics, every periodic trajectory is non-degenerate. In particular, this implies the
finiteness of the set of closed geodesics of any bounded length.
Since we are interested in a quantitative version of non-degeneracy for both periodic geodesics
and nearly periodic geodesics, we need to introduce a few concepts to make precise statements
about non-degeneracy.
Definition 1.4 (returning points). Let g ∈ G ν and β > 0. For t ∈ R and ρ ∈ S ∗M we write
ρ ∈ R(t, β, g) if d(ϕgt (ρ), ρ) < β.
In this case, we say that ρ is β-returning for g at time t.

We also recall that the geodesic flow is a contact flow on S ∗M and thus there is a natural smooth
decomposition of TS ∗M preserved by the geodesic flow (see e.g. [Pat99]). In particular, ξdx|T S ∗ M
is a contact form for S ∗M with the geodesic flow as its Reeb flow. Thus, for all ρ ∈ S ∗M ,
Tρ S ∗M = H (ρ) ⊕ RH|ξ|g (ρ),
where H (ρ) := ker(ξdx|Tρ S ∗M ) and ⊕ denotes the direct sum, and we have
dϕgt (H (ρ)) = H (ϕgt (ρ)).
Since we work with nearly periodic orbits, we need to identify the tangent spaces at ρ and ϕgt (ρ)
when they are close. Let g ∈ G ν and U ⊂ S ∗M ×S ∗M . We say that W U = {Wρ2 ,ρ1 : (ρ2 , ρ1 ) ∈ U }
is a family of transition maps for g on U if for each (ρ2 , ρ1 ) ∈ U the map Wρ2 ,ρ1 is an invertible
linear transformation,
Wρ2 ,ρ1 : Tρ1 S ∗M → Tρ2 S ∗M, (ρ2 , ρ1 ) 7→ Wρ2 ,ρ1 is Lipschitz
(1.3)
Wρ1 ,ρ1 = I, Wρ2 ,ρ1 H|ξ|g (ρ1 ) = H|ξ|g (ρ2 ), Wρ2 ,ρ1 H (ρ1 ) = H (ρ2 ).
Here, by asking that (ρ2 , ρ1 ) 7→ Wρ2 ,ρ1 be Lipschitz, we mean that for any choice of coordinates
ψi : Wi → Vi ⊂ R2d−1 near ρi the map
W1 × W2 ∋ (x1 , x2 ) 7→ dψ2 |ρ=ψ−1 (x2 ) ◦ Wψ−1 (x2 ),ψ1 (x1 ) ◦ d(ψ1−1 (x))|x=x1 ∈ GL(2d − 1)
2 2

is a Lipschitz family of matrices.

We say that a collection W = {W Ui }N ∗ ∗
i=1 is a family of transition maps for g if Ui ⊂ S M × S M
is open,
[N
∗
{(ρ, ρ) : ρ ∈ S M } ⊂ Ui ,
i=1
and, for each i, W Ui is a family of transition maps for g on Ui . We say W is a β0 -family of
transitions maps if for each pair (ρ2 , ρ1 ) with d(ρ2 , ρ1 ) < β0 , there is i such that (ρ2 , ρ1 ) ∈ Ui .
Remark 1.5. It will be convenient throughout the text to have a fixed reference metric on M .
For this, we choose some gf ∈ G ν and whenever we refer to a norm | · |, on T ∗M or T M , it is the
′
one induced by gf . We will, in particular, use this metric to define C ν norms for ν ′ ≤ ν.
6 YAIZA CANZANI AND JEFFREY GALKOWSKI

We are now in a position to define non-degeneracy of a nearly periodic orbit.

Definition 1.6 (non-degenerate points). Let g ∈ G ν , β0 > 0 and W = {W Ui }N i=1 be a β0 -
family of transition maps for g. Let 0 < β < β0 , t ∈ R, ρ ∈ R(t, β, g), and α > 0. We write
ρ ∈ N (t, α, (g, W)) if for every i ∈ {1, . . . N } such that (ϕgt (ρ), ρ) ∈ Ui ,


|Πϕg (ρ) WϕUgi (ρ),ρ − dϕgt v| ≥ α|Πρ v|, v ∈ Tρ S ∗M.
t t

Here, Πρ : Tρ S ∗M → Tρ S ∗M/RH p (ρ) denotes the natural projection map and, by an abuse of
notation, |·| denotes the norm induced by the metric gf fixed above in Remark 1.5. In this case we
say ρ is α non-degenerate for (g, W) at time t. (See Figure 2 for an example of a non-degenerate
orbit.)
Remark 1.7. Although Definition 1.6 depends on the choice of the metric gf , note that the norm
induced by any other gf′ is comparable to that induce by gf .

The main theorem of this article shows that there is a predominant set of metrics such that every
sufficiently returning geodesic is non-degenerate with the degree of non-degeneracy depending
explicitly on the length of the trajectory. As far as the authors are aware, this theorem is the first
quantitative estimate on non-degeneracy of orbits for typical metrics.
Theorem 1.3. Let ν ≥ 5, M be a compact C ν manifold of dimension d without boundary, and
Ων as in (1.1). Then for every Ω > Ων there is a predominant set GΩ ⊂ G ν such that for all
g∈ GΩ and every family of transition maps W for g there are C, c > 0 such that



R t, β(t), g ⊂ N t, β(t), (g, W) for t > c
Ω −1
where β(t) := C −C(t+1) , and
d(ϕgt (ρ), ρ) ≥ c|t| for |t| ≤ c.
In particular, GΩ is dense in G ν .
Remark 1.8. The reason for the growth of Ων as ν → ∞ in Theorems 1.1, 1.2, and 1.3 comes
from the fact that we make perturbations to the metric at increasingly small scales as the length of
trajectories goes to infinity. Because of this, the size of these perturbations in C ν grows as ν → ∞
and this in turn results in weaker non-degeneracy statements. Moreover, with f (t) : [0, ∞) →
[0, ∞) growing faster than any polynomial in t, if one replaces β(t) by CeCf (t) in Theorem 1.3,
then one can work in C ∞ . That is, we obtain predominance in the G ∞ topology.
Remark 1.9. The notion of predominance (see Section 2) involves using certain families of
perturbations to probe the space of metrics. In this article, the predominance in Theorem 1.3
involves probing with the families of perturbations described in Section 9. However, one may
wonder whether one can probe with other families of perturbations (e.g. conformal perturbations)
and still obtain Theorem 1.3 for that family of probes. In fact, in Sections 7 and 8, we prove
Proposition 8.9 which gives a result analogous to Theorem 1.3 under the assumption that a family
of perturbations of metrics satisfies some abstract assumptions (see Definitions 7.2 and 7.3). The
type of perturbations used to probe the space of metrics is then tied to the family of perturbations
satisfying our abstract assumptions. It is only in Section 9 that we construct such a family of
metric perturbations and it is likely that many other families of perturbations suffice. However,
we do not pursue this here.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 7

Outline of the paper. In Section 2, we define predominance and discuss the reasons for in-
troducing this notion: predominance is in some sense more ‘typical’ than Baire generic and our
results are unlikely to hold for a Baire generic set of metrics. Then, in Section 3, we use The-
orem 1.3 to prove Theorems 1.1 and 1.2. Before starting the proof of Theorem 1.3, we give a
detailed description of the ideas used in Section 4.
To begin the proof of Theorem 1.3, we study volumes of relevant sets of symplectic matrices
in Section 5. In Section 6, we review some basic estimates for returning points and introduce the
notion of a chain of symplectomorphisms associated to a flow, as well as that of a well-separated
set. The notion of a well-separated set replaces that of a Poincaré section when a global section
for the flow is not available. Although this requires some technical work, it does not substantially
change the proof of the main result and the reader may wish to first assume that there is a global
section and replace chains of symplectomorphisms by the standard Poincaré map for that section.
Section 7 defines sufficient assumptions on a family of metric perturbations to guarantee Theo-
rem 1.3. Under these assumptions, we study the volume of perturbations that produce degenerate
periodic points of a given length. Section 8 then proves an analog of Theorem 1.3 by implementing
a delicate induction argument.
Finally, in Sections 9 and 10, we construct a family of metric perturbations which satisfies our
technical assumptions and we prove Theorem 1.3 in Section 11.
Appendix A contains some elementary control estimates from ODE theory used to construct
the perturbations of metrics in Section 9.

Index of Notation.
R(t, β, g) Def. 1.4 Wρ2 ,ρ1 (1.3) N (t, α(g, W)) Def. 1.6
probing map Def. 2.2 predominant Def. 2.4 (β, q) non-degenerate (4.1)
MY (V, s) Def. 5.4 R̥ (n, δ, g) Def. 6.1 S̥ (n, α, g) Def. 6.2
well-separated Def. 6.5 PI(n) [g] Def 6.8 TI(n) [g](ρ) Def. 6.9
N ̥ (n, β, g) Def. 6.14 N q,̥ (n, β, g) Def. 6.15 L (7.1)
Good perturbationDef. 7.2 Admissible pairs Def. 7.3 FJRε ,δε (8.8)
Rε ,δε
F∞ (8.9) b (8.18) γj (8.19)
βj,j (8.21) αj , αj,ℓ , βi,j,ℓ , β̃j,ℓ , sℓ (8.22) Φgρ⋆ (9.1)
gσ (9.4) ∆σ , ∆˜σ (9.9) Ψgζ00 (σ), Ξi (σ, ζ0 , g0 ) (9.10)

Acknowledgements. The authors would like to thank Steve Zelditch, Maciej Zworski, and
Leonid Parnovski for comments on an early draft of this paper, Gabriel Paternain for pointing
them to the work of Kaloshin–Hunt [KH07] and Kaloshin [Kal00] at the beginning of their work
on this article, and Peter Sarnak for helpful discussions on the problem. The authors are grateful
to Gerhard Knieper, Amie Wilkinson, Keith Burns, and Nancy Hingston for comments on the
existing literature for bounds on numbers of closed geodesics. Y.C. was supported by the Alfred
P. Sloan Foundation, NSF CAREER Grant DMS-2045494, and NSF Grant DMS-1900519. J.G.
is grateful to the EPSRC for partial funding under Early Career Fellowship EP/V001760/1 and
Standard Grant EP/V051636/1.
8 YAIZA CANZANI AND JEFFREY GALKOWSKI

2. Predominance in Banach Spaces

The main goals of this article are to give upper bounds on the number of closed geodesics of
length T and upper bounds for remainders in the Weyl law for a predominant set of metrics on a
compact manifold M . Before proceeding to define our notion of predominance on Banach spaces
in Section 2.2, we discuss several other available notions and motivate our choice of definition in
Section 2.1. Finally, in Section 2.3 we explain why the notion of Baire genericity is not well-suited
for our purposes.

2.1. Existing notions of ‘full measure’. The main difficulty in defining a concept analogous
full measure in an infinite dimensional space, like the space of Riemannian metrics over a given
manifold, is that there are no non-trivial, translation invariant, Borel measures. Several possible
fixes for this problem have been introduced in the literature. We mention here the concepts of
prevalence [HSY92] which uses an underlying linear structure and metric prevalence [Kal97] which
does not require such a structure. These two notions have three important properties
(1) A prevalent set is dense.
(2) The intersection of prevalent sets is prevalent. (2.1)
n
(3) If G ⊂ R is prevalent, then G has full Lebesgue measure.

Although quite flexible, as far as we are aware, the notion of metric prevalence has not proved
useful in studying quantitative statements such as the growth of the number of periodic orbits of
length T . Because of this, we focus on the notion of prevalence from [HSY92] which has appeared
before in this type of application. In fact, for ν ≥ 1, [KH07] proves that there is a prevalent set
G ⊂ C ν (I; I) of diffeomorphisms on the interval I = [0, 1], such that for all ε > 0 and f ∈ G there
is C > 0 such that
1+ε
#{x ∈ I : f n (x) = x} ≤ CeCn .

Given a Banach space G , a Borel set G ⊂ G is said to be prevalent if there is a Borel measure
µ and a compact set K ⊂ G such that

0 < µ(K) < ∞, G + g has full µ measure for all g ∈ G .

Usually, when one shows that a set G is prevalent, it is convenient to construct a probe Σ ⊂ G
which carries the measure µ. In other words, if G is prevalent, there are Σ ⊂ G , a smooth map

F : G × Σ → G, F (g, σ) := g + σ (2.2)

and a Borel measure, µΣ , on Σ such that for all g ∈ G

µΣ (σ ∈ Σ : F (g, σ) ∈ Gc ) = 0. (2.3)

Because we will be working in an open subset of a Banach space (the space metrics inside the
space of symmetric 2-tensors) we would like a notion which does not rely on the fact that the space
is linear. To do this, we generalize the type of functions allowed in (2.2) and slightly weaken (2.3).
We now define the notion of predominance on an open subset of a Banach space G .
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 9

2.2. Predominant sets. In this section we introduce the notion of predominance. Let G and
G ′ be open subsets of the Banach spaces (B, k kB ) and (B ′ , k kB′ ) respectively,
G ⊂ B, G ′ ⊂ B ′, (2.4)
and such that G ⊂ G ′ and B ⊂ B ′ via a continuous embedding, with B dense in B ′ and G dense
in G ′ . Let ι : G → G ′ be the natural inclusion map.
Remark 2.1. When O ⊂ B is open, we will say that a subset K ⊂ O is bounded if it is bounded
as a subset of B and dB (K, ∂O) > 0, where dB is the distance induced by k · kB .

The space G will be probed by perturbations indexed by a parameter σ ∈ ΣNε for ΣNε as
follows. Given L ∈ N and {Nε }ε>0 ⊂ N ∪ {∞} let
Nε
Y
ΣNε := BRL (0, 1). (2.5)
j=1

We endow this space with the sup-norm k · kℓ∞ .

Definition 2.2 (probing maps). We say that a collection F = {(Fε , Nε )}ε , with ε ∈ (0, 1), is
a family of G ′ -probing maps for G if Nε ≤ dim B, there is a collection of closed bounded sets
Gε ⊂ G with ∪ε Gε = G , Gε2 ⊂ Gε1 for ε1 < ε2 , and such that Fε : Gε × ΣNε → G is a continuous
map with
Fε (g, 0) = g, g ∈ Gε , (2.6)
and the following hold. For all K ⊂ G bounded,
lim sup sup kFε (g, σ) − gkB = 0, (2.7)
ε→0+ g∈K σ∈Σ
Nε

kFε (g, σ1 ) − Fε (g, σ2 )kB′

lim sup sup = 0.
ε→0+ g∈K σ1 ,σ2 ∈ΣNε kσ1 − σ2 kℓ∞ (2.8)
σ1 6=σ2

In addition, the map F̃ε := ι ◦ Fε : Gε × ΣNε → G ′ is Lipschitz, and satisfies that for all g ∈ Gε
the Frechet derivative, Dg F̃ε , of F̃ε in g exists and for all K ⊂ B bounded
lim sup kDg F̃ε |(g,σ) − IkB′ →B′ = 0, (2.9)
ε→0+ g∈K,σ∈Σ
Nε

Remark 2.3. We typically imagine that Nε is non-decreasing as ε → 0+ so that probing maps

become more dispersed as ε → 0. In fact, in our applications, Nε will often be identically equal
to ∞. However, when G is finite dimensional it is reasonable to assume that Nε ≤ dim G since,
otherwise, one would be putting a very diffuse measure on a finite dimensional space.
In addition, the spaces Gε allow us to define probes only in bounded subsets of G provided
that, as ε → 0+ , these subsets exhaust G .

In our treatment, G and G ′ will be the spaces of C ν and C ν−1 Riemannian metrics on a given
manifold, while B and B ′ will be the spaces of C ν and C ν−1 symmetric two-tensors. We need to
refer to G ′ because our probing maps will typically not be Frechet differentiable as a map from
10 YAIZA CANZANI AND JEFFREY GALKOWSKI

G
Fε (g0 , Σε )
Fε (g2 , Σε )
Fε (g1 , Σε )
g0
g1 g2

Figure 1. An example of a probing family Fε at a fixed ε and a thin set L. One

way to think of a probing map is that, to each point g ∈ G , we attach a probe
Fε (g, ΣNℵε ). These probes are asymptotically translates of one another in the limit
ε → 0 and, moreover, are contained in a small ball around g. A set L is thin if its
intersection with each of these probes has vanishing measure in the limit ε → 0.

C ν to C ν and instead, the Frechet derivative of the map will make sense as a map from C ν to C ν−1
and will extend as in (2.9) to a map from C ν−1 to C ν−1 .
We next discuss briefly the roles of each piece of the definition of a family of probing maps.
The assumption (2.6) is crucial to know that Fε probes all of Gε and does not avoid any open
sets. To understand (2.9) and (2.8), recall [Mil97] that one can construct an example of a foliation
of the unit square by analytic leaves {Wg }g∈[0,1] , such that the map g 7→ Wg is continuous (but
not differentiable), and there is a set E ⊂ [0, 1]2 with full measure such that #{Wσ ∩ E} ≤
1. Thus, assumptions (2.9) and (2.8), which imply that the map Fε has reasonable regularity
properties (both as a function of g and ε), are crucial in proving that item (3) in list (2.1) holds
for predominant sets.
In what follows we work with the measure mΣ on ΣNε defined to be the product measure
Nε

mRL |B(0,1)
mΣ := ⊗N ε
j=1 m, m := ,
Nε mRL(B(0, 1))

where mRL denotes the Lebesgue measure on RL . Note that mΣε (Σε ) = 1.

Definition 2.4 (predominant sets). Let G be an open subset of a Banach space B, and F :=
{(Fε , Nε )}ε>0 be a family of G ′ probing maps for G . We say a set L ⊂ G is F -thin if for all
K ⊂ G bounded there is a Borel subset L0 ⊂ G such that L ⊂ L0 and for every g ∈ K, and ε > 0,
there exists an mΣε -measurable set Sg,ε ⊂ ΣNε such that
{σ ∈ ΣNε : Fε (g, σ) ∈ L0 } ⊂ Sg,ε , lim sup mΣ (Sg,ε )= 0. (2.10)
ε→0+ g∈K Nε

We say G ⊂ G is F -predominant if G \G is F -thin. We say G ⊂ G is respectively predominant

or thin if there exists G ′ as above and a family of G ′ probing maps for G such that G is F -
predominant or F -thin respectively. Figure 1 shows a schematic of a family of probing maps and
a thin set.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 11

Remark 2.5. We note that if a set G ⊂ G is predominant the family of submanifolds endowed
with Borel measures, {(Γε,g , µε,g )}g∈G , with Γε,g := Fε (g, ΣNε ) and µε,g := (Fε (g, ·))∗ mΣε satisfies
that g ∈ Γε,g (by (2.6)), µε,g assigns a positive measure to any neighborhood of g (by (2.7)), the
map g 7→ Γε,g is a C 1 family of Lipschitz submanifolds (by (2.8), (2.9)), and µε,g (Gc ∩ Γε,g ) ≤ ε
for every g ∈ G (by (2.10)). Indeed, by (2.9), we have that locally Γε,g is almost a translate of
Γε,g0 .

The direct analogy to (2.3) would replace the condition (2.10) with
sup mΣ (Sg,ε )= 0.
Nε
g∈K

We are, however, not able to show this in our applications and, instead, relax the condition
to (2.10).

2.2.1. Verification of properties (2.1) for predominance. We now check that the notion of F -
predominance satisfies the properties listed in (2.1). We first prove that predominant sets are
dense.
Lemma 2.6. Suppose F is a family of B ′ probing maps for G and that G ⊂ G is F -predominant.
Then, G is dense in G .

Proof. Let G be F -predominant. Fix g0 ∈ G and δ > 0. We will prove that BG (g0 , δ) ∩ G 6= ∅.
Let F := {(Fε , Nε )}ε>0 with Fε : Gε × ΣNε → G , ∪ε Gε = G , and Gε2 ⊂ Gε1 for ε1 < ε2 . By (2.7),
there is ε0 > 0 such that for 0 < ε < ε0 , g0 ∈ Gε and
sup kFε (g0 , σ) − g0 kB < δ. (2.11)
σ∈ΣNε

Let Gc ⊂ L0 with L0 Borel and satisfying (2.10). Then, since mΣε (Σε ) = 1, there is ε1 > 0
such that {σ ∈ ΣNε : Fε (g0 , σ) ∈ L0 } =
6 Σε for 0 < ε < min(ε0 , ε1 ). In particular, there is σ ∈ Σε
such that Fε (g0 , σ) ∈ G \ L0 ⊂ G. By (2.11), this implies that BG (g0 , δ) ∩ G 6= ∅. 

Next, we check that finite intersections of predominant sets are predominant.

Lemma 2.7. Suppose F is a family of B ′ probing maps for G and that Gj ⊂ G , j = 1, 2, . . . , J
T
are F -predominant. Then, Jj=1 Gj is F -predominant.

Proof. Let K ⊂ G bounded, let {Lj,0 }Jj=1 ⊂ G be a collection of Borel sets with Gcj ⊂ Lj,0 , and
S
for ε > 0 and g ∈ K let Sj,g,ε⊂ ΣNε be mΣε -measurable sets satisfying (2.10). Let L0 := Jj=1 Lj,0
S
and for each ε > 0 and g ∈ K set Sg,ε := Jj=1 Sj,g,ε. Then, for all g ∈ K and ε > 0
J
[
{σ ∈ ΣNε : Fε (g, σ) ∈ L0 } = {σ ∈ ΣNε : Fε (g, σ) ∈ Lj,0 } ⊂ Sg,ε .
j=1

P

Finally, limε→0+ supg∈K mΣ Sg,ε ≤ Jj=1 limε→0+ supg∈K mΣ Sj,g,ε = 0, as claimed. 
Nε Nε
12 YAIZA CANZANI AND JEFFREY GALKOWSKI

We end this section by checking that in finite dimensions predominant sets have full measure.
Since we are working in finite dimensions, we assume that Nε < ∞ and G ⊂ Rn for some
n < ∞. We also take B = Rn . Now, the topology induced on Rn from B ′ for any Banach space
B ′ is identical. Therefore, we may assume, without loss of generality, that B ′ = B = Rn .
Lemma 2.8. Let n < ∞, N0 < ∞, and suppose F = {(Fε , Nε )}ε>0 is a family of Rn probing
maps for G := Rn with supε>0 Nε < N0 . If G ⊂ G is F -predominant, then G has full Lebesgue
measure.

Proof. We will show that if L is F -thin, then L has zero Lebesgue measure. Let K0 ⊂ G = Rn
be closed and bounded. Next, observe that by (2.7), for ε > 0 small enough,
πG (Fε−1 (K0 )) ⊂ {g ∈ G : d(g, K0 ) ≤ 1} =: K.
Since K ⊂ G is bounded, we let L0 ⊂ G be a Borel set with L ⊂ L0 and L0 satisfying (2.10) and
we take ε small enough such that K ⊂ Gε . Note that Fε−1 (L0 ∩ K0 ) is measurable for all ε > 0.
Fubini’s theorem yields


ˆ ˆ
1Fε−1 (L0 ∩K0 ) d(mRn×mΣ ) = mΣ {σ : Fε (g, σ) ∈ L0 ∩ K0 } dmRn (g). (2.12)
Nε Nε
G ×ΣN K
ε

Next, by (2.9), DFε |(g,σ) : T(g,σ) (Rm ×Σε ) → Tg Rm is surjective for ε > 0 small enough, g ∈ K,
and σ ∈ Σε . Therefore, by the coarea formula,
ˆ ˆ ˆ 1 
1Fε−1 (L0 ∩K0 ) d(mRn×mΣ ) = 1L0 ∩Fε (G ×ΣN ) dHg,ε dmRn (g), (2.13)
G ×Σε Nε
K0 ε
Fε−1 (g) |JFε |
p
where Hg,ε denotes the dim(ΣNε )-Hausdorff measure on Fε−1 (g) and |JFε | := det DFε (DFε )∗ .
Next, we will prove that there is ε0 > 0 such that for all 0 < ε < ε0 and g0 ∈ K0
πΣ (Fε−1 (g0 )) = ΣNε . (2.14)
Nε

Once we have (2.14), observe that

Hg,ε (Fε−1 (g)) ≥ Hg,ε (πΣ (Fε−1 (g)) ≥ c > 0.
Nε

Hence, since |JFε | ≤ C < ∞, we have for all g

1
ˆ
dHg,ε > c > 0. (2.15)
Fε (g) |JFε |
−1

To see (2.14), fix (g0 , σ) ∈ K0 × ΣNε , and define Ψε : Rn → Rn by

Ψε (g1 ) := g1 − (Dg Fε |(g0 ,σ) )−1 (Fε (g1 + g0 , σ) − g0 ).
We claim that there is ε0 > 0 such that Ψε : B̄(0, 1) → B̄(0, 1) is a contraction for 0 < ε < ε0 ,
where B̄(0, 1) is the closed unit ball. To prove the claim note that by (2.7), (2.8) and (2.9) there
is ε0 > 0 small enough such that for 0 < ε ≤ ε0 , g0 ∈ K0 , and g1 ∈ B̄(0, 1),


k(Dg F |(g0 ,σ) )−1 (Fε (g0 , σ) − g0 )k < 12 , k(Dg Fε |(g0 ,σ) )−1 Dg Fε |(g1 +g0 ,σ) − Dg Fε |(g0 ,σ) k < 14 .
(2.16)
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 13

Now, for g1 , g2 ∈ B̄(0, 1),

 
(Dg Fε |(g0 ,σ) )−1 Fε (g0 + g1 , σ) − Fε (g0 + g2 , σ) − Dg Fε |(g0 +g2 ,σ) (g1 − g2 )
ˆ 1


= (Dg Fε |(g0 ,σ) )−1 Dg Fε |(g0 +(1−t)g2 +tg1 ,σ) − Dg Fε |(g0 +g2 ,σ) (g1 − g2 )dt ≤ 12 kg1 − g2 k. (2.17)
0

Therefore, using (2.16) and letting g2 = 0 in (2.17), we have that for g1 ∈ B̄(0, 1)
 
kΨε (g1 )k = k(Dg Fε |(g0 ,σ) )−1 Dg Fε |(g0 ,σ) g1 − Fε (g0 + g1 , σ) + Fε (g0 , σ) − Fε (g0 , σ) + g0 k < 1,

and so Ψε : B̄(0, 1) → B̄(0, 1) for ε < ε0 . Next, again by (2.16) and (2.17), for g1 , g2 ∈ B̄(0, 1)


kΨε (g1 ) − Ψε (g2 )k = k(Dg Fε |(g0 ,σ) )−1 Dg Fε |(g0 ,σ) (g1 − g2 ) − Fε (g1 + g0 , σ) + Fε (g2 + g0 , σ) k
≤ 34 kg1 − g2 k.

Hence, Ψε : B(0, 1) → B(0, 1) is a contraction. For each (g0 , σ) ∈ Rn × ΣNε let g1 = g1 (g0 , σ) be
the fixed point of Ψε . Then, for each (g0 , σ) ∈ Rn × ΣNε there is g1 with Fε (g1 + g0 , σ) = g0 as
claimed in (2.14).
Finally, by (2.8) there is C > 0 with supε>0 sup(g,σ)∈K×Σε |JFε (g, σ)| < C, (2.12), (2.13),
(2.10), (2.15) yield
ˆ ˆ
1
0 = lim 1Fε−1 (L0 ∩K0 ) d(mRn × mΣ ) ≥ C 1L0 ∩Fε (G ×ΣN ) dmRn = C1 mRn (L0 ∩ K0 ).
ε→0+ G ×ΣNε K0 ε

Hence, mRn (L0 ∩ K0 ) = 0 and, since K0 is an arbitrary closed bounded set, mRn (L0 ) = 0.
In particular, by the completeness of the Lebesgue measure, L is Lebesgue measurable with
mRn (L) = 0 as claimed. 

2.3. Heuristic explanation for lack of Baire-genericity. A set is Baire generic if it contains
the intersection of countably many open dense sets. In order to explain why we do not pursue
this notion of genericity, we discuss one of the key features we require at periodic points for the
geodesic flow. Let Γ ⊂ S ∗M be a Poincaré section through the point ρ and P : Γ → Γ the
corresponding Poincaré map. For simplicity, we will assume that P(ρ) = ρ. Theorem 1.3 has
consequences for the eigenvalues of dP|ρ : Tρ Γ → Tρ Γ. Indeed, we have that there is C > 0 such
that for all n
αν +ε +1
k(dP|ρ )n − I)−1 k = k(d(P n )|ρ − I)−1 k ≤ (Cn)Cn .
Therefore, the eigenvalues, {λj }2d−2
j=1 of dP|ρ must satisfy

n−1 2πp αν +ε −1
inf 1 + |λj | |λj − ei n | ≥ inf |λnj − 1| ≥ (Cn)−Cn , n ≥ 1.
j,p j

Since dP is a symplectic transformation, eigenvalues may be confined to the unit circle (see
Section 5.1) and this becomes analogous to understanding the structure of the set real numbers
which are poorly approximable by rational numbers.
To discuss the issues of genericity in a simpler setting, we forget now about Poincaré maps
and instead discuss them in the context of approximation of real numbers. As explained above,
14 YAIZA CANZANI AND JEFFREY GALKOWSKI

α
letting hℓ (q) = (ℓq)−ℓq −1 we want to investigate the set
[
D := Dℓ , Dℓ := {s ∈ [0, 1] : |s − pq | > hℓ (q), for all p, q ∈ N}. (2.18)
ℓ>0

To study D we consider
∞ [
\ ∞ q−1
[
U := Uq , Uq := ( pq − hq (q), pq + hq (q)).
q0 =1 q=q0 p=0

Observe that U is Baire generic. However, U ∩ Dℓ = ∅ for all ℓ > 0 which implies that D ∩ U = ∅
and hence that D is not Baire generic. Since even the property we want for the eigenvalues of
dP is non-generic, it is unlikely that the set of metrics which produce such P is generic.
For further evidence of lack of Baire genericity, we consider the space Diffν (Γ) of C ν diffeomor-
phisms on a smooth manifold Γ for ν ≥ 2. Kaloshin [Kal00] showed that for any {an }∞ n=1 ⊂ [1, ∞),
the set
n Pn (f ) o
f ∈ Diffν (Γ) : lim sup <∞
n an
is not Baire generic, where Pn (f ) := #{ρ ∈ Γ isolated : f n (ρ) = ρ}. Indeed, when Γ is a
3-manifold, this set is not Baire generic even in the space of volume preserving maps [KS06].
While D in (2.18) is not Baire generic, one can see that, since limn→∞ |[0, 1] \ Dn | = 0, D has
full Lebesgue measure and hence is, in a much stronger sense, typical. Indeed, for many purposes,
the notion of full Lebesgue measure is a better version of ‘typical’ than Baire genericity. For
example, a randomly chosen element of [0, 1] is almost surely in D but may not be in a given
Baire generic set. Also, a full Lebesgue measure set in R has Hausdorff dimension 1, while a
Baire generic set in R may have 0-Hausdorff dimension. Motivated by this discussion, the notion
of ’typicality’ that we use, i.e. that of predominance (see Definition 2.4), is an analog of a full
Lebesgue measure set of metrics g ∈ G ν .

3. Counting closed geodesics and improvements for Weyl laws:

Proof of Theorems 1.1 and 1.2

In this Section, we use Theorem 1.3 to prove Theorems 1.1 and 1.2. Both of these theorems rely
on volume estimates on the set of nearly closed geodesics. We obtain these estimates in Section
3.1 and prove Theorems 1.1 and 1.2 in Sections 3.2 and 3.3 respectively.
We start by letting M be a C ν manifold, g ∈ G ν , β0 > 0, and W be a β0 -family of transition
maps for g. Let β : [0, +∞) → (0, +∞) be a continuous, decreasing function. Throughout the
section we will suppose that

R(t, β(t), g) ⊂ N (t, β(t), (g, W)), for t > c, d(ϕgt (ρ), ρ) ≥ c|t|, for |t| ≤ c. (3.1)

In Theorem 1.3, we show that there is a predominant set of metrics G ⊂ G ν such that (3.1) holds
Ω
for g ∈ G with β(t) = C −C(t+1) −1 and some C > 0.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 15

3.1. Volume of nearly closed geodesics. We start by following [DZ16, Appendix A], replacing
their assumption of hyperbolicity of the flow with (3.1) to estimate the volume of nearly closed
geodesics. Recall that there are C̃, L > 0 such that for all f ∈ C 2 (S ∗M ) and t ∈ R

kf ◦ ϕgt kC 2 (S ∗M ) ≤ C̃eL|t| kf kC 2 (S ∗M ) . (3.2)

In particular, d(ϕgt (ρ), ϕgt (ρ′ )) ≤ C̃eL|t| d(ρ, ρ′ ) for all ρ, ρ′ ∈ S ∗M and t ∈ R.
The following lemma shows that two nearby orbits that return to their starting point after
similar times are almost iterates of one another.
Lemma 3.1. Let M, g, β0 , W and β be such that (3.1) and (1.3) hold. Given t0 > 0 there are
C, δ > 0 such that the following holds. Let t ≥ t0 , t′ ≥ t0 with |t − t′ | ≤ δ, and for 0 ≤ ε < β(t)
let ρ, ρ′ ∈ S ∗M be such that
d(ρ, ϕgt (ρ)) ≤ ε, d(ρ′ , ϕgt′ (ρ′ )) ≤ ε, d(ρ, ρ′ ) ≤ δβ(t)e−Lt .

Then, |t − t′ | ≤ Cε and there exists s ∈ [−1, 1] such that β(t)d(ρ, ϕgs (ρ′ )) ≤ Cε.

Proof. First, observe that for any ε0 > 0, we may increase C enough so that the statement becomes
trivial for ε ≥ ε0 . Therefore, we need only work with ε < ε0 < min( 41 β0 , 12 ) small enough and we
do so from now on. Next, observe that we may shrink δ so that d(ϕgt (ρ), ϕgt (ρ′ )) ≤ 14 β0 whenever
d(ρ, ρ′ ) ≤ δe−Lt , t ≥ t0 . We may also assume that all of the relevant points are contained in a
single coordinate chart and hence we may work in a small ball in R2d−1 .
We also assume that δ is small enough such that whenever d(ρ, ρ′ ) ≤ δ there exists |s| ≤ 1 with
ρ − ϕgs (ρ′ ) ∈ H (ρ) and such that d(ρ, ϕgs (ρ′ )) ≤ 2d(ρ, ρ′ ). Set ρ0 := ϕgs (ρ′ ) and note
ρ − ρ0 ∈ (Hp (ρ))⊥ , d(ρ, ρ0 ) ≤ 2δβ(t)e−Lt , d(ϕgt′ (ρ0 ), ρ0 ) ≤ C̃eL ε. (3.3)
By Taylor expanding in ρ and using (3.2), we have
kϕgt (ρ0 ) − ϕgt (ρ) − dϕgt (ρ)(ρ0 − ρ)k ≤ C̃eLt d(ρ, ρ0 )2 ≤ 2C̃δβ(t)d(ρ, ρ0 ).
Next, Taylor expanding in t and using that there is C > 0 such that k∂t2 ϕgt k ≤ C,
kϕgt′ (ρ0 ) − ϕgt (ρ0 ) − Hp (ϕgt (ρ0 ))(t′ − t)k ≤ C|t′ − t|2 ≤ Cδ|t′ − t|.
Thus, we have that there is C > 0 with
 
kϕgt′ (ρ0 ) − ϕgt (ρ) − dϕgt (ρ)(ρ0 − ρ) − Hp (ϕgt (ρ0 ))(t′ ′
− t)k ≤ Cδ β(t)d(ρ0 , ρ) + |t − t | .

Next, since d(ϕt (ρ), ρ) ≤ ε, by (3.3) there is C > 0 such that

   
k I − dϕgt (ρ) (ρ0 − ρ) − Hp (ϕgt (ρ0 ))(t′ − t)k ≤ Cδ β(t)d(ρ0 , ρ) + |t − t′ | + Cε.

Now, by (1.3) there is C > 0 such that kWϕgt (ρ),ρ − Ik ≤ Cd(ϕgt (ρ), ρ) ≤ Cε. Therefore, there is
C > 0 such that
   
k Wϕgt (ρ),ρ − dϕgt (ρ) (ρ0 − ρ) − Hp (ϕgt (ρ))(t′ − t)k ≤ Cδ β(t)d(ρ0 , ρ) + |t − t′ | + Cε. (3.4)
16 YAIZA CANZANI AND JEFFREY GALKOWSKI

Now, since ρ0 − ρ ∈ H (ρ), we have by (3.1), Definition 1.6, and the fact that that ε ≤ β(t), that
there is C > 0 such that
 
β(t)d(ρ0 , ρ) ≤ kΠϕt (ρ) Wρ,ϕgt (ρ) − dϕgt (ρ) (ρ0 − ρ)k
 
≤ Ck Wρ,ϕgt (ρ) − dϕgt (ρ) (ρ0 − ρ) − Hp (ϕgt (ρ))(t′ − t)k. (3.5)
On the other hand, since Wϕgt (ρ),ρ (Hp (ρ)) = Hp (ϕgt (ρ)), Wϕt (ρ),ρ (H (ρ)) = H (ϕgt (ρ)), and H (ρ)
is transverse to Hp (ρ) (uniformly in ρ), we have
 
|t − t′ | ≤ Ck Wρ,ϕgt (ρ) − dϕgt (ρ) (ρ0 − ρ) − Hp (ϕgt (ρ))(t′ − t)k. (3.6)
1
Combining (3.4), (3.5), (3.6), and choosing δ ≪ 1 and ε0 < 2C , we have β(t)d(ρ, ρ0 )+ |t − t′ | ≤ Cε
as claimed. 

We proceed to bound the volume of nearly closed trajectories in S ∗M .

Lemma 3.2. Let M, g, β0 , W, and β be such that (3.1) holds. Let t0 > 0, then there is C > 0
such that for T > t0 and 0 < ε ≤ β(T ), we have
 
volS ∗M ρ : ∃t ∈ [t0 , T ] such that d(ϕgt (ρ), ρ) ≤ ε ≤ Cε2d−2 eCT β(T )−(4d−3) .

Proof. Let t0 > 0 and T > t0 . Then, let δ as in Lemma 3.1 and fix 0 < ε ≤ β(T ). Next divide
NT NT
[t0 , T ] into intervals, {Ii }i=1 with right endpoints at {Ti }i=1 satisfying |Ii | ≤ 2δ . Next, for each
KT
i = 1, . . . , NT let {ρj }j=1i ⊂ S ∗M be a maximal 2δ e−LTi β(Ti ) separated set. Note that
NT KTi
 [[
ρ : ∃t ∈ [t0 , T ] such that d(ϕgt (ρ), ρ) ≤ε ⊂ πS ∗M (Pi,ρj ),
i=1 j=1 (3.7)
n o
Pi,ρj := (ρ, t) : t ∈ Ii , d(ϕgt (ρ), ρ) ≤ ε, d(ρ, ρj ) ≤ 2δ e−LTi β(Ti ) .

To estimate volS ∗M (πS ∗M (Pi,ρj )), fix (ρ′ , t′ ) ∈ Pi,ρj and let (ρ, t) ∈ Pi,ρj with (ρ, t) 6= (ρ′ , t′ ). By
Lemma 3.1,  
[
|t − t′ | ≤ Cε, d ρ, ϕgs (ρ′ ) ≤ Cεβ(Ti )−1 .
|s|≤1

Then, volS ∗M (πS ∗M (Pi,ρj )) ≤ Cε2d−2 β(Ti )−(2d−2) , and

K
 [
Ti 
volS ∗M πS ∗M πS ∗M (Pi,ρj ) ≤ KTi Cε2d−2 β(Ti )−(2d−2) ≤ Cε2d−2 e(2d−1)LTi β(Ti )−(4d−3) ,
j=1
(3.8)
since there is Cδ > 0 such that KTi ≤ Cδ [e−LTi β(Ti )]−(2d−1) .
The claim follows from combining
(3.8) with (3.7), and using that β is decreasing and NT ≤ CT . 
Remark 3.3. We note that if one is only interested in obtaining a volume estimate of the form
 
volS ∗M ρ : ∃t ∈ [t0 , T ] such that d(ϕgt (ρ), ρ) ≤ ε ≤ Cε2d−2−k eCT β(T )−(4d−3−k) ,
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 17

for some k ≥ 1, then it is possible to use a weaker notion of non-degeneracy. Indeed, one can
replace the definition of N (t, ρ, (g, W)) by saying that ρ ∈ N k (t, ρ, (g, W)) if there is a 2d − 2 − k
dimensional subspace of v ∈ Tρ S ∗M such that |Πϕg (ρ) (Wρ,ϕg (ρ) − dϕgt )v| > α|Πρ v|. If implemented
t t
throughout the article, this would in turn lead to a smaller value of αν in Theorem 1.1 but would
not be sufficient for counting closed geodesics.

3.2. Proof of Theorem 1.2. Let Ω > Ων . By Theorem 1.3 there is a predominant set GΩ ⊂ G ν
such that for all g ∈ G and W a family of transition maps for g there is C > 0 such that (3.1)
Ω
holds with β(t) = C −C(t+1) −1 .
Let g ∈ GΩ and 0 < t0 < injg (M ) so that there are no closed geodesics with length ≤ t0 .
Let δ be as in Lemma 3.1. First, we claim that c(T, g) < ∞ for each T > 0. Indeed, suppose
g
that {ρi }∞ ∗ ∞
i=1 ⊂ S M and {ti }i=1 ⊂ [t0 , T ] are sequences satisfying ϕti (ρi ) = ρi and such that
g
/ {ϕs (ρi ) : |s| ≤ 4} for all i 6= j. Then, we may assume that there are ρ ∈ S ∗M and t ∈ [t0 , T ]
ρj ∈
such that ρi → ρ and ti → t. By continuity, we conclude ϕgt (ρ) = ρ. Next, let i0 > 0 be such
that d(ρi , ρ) ≤ δβ(T )e−LT and |ti − t| ≤ δ for i ≥ i0 . Then, by Lemma 3.1 we obtain that for
i > i0 there is si ∈ [−1, 1] such that ϕgsi (ρi ) = ρ and ti = t. In particular, for i, j > i0 , we have
ρj = ϕgsi −sj (ρi ) which is a contradiction since |si − sj |≤ 3.
c(T,g)
Since c(T, g) < ∞, we let {γi }i=1 be the finite collection of primitive closed geodesics of length
≤ T . Then, there is δ0 = δ0 (T ) > 0 such that if γi (δ) ⊂ S ∗M denotes the δ neighborhood of γi ,
γi (δ) ∩ γj (δ) = ∅, i 6= j, 0 ≤ δ < δ0 . (3.9)
Letting Ti be the length of γi , we have by (3.2) that supρ∈γi (δ) d(ρ, ϕgT (ρ)) ≤ C̃δeLTi . Therefore,
i
using Lemma 3.2 with C̃δeLT in place of ε, there is C > 0 such that,
X  [ 
volS ∗M (γi (δ)) = volS ∗M γi (δ) ≤ Cδ2d−2 eCT +(2d−2)LT β(T )−(4d−3) . (3.10)
1≤i≤c(T,g) 1≤i≤c(T,g)

In addition, since there is c > 0 such that volS ∗M (γi (δ)) ≥ cδ2d−2 t0 for all 1 ≤ i ≤ c(T, g), we
obtain, using (3.9) and (3.10) that there is C > 0 such that
c(T, g)t0 ≤ CeCT +(2d−2)LT β(T )−(4d−3) .
Together with the fact that c(T, g) = 0 for 0 < T < t0 , this implies Theorem 1.2. 

3.3. Proof of Theorem 1.1. Let g be a C ν -Riemannian metric on a manifold M of dimension

d. For R > 0 let P R (t0 , T ) ⊂ S ∗M be the set of directions that yield trajectories that are R close
to being periodic at some time with t0 ≤ |t| ≤ T . That is,
 [ 
P R (t0 , T ) := ρ ∈ S ∗M : ϕgt (BS ∗M (ρ, R)) ∩ BS ∗M(ρ, R) 6= ∅ . (3.11)
t0 ≤|t|≤T

In [CG20, Theorem 2] it is proved that if ν ≥ ν0 and T(R) is a sub-logarithmic resolution function

and there exist C > 0 and t0 such that


lim sup volS ∗M B(P R (t0 , T(R)), R) T(R) ≤ C, (3.12)
R→0+
18 YAIZA CANZANI AND JEFFREY GALKOWSKI

then, as λ → ∞,
volRd (B d ) volg (M ) d  λd−1 
#{j : λj (g) ≤ λ} = λ + O . (3.13)
(2π)d T(λ−1 )
Remark 3.4. The application of [CG20, Theorem 2] deserves a brief comment. Note that, as
stated, the theorem is valid for C ∞ metrics. However, it is clear that all of the proofs rely only
on the C ν norm of the metric for some sufficiently large ν. Therefore, provided that we work with
sufficiently smooth metrics, we may apply [CG20, Theorem 2].

The proof of Theorem 1.1, given in subsection 3.4, is then reduced to proving that for any
Ω > Ων and some C0 > 0
T(R) = f −1 (Rγ ), (3.14)
with
2d − 2 β(t) −C0 t Ω
γ := , f (t) = e , β(t) := C −C(t+1) −1
4d − 3 β(0)
is a sub-logarithmic resolution function and that (3.12) holds. Indeed, Theorem 1.1 follows since
1 1 1
C0 (log λ) Ω ≥ T(λ−1 ) = f −1 (λ−γ ) ≥ (log λ) Ω , λ ≥ λ0 (3.15)
C0
for some C0 > 0 and λ0 > 0 large enough.
Before we proceed to the proof of Theorem 1.1, we recall the notion of a sub-logarithmic
resolution function from [CG20, Definition 1.1]. We say that T : (0, 1) → (0, ∞) is a resolution
function if it is continuous and decreasing. We say that T is sub-logarithmic if it is differentiable
and
1
(log T(R))′ ≥ − 0 < R < 1. (3.16)
R log R−1
Next, we introduce a convenient class of sub-logarithmic rate functions.
Lemma 3.5. Suppose that A(t) : R → R is twice differentiable and satisfies A(0) = 0, A′ (t) > 0,
A′′ (t) ≥ 0 for t ≥ 0. Then, with f (t) = e−A(t) , we have that T(R) := f −1 (Rγ ) is a sub-logarithmic
rate function for γ > 0.

Proof. First, observe that T(R) is decreasing and differentiable. Moreover,

γRγ γ
(log T)′ (R) = ′ −1 γ −1 γ
=− .
Rf (f (R ))f (R ) RA (f (Rγ ))f −1 (Rγ )
′ −1

Next, note that

ˆ t ˆ t
′ ′′ ′
A (t)t = A (s)s + A (s)ds ≥ A′ (s)ds = A(t).
0 0
Therefore, A′ (f −1 Rγ )f −1 (Rγ ) ≥ A(f −1 (Rγ )) =γ −1
log R , and hence (3.16) holds as claimed. 
Remark 3.6. By Lemma 3.5, the function T(R) as defined in (3.14) is a sub-logarithmic resolu-
tion function. Indeed, f (t) = e−A(t) with A(t) = C0 t + C((t + 1)Ω − 1). Note that A(0) = 0, and,
since Ω ≥ 1, it is easy to check that A′ (t) > 0, and A′′ (t) ≥ 0 for t ≥ 0 as claimed.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 19

3.4. Proof of Theorem 1.1. Let Ω > Ων . By Theorem 1.3 there is a predominant set GΩ ⊂ G ν
such that for all g ∈ GΩ and W a family of transition maps for g there is C > 0 such that
Ω
(3.1) holds with β(t) = C −C(t+1) −1 . Without loss of generality we assume C > 1. In Lemma
3.5 we proved that T as defined in (3.14) is a sub-logarithmic resolution function. Therefore, as
explained above, the proof of Theorem 1.1 would follow from combining (3.13) and (3.15) once
we prove (3.12).
Let 0 < t0 < T . In order to prove (3.12), we first claim that there is C > 0 such that
n o
B(P R (t0 , T ), R) ⊂ q ∈ S ∗M : ∃t ∈ (t0 , T ) ∪ (−T, −t0 ) s.t. d(ϕt (q), q) ≤ CeCT R . (3.17)

Indeed, for q ∈ B(P R (t0 , T ), R), there is ρ ∈ P R (t0 , T ) such that d(ρ, q) < R. Then, there
are t ∈ (t0 , T )∪(−T , −t0 ) and ρ1 ∈ S ∗M such that d(ρ, ρ1 ) < R and d(ϕt (ρ1 ), ρ) < R. Since
ϕt ((x, ξ)) = ϕ−t (x, −ξ), we may assume without loss of generality that t > 0. Therefore, by (3.2)
there is C > 0 such that
d(ϕt (q), q) ≤ d(ϕt (q), ϕt (ρ1 )) + d(ϕt (ρ1 ), ρ) + d(ρ, q) ≤ C̃eLt d(q, ρ1 ) + R + R ≤ CeCt R,
proving the claim in (3.17).
By (3.17) and Lemma 3.2, for R ≤ β(T )(CeCT )−1 , we have


volS ∗M B(P R (t0 , T ), R) T ≤ CT (CeCT R)2d−2 eCT β(T )−(4d−3) ≤ CeC1 T R2d−2 β(T )−(4d−3) .
C1
Therefore, we may choose C0 = 4d−3 and use the definition of T(R) to obtain that R ≤
CT(R)


β(T(R))(Ce −1
) for R small enough and volS ∗M B(P R (t0 , T(R)), R) T(R) ≤ C as claimed.


4. Outline of the proof of Theorem 1.3

To simplify the exposition in this outline, we will first imagine that it is only necessary to
understand exactly periodic points rather than returning points. We also assume that one can
construct a global Poincaré section (see Section 4.5). In addition, rather than providing the details
to prove the result for a predominant set of metrics, we will outline a proof of the theorem for a
dense set of metrics. See the end of the section for remarks on how to drop these assumptions
and prove predominance.

4.1. The perturbation. The first key point to understand when proving Theorem 1.3 is how
to perturb away the degeneracy of a single periodic orbit with quantitative control on how large
a perturbation is needed to produce non-degeneracy. When doing this, it is important to use a
perturbation of the metric which interacts with the periodic geodesic exactly once. That is, the
perturbation must be supported in a ball, B ⊂ M , over which the geodesic passes exactly once.
Because of this, we will only be able to directly perturb away degeneracy for primitive closed
geodesics. The construction of one such the perturbation is done in Section 9.
Remark 4.1. Our main inductive argument is actually proved under some general assumptions
on the perturbation (see Definitions 7.2 and 7.3). Because of this, we postpone this construction
until after our main inductive argument, which appears in Section 8.
20 YAIZA CANZANI AND JEFFREY GALKOWSKI

Restricting our attention to a single geodesic, γg0 , for a metric g0 , we find a finite dimensional
family of perturbations gσ of the metric g0 . When γg0 is a primitive periodic geodesic, these
perturbations should produce non-degeneracy of the orbit γgσ . The perturbations are a combina-
tion of those considered by Anosov [Ano82] and Klingenberg [Kli78, §3.3] (see also Klingenberg–
Takens [KT72]). In our case, when γg0 is primitive, we give careful estimates on how these
perturbations affect the Poincaré map, Pγgσ , associated to γgσ and its derivative, dPγgσ . In
particular, estimating the size of the inverse of the derivative of the map σ 7→ (Pγgσ , dPγgσ ).
Once we have a family of perturbations for any given geodesic, we cover S ∗M by finitely many
small balls, Bi , of radius r and center ρi and associate to each ball a family of perturbations
modelled on those above; replacing the geodesic γg0 by the geodesic, γg0 ,ρi for g0 through ρi .
After an approximation argument, this gives a finite (albeit very large) dimensional family of
perturbations which can be used to perturb away degeneracy for primitive geodesics of a given
length. In particular, for every such geodesic γg0 , we can find a ball, Bi , such that, with gi,σ
the perturbation of g0 associated to γg0 ,ρi , the derivative of the map σ 7→ (Pγgi,σ ,ρi , dPγgi,σ ,ρi )
is invertible at σ for all σ ∈ BRL (0, 1). Once we have this in place, it will be possible to control
the volume of the set of perturbations for which there is a degenerate primitive closed geodesic
of some length, with some quantitative control on how degenerate such an orbit may be. In
particular, once this volume is small enough, there is at least one perturbation of g0 for which all
primitive orbits of a certain length are non-degenerate. Since g0 is arbitrary, and we make our
perturbations arbitrarily small, we will eventually obtain density after an induction on the length
of the closed geodesics.

Remark 4.2. In order to obtain predominance in our main theorem, we use the control on the
volume of bad perturbations more seriously. In particular, making it smaller than ε > 0 for any
chosen ε.

4.2. The Induction. The proof of Theorem 1.3 relies on an induction on the length of an orbit.
In order to do this, we follow a strategy motivated by that of Yomdin [Yom85]. The work in
[Yom85] shows that every diffeomorphism f0 on M can be perturbed to a diffeomorphism fε in
which every n-periodic point x is γε -hyperbolic in the sense that the eigenvalues λj of dfεn (x)
satisfy ||λj | − 1| ≥ γε where γε is a power of ε depending on n, ε and the regularity of f0 . The
idea there is to first use a perturbation to make primitive orbits hyperbolic and then to use the
fact that hyperbolicity of an orbit passes to its multiples and, moreover, that every multiple is
more hyperbolic than the previous multiple.
In contrast to what happens when working with diffeomorphisms on M , one of the main
difficulties to overcome in the case of geodesic flows (or indeed symplectic maps) is that, in
general, it is not possible to perturb a closed orbit to turn it into a hyperbolic closed orbit.
Indeed, eigenvalues of the Poincaré map may be ‘trapped’ on the unit circle (see Section 5.1) and
hence there are so-called stable closed orbits which cannot be perturbed away. Because of this
structure, one must find a different property which guarantees non-degeneracy and can be passed
from a primitive closed orbit to its multiples.
Given a metric g0 , our objective is to find a metric g∞ , that is arbitrarily close to g0 , with the
property that for each ℓ every periodic trajectory of length ≤ 2ℓ is βℓ non-degenerate in the sense
of Definition 1.6. Here, {βℓ }ℓ is a decreasing sequence of numbers.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 21

We build g∞ by induction on the parameter ℓ which represents (the logarithm of) the maximum
length of the orbits up to which non-degeneracy is controlled. Note that we start with the most
naive possible version of the induction and gradually add the details necessary to handle our
situation.
A. Hypothesis. Assume gℓ is a metric such that for k ≤ ℓ, all closed trajectories of length
n ∈ (2k−1 , 2k ] are βk non-degenerate in the sense of Definition 1.6.
B. Perturbation: Find a perturbation, gℓ+1 of gℓ which satisfies
• primitive closed trajectories of length n ∈ (2ℓ , 2ℓ+1 ] are βℓ+1 non-degenerate,
• closed trajectories of length n ∈ (2k−1 , 2k ] with k ≤ ℓ remain βk non-degenerate.
C. Deal with non-primitive closed trajectories: A non-primitive closed trajectory γ for g ℓ+1
of length n ∈ (2ℓ , 2ℓ+1 ] is a multiple of a primitive closed trajectory γ̃ of length m ∈ (2k−1 , 2k ]
with k ≤ ℓ. By the inductive hypothesis γ̃ is βk non-degenerate. Use this to show that the
βk non-degeneracy of γ̃ implies βℓ+1 non-degeneracy of γ.

Why it fails: The βk non-degeneracy of the orbit γ̃ does not imply βℓ+1 non-degeneracy of its
multiple γ. Indeed, the Poincaré map associated to γ may have a root of unity as an eigenvalue.
That is, Step C cannot be completed.
Solution: We define the notion of (β, q)-nondegeneracy. For ℵ > 0 and β, q > 0 we say a matrix
A ∈ M(ℵ), the set of ℵ × ℵ matrices, is (β, q) non-degenerate if
ℵ
(I − Aq )−1 ≤ ( 25 q 2 β −1 ) (1 + kAkq )ℵ−1 . (4.1)

The key observation here is that, if the derivative, A = dPγ , of the Poincaré map associated to a
closed orbit γ is (β, q)-nondegenerate, then the q iterate of the orbit is ( 52 q 2 β −1 )−ℵ non-degenerate
(see Figure 2 for a schematic of a (β, 2) non-degenerate orbit). Therefore, our goal will be to make
perturbations of the metric, gℓ , so that primitive orbits become (β(q), q)-non-degenerate for all q
and some sequence {β(q)}∞ q=1 . This will be possible provided that β(q) = O(q
−2−0 ) as q → ∞.

The main difficulty with the notion of (β(q), q) non-degeneracy is that, although for each fixed
(β(q), q) the property of (β(q), q) non-degeneracy is stable under small perturbations, the property
of being (β(q), q) non-degenerate for all q is not. Therefore, we only look for non-degeneracy for
q < Q.
We next present the modified induction argument. The induction is done on the parameter ℓ.
One can find sequences of numbers {βk }k , {βk,ℓ }k,ℓ , and {Qk,ℓ }k,ℓ so that the induction below can
be completed.
A. Hypothesis. Assume gℓ is a metric such that for k ≤ ℓ
– closed trajectories of length n ∈ (2k−1 , 2k ] are βk non-degenerate,
– primitive closed trajectories of length n ∈ (2k−1 , 2k ] are (βk,ℓ q −3 , q) non-degenerate for
all 0 < q < Qk,ℓ .
B. Perturbation: Find a perturbation, gℓ+1 , of gℓ which satisfies
– primitive closed trajectories of length n ∈ (2ℓ , 2ℓ+1 ] for gℓ+1 are βℓ+1 non-degenerate
– closed trajectories of length n ∈ (2k−1 , 2k ] for gℓ+1 remain βk non-degenerate for all
k ≤ ℓ.
22 YAIZA CANZANI AND JEFFREY GALKOWSKI

v ∼β
g
dϕT v
dϕg2T v

Figure 2. In the figure, we show a primitive periodic orbit of length T such that
the derivative dP of the Poincaré map is (β, 2) non-degenerate. The primitive
orbit is shown in the blue dotted line. Note that dP satisfies (dP)k = dϕgkT .
Thus, the property of (β, 2) non-degeneracy implies that kdϕg2T v − vk ∼ β.

– primitive closed trajectories of length n ∈ (2k−1 , 2k ] for gℓ+1 are (βk,ℓ+1 q −3 , q) non-
degenerate for 0 < q < Qk,ℓ+1 and k ≤ ℓ + 1,
C. Deal with non-primitive closed trajectories: Let γ be a closed trajectory for gℓ+1 of length
n ∈ (2ℓ , 2ℓ+1 ] that is non-primitive. Then γ is a multiple of a primitive closed trajectory γ̃
of length m ∈ (2k−1 , 2k ] with k ≤ ℓ. Note that then γ̃ satisfies the hypothesis in Step A. In
particular, to show that γ is βℓ+1 non-degenerate, we would like to use that γ̃ is (βk,ℓ q −3 , q)
non degenerate with q = n/m, and for that we need n/m < Qk,ℓ . We then show that
2ℓ+1−k < Qk,ℓ and that by the inductive hypothesis this implies βℓ+1 non-degeneracy of γ.
After the steps A-C above (modulo the fact that we have only consider exactly closed orbits),
the inductive argument yields a metric g∞ which is arbitrarily close to g0 and has the desired
non-degeneracy property that, for each ℓ, every closed trajectory for g∞ of length ≤ 2ℓ is βℓ
non-degenerate. This yields a discretized version of the statement in Theorem 1.3.
It is important to note that one cannot do a simplified version of this induction in which a
sequence {Qk }k is used in place of {Qk,ℓ }k,ℓ and {βk }k is used in place of {βk,ℓ }k,ℓ . The problem
would be that, when applying the inductive hypothesis to carry out Step C, one would need
n/m < Qk so that the non-degeneracy of γ̃ may pass to that of γ. However, this may not be
possible to arrange. For example, for a closed trajectory of length n that is a multiple of a
primitive trajectory of length m = 1. One might wonder, at this point, why we do not simply
take Qk,ℓ = Qℓ = 2ℓ+2 . We will see below, in sections 4.3 and 4.4, that this is indeed not possible.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 23

4.3. How to deal with almost closed trajectories. Next, we explain how to modify the
argument to deal with trajectories that are not necessarily closed and primitive but are very
close to being such. First of all, instead of working with closed trajectories, one works with (n, α)
returning trajectories. In this discussion a point ρ is said to be (n, α) returning if d(ρ, P n (ρ)) ≤ α,
where we are still pretending that we can work with a globally defined Poincaré section P (see
Section 4.5).
There is one serious difficulty which is hidden by imagining that we only deal with exactly closed
trajectories. This difficulty arises from the fact that the notion of primitiveness for non-closed
trajectories is not a well-defined condition. One should instead think of a degree of simplicity.
We will say that a point ρ is (n, α) simple if d(ρ, P k (ρ)) > α for 0 < k < n. Instead of working
with primitive closed trajectories of length n one works with trajectories that are (n, α) simple
for some α.
It is in fact only possible to make an (n, α) simple point be (βq −3 , q) non-degenerate whenever
β is small depending on α. Therefore, there are sequences {βk,ℓ }k,ℓ and {αk,ℓ }k,ℓ such that in the
ℓ-th step of the induction argument, orbits of length n ∈ (2k−1 , 2k ) which are βk,ℓ returning can
only be made to be (βk,ℓ q −3 , q) non-degenerate for 0 < q < Qk,ℓ provided they are (n, αk,ℓ )-simple
and βk,ℓ is small in terms of αk,ℓ .
The induction argument is the same as the one outlined before but with each instance of the
word ‘closed’ replaced by (n, β) returning for some parameter β depending on (k, ℓ) and each
instance of the word ‘primitive’ replaced by (n, α) simple for some parameter α depending on
(k, ℓ).
A. Hypothesis. Assume gℓ is a metric such that for k ≤ ℓ
– βk,ℓ returning trajectories of length n ∈ (2k−1 , 2k ] are βk non-degenerate,
– βk,ℓ returning trajectories that are αk,ℓ simple and have length n ∈ (2k−1 , 2k ] are
(βk,ℓ q −3 , q) non-degenerate for all 0 < q < Qk,ℓ .
B. Perturbation: Find a perturbation, gℓ+1 , of gℓ which satisfies
– βℓ+1,ℓ+1 returning trajectories that are αℓ+1,ℓ+1 simple and have length n ∈ (2ℓ , 2ℓ+1 ]
are βℓ+1 non-degenerate
– βk,ℓ+1 returning trajectories of length n ∈ (2k−1 , 2k ] will remain βk non-degenerate for
all k ≤ ℓ.
– βk,ℓ+1 returning trajectories that are αk,ℓ+1 simple and have length n ∈ (2k−1 , 2k ] will
be (βk,ℓ+1 q −3 , q) non-degenerate for 0 < q < Qk,ℓ+1 and k ≤ ℓ + 1.
C. Deal with non-simple trajectories: This step becomes substantially more complicated as
a result of needing to handle non-closed trajectories. One consequence of the new difficulties,
is that, instead of using 2-adic intervals, we must use a-adic intervals for some 1 < a < 2.
We include a sketch of how to handle Step C below in Section 4.4, in order to present a
more or less complete picture of the induction. However, the reader may wish to skip the
next section on first reading.

4.4. Detailed sketch of Step C. Suppose that ρ yields a trajectory of length n ∈ (2ℓ , 2ℓ+1 ] that
is (n, βℓ+1,ℓ+1 ) returning but not (n, αℓ+1,ℓ+1 ) simple.
24 YAIZA CANZANI AND JEFFREY GALKOWSKI

It is in general not possible to find an m ≤ n/2 such that ρ is (m, αℓ+1,ℓ+1 ) returning and
(m, αℓ+1,ℓ+1 ) simple. Instead, one only obtains (m1 , α̃1 ) returning with α̃1 ∼ C n αℓ+1,ℓ+1 and
m1 ≤ 2−1 n. Therefore, in Step C of the induction, if m1 ∈ (2k1 −1 , 2k1 ), we would need α̃1 < βk1 ,ℓ
and that the trajectory be at least αk1 ,ℓ simple to apply the inductive hypothesis and obtain the
(βk1 ,ℓ q −3 , q) non-degeneracy of the orbit for q < Qk1 ,ℓ . In fact, the first condition that we impose
on αℓ+1,ℓ+1 is that α̃1 < βk1 ,ℓ .
Then, we need to ask whether the trajectory is (m, αk1 ,ℓ ) simple. If it is, we can apply the
inductive hypothesis to obtain non-degeneracy. However, if it is not, we need to repeat this process,
obtaining that ρ is (m2 , α̃2 ) returning, with α̃2 ≤ C m1 αk1 ,ℓ and m2 ≤ m1 /2 with m2 ∈ (2k2 , 2k2 +1 ].
As in the case of treating ρ as (m1 , α˜1 ) returning, we compare α̃2 and βk2 ,ℓ−1 and ask whether
ρ is (m2 , αk2 ,ℓ−1 ) simple. The requirement that α̃2 ≤ βk2 ,ℓ−1 puts an additional upper bound on
αk1 ,ℓ . If the trajectory is not simple enough, we repeat once again finding (mi , ki , α̃i ) until ρ is
(mi , αki ,ℓ−i+1 ) simple. Note that this will happen, since every 1 returning trajectory is simple
and hence, if mi = 1, then, provided we have chosen the αk,ℓ correctly, ρ is β0,0 returning and
α0,0 simple and hence we may apply the inductive hypothesis.
Since mi ≤ mi−1 /2, we have mi ≤ 2−i n ≤ 2ℓ+1−i . In some circumstances, we may be working
with n = 2ℓ+1 and the iteration may not terminate until mi = 1. If this is the case, then we
want to use the inductive hypothesis on (1, β1,1 ) returning points to obtain non-degeneracy of γ.
However, this requires that every (1, β0,0 ) returning point be (β0,0 2−3(ℓ+1) , 2ℓ+1 ) non-degenerate.
This will eventually fail, since Q1,1 < ∞ and ℓ is unbounded above. This can be remedied by
putting 1 < a < 2 and replacing 2-adic intervals by a-adic intervals. By doing this, we guarantee
that 1 ≤ mi ≤ 2−i aℓ and hence, the above iteration terminates in i < cℓ steps for some c < 1.

4.5. Comments on the lack of a global Poincaré section. In the outline above, we have
worked as though one can find a single compact symplectic manifold without boundary, Γ ⊂ S ∗M ,
to serve as a Poincaré section. In general, this may not be the case and one may need to work
with Poincaré sections with boundaries. This leads to a number of technical complications in
the argument which are handled by introducing the notion of chains of symplectomporphisms
and well-separated sets (see Section 6.2). However, these technical complications do not lead to
any fundamental change in the proof and the reader may wish to imagine that there is a global
Poincaré section; in particular, the reader can then safely ignore the discussion about chains
of symplectomorphisms and well-separated sets, instead thinking of the Poincaré map and its
iterates.

4.6. Comments on proving predominance. In order to prove predominance instead of den-

sity, at each
Q step of the induction, rather than finding a single metric, gℓ , we produce a family gℓ,σ
for σ ∈ ∞ j=1 BRL (0, 1) in a set with almost full measure. This is possible since we actually showed
that, at each step, all but a small measure of our perturbations have the required properties.

5. Volume of the set of (β, q)-degenerate symplectic matrices

In this section, we do some preliminary work on volumes and covering numbers (see Definition
5.4) in the space of symplectic matrices with real coefficients in 2ℵ dimensions, Sp(2ℵ). As
explained in Section 4, it is necessary to work with (β, q)-non-degenerate matrices (see (4.1) for
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 25

the definition). The goal of this section is to understand the covering number of the complement
of these matrices in Sp(2ℵ).

5.1. Properties of eigenvalues of symplectic matrices. We recall here a few properties of

the spectrum of a symplectic matrix A ∈ Sp(2ℵ). A 2ℵ × 2ℵ matrix A with real coefficients is in
Sp(2ℵ) provided
 
0 I
AT JA = J, J = −I 0ℵ , (5.1)
ℵ

where Iℵ ∈ M(ℵ) is the identity matrix. Symplectic matrices are invertible with kAk = kA−1 k.
They also have det A = 1.
We note that if λ is in the spectrum of A and has multiplicity m, then λ−1 , λ̄ and λ̄−1 are
also in the spectrum and have multiplicity m. However, λ, λ−1 , λ̄ and λ̄−1 may not be distinct.
/ R and |λ| =
First, if λ ∈ 6 1, then all four are distinct. Thus, each non-real eigenvalue off the unit
circle comes in a quadruplet {λ, λ−1 , λ̄, λ̄−1 }. Next, if λ ∈
/ R with |λ| = 1, then λ−1 = λ̄ and such
−1
eigenvalues come in pairs of the form {λ, λ }. Third, if λ ∈ R with |λ| = 6 1, then λ = λ̄ and the
eigenvalues again come in pairs of the form {λ, λ−1 }. Finally, if λ = ±1, then λ = λ̄ = λ−1 = λ̄−1 ,
and it is known that ±1 always occurs with even multiplicity.
We say that an invertible matrix is hyperbolic if none of its eigenvalues lie on the unit circle.
Note that if A is hyperbolic then the same is true for Aq for all q. Moreover, the distance from
the eigenvalues to the unit circle is increasing with q.
One important consequence of the facts we recalled about eigenvalues of symplectic matrices is
that every small enough perturbation of a symplectic matrix with an eigenvalue λ0 ∈ {λ : |λ| =
1, λ 6= ±1} having odd multiplicity must have an eigenvalue on the unit circle since eigenvalues off
the unit circle come in quadruplets. In other words, certain eigenvalues are ‘trapped’ on the unit
circle. Thus, unlike general matrices, symplectic matrices cannot be perturbed to be hyperbolic.
This makes perturbing a matrix to one which is (β(q), q)-non-degenerate for all q much more
delicate in the symplectic category than in the category of matrices and requires the delicate
induction argument carried out in Section 8.

5.2. (β, q)-non-degeneracy and the location of eigenvalues. We start by discussing the
relationship between (β, q)-non-degeneracy and the location of the eigenvalues {λj (A)}2ℵ
j=1 of a
matrix A.
Lemma 5.1. For all A ∈ M(2ℵ), q ∈ N, s ∈ (0, 1),

2ℵ
if min |λj (A) − e2πip/q | ≥ s, then | det(I − Aq )| > min( 25 qs , 51 ) .
1≤j≤2ℵ
0≤p≤q−1
Q2ℵ
Proof. First, note that λj (I − Aq ) = 1 − λj (A)q . Therefore, | det(I − Aq )| = j=1 (1 − λj (A)q ) .
The lemma will follow once we prove that for all q ∈ N, s ∈ (0, 1),
min |λj (A) − e2πip/q | ≥ s, ⇒ min |1 − λj (A)q | ≥ min( 25 qs, 51 ). (5.2)
j,p j
Fix 1 ≤ j ≤ 2ℵ, and set λj (A) = reiθ for some r > 0 and θ ∈ [0, 2π). First, notice that if
|r − 1| ≥ 12 , then |1 − λj (A)q | ≥ 12 for all q ∈ N, and hence (5.2) holds in this case. Therefore, we
assume that |r − 1| < 21 from now on.
26 YAIZA CANZANI AND JEFFREY GALKOWSKI

Next, fix q ∈ N, s ∈ (0, 1), and for p ∈ N let αp = θ − 2π pq . We claim that

either |αp | ≥ 21 s for all p, or |r − 1| ≥ 12 s. (5.3)

To prove the claim suppose there exists p such that |αp | < 2s . First, note that by the assumption
in (5.2)
s2 ≤|reiθ − e2πip/q |2 = |1 − reiαp |2 = (r − cos αp )2 + (sin αp )2 .
2
Therefore, 3s4 ≤ (r − cos αp )2 since | sin αp | ≤ |αp | < s/2. In particular, since |r − 1| < 1
2 and
α 2
|1 − cos αp | = 2| sin2 ( 2p )| ≤ s8 ,
3s2

2 2
4 ≤ r − 1 + (1 − cos αp ) ≤ (r − 1)2 + s4 ,
and hence |r − 1| ≥ √1 s ≥ 21 s as claimed in (5.3).
2
s
With (5.3) in place, we first consider the case |r − 1| ≥ 2 and divide it into two sub-cases.
Case 1: Suppose that r ≥ 1 + 2s . Then, |1 − λj (A)q | ≥ r q − 1 ≥ q 2s as claimed in (5.2).
Case 2: Suppose r ≤ 1 − 2s . Then, r −q ≥ 1 + q 2s (1 − 2s )−1 since s ∈ (0, 1). Thus, r q ≤
1 − q 2s 1+(q−1)
1
s . In particular, as claimed in (5.2),
2

|1 − λj (A)q | ≥ 1 − r q ≥ qs(2 + (q − 1)s)−1 ≥ min(qs 25 , 15 ).

Finally, suppose that |r − 1| < 2s . Then, |αp | ≥ s

for all p by (5.3) and hence
2
(
q 2 q iqθ 2 2 2 2 sin2 (qθ) inf k∈Z |qθ − 2πk| < π2 ,
|1−λj (A) | = |1−r e | ≥ inf (1 − a cos(qθ)) + a sin (qθ) =
a≥0 1 else.


In particular, since 2s ≤ minp |αp | ≤ πq , we have |1 − λj (A)q | ≥ | sin q 2s | ≥ min( 25 qs, 15 ).


In what follows, it will be necessary to avoid symplectic matrices which have eigenvalues close
to a given number. To this end, for θ ∈ R and s > 0 let
D(θ, s) := {A ∈ Sp(2ℵ) : BC (eiθ , s) ∩ spec(A)6=∅}. (5.4)

We now record a corollary of Lemma 5.1 which we use in the next section. It controls the size
of the inverse of I − Aq in terms of the distance between the spectrum of A and the roots of unity
{e2πip/q }0≤p<q .
Sq−1
Lemma 5.2. Let q ∈ N, 0 < s ≤ 2q 1
, and A ∈ M(2ℵ). If A ∈/ p=0 D( 2πp
q , s), then,

k(I − Aq )−1 k ≤ ( 25 )2ℵ (qs)−2ℵ (1 + kAkq )2ℵ−1 .

Proof. Note that by assumption minj,p |λj (A) − e2πip/q | ≥ s and hence | det(I − Aq )| ≥ ( 25 qs)2ℵ
1
by Lemma 5.1 together with s ≤ 2q . In particular, the corollary follows from the fact that
kB−1 k ≤ | det(B)|−1 kBk2ℵ−1 for all B ∈ M(2ℵ). Indeed, let σ12 ≤ σ22 ≤ · · · ≤ σ2ℵ 2 be the
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 27

Q2ℵ
eigenvalues of B∗ B. Then, det(B∗ B) = | det(B)|2 = 2
j=1 σj , kB−1 k2 = σ1−2 , and kBk2 = σ2ℵ
2 .

Therefore, as claimed,
Y
2ℵ −1
2(2ℵ−1)
kB−1 k2 = σ1−2 ≤ σ2ℵ σj2 = kBk2(2ℵ−1) | det(B)|−2 .
j=1

An immediate consequence of the previous lemma is the following.

Corollary 5.3. Let q ∈ N and β > 0 with 2β < q 2 . If A is not (β, q) non-degenerate, then
Sq−1
A ∈ p=0 D( 2πp −3
q , βq ).

Corollary 5.3 will be used to estimate the volume of the set of matrices which are not (β, q)
non-degenerate using covering numbers.

5.3. Covering numbers and (β, q) non-degeneracy. Throughout the article, the notion of
the covering number of a set will play an important role. Let (Y, d) be a metric space.
Definition 5.4. The covering number of a set V ⊂ Y and radius s is defined by
n N
[ o
MY (V, s) := inf N ∈ N : ∃{yi }N
i=1 ⊂ Y such that V ⊂ B(yi , s) .
i=1

5.3.1. Covering numbers in M(2ℵ) and Sp(2ℵ). In this subsection, we will relate covering numbers
in M(2ℵ) to those in Sp(2ℵ). The goal is to compare covering numbers of a given set as measured
by balls in Sp(2ℵ) and in M(2ℵ). The following lemma controls how Sp(2ℵ) sits inside M(2ℵ).
Lemma 5.5. There exists ε0 > 0 such that for all A ∈ Sp(2ℵ) and 0 < r ≤ ε0 kAk,
r
BM(2ℵ) (A, 2kAk2 ) ∩ Sp(2ℵ) ⊂ BSp(2ℵ) (A, r).

Proof. Let A ∈ Sp(2ℵ) and set V := A−1 (BSp(2ℵ) (A, r))⊂ Sp(2ℵ). We have that V is a neighbor-
hood of I in Sp(2ℵ) and BSp(2ℵ) (I, kAk−1 r) ⊂ V. Let ε0 = ε0 (ℵ) be such that for 0 < r̃ ≤ ε0

BSp(2ℵ) (I, 14 r̃) ⊂ BM(2ℵ) (I, 12 r̃) ∩ Sp(2ℵ) ⊂ BSp(2ℵ) (I, r̃)
r
Therefore, BM(2ℵ) (I, 2kAk )∩ Sp(2ℵ) ⊂ V for r ≤ ε0 kAk. Hence,
     
r r
BM(2ℵ) A, 2kAkkA −1 k ∩ Sp(2ℵ) ⊂ A BM(2ℵ) (I, 2kAk )∩ Sp(2ℵ) ⊂ AV = BSp(2ℵ) (A, r).

Since A ∈ Sp(2ℵ), we have kAk = kA−1 k and the result follows. 

With Lemma 5.5 in place, we are ready to study the relationship between the covering numbers
of a given set as measured by balls in Sp(2ℵ) and in M(2ℵ). To ease notation, for r > 0 we also
introduce the ball
BA0(r) := {A ∈ M(2ℵ) : kA − A0 k ≤ r}.
28 YAIZA CANZANI AND JEFFREY GALKOWSKI

Lemma 5.6. There exists ε0 > 0 such that the following holds. If Z ⊂ B0(r) ∩ Sp(2ℵ) for some
r > 0, A0 ∈ Sp(2ℵ), 0 < s ≤ ε0 r, and r0 > 0, then
MSp(2ℵ) (Z ∩ BA0(r0 ), s) ≤ MM(2ℵ) (Z ∩ BA0(r0 ), 14 sr −2 ).

SN
Proof. Fix r0 > 0. Let s̃ > 0 and {Ai }N i=1 ⊂ M be such that Z ∩ BA0(r0 ) ⊂ i=1 BAi (s̃).
Without loss of generality, we may assume that BAi (s̃) ∩ Z = N
6 ∅. Let {Zi }i=1 ⊂ Z be such that
SN
Z ∩ BA0(r0 ) ⊂ i=1 BM(2ℵ) (Zi , 2s̃). Then, since Z ⊂ B0(r) ∩ Sp(2ℵ), by Lemma 5.5 there is ε0 > 0
such that
N
[ N
[
2
Z ∩ BA0(r0 ) ⊂ BSp(2ℵ) (Zi , 4s̃kZi k ) ⊂ BSp(2ℵ) (Zi , 4s̃r 2 ),
i=1 i=1

provided 4s̃r ≤ ε0 . In particular, MSp(2ℵ) (Z ∩ BA0(r0 ), 4s̃r 2 ) ≤ MM(2ℵ) (Z ∩ BA0(r0 ), s̃). The
Lemma follows by putting s = 4s̃r 2 . 

5.3.2. Covering numbers for matrices that are not (β, q) non-degenerate. We now study the cov-
ering numbers of the set of matrices in Sp(2ℵ) with a given eigenvalue of unit length: for θ ∈ R
let
D(θ) := {A ∈ Sp(2ℵ) : eiθ ∈ spec(A)}. (5.5)
In particular, D(θ) = ∩s D(θ, s) with D(θ, s) as in (5.4). We start by bounding the covering
number of D(θ) in M(2ℵ).
Lemma 5.7. There is C > 0 such that for all A0 ∈ M(2ℵ), θ ∈ R, r0 > 0, and 0 < s < r0 ,
MM(2ℵ) (D(θ) ∩ BA0(r0 ), s) ≤ C(r0 /s)L−2ℵ−1 , L = ℵ(2ℵ + 3).

Proof. Note that Sp(2ℵ) ⊂ M(2ℵ) is defined by ℵ(2ℵ − 1) algebraically independent polynomial
equations. Now, having eigenvalue eiθ is equivalent to asking both ℜ det(A − eiθ I) = 0 and
ℑ det(A − eiθ I) = 0. In particular, D(θ) ⊂ R(θ) := {A ∈ Sp(2ℵ) : ℜ det(A − eiθ I) = 0}.
We claim that dim R(θ) = dim M(2ℵ) − (ℵ(2ℵ − 1) + 1) = L − 2ℵ − 1. Once we prove this,
we would have that D(θ) is an algebraic set with dim D(θ) ≤ L − 2ℵ − 1. Then, the bound on
the covering number for D(θ) ∩ BA0(r0 ) will follow from the estimates on covering numbers for
algebraic sets given in [YC04, Corollary 5.7].
To prove that dim R(θ) = L − 2ℵ − 1, it suffices to show that ℜ det(A − eiθ I) does not just
take the value 0 in Sp(2ℵ). Indeed, one may find matrices A in Sp(2ℵ) such that ℜ det(A − eiθ I)
takes all values in an open set in R. To see this, let A be the diagonal matrix with entries
(λ1 , λ2 , . . . , λℵ , λ−1 −1
1 , . . . , λℵ ) with λi ∈ R \ {0} and adjust λi as needed. (In fact, it is enough to
take λ3 , . . . , λℵ = 1 except when θ = 0 in which case one can take λ3 , . . . , λℵ = −1.) 

Our next goal is to bound the covering number of D(θ, s). To do this, in Lemma 5.9 below, we
first control how D(θ) sits inside D(θ, s). The proof of Lemma 5.9 hinges on the following result.
Lemma 5.8. The set {A ∈ Sp(2ℵ) : A has simple eigenvalues} is dense in Sp(2ℵ).
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 29
Q
Proof. Given A ∈ Sp(2ℵ) with eigenvalues {λi }2ℵ i=1 define the polynomial p(A) = i6=j (λi − λj )
and notice that A has simple eigenvalues if and only if p(A) 6= 0. Next, recall that any symmetric
polynomial in the eigenvalues of a matrix can be expressed as Q a polynomial in the coefficients of
that matrix [OCV11, Proposition 7.1.10]. In particular, since i6=j (λi − λj ) is symmetric, there
is q : Sp(2ℵ) → C such that q(A) is a polynomial in the (2ℵ)2 coefficients of the matrix with
q(A) 6= 0 if and only if A has simple eigenvalues (see also [dlCS21, Example ii]).
Let A ∈ Sp(2ℵ). Then A can be written as A = eX eY for two elements X, Y ∈ sp(2ℵ). (To see
this, we combine [Hal15, Corollary 11.10], which states that the exponential map is surjective on
compact, connected Lie groups, with the polar decomposition A ∈ Sp(2ℵ) = BeY with Y ∈ sp(2ℵ)
and B ∈ SO(2n)∩Sp(2n) [HN12, Proposition 4.3.3], noting that SO(2n)∩Sp(2n) is a compact Lie
group.) Furthermore, it is easy to see that any diagonal matrix D = diag(λ1 , . . . , λℵ , λ−1 −1
1 , . . . λℵ ),
with λi ∈ (0, ∞)\{1}, λi 6= λj , for i 6= j has simple eigenvalues and lies in Sp(2ℵ). Fix one such
D and for t ∈ R let
F (t) = D 1−t etX etY ∈ Sp(2ℵ).
We claim that there is {tn }∞
n=1 ⊂ R such that F (tn ) → F (1) = A, and F (tn ) has simple
eigenvalues. This would prove the density statement. To prove the claim note that q(F (t))
extends to an analytic function of t ∈ C and q(F (0)) = q(D) 6= 0. In particular, q(F (t)) has
a discrete set of zeros and hence, there is tn ∈ R such that tn → 1 and q(F (tn )) 6= 0. Since
F : R → Sp(2ℵ) is continuous, F (tn ) → A as claimed. 
Lemma 5.9. Let a > 0 and suppose that A ∈ Sp(2ℵ) has an eigenvalue λ with |λ| > a. Then,
for all θ ∈ R there is B ∈ D(θ) and
kA − Bk ≤ 3(1 + a−1 )|λ − eiθ |. (5.6)

Proof. Let θ ∈ R. Without loss of generality, we may assume that λ 6= eiθ since otherwise the
claim is trivial. By Lemma 5.8, for any ε > 0, there is à ∈ Sp(2ℵ) with simple eigenvalues, one
of which we label µ, such that
kA − Ãk ≤ ε|λ − eiθ |, |µ − λ| ≤ ε min(|λ − eiθ |, 1). (5.7)
Without loss of generality we assume that ε is small enough that
ε + 2(1 + ε)(1 + |a − ε|−1 ) ≤ 3(1 + a−1 ). (5.8)
/ S 1 ∪ R,
Let Eµ (Ã) denote the eigenspace of µ for Ã. Then we need to consider three cases µ ∈
µ ∈ S 1 , and µ ∈ R.
Case 1: µ ∈ / S 1 ∪ R. In this case, one has that µ, µ−1 , µ̄, µ̄−1 are distinct, and hence we have
Eµ , Eµ̄ , Eµ−1 , Eµ̄−1 are distinct subspaces of dimension 1. Note that Eµ = Eµ̄ and Eµ−1 = Eµ̄−1 .
Consider now the matrix
B = Ã + (eiθ − µ)Πµ + (e−iθ − µ̄)Πµ̄ + (e−iθ − µ−1 )Πµ−1 + (eiθ − µ̄−1 )Πµ̄−1 ,
where Πµ denotes the projector which gives the component of a vector in the Eµ direction i.e.
if {µi }2ℵ are the eigenvalues of Ã, then Πµi : C2ℵ → Eµi are the unique operators such that
Pi=1
v = i Πµi v for v ∈ C2ℵ .
30 YAIZA CANZANI AND JEFFREY GALKOWSKI

Note that, (B − eiθ )Πµ = 0 since AΠµ = µΠµ . That is, B has eigenvalue eiθ . Then, by (5.7)
kB − Ãk ≤ |eiθ − µ| + |e−iθ − µ−1 | + |e−iθ − µ̄| + |eiθ − µ̄−1 |
≤ 2(1 + |µ|−1 )|µ − eiθ | ≤ 2(1 + |µ|−1 )(1 + ε)|λ − eiθ |
≤ 2(1 + |a − ε|−1 )(1 + ε)|λ − eiθ |.
The bound on (5.6) then follows from (5.8) and (5.7).
To see that B ∈ D(θ), it remains to check that B ∈ Sp(2ℵ). First, to see that B has real
coefficients one must check that Bv = Bv for all v real valued vectors. This claim follows from
the facts that
Ãv = Ãv, Πµ v̄ = Πµ̄ v, Πµ̄ v̄ = Πµ v, Πµ−1 v̄ = Πµ̄−1 v, Πµ̄−1 v̄ = Πµ−1 v.
To see that B is symplectic, note that if µi µj 6= 1, then hJvµi , vµj i = 0 for vµi ∈ Eµi since
hJvµi , vµj i = hJ Ãvµi , Ãvµj i = µi µj hJvµi , vµj i. (5.9)
Therefore, since B preserves Eµi for all i and R2ℵ = ⊕2ℵ
i=1 Eµi it suffices to note

hJBvµ , Bvµ−1 i = hJ(Ã + (eiθ − µ))vµ , (Ã + (e−iθ − µ−1 ))vµ−1 i = hJvµ , vµ−1 i.

Case 2: µ ∈ S 1 . Since µ ∈ S 1 , then µ = µ̄−1 . We consider

B = Ã + (eiθ − µ)Πµ + (e−iθ − µ−1 )Πµ−1 .

As before, (B − eiθ )Πµ = 0, and kà − Bk ≤ (1 + ε)(1 + |a − ε|−1 )|λ − eiθ |. The bound on (5.6)
then follows from (5.8) and (5.7). Therefore, to see that B ∈ D(θ) we need to check that B is
real and symplectic. The computation is identical to that in Case 1.
Case 3: µ ∈ R. In the case µ ∈ R, we let vµ ∈ R2ℵ with kvµ k = 1 and (Ã − µ)vµ = 0, vµ−1 ∈ R2ℵ
with kvµ−1 k = 1, and (Ã − µ−1 )vµ−1 = 0. Observe that R2ℵ = Eµ ⊕ Eµ−1 ⊕2ℵ i=3 Eµi , with
/ {µ, µ−1 }, and dim Eµi = 1 for i = 3, . . . , 2ℵ. Note that by (5.9) we have hJvµi , vµ i = 0 and
µi ∈
hJvµi , vµ−1 i = 0 for vµi ∈ Eµi and i = 3, . . . , 2ℵ. We then set
B = Ã + (cos θ − µ)vµ vµt Πµ + (cos θ − µ−1 )vµ−1 vµt −1 Πµ−1 − sin(θ)vµ vµt −1 Πµ−1 + sin(θ)vµ−1 vµt Πµ .
One can check as before that B ∈ Sp(2ℵ). Then, B ∈ D(θ) since (B − eiθ )(vµ − ivµ−1 ) = 0.
Finally,
q q
kB − Ãk ≤ 2 | cos θ − µ| + sin θ + 2 | cos θ − µ−1 |2 + sin2 θ = |µ − eiθ | + |µ−1 − eiθ |
2 2

≤ 2(1 + ε)(1 + (a − ε)−1 )|λ − eiθ |.

The bound on (5.6) then follows from (5.8) and (5.7). 

We proceed to study the covering number for D(θ, s).

Lemma 5.10. There exists ε0 > 0 so that if r0 > 0, r≥ 1, 0 < s ≤ ε0 r, θ ∈ R, and A0 ∈ Sp(2ℵ),
then



MSp(2ℵ) D(θ, s) ∩ B0(r) ∩ BA0(r0 ), s ≤ MM(2ℵ) D(θ) ∩ B0(r) ∩ BA0(r0 ), 12 1
sr −2 .
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 31

Proof. Let r0 > 0, r > 0, s > 0, θ ∈ R, A0 ∈ Sp(2ℵ). By Lemma 5.6 there is ε0 > 0 such that
MSp(2ℵ) (D(θ, s) ∩ B0(r) ∩ BA0(r0 ), s) ≤ MM(2ℵ) (D(θ, s) ∩ B0(r) ∩ BA0(r0 ), 4rs2 ),
1
for 0 < s ≤ ε0 r. We claim that for all s̃ > 0, 0 < s′ < 2
MM(2ℵ) (D(θ, s′ ) ∩ B0(r) ∩ BA0(r0 ), 6s′ + s̃) ≤ MM(2ℵ) (D(θ) ∩ B0(r) ∩ BA0(r0 ), s̃). (5.10)
S N
Let {Ai }Ni=1 ⊂ M(2ℵ) be so that D(θ) ∩ B0(r) ∩ BA0(r0 ) ⊂ i=1 BM(2ℵ) (Ai , s̃). Suppose A ∈
D(θ, s ) ∩ B0(r) ∩ BA0(r0 ). Then, there exists λ ∈ spec A with |λ − eiθ | ≤ s′ . In particular, by
′

Lemma 5.9 d(A, D(θ)) ≤ 6s′ , so

  N
[
D(θ, s′ ) ∩ B0(r) ∩ BA0(r0 ) ⊂ BM(2ℵ) (D(θ) ∩ B0(r) ∩ BA0(r0 ), 6s′ ) ⊂ BM(2ℵ) (Ai , 6s′ + s̃).
i=1
s s
The result follows from (5.10) after setting = s′and s̃ =
36r 2
which yield s̃ + 6s′ = 41 sr −2 ,
12r 2
,
ε0 r 1
and observing that, since r ≥ 1 and s ≤ ε0 r, shrinking ε0 if necessary s′ ≤ 36r 2 ≤ 2. 

Next, we use the results above to control the covering numbers of (β, q)-degenerate matrices.
Sq−1
Recall that by Corollary 5.3, if A is not (β, q) non-degenerate, then A ∈ p=0 D( 2πp −3
q , βq ). In
particular, the following result bounds the covering number for such matrices.
Corollary 5.11. There exists ε0 > 0 C > 0 such that for all A0 ∈ Sp(2ℵ), q ∈ N, r0 > 0, r≥ 1,
0 < s ≤ min(ε0 r, 12r 2 r0 ),
 q−1
[ 
MSp(2ℵ) D( 2πp
q , s) ∩ B 0(r) ∩ B (r
A0 0
) , s ≤ Cq (r 2 r0 /s)L−2ℵ−1 .
p=0

Proof. The Corollary follows from combining Lemmas 5.7 and 5.10. 

6. Returning points, simple points, and their iterates

In Section 6.1, we introduce a version of returning points with discretized time as well as the
notion of a simple point. Next, in Section 6.2 we introduce the concept of a well separated set
for the geodesic flow and the corresponding chain of symplectomorphisms. These concepts are
replacements for, respectively, a global Poincaré section and the Poincaré map associated to the
global section and, while they require some technical work, do not substantially change the main
idea of the proof. Because of this, the reader may first wish to replace the concepts of chains of
symplectomorphisms and well separated sets by the simpler notions of Poincaré map and section.
In the next sections of this paper, we will be varying the metric g. Because of this, it will be
useful to have a single space on which the geodesic flow for any g is defined. This space will be
canonically isomorphic to S ∗M for any g and will be defined as follows. Let
]
S ∗M := (T ∗ M \ {0})/ ∼, (x, λξ) ∼ (x, ξ) for all λ > 0.
Then, since the geodesic flow (the Hamiltonian flow for p(x, ξ) = |ξ|g ), is homogeneous of degree
]
0 in ξ, the flow ϕgt passes naturally to the quotient S ∗M . We also endow S ]∗M with the distance
∗
inherited from Sg M for some fixed reference metric gf .
f
32 YAIZA CANZANI AND JEFFREY GALKOWSKI

6.1. Returning points and simple points. In this section, we define the notions of returning
trajectories of length n and simple trajectories of length n, and show that non-simple returning
points can be seen as iterates of shorter returning trajectories. Because we will pass from the
continuous time flow to a discrete time object below, these notions will be defined relative to a
fixed number ̥ such that we can guarantee there are no periodic (or near periodic) trajectories
of length < ̥. In Section 6.2 we will also insist that the flow can be effectively reduced to a
discrete time map by cutting in time at intervals of length ∼ ̥. We will be working with metrics
in small balls B ⊂ G 3 and will show in Section 10 that one can take ̥ to be a small multiple of
the injectivity radius for a fixed metric in B.
The notion of a returning point generalizes that of periodicity.
Definition 6.1. For ̥ > 0, g ∈ G 2 , n ∈ N, and δ > 0, we write
ρ ∈ R̥ (n, δ, g) if inf d(ϕgt (ρ), ρ) < δ.
(n−1)̥<t≤n̥
t>̥/4

In this case we say that ρ is ̥-(n, δ, g) returning.

It will be important to have a notion of a simple returning point of length n, i.e. one which is
not returning for any smaller n.
Definition 6.2. For ̥ > 0, g ∈ G 2 , n ∈ N, and α > 0, we write
ρ ∈ S̥ (n, α, g) if inf d(ϕgt (ρ), ρ) > α.
1
2
̥<t<(n− 12 )̥

In this case we say that ρ is ̥-(n, α, g) simple.

The importance of simple points comes from the fact that the effect of perturbations of g on
simple trajectories can be understood. This is much harder to do when considering non-simple
trajectories because the trajectory will interact with a given perturbation many times.
The following two lemmas are similar to [Yom85, Lemma 3.1] and show that non-simple re-
turning trajectories are multiples of shorter returning trajectories.
Lemma 6.3. Let K ⊂ G 3 (Γ) be bounded and ̥ > 0. Then there is C > 0 such that the following
∗M with d(ϕg (ρ), ρ) < α. If d(ϕg (ρ), ρ) < α where ̥ ≤ |s| < |t|
holds. Let g ∈ K, α > 0, and ρ ∈ Sg t s 2
and q ∈ Z such that t = qs + r with |r| < |s|, then
d(ϕgr (ρ), ρ) < C |t| α.

Proof. By (3.2) there is C = C(K) ≥ 1 such that, for all m ∈ Z

d(ϕgms (ρ), ϕg(m−1)s (ρ)) ≤ C |s|d(ϕg(m−1)s (ρ), ϕg(m−2)s (ρ)).
P
Using that d(ϕgs (ρ), ρ) < α and |s| ≥ ̥2 > 0, we have d(ϕgms (ρ), ρ) ≤ m−1 j=0 C
|js|α ≤ 2C |ms| α.

Similarly, since d(ϕgt (ρ), ρ) ≤ α, we have d(ϕgt−qs (ρ), ϕg−qs (ρ)) ≤ C |qs|α. It follows that
d(ϕgr (ρ), ρ) ≤ d(ϕgt−qs (ρ), ϕg−qs (ρ)) + d(ϕg−qs (ρ), ρ) ≤ 3C |qs|α ≤ C |t| α,
where in the last inequality we used that |qs| = |t − r| ≤ |t| + |r| ≤ 2|t|. 
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 33

Lemma 6.4. Let K ⊂ G 3 be bounded and ̥ > 0. Suppose there are c > 0, C̥ > 0 such that for
all g ∈ K and ρ ∈ Sg ∗M we have d(ϕg (ρ), ρ) ≥ c|t| if |t| ≤ C ̥. Then there is C > 0 such that for
t ̥

all g ∈ K, δ > 0, t0 > ̥, α ≥ δ, and ρ ∈ Sg ∗M satisfying

d(ϕgt0 (ρ), ρ) < δ,

one of the following two statements hold:
(1) d(ϕgs (ρ), ρ) ≥ α for all ̥2 ≤ s ≤ t0 − ̥2 .
(2) there is s0 ∈ [C̥ ̥ − C t0 α, t20 ] such that d(ϕgs0 (ρ), ρ) ≤ C t0 α and there is q ∈ N, q > 1,
such that qs0 = t0 . Moreover, if t1 = q1 s0 + r ≤ t0 for q1 ∈ N and 0 ≤ r < s0 , then
d(ϕgt1 (ρ), ϕgr (ρ)) ≤ C t0 α.

Proof. Suppose that statement (1) is false. Then there is s such that ̥2 ≤ s ≤ t0 − ̥2 and
d(ϕgs (ρ), ρ) < α. Set t1 = s and let q1 ∈ Z such that t0 = q1 t1 + t2 for |t2 | ≤ 2s . Then, by
Lemma 6.3, d(ϕgt2 (ρ), ρ) ≤ C t0 α. If |t2 | ≤ C̥ ̥, we put s̃0 = t1 .
If |t2 | > C̥ ̥, we continue the process and suppose we have found {qi }m+1 m+2
i=1 ⊂ Z, and {ti }i=0
such that for i = 2, . . . , m,
Pi−2
ti = qi+1 ti+1 + ti+2 , |ti+2 | ≤ 21 |ti+1 |, |ti | > C̥ ̥, d(ϕgti (ρ), ρ) ≤ C j=0 |tj |
α.
Then, letting qm+2 ∈ Z, |tm+3 | ≤ 21 |tm+2 |, such that tm+1 = qm+2 tm+2 + tm+3 , Lemma 6.3 yields
Pm+1
d(ϕgtm+3 (ρ), ρ) ≤ C j=0 |tj | α. Then, if |tm+3 | ≤ C̥ ̥, we set s̃0 = tm+2 . In particular, since
P
|tj | ≤ 21 |tj−1 | for j = 2, . . . , we have ∞j=0 |tj | ≤ 3|t0 |. Therefore, we have found s̃0 such that

d(ϕgs̃ (ρ), ρ) ≤ C 3|t0 | α.

0

Let this process terminate with |tm+3 | ≤ C̥ ̥. We claim that |tm+3 | ≤ C 4|t0 | α. Indeed, since
d(ϕgt (ρ), ρ) ≥ c|t| for |t| ≤ C̥ ̥ and |tm+3 | ≤ C̥ ̥, by Lemma 6.3 we know
c|tm+3 | ≤ d(ϕgtm+3 (ρ), ρ) ≤ C 3|t0 | α. (6.1)
tm+3
Finally, set s0 = s̃0 + qm+2 . Then, after possibly modifying C, and using the bound on |tm+3 |
from (6.1) we conclude
d(ϕgs0 (ρ), ρ) ≤ C |tm+3 | d(ϕgs̃0 (ρ), ρ) + d(ϕgtm+3 (ρ), ρ) ≤ 2C 4t0 α
qm+2

tm+3
as claimed, since |qm+2 | ≥ 1. Note that, since tm+1 = qm+2 tm+2 + tm+3 = qm+2 (tm+2 + qm+2 ) =
qm+2 s0 , we have by construction t0 = qs0 for some q ∈ Z. In addition,
|s0 | ≥ |tm+2 | − | qtm+3
m+2
| ≥ C̥ ̥−c−1 C 4|t0 | α.

If s0 > 0, we have the claimed properties of s0 . If, on the other hand, s0 < 0, we have,
modifying C if necessary,
d(ϕg−s0 (ρ), ρ) ≤ C |s0 | d(ρ, ϕgs0 (ρ)) ≤ C 4|t0 | α.
34 YAIZA CANZANI AND JEFFREY GALKOWSKI

Finally, observe that for qs0 ≤ t0 ,

q−1
X q−1
X
d(ϕgqs0 (ρ), ρ) ≤ d(ϕgjs0 (ρ), ϕg(j+1)s0 (ρ)) ≤ C j|s0| d(ϕgs0 (ρ), ρ) ≤ C |t0 | α.
j=0 j=0

Therefore, if t1 = qs0 + r ≤ t0 for some q ∈ N, 0 ≤ r < s0

d(ϕgt1 (ρ), ϕgr (ρ)) ≤ C r d(ϕgqs0 (ρ), ρ) ≤ C |t0 | α.


6.2. Non-degeneracy and chains of symplectomorphisms. In general, it is not possible to

find a global Poincaré section for the geodesic flow; i.e. a connected, closed submanifold, Γ, of
co-dimension 1 in S ]∗M that is everywhere transverse to H such that every geodesic passes
|ξ|g
through Γ. Because of this, we need to replace the notion of a global Poincaré section and its
associated Poincaré map with a more complicated submanifold which captures all the dynamical
information for ϕgt . To this end, we introduce below, the notion of a well separated set and its
associated chains of symplectomorpshims; replacing, respectively, the global Poincaré section and
its associated Poincaré map. The idea of a well-separated set will be to find, Γ̃, a disjoint union
of open 2d − 2 dimensional submanifolds, and a compactly embedded open subset, Γ ⊂ Γ̃, such
that 1) every geodesic passes through Γ in a controlled time and 2) no geodesic passes in a very
short time from Γ̃ to itself. These properties will guarantee that understanding geodesics which
pass from Γ to itself allows one to understand the full dynamics for the geodesic flow.
To understand the need for Γ̃ note that since the connected components of Γ have boundaries,
there are points ρ ∈ Γ such that small perturbations of ρ will cause large jumps in the first
impact point PΓ (ρ) of the geodesic through ρ with Γ; i.e there are sequences ρn → ρ such that
d(PΓ (ρn ), PΓ (ρ)) ≥ c > 0. This forces us to include the slightly larger Γ̃ and introduce the notion
of Poincaré chains below. (See Figure 3 for a schematic of a well-separated set and a Poincaré
chain.)
Let Wi ⊂ S] ∗M , i = 1, . . . , N be disjoint, open symplectic submanifolds of dimension 2d−2 = 2ℵ

and Vi ⋐ Wi open subsets. We write

N
G N
G
Γ := Vi , e :=
Γ Wi . (6.2)
i=1 i=1

Definition 6.5. For ̥ > 0, and G ⊂ G2

bounded, we say that {(Wi , Vi )}N
i=1 are ̥-well-separated
for G if the Wi are uniformly transverse to H|ξ|g for all g ∈ G and there are cΓ , CΓ > 1 such that

sup sup inf{t > 0 : ϕgt (ρ) ∈ Γ} < CΓ ̥, inf inf inf{t > 0 : ϕgt (ρ) ∈ Γ̃} ≥ cΓ ̥,
g∈G ρ∈S]
∗M g∈G ρ∈Γ̃
(6.3)
inf min inf inf{t > 0 : ϕgt (ρ) ∈ W i }>CΓ ̥.
g∈G i ρ∈Wi

When it will not lead to confusion, we will say that Γ is ̥-well-separated.

In words, if {(Wi , Vi )}N

i=1 is ̥-well-separated for G, it means that (1) every trajectory hits Γ
in time less than CΓ ̥; (2) trajectories starting from Γ̃ always take at least cΓ ̥ time to return to
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 35

Γ̃; (3) trajectories that start within some Wi always take at least CΓ ̥ to return to Wi no matter
the choice of i.
Before we move on to the definition of chains of symplectomorphisms associated to well sepa-
rated sets, we show that Lemma 6.4 applies when there is an ̥-well separated set.
Lemma 6.6. Let G ⊂ G 2 bounded. Suppose that ̥ > 0 and Γ is ̥-well separated for G. Then,
for all K ⊂ G bounded in G 3 , there is c > 0 such that for all g ∈ K and |t| ≤ CΓ ̥,
d(ϕgt (ρ), ρ) ≥ c|t|, ρ ∈ Sg
∗M .

Proof. First, notice that, since K is bounded in G 3 , there are c1 , c2 > 0 such that, d(ϕgt (ρ), ρ) ≥
]
c2 |t| for all g ∈ K, ρ ∈ S ∗M , and |t| < c . Thus, it remains only to check that there is c > 0 such
1
g
that d(ϕt (ρ), ρ)) ≥ c for all g ∈ K, ρ ∈ S ] ∗M , and c ≤ |t| ≤ C ̥.
1 ̥

]
Suppose by contradiction there are gn ∈ K, ρn ∈ S ∗M , and t with |t | ∈ [c , C ̥], such that
n n 1 ̥

d(ϕgtnn (ρn ), ρn ) → 0.
2
G ]
Then, without loss of generality, we may assume gn → g ∈ G, ρn → ρ ∈ S ∗M , and t → t with
n
|t| ∈ [c1 , C̥ ̥]. Now,
d(ϕgt (ρ), ρ) ≤ d(ϕgt (ρ), ϕgt n (ρ))+d(ϕgt n (ρn ), ϕgt n (ρ))+d(ϕgtnn (ρn ), ϕgt n (ρn ))+d(ϕgtnn (ρn ), ρn )+d(ρn , ρ).
Since gn → g in G 2 , the right hand side of the above inequality tends to 0 as n → ∞, and we
have ϕgt (ρ) = ρ. Let T0 := inf{s > 0 : ϕgs (ρ) ∈ Γ} and ρΓ := ϕgT0 (ρ) (note that 0 < T0 < ∞ by
the first inequality in (6.3)). Then, ϕgt (ρΓ ) = ϕgt+T0 (ρ) = ϕgT0 (ρ) = ρΓ , with c1 ≤ |t| ≤ C̥ ̥. This
contradicts the last inequality in (6.3). 

We next use well-separated sets to reduce the continuous flow ϕgt to a discrete time system.
Let ν ≥ 3, ̥ > 0, G ⊂ G ν bounded,, and {(Wi , Vi )}N
i=1 be ̥-well-separated for G. Then define
g g
Tj,i : Wi → R, Tj,i (ρ) := inf{t > 0 : ϕgt (ρ) ∈ Wj }, (6.4)
g g −1
and, with Uj,i := (Tj,i ) ((0, CΓ ̥)) ⊂ Wi , let κgj,i : g
Uj,i → Wj be the function
κgj,i (ρ) = ϕgT g (ρ) (ρ).
j,i

Lemma 6.7. For all g ∈ G, j, i, the map κgj,i : Uj,i

g
→ Wj is a C ν−2 symplectomorphism onto its
image.

Proof. Once we show that κgj,i is C ν−2 , the fact that κgj,i is a symplectomorphism onto its image
is inherited from the facts that ϕgt is a symplectomorphism and that Wj is transverse to H|ξ|g .
We now show that κgj,i is C ν−2 . To do this, note that Wj is open and transverse to H|ξ|g .
Therefore, since (t, ρ) 7→ φgt (ρ) is C ν−2 with ν − 2 ≥ 1 and ϕgT g (ρ) (ρ) ∈ Wj , the implicit function
j,i
g
theorem implies that there is a neighborhood V ⊂ Wi of ρ and a C ν−2 function T̃j,i (ρ′ ) : V → R
g
such that ϕgT̃ g (ρ′ ) (ρ′ ) ∈ Wj and T̃j,i g
(ρ) = Tj,i g
(ρ). Since T̃j,i g
is continuous and Tj,i (ρ) < CΓ ̥,
j,i
g g
shrinking V if necessary, we may assume T̃j,i (ρ′ ) < CΓ ̥. Finally, we need to check that T̃j,i (ρ′ ) =
36 YAIZA CANZANI AND JEFFREY GALKOWSKI

g
Tj,i (ρ′ ). For this, it suffices to show that ϕgs (ρ′ ) ∈ / Wj for 0 < s < T̃j,i g
(ρ′ ). Indeed, suppose
g
that there is 0 < s < T̃j,i (ρ′ ) such that ϕgs (ρ′ ) ∈ Wj . Then, ϕgT̃ g (ρ′ )−s (ϕgs (ρ′ )) ∈ Wj , and
j,i
g
0 < T̃j,i (ρ′ ) − s < CΓ ̥. This contradicts the third part of (6.3). 

Definition 6.8. Let ν ≥ 3. Let G ⊂ G ν be bounded in G 2 and {(Wi , Vi )}N i=1 be ̥-well sep-
N
arated for G. For I ∈ {1, . . . , N } , I = (i0 , i1 , i2 , . . . ), g ∈ G, and n ∈ N we define the C ν−2
symplectomorphism
PI(n) [g] : DI(n) [g] → Win , PI(n) [g] := κgin ,in−1 ◦ . . . κgi2 ,i1 ◦ κgi1 ,i0 ,
(0)
where the DI(n) [g] ⊂ Wi0 is the domain of the composition. We also define PI [g] = I and define
PI [g] := (PI(0) [g] , PI(1) [g] , . . . )
and call it the Poincaré chain associated to I.
Definition 6.9. We define the time sequence TI associated to the chain PI . For each n ∈ N and
g ∈ G let TI(n) [g] : DI(n) [g] → [0, ∞) be defined by TI(0) [g] ≡ 0 and for all ρ ∈ DI(n) [g]
TI(n) [g](ρ) = TI(n−1) [g](ρ) + Tign ,in−1 (PI(n−1) [g](ρ)).
for Tign ,in−1 as defined in (6.4).

By construction, for ρ ∈ DI(n) [g] we have

PI(n) [g](ρ) = ϕg (n) (ρ). (6.5)
TI [g](ρ)

Remark 6.10. If ρ ∈ Γ̃ and ϕgt0 (ρ) ∈ Γ̃, then there are n and I such that ρ ∈ DI(n) [g], PI(n) [g](ρ) =
ϕgt0 (ρ) and TI(n) [g](ρ) = t0 .

Below, we will be varying the metric and need to control how Poincaré chains vary with the
metric.
Lemma 6.11. Let ν ≥ 2, 0 ≤ k ≤ ν − 2, G ⊂ G ν bounded, and g : BRk (0, 1)σ → G a C 2 map.
Then there are C > 0, Ck > 0, such that
dσ ϕgt : Rk → T C ν−3 (Sg
∗M ; Sg
∗M ), d2σ ϕgt : Rk × Rk → T C ν−4 (Sg
∗M ; Sg
∗M )

are well defined and satisfy that for all t ∈ R, v, w ∈ Rk , f ∈ C k (S ∗M ; R),

• kdσ ϕgt |σ=0 vkC ν−3 ≤ C |t| kvkk∂σ g|σ=0 kC ν−1 , if ν ≥ 4
• kd2σ ϕgt |σ=0 (v, w)kC ν−4 ≤ C |t| kvkkwk(k∂σ2 g|σ=0 kC ν−2 + k∂σ g|σ=0 k2C ν−1 ), if ν ≥ 5
• kf ◦ ϕgt kC k ≤ Ck C |k||t|kf kC k .
Moreover, for all n ∈ N,
• kdσ PI(n) [gσ ]vkC ν−3 ≤ C nCΓ ̥ kvkk∂σ g|σ=0 kC ν−1 , if ν ≥ 4
• kd2σ PI(n) [gσ ](v, w)kC ν−4 ≤ C nCΓ ̥ kvkkwk(k∂σ2 g|σ=0 kC ν−2 + k∂σ g|σ=0 k2C ν−1 ), if ν ≥ 5
• kf ◦ PI(n) [g]kC k ≤ Ck C |k|nCΓ ̥ kf kC k .
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 37

Proof. The proof of this lemma is a tedious calculation. We sketch here the proof.
Recall that the geodesics flow can be written as the solution (x(t), ξ(t)) to
ẋj (t) = 1 ij
|ξ(t)|g(x(t)) g (x(t))ξi (t), ξ˙k (t) = − 2|ξ(t)|1 ∂xk gij (x(t))ξi (t)ξj (t).
g(x(t))

Letting gσ be a family of metrics with g0 = g and differentiating with respect to σ, working always
with initial data so that |ξ0 (gσ )|gσ (x0 (gσ )) = 1, we obtain equations of the form

∂σ ẋ = (∂x gξ)∂σ x + g∂σ ξ + ∂σ gξ, ∂σ ξ̇ = (∂x2 g)ξ 2 ∂σ x + ∂x gξ∂σ ξ + (∂xσ

2
g)ξ 2 .
In particular, provided that ∂σ g ∈ C ν−1 , this is an equation with C ν−2 coefficients (bounded by
the C ν norm of g and the C ν−1 norm of h) and hence results in a C ν−3 flow with the claimed
bounds.
Differentiating again in g, we obtain equations of the form
∂σ2 ẋ = (∂x2 gξ)(∂σ2 x) + g∂σ2 ξ + (∂x2 gξ)(∂σ x)2 + ∂x g∂σ x∂σ ξ + ∂xσ 2
gξ∂σ x + ∂σ g∂σ ξ + ∂σ2 gξ
∂σ2 ξ˙ = (∂x2 g)ξ 2 ∂σ2 x + ∂x gξ∂σ2 ξ + (∂x3 g)ξ 2 (∂σ x)2 + ∂σxx
3
gξ 2 ∂σ x + ∂x2 gξ∂σ ξ∂σ x
2
+ ∂xσ gξ∂σ ξ + ∂x g(∂σ ξ)2 + ∂xσσ
3
gξ 2 .

Now, since ∂σ (x, ξ) ∈ C ν−3 if ∂σ g ∈ C ν−1 , we require ∂σ2 g ∈ C ν−2 , and ∂σ g ∈ C ν−1 to obtain a
solution in C ν−4 .
The estimates on the derivative in time now follow from estimating the coefficients in these
equations in L∞ by the relevant C ν norms of the metric and its derivatives. 

The following lemma shows that if ρ ∈ Γ is a returning point, say d(ϕgt0 (ρ), ρ) < δ, then we can
associate to it a Poincaré chain under which PI(m) [g](ρ) = ϕgt0 +s (ρ) for some m ∈ N and s small.
Morover, d(PI(m) [g](ρ), ρ) . δ.

Lemma 6.12. Let G ⊂ G ν be bounded in G 2 and {(Wi , Vi )}N i=1 be ̥-well separated for G. Then
there are CG > 0 and δG > 0 such that for all g ∈ G, 0 < δ < δG , j ∈ {1, . . . , N }, and
ρ ∈ R̥ (n, δ, g) ∩ Vj the following holds. There exist m ∈ N and I such that
 
ρ ∈ DI(m) [g], TI(m) [g](ρ) ∈ (n − 1)̥ − CG δ , n̥ + CG δ , (6.6)

and d(PI(m) [g](ρ), ρ) < CG δ. Moreover, for any m1 , m2 ∈ N and I1 , I2 such that
 
ρ ∈ DI(mj ) [g], TI(mj ) [g](ρ) ∈ (n − 1)̥ − CG δG , n̥ + CG δG , j = 1, 2, (6.7)
j j

PI(m1 ) [g](ρ) = PI(m2 ) [g](ρ), TI(m1 ) [g](ρ) = TI(m2 ) [g](ρ).

1 2 1 2

Proof. Since Γ̃ is uniformly transverse to H|ξ|g for g ∈ G, and for all j we have Vj ⋐ Wj , there
]
are CG > 0 and δG > 0 such that for all g ∈ G, j ∈ {1, . . . , N }, 0 < δ < δG , and ρ ∈ S ∗M with

d(ρ, Vj ) < δ, there is |s| ≤ CG δ such that

ϕgt0 +s (ρ) ∈ Wj . (6.8)
38 YAIZA CANZANI AND JEFFREY GALKOWSKI

Wi1

Wi0
PI(1) (ρ) Vi1

ρ Vi0

Wi3
Vi3

Vi2

PI(3) (ρ)
PI(2) (ρ)

Wi2

Figure 3. A portion of a well-separated set and the chain of symplectomorphisms

evaluated at a point ρ ∈ Γ. The chain pictured begins I = {i0 , i1 , i2 , i3 }. Notice
that the geodesic encounters the boundary of Wi1 between Wi2 and Wi3 . It is
because of geodesics like this that we must define a Poincaré chain. Furthermore,
this geodesic encounters the boundary of Vi3 and, to make the map PI(3) [g] smooth
near ρ, one must refer to the subset Γ̃ in addition to Γ.

Next, let ρ ∈ R̥ (n, δ, g) ∩ Vj with 0 < δ < δG . Then, there is (n − 1)̥ < t0 ≤ n̥ such
that d(ϕgt0 (ρ), ρ) < δ. Therefore, there is |s| ≤ CG δ such that ϕgt0 +s (ρ) ∈ Wj . In particular, by
Remark 6.10 there are m and I such that (6.6) holds and PI(m) [g](ρ) = ϕgt0 +s (ρ).
Note that, increasing CG if necessary (uniformly for g ∈ G),
d(ϕgt0 +s (ρ), ρ) ≤ d(ϕgt0 +s (ρ), ϕgs (ρ)) + d(ϕgs (ρ), ρ) < CG δ. (6.9)

Therefore, it only remains to check that if (m1 , I1 ) are such that (6.7) holds, then TI(m1 ) [g](ρ) =
1
t0 + s. Suppose this is not the case for some (m1 , I1 ). Then, since ϕg (m1 ) (ρ) ∈ Γ̃, by (6.5), the
TI [g](ρ)
1
second equation in (6.3) yields |TI(m1 ) [g](ρ) − (t0 + s)| ≥ cΓ ̥. In particular, either TI(m1 ) [g](ρ) ≥
1 1
(n − 1)̥ − CG δ + cΓ ̥, or TI(m1 ) [g](ρ) ≤ n̥ + CG δ − cΓ ̥. Since δ < δG , this contradicts (6.7)
1
provided that 2CG δG < (cΓ − 1)̥. Shrinking δG if necessary completes the proof.

 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 39

e The following lemma shows that

Recall that the image of PI(m) [g0 ] lands in Γ̃ and that Γ ⋐ Γ.
if the iterates of ρ land sufficiently close to Γ, then there is a neighborhood of ρ that lies in the
domain of the same Poincaré chain.
Lemma 6.13. Let G ⊂ G 3 bounded and Γ be ̥ well-separated for G. There are c > 0, δ0 > 0
such that the following holds. Let g0 ∈ G, ρ ∈ Γ, n ∈ N, and I be such that
sup d(PI(m) [g0 ](ρ), Γ) < δ0 .
m≤n

Then, BΓ̃ (ρ, cn ) ⊂ DI(m) [g] for all m ≤ n provided kg − g0 kC 3 ≤ cn .

Proof. Let δ0 > 0 such that {ρ ∈ Γ̃ : d(ρ, Γ) < 3δ0 } ⋐ Γ̃, and
sup sup inf{t > 0 : ϕgt (ρ) ∈ Γ} < CΓ ̥ − 3δ0 .
g∈G ρ∈S]
∗M

First, note that there is C > 0 such that d(ϕgt (ρ), ϕgt 0 (ρ)) ≤ εC t+1 , for all g0 ∈ G, ε > 0, t ∈ R,
and g ∈ G 3 with kg − g0 kC 3 ≤ ε. In particular, since TI(m) [g0 ](ρ) ≤ mCΓ ̥ by (6.3),
 
d ϕg (m) (ρ) , ϕg0(m) (ρ) ≤ εC mCΓ ̥+1 .
TI [g0 ](ρ) TI [g0 ](ρ)

Thus, since d(ϕg (m) (ρ), Γ) < δ0 and H|ξ|g is uniformly transverse to Γ̃, for g ∈ G, the
TI [g0 ](ρ)
implicit function theorem implies that for ε small enough such that εC mCΓ ̥+3 ≪ δ0 , there are tj ,
j = 1, . . . , m such that |tj − TI(j) [g0 ](ρ)| ≤ εC jCΓ ̥+2 ≤ δ0 and
ϕgtj0 (ρ) ∈ Γ̃, d(ϕgtj0 (ρ), ϕg (j) (ρ)) ≤ εC jCΓ ̥+3 ≤ 2δ0 .
TI [g0 ](ρ)

In particular, since TI(j) [g0 ](ρ) − TI(j−1) [g0 ](ρ) < CΓ ̥ − 3δ0 for all 1 ≤ j ≤ m, we have |t1 | < CΓ ̥,
and |tj − tj−1 | < CΓ ̥ for 1 ≤ j ≤ m. Thus, ρ ∈ DI(m) [g] and
|TI(m) [g0 ](ρ) − TI(m) [g](ρ)| ≪ 1,
provided ε ≪ C −mCΓ ̥−1 . Thus, we have
d(ϕg (m) (ρ), Γ) < δ0 + εC mCΓ +1 < 2δ0 .
TI [g](ρ)

Arguing as above, since there is C > 0 such that for all g with dC 3 (g, G) < c,
d(ϕgt (ρ), ϕgt (ρ1 )) < C t d(ρ, ρ1 ),
we have for d(ρ, ρ1 ) < ε,
|TI(m) [g](ρ1 ) − TI(m) [g](ρ2 )| ≪ 1,
and hence ρ1 ∈ DI(m) [g]. 

We now define the notion of a non-degenerate returning point for a Poincaré chain PI . Let
G ⊂ G ν be bounded in G 3 and {(Wi , Vi )}N i=1 be ̥-well separated for G. To define non-degeneracy,
we fix a finite atlas of coordinates charts on Γ

A := (ψi , Ui ) : ψi : Ui → ψi (Ui ), i = 1, . . . , N , (6.10)
40 YAIZA CANZANI AND JEFFREY GALKOWSKI

with Ui ⊂ Γ and ψi (Ui ) ⊂ R2ℵ . Let CG and δG be as in Lemma 6.12 and let δ0 be as in Lemma
6.13. Without loss of generality we assume that δG is small enough that
δG < min(1, δ0 ),
(6.11)
for every ρ ∈ Γ there is (ψ, U ) ∈ A such that B(ρ, CG δG ) ⊂ U.

Next, observe that by Lemma 6.12, if ρ ∈ Γ∩R̥ (n, δ, g) for 0 < δ < δG , then there is (ψ, U ) ∈ A
such that ρ ∈ U for every (m, I) satisfying (6.6), PI(m) [g](ρ) ∈ U , and we may work in R2ℵ by
setting
(PI(m) [g])ψ := ψ ◦ PI(m) [g] ◦ ψ −1 . (6.12)
Abusing notation slightly, we write d((PI(n) [g])ψ )(ρ) := d((PI(n) [g])ψ )(ψ(ρ)).
We now define the notion of non-degeneracy of a trajectory for a chain of symplectomorphisms.
Definition 6.14. Let CG and δG be as in Lemma 6.12 and (6.11). Let n ∈ N, and β > 0. We
say that ρ ∈ R̥ (n, δG , g) is (n, β, g) non-degenerate, and write ρ ∈ N ̥ (n, β, g), if there exist I
 
and m such that TI(m) [g](ρ) ∈ (n − 1)̥ − CG δG , n̥ + CG δG and

−1
sup I − d((PI(m) [g])ψ )(ρ) < β −1 . (6.13)
{(ψ,U )∈A: B(ρ,CG δG )⊂U }

We next strengthen the notion of non-degeneracy of a trajectory to one that can pass from a
trajectory to its iterates by introducing a type of (β, q)-non-degeneracy for returning points.
Definition 6.15. Let CG and δG be as in Lemma 6.12 and (6.11). Let q ∈ N, n ∈ N and β ∈ (0, 1).
We say that ρ ∈ R̥ (n, δG , g) is (n, β, g) q-non-degenerate, and write ρ ∈ Nq,̥ (n, β, g), if there
 
exist I and m satisfying TI(m) [g](ρ) ∈ (n − 1)̥ − CG δG , n̥ + CG δG and
(I − [(d((PI(m) [g])ψ )(ρ)]q )−1
sup (m)
≤ f(β, q), f(β, q) = ( 25 q 2 β −1 )2ℵ . (6.14)
{(ψ,U )∈A: B(ρ,CG δG )⊂U } (1 + kd((PI [g])ψ )(ρ)kq )2ℵ−1

That is, ρ ∈ N q,̥ (n, β, κ) provided there are I and m satisfying (6.7) such that the matrix
A := d(ψ ◦ PI(m) [g] ◦ ψ −1 )(ψ(ρ)) is (β, q) non-degenerate (as defined in (4.1)) for every chart
(ψ, U ) ∈ A with B(ρ, CG δG ) ⊂ U .
Remark 6.16. The reason for the choice of f comes from Corollary 5.3 and the fact that we need
to be able to sum measures of bad sets over all q in Lemma 7.6. We will make perturbations
which guarantee certain properties of the eigenvalues of dPI(n) [g] and will use Corollary 5.3 to
bound the inverse of I − (dPI(n) [g])q . In fact, q 2 could be replaced by q γ for any γ > 1.
Remark 6.17. In Definitions 6.14 and 6.15, the existence of I and m satisfying (6.6) and, respec-
tively (6.13) or (6.14) could be replaced by the requirement that for all I and m satisfying (6.7),
respectively, the estimates (6.13) or (6.14) hold. To see this, we observe that by the second part of
Lemma 6.12, all I and m satisfying (6.7) in fact satisfy (6.6) and result in the same map PI(m) [g].

The next lemma demonstrates how the notion of (n, β, g) q non-degeneracy can be passed to
(nq, β, g) non-degeneracy with the trajectory of length nq being a multiple of that of length n.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 41

Lemma 6.18. Let G ⊂ G 4 (Γ) bounded and Γ be ̥ well separated for G. let CG and δG be as in
Lemma 6.12 and (6.11). There is C > 0 such that the following holds. Let g ∈ G, k > 0, β > 0,
q0 ∈ N and 0 < δ ≤ min{δG , (2f(β, q0 )C kq0 )−1 }. Let

ρ ∈ Γ ∩ R̥ (k, δ, g) ∩ N q0 ,̥ (k, β, g),

 
and I and m such that TI(m) [g](ρ) ∈ (k − 1)̥ − CG δG , k̥ + CG δG ,
 (qm) 
q 
q0
PI [g](ρ) q=0 ⊂ DI(m) [g], max d PI(m) [gℓ ] (ρ), ρ ≤ δ,
0≤q≤q0

q
PI(qm) [g](ρ′ ) = PI(m) [g] (ρ′ ), 0 ≤ q ≤ q0 , ρ′ in a neighborhood of ρ.

Then,


ρ ∈ N ̥ n, (2f(β, q0 )C kq0 )−1 , g
for all n ∈ N such that TI(q0 m) (ρ) ∈ [(n − 1)̥ − CG δG , n̥ + CG δG ].

Proof. For 0 ≤ q ≤ q0 , set ρq := PI(qm) [g](ρ) and ∆q := dPI(m) [g](ρq ) − dPI(m) [g](ρ). Then, since
there is C > 0 such that kd2 ϕgt k ≤ C t for all t ≥ 0, by assumption we conclude that there is C > 0
such that maxq k∆q k ≤ C k δ. Therefore,

dPI(mq0 ) [g](ρ) = (dPI(m) [g](ρ) + ∆q0 −1 ) ◦ · · · ◦ (dPI(m) [g](ρ) + ∆1 ) ◦ dPI(m) [g](ρ)

= [dPI(m) [g](ρ)]q0 + ∆′

with k∆′ k ≤ 2q0 kdPI [g](m) (ρ)kq0 C k δ. We conclude that there is C0 > 0 depending only on G such
that
kdPI(q0 m) [g](ρ) − [dPI(m) [g](ρ)]q0 k ≤ C0kq0 δ. (6.15)
Since ρ ∈ N q0 ,̥ (k, β, g), there is C > 0 depending only on G such that for all (ψ, U ) ∈ A with
B(ρ, CG δG ) ⊂ U ,
 q
−1
I − d(PI(m) )ψ (ρ) 0 ≤ C kq0 f(β, q0 ).

In particular, provided that C02kq0 δf(β, q0 ) < 12 , (6.15) yields

k[I − d(PI(q0 m) )ψ (ρ)]−1 k ≤ C kq0 f(β, q0 )(1 − C02kq0 δf(β, q0 ))−1 ≤ 2C kq0 f(β, q0 ).

The claim now follows from the definition of non-degeneracy. 

7. Perturbing away q-degeneracy for simple points

In this section, we first introduce general assumptions on a family of perturbations which will
allow us to change the degeneracy properties of a given simple point. We then use a quantitative
analog of Sard’s theorem to show that the collection of perturbation parameters which produce
degeneracy is small.
42 YAIZA CANZANI AND JEFFREY GALKOWSKI

7.1. General assumptions. Let G ν denote the space of C ν metrics on M with the topology
induced by the C ν topology on symmetric tensors, and ιCν : C ν → C ν−1 be the natural inclusion
map.
Given a function N : (0, 1) → N, R > 0, and δ > 0, we will work with perturbation maps
QR,δ : G ν × ΣN(R) → G ν where
N (R)
Y
ΣN(R) := BRL (0, 1), L := ℵ(2ℵ + 3), (7.1)
j=1

Remark 7.1. Recall that 2ℵ = 2d − 2 = dim Γ, where Γ is a section of S ∗M . Thus, the dimension
L = ℵ(2ℵ + 3) appears because we want the parameter σ ∈ BRL (0, 1) to be in correspondence with
points in the space (Γ, Sp(2ℵ)), and the latter has dimension L. Here Sp(2ℵ) will parametrize the
possible derivatives of a symplectic map at a point in Γ.
N (R)
We write σ = (σj )j=1 where each σj is a component in BRL (0, 1), and we refer to QR,δ (g, σ)
as the perturbation of g by the parameter σ. When we construct our family of perturbations in
Section 9, R will control the scale on which our metric perturbations take place and δ will control
the size of the perturbation takes place. We ask that the perturbation maps satisfy the following
assumptions.
Definition 7.2 (Good perturbations). Let ν0 ≥ 0 and N : (0, 1) → N. For each R > 0, δ > 0
consider a map
QR,δ : G ν0 × ΣN(R) → G ν0
such that for each g ∈ G ν0 , the map σ 7→ QR,δ (g, σ) with σ ∈ ℓ∞ (BRL (0, 1)) is Frechet differen-
tiable, and for all σ ∈ ΣN(R) the map g 7→ ιCν0 −1 (QR,δ (g, σ)) is Frechet differentiable.
We say that {QR,δ }R,δ is a (ν0 , N )-good family of perturbations if for each K ⊂ G ν0 bounded
and each 0 ≤ ν ≤ ν0 there exist ϑν ∈ R, ϑ̃ν ∈ R and C > 0 such that the following holds.
For δ > 0, R > 0,
• QR,δ (g, 0) = g, g ∈ G ν0 . (7.2)
• kQR,δ (g, σ) − QR,δ (g, σ̃)kCν ≤ CδR−ϑν kσ − σ̃kℓ∞ , σ, σ̃ ∈ ΣN(R) , g ∈ K. (7.3)
• kDg QR,δ − IkCν−1 →Cν−1 ≤ CδR−ϑν , g ∈ K, σ ∈ ΣN(R) . (7.4)

Furthermore, for all α ∈ NL there is Cα > 0 such that for all g ∈ K, δ > 0, and R > 0

max k∂σαj QR,δ |(g,σ) kB (0,1)→C ν ≤ Cα R−|α|ϑ̃ν δ|α| , g ∈ K, σ ∈ ΣN(R) . (7.5)

1≤j≤N (R) RL

We also need an assumption which guarantees that the perturbation QR,δ explores a sufficiently
large set of possible metrics.
Definition 7.3 (Admissible pairs). Let ν ≥ 5, b > 0 and y ≥ 1. Let N : (0, 1) → N and {QR,δ }R,δ
be a (ν, N )-good family of perturbations.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 43

We say that (Γ, G) is a (̥, b, y)-admissible pair for {QR,δ }R,δ if G ⊂ G ν is bounded, Γ is ̥-well
separated for G, and for all K ⊂ G bounded in G ν there are c > 0 and ε0 > 0 such that for all
R > 0, δ > 0, g ∈ K, α ∈ (0, c), n ∈ N, and ρ0 ∈ Γ satisfying
R < cn α, δ ≤ Rb cn+1 , ρ0 ∈ S̥ (n, α, g) ∩ R̥ (n, ε0 , g),
there exist j0 ∈ {1, . . . , N (R)}, I0 , m0 such that BΓ (ρ0 , Ry ) ⊂ DI(m0 ) [QR,δ (g, σ)] for all σ ∈ ΣN(R) ,
  0
TI(m0 ) [g](ρ) ∈ (n − 1)̥ − CG δG , n̥ + CG δG , and for all ρ ∈ BΓ (ρ0 , Ry ) and σ ∈ ΣN(R) ,
0

|dσj0 ΨR,δ
g,I (m0 , σ, ρ)v| ≥ δc
n+1
|v|, for all v ∈ Tσj0 BRL (0, 1), (7.6)
0

where 

ΨR,δ
g,I (m 0 , σ, ρ) := PI
(m0 ) R,δ
[Q (g, σ) ](ρ) , dρ PI
(m0 ) R,δ
[Q (g, σ) ] (ρ) . (7.7)
0 0 0

The constants CG and δG are the ones given in Lemma 6.12.

To understand why we are interested in the map ΨR,δ

g , observe that knowing

(PI(m0 ) [g](ρ), dρ PI(m0 ) [g](ρ))

is exactly the information need to understand (1) whether ρ ∈ R̥ (n, β, g) and (2) if ρ ∈
R̥ (n, β, g), whether ρ ∈ N q,̥ (n, β, g). Thus, the dimension of BRL (0, 1) is chosen so that we
are able to use the j0 -th factor in ΣN(R) to locally explore an open set in the image of the maps
ΨR,δ
g,I (whose dimension is L = ℵ(2ℵ + 3)) by varying σj0 ∈ BRL (0, 1). See also Remark 7.1.

7.2. Perturbing away (β, q)-degeneracy for simple points. In order to control the volume of
the parameters in ΣN(R) which may produce degeneracy we apply a quantitative version of Sard’s
theorem ([Yom85, Theorem 4.2]). We recall the theorem here for convenience. Below, MRk (Y, s)
denotes the covering number of Y ⊂ Rk by balls in Rk of radius s (see Definition 5.4).
Proposition 7.4 ([Yom85] Theorem 4.2). Let Ω ⊂ Rℓ and L > 0 such that for all ρ1 , ρ2 ∈ Ω, there
is a path in Ω of length smaller than Lkρ1 − ρ2 k connecting ρ1 and ρ2 . Let Ψ : BRk (0, 1) × Ω → Rk
be a continuously differentiable mapping such that there exist c1 > 0, c2 > 0 so that for all ρ ∈ Ω,
and σ, σ̃ ∈ BRk (0, 1)
kdρ Ψ(σ, ρ)k ≤ c1 , kΨ(σ̃, ρ) − Ψ(σ, ρ)k ≥ c2 kσ − σ̃k.
Then, for any s > 0 and any X ⊂ Ω, Y ⊂ Rk ,
MRk (∆Ψ (X, Y ), s̃) ≤ MRℓ (X, s)MRk (Y, s)
where s̃ := 2(Lc−1
2 (1 + c1 ) + 1)s and
∆Ψ (X, Y ) := {σ ∈ BRk (0, 1) : there is ρ ∈ X such that Ψ(σ, ρ) ∈ Y }. (7.8)

Let mRL be the Lebesgue measure on RL normalized so that

mRL (BRL (0, 1)) = 1 (7.9)
and let mΣ be the product measure on ΣN(R) . We write mΓ for the volume measure on Γ.
N(R)

Next, we show that there is c > 0 such that if ρ0 ∈ Γ is (n, c) returning and (n, α) simple
for some metric g, then there is only a small measure set of σ that yield a perturbed metric
44 YAIZA CANZANI AND JEFFREY GALKOWSKI

′
QR,δ (g, σ) for which there are points ρ ∈ B(ρ0 , Ry ) that are (n, βq −3 ) returning and are not
′
(n, β) q-non-degenerate for some q ∈ N. This can be done for α < c and β < cn Ry provided that
we work with metrics that are at least C 5 .
Remark 7.5. Lemma 7.6 below is the only place where we require g to have 5 derivatives.
Lemma 7.6. Let b > 0, ν ≥ 5, y ≥ 1, and N : (0, 1) → N. Let {QR,δ }R,δ be a (ν, N )-good family
of perturbations and (Γ, G) be a (̥, b, y)-admissible pair for {QR,δ }R,δ .
Let K ⊂ G bounded in G 5 and y′ ≥ y. Let δG be as in Lemma 6.12 and (6.11). There exist
c ∈ (0, 1), C > 0 such the following holds. Let n ∈ N, α ∈ (0, c),
0 < R < min(cn α, 12 cδG ), 0 < δ ≤ min(Rb cn+1 , R2ϑ̃4 cn+1 , cRϑ5 ).
′
If g ∈ K, ρ0 ∈ Γ, and BΓ (ρ0 , Ry ) ∩ S̥ (n, α, g) ∩ R̥ (n, c, g) 6= ∅, then for all 0 < β <
′
min ( 12 δG , cn Ry )
′ ′
mΣ (Sgρ0 ,R,δ (n, α,β)) ≤ mΓ (BΓ (ρ0 , Ry ))(C n Ry )L−2ℵ−1 δ−L β,
N(R)

where
n ′
Sgρ0 ,R,δ (n, α,β) := σ ∈ ΣN(R) : there are ρ ∈ BΓ (ρ0 , Ry ) and q ∈ N such that


o
ρ ∈ R̥ n, βq −3 , QR,δ (g, σ) N q,̥ n, β, QR,δ (g, σ) .

Proof. Let ε0 as in Definition 7.3. Using the estimate on (ϕgt )∗ : C 1 → C 1 in Lemma 6.11, we first
note that there exists c = c(K) > 0 such that if g ∈ K, ρ0 ∈ Γ, 0 < α < 1, n ∈ N, and 0 < R ≤
′
cn α, then B(ρ0 , Ry ) ∩ S̥ (n, α, g) ∩ R̥ (n, c, g) 6= ∅ implies ρ0 ∈ S̥ (n, 13 α, g) ∩ R̥ (n, ε0 , g). By
Definition 7.3, this allows us to work with indices j0 , m0 , and I0 such that the lower bound on
the differential of ΨR,δ
g,I0 in (7.6) holds.
The result will follow once we find c, C such that, under the assumptions of the lemma, for all
QN (R)−1 ′
σ̂ ∈ j=1 BRL (0, 1) and 0 < β < min( 12 δG , cn Ry )
′ ′
mRL (Sgρ0 ,R,δ (n, α, β; j0 , σ̂)) ≤ mΓ (BΓ (ρ0 , Ry ))C(C n Ry )L−2ℵ−1 δ−L β, (7.10)
QN (R)−1 j0 −1 N (R)−1
where for σ̂ ∈ j=1 BRL (0, 1), we let σj−0 := (σ̂i )i=1 , σj+0 := (σ̂i )i=j0 , and
n
o
Sgρ0 ,R,δ (n, α,β; j0 , σ̂) = σ ∈ BRL (0, 1) : σj−0 , σ, σj+0 ∈ Sgρ0 ,R,δ (n, α, β) .

To prove the claim in (7.10) we restrict the map ΨR,δ th

g,I0 to its j0 -component in σ. For
QN (R)−1

σ̂ ∈ j=1 BRL (0, 1) define the restricted map ΨR,δ R,δ − +
g,I ,j0 ,σ̂ (n, σ, ρ) = Ψg,I0 n, (σj0 , σ, σj0 ), ρ , with
0
′
domain (n, σ, ρ) ∈ N × BRL (0, 1) × BΓ (ρ0 , Ry ).
By assumption (7.3), after possibly shrinking c in a way that depends only on K, the set
{QR,δ (g, σ) : g ∈ K, σ ∈ ΣN(R) } ⊂ G 5 is bounded provided δ < cRϑ5 . Then, combining (7.5)
with Lemma 6.11 we obtain that there is C > 0 depending only on (K, Γ, ̥) such that
sup sup kd2σ ΨR,δ n −2ϑ̃4 2
g,I ,j,σ̂ (n, σ, ρ)k ≤ C R δ . (7.11)
ρ,σ̂,σ,j (n)
{I:ρ∈DI [QR,δ (g,σ)]}
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 45

Therefore, since ρ0 ∈ Γ ∩ S̥ (n, 13 α, g) ∩ R̥ (n, ε0 , g), assumption (7.6) and (7.11) yield that for
each σ̂, and σ, σ̃ ∈ BRL (0, 1),

inf kΨR,δ R,δ

g,I ,j0 ,σ̂ (m0 , σ, ρ) − Ψg,I ,j0 ,σ̂ (m0 , σ̃, ρ)k ≥ δc
n+1
kσ − σ̃k, (7.12)
ρ∈BΓ (ρ0 ,Ry ) 0 0

provided that
R < cn α, δ ≤ Rb cn+1 , C n R−2ϑ̃4 δ2 ≤ 1 n+1
10 c δ, (7.13)
after possibly shrinking c in a way that only depends on (K, Γ, ̥).
Note that if σ ∈ Sgρ0 ,R,δ (n, α, β; j, σ̂), then by Lemma 6.12 there are ρ ∈ BΓ (ρ0 , Ry ) and q ∈ N
′

such that



d ρ , PI(m0 ) [QR,δ (g, σ)](ρ) < CG βq −3 , ρ∈/ N q,̥ n, β, QR,δ (g, σ) . (7.14)
0
δ
After requiring that Ry < CG δG and β < 2G , there exists a coordinate chart (U, ψ) ∈ A (see
′
(6.11)) with BΓ (ρ0 , Ry ) ⊂ ψ(U ) such that (6.14) does not hold with I = I0 and m = m0 .
′
Furthermore, PI(m0 ) [QR,δ (g, σ))](ρ) ∈ ψ(U ) for all ρ ∈ R̥ (n, βq −3 , QR,δ (g, σ)) ∩ BΓ (ρ0 , Ry ).
0

Next, consider the map Ψ̃R,δ

′
g : N × ΣN(R) × ψ(BΓ̃ (ρ0 , Ry )) → ψ(U ) × Sp(2ℵ)
 
Ψ̃R,δ
g (n, σ, x) := (PI
(m0 ) R,δ
[Q (g, σ) ]ψ (x) − x , dx ((PI
(m0 ) R,δ
[Q (g, σ) ])ψ (x) .
0 0

and for each n ∈ N consider its restriction

 
Ψ̃n : BRL (0, 1) × ψ(BΓ (ρ0 , Ry )) → ψ(U ) × Sp(2ℵ), Ψ̃n (σ, x) := Ψ̃R,δ
g n, (σj−0 , σ, σj+0 ), x .

It follows from (7.14), our choice of ψ, and the comment following Definition 6.15, that A :=
dx (ψ◦(PI(m0 ) [QR,δ (g, σ)])◦ψ −1 )(ψ(ρ)) is not (β, q) non-degenerate. Therefore, Corollary 5.3 applied
0
to A combined with the first statement in (7.14) yield, that, with sq := βq −3 , D( 2πp
q , sq ) as in
(5.4), and ∆Ψ̃ as in (7.8),
n

Sgρ0 ,R,δ (n, α,β; j, σ̂) ⊂ ∆Ψ̃n (X, ∪∞

q=1 Yq ), (7.15)
q−1
[
′
X := ψ(BΓ (ρ0 , Ry )), Yq := BR2ℵ (0, CG sq ) × D( 2πp
q , sq )∩BA0(r0 ).
p=0
′
Here, we let A0 := dx (PI(m0 ) [g])ψ (x) and r0 := C1n Ry with C1 chosen so that the image of the
0
second factor in Ψ̃n lies within BA0 (r0 ). This is possible to arrange since there is C0 > 0 depending
only on (K, ψ, U ) such that kA0 k ≤ C0n . Also, by Lemma 6.11 with ν = 4 and Definition 7.2 with
ν = 3, there is C1 > 0 depending only on (K, ψ, U ) such that
′
kdx ((PI(m0 ) [QR,δ (g, σ)])ψ (x) − A0 k ≤ (C1 δR−ϑ3 )n Ry ≤ r0 .
0

Note that the last inequality holds provided δ < Rϑ3 .

With (7.15) in place we proceed to check that the hypothesis in Proposition 7.4 hold. Note that
by Lemma 6.11 there is C > 0, depending only on (K, ψ, U ), with supx∈ψ(B (ρ0 ,Ry′ )) kdx Ψ̃n (σ, x)k ≤
Γ
46 YAIZA CANZANI AND JEFFREY GALKOWSKI

C n . In addition, by (7.12), there is c > 0 depending only on (K, ψ, U ) such that for σ, σ̃ ∈ BRL (0, 1)

inf kΨ̃n (σ, x) − Ψ̃n (σ̃, x)k ≥ δcn+1 kσ − σ̃k.

x∈ψ(BΓ (ρ0 ,Ry ))

It follows that, for each n ∈ N, Proposition 7.4 applies to Ψ̃n with k := L = dim(ψ(U ) ×
Sp(2ℵ)), ℓ := 2ℵ, c1 = C n , c2 = δcn+1 , and some L depending only on (ψ, U ). Note that we are
abusing notation slightly here to identify Sp(2ℵ) with Rℵ(2ℵ+1) in the codomain. Finally, since
mRL (∆Ψ̃n (X, Yq )) ≤ cL MRL (∆Ψ̃n (X, Yq ) , s) sL for all s > 0, Proposition 7.4 and (7.15) yield that,
with s̃q := 2(Lδ−1 c−n−1 (1 + C n ) + 1)sq ,
∞
X ∞
X
mRL (Sgρ0 ,R̃,δ (n, α,β; j, σ̂)) ≤ mRL (∆Ψ̃n (X, Yq )) ≤ cL MR2ℵ (X, sq )MRL (Yq , sq ) s̃Lq . (7.16)
q=1 q=1

To control the right hand side of (7.16) first observe that

′
MR2ℵ (X, sq ) ≤ CmΓ (BΓ (ρ0 , Ry ))s−2ℵ
q (7.17)
′
provided 0 < sq ≤ Ry , where C > 0 depends only on (Γ, ψ, U ). Next, by Corollary 5.11 with
r = C0n + r0 , there are C, ε0 > 0 such that if 0 < sq ≤ min(ε0 (C0n + r0 ), 12r0 (C0n + r0 )2 ), then

MRL (Yq , sq ) ≤ Cq((C0n + r0 )2 r0 sq −1 )L−2ℵ−1 . (7.18)

δG n y′
By possibly adjusting c > 0 one last time, asking β ≤ min( 2 ,c R ) makes sq be sufficiently
small, and so combining (7.16), (7.17) and (7.18) yields
∞
X ′
mRL (Sgρ0 ,R̃,δ (n, α, β; j, σ̂)) ≤ CmΓ (BΓ (ρ0 , Ry ))s−2ℵ
q Cq((C0n + r0 )2 r0 sq −1 )L−2ℵ−1 s̃Lq
q=1
∞
X
′ ′
≤ C(C n Ry )L−2ℵ−1 mΓ (BΓ (ρ0 , Ry ))δ−L β q −2 .
q=1

δG n η′
This proves the claim in (7.10) provided R, α, δ, n satisfy (7.13), β < min( 2 , c R ), and δ <
cRϑ5 . 

Corollary 7.7. Let b > 0, ν ≥ 5, y ≥ 1, and N : (0, 1) → N. Let {QR,δ }R,δ be a (ν, N )-good
family of perturbations and (Γ, G) be a (̥, b, y)-admissible pair for {QR,δ }R,δ .
Let K ⊂ G bounded in G 5 and y′ ≥ y. Let δG be as in Lemma 6.12 and (6.11). There exist
c ∈ (0, 1), C > 0 such the following holds. Let n ∈ N, α ∈ (0, c),

0 < R < min(cn α, 12 cδG ), 0 < δ ≤ min(Rb cn+1 , R2ϑ̃4 cn+1 , cRϑ5 +1 ). (7.19)
′
If g ∈ K, then for all 0 < β < min ( 12 δG , cn Ry )
′
mΣ (SgR,δ (n, α, β)) ≤ C(C n cRy )L−2ℵ−1 δ−L β,
N(R)
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 47

where
n
SgR,δ (n, α, β) := σ ∈ ΣN(R) : ∃q ∈ N s.t.


o
Γ ∩ S̥ (n, α, QR,δ (g, σ)) ∩ R̥ n, βq −3 , QR,δ (g, σ)) N q,̥ n, β, QR,δ (g, σ) 6= ∅ .
′
Proof. Let {ρi }i ⊂ Γ be an Ry maximal separated set in Γ. We claim that
[
SgR,δ (n, α,β) ⊂ ZgR,δ (n, α,β) ⊂ Sgρi ,R,δ (n, α,β), (7.20)
i∈I
′ 1
where I := {i : BΓ (ρi , Ry )
∩ S̥ (n, ∩ R̥ (n, c, g) 6= ∅} and
2 α, g)
n
ZgR,δ (n, α,β) := σ ∈ ΣN(R) : ∃q ∈ N s.t.


o
Γ ∩ S̥ (n, 12 α, g) ∩ R̥ n, βq −3 , QR,δ (g, σ)) N q,̥ n, β, QR,δ (g, σ) 6= ∅ .
The claim in (7.20) follows from observing that
S̥ (n, α, QR,δ (g, σ)) ⊂ S̥ (n, 21 α, g), R̥ (n, βq −3 , QR,δ (g, σ)) ⊂ R̥ (n, c, g).
Indeed, these two inclusions follow from combining the conclusion from Lemma 6.11 that for all
t∈R
QR,δ (g,σ)
d(ϕgt (ρ), ϕt (ρ)) ≤ C |t| kg − QR,δ (g, σ)kC 3
with (7.2), (7.3), and the requirement that δR−ϑ3 ≤ cn α (the latter holds since ϑ5 ≥ ϑ3 and
δ < cRϑ5 +1 ).
′
Finally, combining (7.20) with Lemma 7.6 applied to each ball BΓ̃ (ρi , Ry ) with i ∈ I yields the
desired bound. 

8. Induction on the length of orbits

Let b > 0. Let ν > 0, N : (0, 1) → N, {QR,δ }R,δ be a (ν, N )-good family of perturbations, and
let (Γ, G) be a (̥, b, y)-admissible pair for {QR,δ }R,δ . In this section we will prove that there is a
predominant set, G of metrics such that for all g in G, there is C > 0 such that
Ων
Γ ∩ R̥ (n, βn , g) ⊂ N ̥ (n, βn , g), βn = C −Cn , (8.1)
We do this using a family of probing maps generated by the (ν, N )-good family of perturbations.
Once this is done (see Proposition 8.9), to prove Theorem 1.3 it will remain to show that there
are (ν, N )-good families of perturbations (we do this in Section 9). We continue to work in the
setting of Section 7.

8.1. Construction of families of probing maps. Let ν ≥ 0, a function N : (0, 1) → N, and let
{QR,δ }R,δ be a (ν, N )-good family of perturbations. Let {Ks }s∈(0,1) ⊂ G ν be a family of nested
closed bounded subsets Ks1 ⊂ Ks2 for s1 > s2 such that G ν = ∪s∈(0,1) Ks . By Definition 7.2 for
every s ∈ (0, 1) bounded there is Cs > 0 such that for all δ > 0, R > 0,
kQR,δ (g, σ) − QR,δ (g, σ̃)kCν ≤ Cs δR−ϑν kσ − σ̃kℓ∞ , (8.2)


kQR,δ (g̃, σ) − QR,δ (g, σ kCν−1 ≤ (1 + Cs δR−ϑν )kg − g̃kCν−1 . (8.3)
48 YAIZA CANZANI AND JEFFREY GALKOWSKI

for σ, σ̃ ∈ ΣN(R) , g ∈ Ks . Next, for ε > 0 let s(ε) > 0 be an increasing function such that
limε→0+ Cs(ε) ε = 0 and Cs(ε) ε is increasing. Next, define the closed and bounded sets
Gεν := {g ∈ Ks(ε) : d(g, G ν \Ks(ε) ) ≥ 2Cs(ε) ε}. (8.4)
Note that Gεν2 ⊂ Gεν1 for ε1 < ε2 and G ν = ∪ε>0 Gεν .
For ε ∈ (0, 1) let δε = {δj (ε)}∞ ∞
j=0 ⊂ [0, 1) and Rε = {Rj (ε)}j=0 ⊂ (0, 1), satisfy
∞
X
δj (ε)(Rj (ε))−ϑν ≤ ε. (8.5)
j=0

We now construct a family of probing maps associated to the perturbation maps QRj (ε),δj (ε) .
Given J = 0, . . . , N define the spaces
J
Y ∞
Y
ΣJ (Rε ) := ΣN(Rj (ε)) , Σ∞ (Rε ) := ΣN(Rj (ε)) . (8.6)
j=0 j=0

Also, given a sequence σ = (σ0 , σ2 , σ3 , . . . ) ∈ Σ∞ (Rε ) we write σ̂J := (σ0 , . . . , σJ ) ∈ ΣJ (Rε ).

We are now ready to define the family of probing maps (see Definition 2.2). We start by letting
F R,δ
0
: Gεν × ΣN(R0 ) → G ν be the map

F0Rε ,δε (g, σ) := QR0 (ε),δ0 (ε) (g, σ) (8.7)

and FJRε ,δε : Gεν × ΣJ (Rε ) → G ν by

FJRε ,δε (g, σ̂J ) := QRJ (ε),δJ (ε) FJR−1ε ,δε (g, σ̂J −1 ), σJ . (8.8)
Finally, we define
F∞Rε ,δε : Gεν × Σ∞ (Rε ) → G ν , F∞Rε ,δε (g, σ) = lim FJRε ,δε (g, σ̂J ). (8.9)
J→∞

Note that, apriori, the limit in (8.9) may not exist, but we prove in Lemma 8.1 that it does under
our assumptions on Rε and δε . When it will not lead to confusion, we will omit the R, δ from
the notation for FJR,δ .
When checking the probing maps definition we work with G := G ν and G ′ := G ν−1 , while B
and B ′ will be the spaces of C ν and C ν−1 symmetric two-tensors.
Lemma 8.1. Let ν ≥ 0, N : (0, 1) → N and {QR,δ }R,δ be a (ν, N )-good family of perturba-
tions. Then for all δε = {δj (ε)}∞ ∞
j=0 ⊂ (0, 1) and Rε = {Rj (ε)}j=0 ⊂ (0, 1) satisfying, (8.5).
{(F∞Rε ,δε , ∞)} ν−1 -family of probing maps for G ν .
ε>0 is a G

Moreover, for all K ⊂ G ν bounded and K̃ ⊂ G ν bounded with d(K̃ c , K) > 0, there is ε1 > 0
such that for all 0 < ε < ε1 and δε and Rε satisfying (8.5), FJRε ,δε (K × ΣJ (Rε )) ⊂ K̃ for all
J ∈ N ∪ {∞}.

Proof. Let 0 < ε < 1 and assume (8.5) is satisfied. We first show that F∞Rε ,δε is well defined.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 49

Given g ∈ Gεν we claim that FJRε ,δε (g, σ̂J ) ∈ Ks(ε) for all J and σ̂J ∈ ΣJ (Rε ). Indeed, we claim
 J
X 
Rε ,δε ν
d(FJ (g, σ̂J ), G \Ks(ε) ) ≥ Cs(ε) 2ε − δJ (ε)RJ (ε)−ϑν . (8.10)
j=0

Note that when J = 0 this follows from the definition of Gεν since (8.2) and (7.2) yield kF0Rε ,δε (g, σ)−
gkC ν ≤ Cs(ε) δ0 (ε)R0 (ε)−ϑν . Next, assume the claim holds for all j ≤ J − 1. Then, again by (8.2)
we obtain
kFJRε ,δε (g, σ̂J ) − FJR−1ε ,δε (g, σ̂J −1 )kC ν = QRJ (ε),δJ (ε) (FJR−1ε ,δε (g, σ̂J −1 ), σJ ) − FJR−1ε ,δε (g, σ̂J −1 ) Cν
−ϑν
≤ Cs(ε) δJ (ε)RJ (ε) kσJ kℓ∞ . (8.11)
By (8.11) and (8.5) the claim in (8.10) holds and so FJRε ,δε (g, σ̂J ) ∈ Ks(ε) . Furthermore, (8.11)
and (8.5) guarantee that F∞Rε ,δε (g, σ) is well defined and F∞Rε ,δε (g, σ) ∈ G ν . Also, since FJRε ,δε is
Rε ,δε .
continuous, so is F∞
Note that (8.11) implies the second statement in the Lemma and we need only check that
Rε ,δε
F∞ is a family of probing maps.
Next, note that (7.2) implies F∞Rε ,δε (g, 0) = g for all g ∈ Gεν so (2.6) holds. To check that
(2.7), (2.8), (2.9) hold, let K ⊂ G ν be bounded. Then, there is ε0 > 0 such that K ⊂ Gεν for all
0 < ε < ε0 . In particular, (8.11) implies that for all 0 < ε < ε0 , g ∈ K, and σ ∈ Σ∞ (Rε ),
∞
X
Rε ,δε
kF∞ (g, σ) − gk Cν ≤ Cs(ε) δJ RJ−ϑν kσJ kℓ∞ ≤ Cs(ε) εkσkℓ∞ .
J=1

Thus, since limt→0+ Cs(ε) ε = 0, (2.7) holds.

We next prove that F∞ Rε ,δε satisfies (2.8). To ease notation, from now on we write F in place
J
of FJRε ,δε and δJ = δJ (ε), RJ = RJ (ε). First, note that (8.2) implies that for g ∈ Kt(ε) and
σ, µ ∈ Σ∞ (R)
kF0 (g, σ̂0 ) − F0 (g, µ̂0 )kCν−1 ≤ Cs(ε) δ0 R0−ϑν kσ0 − µ0 kℓ∞ .
Our induction hypothesis is that for some J ≥ 1 and all g ∈ Gεν , σ, µ ∈ Σ∞ (R),

kFJ −1 (g, σ̂J −1 ) − FJ −1 (g, µ̂J −1 )kCν−1

J−2
X J−1
Y −ϑν−1
≤ Cs(ε) RJ−ϑ
−1
ν−1
δJ −1 kσJ −1 − µJ −1 kℓ∞ + Cs(ε) (1 + Cs(ε) δk Rk−ϑν )Rj δj kσj − µj kℓ∞ .
j=0 k=j+1

Next, note that by (8.2),

kFJ (g, σ̂J ) − FJ (g, (σ̂J −1 , µJ ))kCν−1 ≤ Cs(ε) RJϑν−1 δJ kσJ − µJ kℓ∞ .
In addition, by (8.3) and the induction hypothesis
kFJ (g, µ̂J ) − FJ (g, (σ̂J −1 , µJ ))kCν−1 ≤ (1 + Cs(ε) δJ RJ−ϑν )kFJ −1 (g, σ̂J −1 ) − FJ −1 (g, µ̂J −1 )kCν−1
J−1
X J
Y −ϑν−1
≤ Cs(ε) (1 + Cs(ε) δk Rk−ϑν )Rj δj kσj − µj kℓ∞ .
j=0 k=j+1
50 YAIZA CANZANI AND JEFFREY GALKOWSKI

Therefore, for all J = 0, 1, . . . , g ∈ Gεν , and σ, µ ∈ Σ∞ (R),

kFJ (g, σ̂J ) − FJ (g, µ̂J )kCν−1
≤ kFJ (g, σ̂J ) − FJ (κ, (σ̂J −1 , µJ )kCν−1 + kFJ (g, µ̂J ) − FJ (g, (σ̂J −1 , µJ )kCν−1
J−1
X J
Y −ϑν−1
≤ Cs(ε) RJ −ϑν−1
δJ kσJ − µJ kℓ∞ + Cs(ε) (1 + Cs(ε) δk Rk−ϑν )Rj δj kσj − µj kℓ∞ .
j=0 k=j+1

Sending J → ∞, we conclude that the inequality in (2.8) holds since limt→0+ Cs(ε) ε = 0 and
∞
X ∞
Y −ϑν−1
kF∞ (g, σ) − F∞ (g, µ)kCν−1 ≤ Cs(ε) (1 + Cs(ε) δk Rk−ϑν )Rj δj kσj − µj kℓ∞
j=0 k=j+1

≤ Cs(ε) ε exp Cs(ε) ε)kσ − µkℓ∞ .

Rε ,δε = ι ◦ F Rε ,δε : G ν × Σ (R) → G ν−1 is Lipschitz in both σ and g follows
The fact that F̃∞ ∞ ∞
after checking by induction that for all σ ∈ Σ∞ (R)
∞
Y
kF∞ (g, σ) − F∞ (g̃, σ)kCν−1 ≤ (1 + Cs(ε) δj Rj−ϑν )kg − g̃kCν−1 ≤ exp(Cs(ε) ε)kg − g̃kC ν−1 .
j=0

We proceed to prove (2.9). We claim that

Dg F∞ (g, σ) = lim HJ (g, σ̂J ) . . . H1 (g, σ̂1 )H0 (g, σ̂0 ), (8.12)
J→∞

where we abbreviate HJ (g, σ̂J ) := Dg QRJ ,δJ (FJ −1 (g, σ̂J −1 ), σJ ) for J ≥ 1 and set H0 (g, σ̂0 ) =
Dg QR0 ,δ0 (g, σ̂0 ). Indeed, observe that for J > K
 
Dg FJ (g, σ̂J ) − Dg FK (g, σ̂K ) = HJ (g, σ̂J ) . . . HK+1 (g, σ̂K+1 ) − I Dg FK (g, σ̂K ). (8.13)
P
Now, fix ε0 > 0 and choose K > 0 such that Cs(ε) ∞ j=K+1 δj Rj
−ϑν
< ε0 . Then, using (7.4),
kHJ (g, σ̂J ) . . . HK+1 (g, σ̂K+1 ) − IkCν−1 →Cν−1
J
X  
≤ HJ (g, σ̂J ) . . . Hj+1 (g, σ̂j+1 ) Hj (g, σ̂j ) − I
C ν−1 →C ν−1
j=K+1
J
X J
Y
≤ (1 + Cs(ε) δℓ Rℓ−ϑν )Cs(ε) δj Rj−ϑν ≤ exp(Cs(ε) ε)ε0 . (8.14)
j=K+1 ℓ=j+1

Therefore, by (8.13), the sequence {Dg FJ (g, σ̂J )}J is Cauchy. This implies that the limit in (8.12)
exists and the convergence is uniform. Since FJ → F∞ , this implies the equality (8.12). Moreover,
setting K = −1 in (8.14) and combining it with (8.12) we see that
∞
X
kDg F∞ − IkCν−1 →Cν−1 ≤ Cs(ε) exp(Cs(ε) ε) δj Rj−ϑν ≤ Cs(ε) ε exp(Cs(ε) ε) → 0,
j=0

and hence the estimate in (2.9) holds.


 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 51

8.2. Induction on the length of orbits. In this section, we construct the probing family for
which (8.1) holds for a predominant set of metrics. The idea is to show that for any 1 < a < 2,
given non-degeneracy of returning points at time n for all n ≤ ℓ (together with some additional
conditions on the eigenvalues of Poincaré maps), most perturbations have non-degeneracy of
returning points for times n with n ≤ aℓ. An outline is given in Section 4.2.
We start by obtaining non-degeneracy of orbits of length 1. In what follows, given δ0 (ε) and
R0 (ε) we write F0 (g, σ) := QR0 (ε),δ0 (ε) (g, σ) as in (8.7).
Lemma 8.2. Let ν > 0 and b > 0, a function N : (0, 1) → N, and a (ν, N )-good family
of perturbations {QR,δ }R,δ . For all ε > 0, there exist positive constants δ0 , R0 , β0,0 such that
δ0 R0−ϑν < 12 ε and the following holds. For every ̥ > 0 and (Γ, G) that is a (̥, b, y)-admissible
pair for {QR,δ }R,δ , and all K ⊂ G bounded in G ν , there are C > 0 and ε0 > 0 such that for
0 < ε < ε0
sup kF0 (g, σ) − gkC ν < 14 Cε, for all g ∈ K, (8.15)
σ∈Σ(R0 )
and


mΣ(R0 ) σ ∈ Σ(R0 ) : F0 (g, σ) ∈ L̥̃,0 ≤ 12 ε, for all g ∈ K (8.16)
where the set L̥̃,0 is defined by

L̥̃,0 := g ∈ G ν : ∃q ∈ N s.t. Γ ∩ R̥ (1, β0,0,0 q −3 , g)\N q,̥ (1, β0,0,0 , g) 6= ∅ , (8.17)
with β0,0,0 := 2β0,0 .

Proof. Let ϑ̃4 , ϑ5 be as in Definition 7.2, and set

max(b,2ϑ̃4 ) max(ϑν ,ϑ5 +1,1)

R0 = 12 ε, δ0 = min ε2 R0 , εR0 , β0,0 = 14 R0y+1 δ0L .

Next, fix K ⊂ G bounded in G ν . By Definition 7.2 there is C0 > 0 such that kg − F0 (g, σ)kC ν ≤
C0 δ0 R0−ϑν ≤ 21 C0 ε for g ∈ K implying (8.15). Let c, C be as in Corollary 7.7 (with y′ = y) and
note that its hypothesis for n = 1 are satisfied with δ := δ0 and R := R0 provided ε < c. Note
that since S̥ (1, α, g) = S ]∗M ⊃ Γ for all g ∈ G and α > 0, then

σ ∈ Σ(R0 ) : F0 (g, σ) ∈ L̥̃,0 = SgR0 ,δ0 (1, α, 2β0,0 ).
Let α = 12 c. Finally, since y ≥ 1, if the conclusion follows from Corollary 7.7 after asking
ε = 2R0 < min(c2 , cδG , C −1 ).


Now that we have obtained non-degeneracy for returning points with orbits of length 1, we
need to induct on the length of the returning orbit. In order to do this, we work with lengths
between aℓ ̥ and aℓ+1 ̥ for some 1 < a < 2 and induct on ℓ. Before stating our lemma, we set
b := 1 − log2 a ∈ (0, 1), (8.18)
and for j ≥ 0 and ε > 0 set
(j+1)/b
γj (ε) := 5−2ℵ ε(2ℵ+1)a a−4ℵ(j+1)/b . (8.19)
52 YAIZA CANZANI AND JEFFREY GALKOWSKI

We are now ready to do the induction on the length of the orbits. At step ℓ in the induction,
we study returning trajectories of length t ≤ ̥aℓ+1 , in particular, making perturbations so that
the returning trajectories with (aℓ − 1)̥ < t ≤ aℓ+1 ̥ are non-degenerate and so that the non-
degeneracy of shorter trajectories is maintained. See the outline of the proof in Section 4 for
a more detailed explanation of the argument. In what follows, we continue to use the notation
Σℓ (R) introduced in (8.6).
Proposition 8.3. Let ν ≥ 5, y ≥ 1, b > 0, N : (0, 1) → N, and {QR,δ }R,δ be a (ν, N )-good family
of perturbations. There exists d > 0 and for all ε > 0 there are δε := {δℓ }∞ ∞
ℓ=0 , Rε := {Rℓ }ℓ=0 ,
and β0,0 > 0, such that δℓ Rℓ−ϑν < 2−ℓ−1 ε and the following holds. For every ̥ > 0, (Γ, G) that is
a (̥, b, y)-admissible pair for {QR,δ }R,δ , and K ⊂ G bounded in G ν , there are C > 0 and ε0 > 0
such that for 0 < ε < ε0 and all ℓ = 0, 1, . . .
sup bℓ )kC ν ≤ C2−ℓ−2 ε,
kFℓ+1 (g, σ) − Fℓ (g, σ g ∈ K. (8.20)
σ∈Σℓ+1 (R)

In addition, for 0 ≤ i ≤ j ≤ ℓ, there exist constants βi,j ∈ (0, 1], with βi+1,j ≤ βi,j ,
c da j/b
βj+1,j+1 = βj,j ε , (8.21)
with c := max( 2ℵ(2ℵ+3)
2ℵ+1 mν , 2ℵy) and mν := max(b, ϑν , ϑ5 + 1, 2ϑ̃4 , 1) such that the following holds.
−1
Let α0 = 4ε and for 0 ≤ j ≤ ℓ set
2ℵ
αj+1 := γj βj,j , αj,ℓ := (1 − 2−(ℓ−j)−1 )αj , βi,j,ℓ := (1 + 2−(ℓ−j) )βi,j ,
(8.22)
2ℵ 2ℵ−1 1+2−ℓ−1
β̃j,ℓ+1 := sℓ β̃j,ℓ , β̃j,j := βj,j ε , sℓ := 1+2−ℓ
with γj as in (8.19). Then for all g ∈ K, ℓ ≥ 0, and 0 < ε < ε0
 
bj ) ∈ L̥,j ≤ (1 − 2−ℓ−1 )ε,
mΣ∞ (R) σ ∈ Σ∞ (R) : ∃ j ≤ ℓ s.t. Fj (g, σ (8.23)
and, for all σ ∈ Σℓ−1 (R) such that Fℓ−1 (g, σ) ∈ G ν \ L̥,ℓ−1
 
mΣ(R ) σℓ : Fℓ (g, (σ, σℓ )) ∈ L̥,ℓ ≤ 2−ℓ−1 ε. (8.24)
ℓ

Here, L̥,−1 = ∅, F−1 = πG ν , and for ℓ ≥ 0 the set L̥,ℓ is defined as follows: g ∈ G ν \ L̥,ℓ
(1) For 0 ≤ i ≤ ℓ, ai−1 < n ≤ ai ,
Γ ∩ R̥ (n, β̃i,ℓ , g) ⊂ N ̥ (n, β̃i,ℓ , g). (8.25)
(2) For all bℓ ≤ j ≤ ℓ, 0 ≤ i ≤ j, ai−1 < n ≤ ai , and all 1 ≤ q ≤ aj/b−i+1 ,
Γ ∩ R̥ (n, βi,j,ℓ q −3 , g) ∩ S̥ (n, αj,ℓ , g) ⊂ N q,̥ (n, βi,j,ℓ , g). (8.26)
Remark 8.4. There are a large number of parameters in Proposition 8.3. One should think
of these parameters as follows. Those with a β control the degree of q-non-degeneracy that
trajectories have while those with a β̃ control simple non-degeneracy of trajectories. Parameters
with an α control the degree of simplicity that a trajectory must have in order to apply the
inductive hypothesis.
The most important constants will turn out to be c and a. While 1 < a < 2 is left free
for the moment, and will be fixed in the proof of Proposition 8.9, c controls how rapidly the
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 53

non-degeneracy of trajectories decays. This constant is determined in Step 3 of the induction

(Section 8.2.4).
Notice that the parameter c depends on mν , which depends on the regularity ν in which
we work. The reason for definition of mν is as follows: 1) ϑν appears because we want our
perturbations to be close in G ν (see Lemma 8.7), 2) b, 2ϑ̃4 , and ϑ5 + 1 appear because we need
to apply Corollary 7.7 (see Step 3 of the induction (Section 8.2.4)

δε ,Rε
Given ℓ, the proposition yields that for all the perturbations Fℓ−1 (g, σ) that satisfy both
δε ,Rε
(8.25) and (8.26) the set of σℓ that would yield a new perturbation Fℓ (g, σ, σℓ ) not satisfying
either (8.25) or (8.26) is bounded in volume by 2 −(ℓ+1) ε. This allows us to show in the proof
δε ,Rε
of Theorem 1.3 that, except for on a measure ε set of σ, the perturbations F∞ (g, σ) satisfy
(8.25) for all n.
We next explain the reason why we need to include (8.26) in the statement (this is also explained
carefully in Section 4). We prove the proposition by induction in ℓ. Suppose we could perturb
δε ,Rε
Fℓ−1 (g, σ) in such a way that orbits that start in Γ and and are (n, βℓ,ℓ ) returning with n ∈
(a , aℓ ] will also be (n, βℓ,ℓ ) non-degenerate. Then, applying (8.25) to get non-degeneracy of
ℓ−1

(n, βℓ,ℓ ) returning orbits with n ∈ (0, aℓ−1 ] would finish the job. However, this cannot always
be done. The issue is that we can only arrange for most of the perturbations to yield (n, βℓ,ℓ )
non-degenerate orbits when the orbits are (n, αℓ,ℓ ) simple with βℓ,ℓ small in terms of αℓ,ℓ (see
Corollary 7.7). If all (n, βℓ,ℓ ) returning orbits were indeed (n, αℓ,ℓ ) simple, then (8.25) applied to
ℓ − 1 would guarantee that orbits that start at Γ and are (n, βℓ,ℓ ) returning with n ∈ [0, aℓ ] will be
(n, βℓ,ℓ ) non-degenerate as desired. In reality, some returning orbits will not be (n, αℓ,ℓ ) simple,
and so those we will view as iterates of shorter ’loops’ of length n/q ∈ (ai−1 , ai−1 ] with q > 1
that are (n/q, C n αℓ,ℓ ) returning (see Lemma 6.4). If the shorter ’loop’ is (n/q, αi,ℓ ) simple and
C n αℓ,ℓ < βi,ℓ q −3 , then we can apply (8.26) to say that such an orbit is (n/q, βi,ℓ ) q non-degenerate
and use that to infer that the long orbit is actually (n, βℓ,ℓ ) non-degenerate (see Lemma 6.18).
If the shorter loop is not (n/q, αi,ℓ ) simple, then we view it as an iterate of a shorter loop and
repeat the process until it terminates. The reason for the third index in βi,j,ℓ is that we will need
some extra non-degeneracy which is ‘used up’ by making additional small perturbations in future
inductive steps.
We divide the proof of Proposition 8.3 into steps. In Step 0 we show that the base case ℓ = 0
holds. In Step 1 we deal with the returning orbits that are not simple enough to get non-degeneracy
by perturbation, instead showing that they inherit the non-degeneracy by decomposing them into
shorter loops as described above. In Step 2 we prove that the non-degeneracy created up to step
ℓ is preserved by the perturbation performed in the ℓ + 1 step. Finally, in Step 3, we prove that
the volume of the set of perturbations that yield orbits whose degeneracy we cannot control is
small.

8.2.1. Step 0: Setting up the induction argument. Let (Γ, G) be a (̥, b, y)-admissible pair for
{QR,δ }R,δ , and K ⊂ G bounded. Let K̃ ⊂ G be bounded and satisfy K ⊂ K̃ with d(K, K̃ c ) > 0.
By Lemma 8.1 there is ε1 > 0 such that K̃ ⊂ Gεν for 0 < ε < ε1 , where Gεν is as in (8.4), and for
54 YAIZA CANZANI AND JEFFREY GALKOWSKI

any choice of δε and Rε satisfying (8.5), we have

FJRε ,δε (K × ΣJ (Rε )) ⊂ K̃, 0 < ε < ε1 , J ∈ N ∪ {∞}.

Since all the metrics we work with will be of the form FJRε ,δε (g, σ̂J ), for some g ∈ K, they will lie
inside K̃ and hence have uniform estimates.
Let ε̃0 be the number ε0 found in Lemma 8.2. From now on we assume that ε0 ≤ min{ε̃0 , ε1 }.
For convenience, in what follows we work with
C0 = 2 max{1, C, C1 , C2 , C3 , C4 , δG−1 , CG }, (8.27)
where C is the constant from Corollary 7.7, C1 is the constant from Lemma 6.4, C2 > 0 is such
that kϕgt − ϕg̃t kC 1 ≤ C2t kg − g̃kC 3 for all g, g̃ ∈ K̃ and t ∈ R, C3 is as in Lemma 6.18, and C4 is
the maximum of the constants C in Definition 7.2 for the choices of ν in {3, 5, ν}, and CG , δG are
from Lemma 6.12 and (6.11). Note that C0 depends only on ν, N, y, b, a, ̥, Γ, G, K. During the
induction argument we will ask that ε0 be small in terms of powers of C0−1 .
We prove the lemma by induction on ℓ. We check the cases 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1. Since
a⌈loga 2⌉−1 < 2, this amounts to considering 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1, n = 1, 0 ≤ i ≤ ℓ in (8.25) and
bℓ ≤ j ≤ ℓ, 0 ≤ i ≤ j, and 1 ≤ q ≤ aj/b−i+1 in (8.26).
By Lemma 8.2, for all ε > 0 there are β0,0 , R0 , and δ0 such that such that (8.15) holds and (8.16)
holds for all g ∈ K and 0 < ε < ε0 (Γ, K). ,Set δℓ (ε) = 0, Rℓ (ε) = 1 for 1 ≤ ℓ ≤ ⌈loga 2⌉ − 1. That
is,
F0Rε ,δε (g, σ0 ) = FℓRε ,δε (g, (σ0 , σ1 , . . . , σℓ )), 1 ≤ ℓ ≤ ⌈loga 2⌉ − 1. (8.28)
Then, (8.15) implies (8.20) for ℓ ≤ ⌈loga 2⌉ − 1.
Set βi,j = βi,i for 0 ≤ i ≤ j ≤ ⌈loga 2⌉ − 1. We note that by (8.28), it is enough to show that
 [ 
mΣ(R0 ) σ : F0 (g, σ) ∈ L̥,ℓ ≤ 2−1 ε.
0≤ℓ≤⌈loga 2⌉−1

We next claim that for ℓ ≤ ⌈loga 2⌉ − 1, and g ∈ K,

n [ o n o
σ : F0Rε ,δε (g, σ) ∈ L̥,ℓ ⊂ σ : F0Rε ,δε (g, σ) ∈ L̥̃,0
0≤ℓ≤⌈loga 2⌉−1

where L̥̃,0 is defined in Lemma 8.2.

2ℵ ε2ℵ−1 and β
Indeed, since β̃0,0 = β0,0 0,0,0 = 2β0,0 for ε > 0 small enough depending on K and
all g ∈ K,
N (1, β̃0,0 , F0 (g, σ)) ⊂ N 1,̥ (1, β0,0,0 , F0 (g, σ)).
Therefore, by (8.17), if F0Rε ,δε (g, σ) ∈
/ L̥̃,0 , then (8.25) holds for 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1 and g
Rε ,δε
replaced by F0 (g, σ).
Next, since βi,j,ℓ ≤ β0,0,0 for i = 0, bℓ ≤ j ≤ ℓ, and 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1, we have that

g ∈ G ν : Γ ∩ R̥ (1, β0,0,0 q −3 , g) ∩ S̥ (n, αj,ℓ , g)\N q,̥ (1, β0,0,0 , g) 6= ∅ ⊂ L̥̃,0
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 55

and hence, if F0Rε ,δε (g, σ) ∈

/ L̥̃,0 , then (8.26) holds for 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1 and g replaced by
Rε ,δε
F0 (g, σ). The claim in (8.23) then follows from (8.16). This concludes the proof of the base
case in the induction argument.
Note that with β0,0 in place, the constants βℓ,ℓ and αℓ are defined as in (8.21) and (8.22) for
all ℓ. For ε > 0 and ℓ = ⌈loga 2⌉, ⌈log a 2⌉ + 1, . . . we set
ℓ
 ℓ 
max(b,2ϑ̃4 ) max(ϑν ,ϑ5 +1,1)
Rℓ (ε) := 18 εa +1 αℓ , δℓ (ε) := min εa +1 Rℓ , ε2−ℓ Rℓ . (8.29)

We note that δℓ (ε)Rℓ (ε)−ϑν < 2−ℓ−1 ε and so δε = {δℓ (ε)}∞ ∞

ℓ=0 and Rε = {Rℓ (ε)}ℓ=0 satisfy (8.5).
We also note that with these definitions there is ε0 small enough that the assumptions (7.19) in
Corollary 7.7 are satisfied with R = Rℓ+1 (ε), δ = δℓ+1 (ε), α = 21 αℓ+1 , and n ≤ aℓ+1 .
We now start the inductive step. Note that the base case 0 ≤ ℓ ≤ ⌈loga 2⌉ − 1 only covers n = 1
since aℓ < n ≤ aℓ+1 and a < 2. Since to get to n = 2 we need 2 ≤ aℓ+1 , we may assume that the
inductive hypotheses hold for some ℓ ≥ ⌈loga 2⌉ − 1. We split the proof of the inductive step with
ℓ + 1 into three steps.

8.2.2. Step 1: Nondegeneracy of returning points that are not simple. The goal of this step is to
study the set of points in Γ that under the perturbed metric Fℓ (g, σ) generate orbits that return to
Γ at some ‘discrete’ time n ∈ (aℓ , aℓ+1 ] but that are not simple enough that the Fℓ+1 perturbations
would make them non-degenerate. As explained before, we decompose these returning orbits into
shorter ‘loops’ and we use the (βq , q) non-degeneracy of shorter orbits given by (8.26) to show
that the original orbits were already non-degenerate.
We note that for this step in the proof the exact powers in the definition of βj,j do not play
a role. Instead, we only use that βj,j ≥ βj+1,j+1 and βi+1,j ≤ βi,j for all i and j. The precise
definition of γj as well as the definition of αj+1 in terms of γj and βj,j do, however, play a role.
Lemma 8.5. There is ε0 > 0 depending only on ν, N, y, b, a, ̥, Γ, G, K so that the following
holds. Suppose the conclusions of Proposition 8.3 hold up to the index ℓ. Then, for all 0 < ε < ε0 ,
aℓ < n ≤ aℓ+1 and (g, σ) ∈ K × Σℓ (Rε ) such that gℓ := Fℓ (g, σ) ∈ G ν \LT,ℓ we have
Γ ∩ R̥ (n, αℓ+1 , gℓ ) \ S̥ (n, αℓ+1 , gℓ ) ⊂ N ̥ (n, αℓ+1 , gℓ ). (8.30)

Proof. Let aℓ < n ≤ aℓ+1 and (g, σ) ∈ K × Σℓ (R) such that gℓ := Fℓ (g, σ) ∈ G ν \L̥,ℓ . Let
ρ ∈ Γ ∩ R̥ (n, αℓ+1 , gℓ ) \ S̥ (n, αℓ+1 , gℓ ). (8.31)

We divide the proof into fours steps. In Step A, we decompose the non-simple orbit associated
to ρ into shorter (km , δm (ℓ)) returning orbits that are (km , αm (ℓ)) simple. The returning times
tm associated to these orbits are such that tm divides tm−1 , and km = ⌈ t̥m ⌉ with k0 = n. In
Step B we show that ρ ∈ N qm ,̥ (km , βm (ℓ), gℓ ) with qm := t0 /tm ∈ Z. In Step C we prove that
there are I and p ∈ Z such that for all q ≤ qm there is a time tq,m sufficiently close to qtm
with the property that (PIp [gℓ ])q (ρ) = PIqp [gℓ ](ρ) = ϕgtq,m
ℓ
(ρ). In particular, one can control the
p q
distance from (PI [gℓ ]) (ρ) to ρ for q ≤ qm . In Step D we check that the above guarantees that
the hypothesis of Lemma 6.18 are satisfied and so ρ is (n, αℓ+1 ) non-degenerate as claimed.
56 YAIZA CANZANI AND JEFFREY GALKOWSKI

Step A. We start the proof by showing that there are 1 ≤ m ≤ ⌊(ℓ + 1) log2 a⌋, {ti }m i=0 ⊂ R, and
{ki }m
i=0 ⊂ Z, such that k0 = n, (ki − 1)̥ < t i ≤ ki ̥, t i divides t i−1 for 1 ≤ i ≤ m, t i > ̥,

ρ ∈ Γ ∩ R̥ (km , δm (ℓ), gℓ ) ∩ S̥ (km , αm (ℓ), gℓ ), d(ϕgtmℓ (ρ), ρ) ≤ αm (ℓ), (8.32)

and
d(ϕgt ℓ (ρ), ϕgr ℓ (ρ)) ≤ δi (ℓ), 0 ≤ t ≤ ti−1 , t = qti + r, 0 ≤ r < ti . (8.33)
Here, for i ≥ 1
t
αi (ℓ) := αℓ−i+1 , δ0 (ℓ) = α0 (ℓ), δi (ℓ) := C0i−1 αi−1 (ℓ),
with C0 as in (8.27).
Note that ρ ∈ Γ ∩ R̥ (k0 , δ0 (ℓ), gℓ ) \ S̥ (k0 , α0 (ℓ), gℓ ) by (8.31). We next show that, by splitting
the orbit into shorter returning ones, we can find {ti }m−1 i=0 such that ti+1 divides ti and ρ ∈
ti
Γ ∩ R̥ (ki , δi (ℓ), gℓ ) \ S̥ (ki , αi (ℓ), gℓ ) with ki := ⌈ ̥ ⌉, and that both (8.32) and (8.33).
To do this, suppose we have found {ti }m−1 ti
i=0 such that ti+1 divides ti , and, with ki := ⌈ ̥ ⌉,

ρ ∈ R̥ (ki , δi (ℓ), gℓ ), (8.34)

ρ∈
/ S̥ (ki , αi (ℓ), gℓ ), (8.35)
and (8.33) hold for 0 ≤ i ≤ m − 1. We prove by induction that there is tm dividing tm−1 such that
(8.34) holds with i = m and (8.33) holds for 1 ≤ i ≤ m. Note that the base case i = 0 follows
from (8.31) after letting t0 be such that n = ⌈ t̥0 ⌉ since k0 = n.
First, note that δi (ℓ) ≤ αi (ℓ) for 1 ≤ i ≤ m − 1. Indeed, since {βj,j }j is decreasing it is
t
straightforward to check that δi (ℓ) ≤ αi (ℓ) provided C0i−1 γℓ−i−1 ≤ γℓ−i−2 . Since a−4ℵ/b ≤ 1, this
t (ℓ−i+1)/b (a1/b −1)
reduces to checking that C0i−1 ε(2ℵ+1)a ≤ 1. Since ti−1 ≤ ki−1 ̥ ≤ aℓ+1 2−(i−1) ̥, the
claim follows from letting ε0 be small.
Next, since (8.34) and (8.35) hold with i = m − 1 and δi (ℓ) ≤ αi (ℓ) for 1 ≤ i ≤ m − 1, using
Lemma 6.6 together with CΓ > 1 we may apply Lemma 6.4 with C̥ := CΓ > 1, δ := δm−1 (ℓ),
t
α := αm−1 (ℓ), t0 := tm−1 , to obtain the existence of tm ∈ [CΓ ̥ − C0m−1 αm−1 (ℓ) , 12 tm−1 ] dividing
gℓ
tm−1 such that d(ϕtm (ρ), ρ) ≤ δm (ℓ) and for 0 ≤ l < tm−1 with l = qtm + r and 0 ≤ r < tm
d(ϕgl ℓ (ρ), ϕgr ℓ (ρ)) = δm (ℓ). (8.36)
Note that, provided ε > 0 was chosen small enough small enough (uniformly in K, ℓ), and using
that CΓ > 1, we have tm > ̥. If ρ ∈ / S̥ (km , αm (ℓ), gℓ ), the inductive step i = m is complete. On
the other hand, if ρ ∈ S̥ (km , αm (ℓ), gℓ ) then we have proved (8.32) and (8.33).
At this point, we have proved either that there is m ≤ ⌊(ℓ+1) log 2 a⌋ such that (8.32) and (8.33)
hold or there are {ti }N i=1 , with N > ⌊(ℓ + 1) log 2 a⌋ such that ti divides ti−1 , ti > ̥, and with
ti
ki = ⌈ ̥ ⌉, (8.33), (8.34), (8.35) hold for all i. We claim that the second alternative is not possible.
Indeed, note that since ti divides ti−1 ,
̥ < ti ≤ 2−i t0 ≤ 2−i aℓ+1 ̥,
and this is not possible for i > ⌊(ℓ + 1) log2 a⌋ since then 2−i aℓ+1 ≤ 1. It follows that the
claims (8.32) and (8.33) are true with m as stated.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 57

Step B. Let qm = t0 /tm and βm (ℓ) = βim ,jm with im = ⌈loga km ⌉, jm = ℓ − m + 1. We claim
ρ ∈ N qm ,̥ (km , βm (ℓ), gℓ ). (8.37)
The objective is to show that we can apply the induction hypothesis (8.26) with (n, q, i, j, ℓ) :=
(km , qm , im , jm , ℓ). To this end, we first observe that by definition aim −1 < km ≤ aim . In addition,
km ≤ 2−m aℓ+1 ≤ aℓ−m+1 and so im ≤ jm . Also, since m ≤ ⌊(ℓ + 1)(1 − b)⌋, we have b(ℓ + 1) ≤
jm
jm ≤ ℓ. In particular, t̥0 ≤ aℓ+1 ≤ a b . Since t̥m ≥ aim −1 , this also yields qm = ttm0 ≤ ajm /b−im +1 .
We first claim that
−3
δm (ℓ) ≤ βm (ℓ)qm .
Indeed, using that βjm ,jm ≤ βim ,jm (since im ≤ jm ) and qm ≤ aℓ+1 , this reduces to checking that
t
a3(ℓ+1) C0m−1 αjm +1 ≤ βjm ,jm . Furthermore, since βjm ,jm ≤ 1, the claim reduces to showing that
t
a3(ℓ+1) C0m−1 γjm ≤ 1, and this is equivalent to
ℓ+1 2−m+1 (ℓ−m+2)/b
5−2ℵ ≤ C0−a a4ℵ(ℓ−m+2)/b−3(ℓ+1) ε−(2ℵ+1)a .
The claim then follows from the facts that ε0 ≤ C0−1 and (ℓ − m + 2)/b ≥ ℓ + 1.
Since δm (ℓ) ≤ βm (ℓ)qm −3 and β −m )β (ℓ), we conclude that R (k , δ (ℓ), g ) ⊂
im ,jm ,ℓ = (1 + 2 m ̥ m m ℓ
−3
R̥ (km , βim ,jm ,ℓ qm , gℓ ). Therefore, the claim in (8.37) follows from combining (8.32) and (8.26)
with (n, i, j) = (km , im , jm ). This can be done since βim ,jm ,ℓ ≥ βm (ℓ) and S̥ (km , αm (ℓ), gℓ ) ⊂
S̥ (km , αjm ,ℓ , gℓ ) because αjm ,ℓ = (1 − 2−m )αm (ℓ).

Step C. Recall that km = ⌈ t̥m ⌉ and t0 = qm tm with k0 = n. Let CG , δG be as in Lemma 6.12 and
(6.11). Let
ℓ+1
δm,ℓ := mC0̥a αm−1 (ℓ). (8.38)
 
We next show there are I and p with TI(p) [gℓ ](ρ) ∈ (km − 1)̥ − CG δG , km ̥ + CG δG ,
 (qp) 
q 
qm
PI [gℓ ](ρ) q=0 ⊂ DI(p) [gℓ ], max d PI(p) [gℓ ] (ρ), ρ ≤ CG δm,ℓ ,
0≤q≤qm (8.39)
(qp) ′ (p)

q ′
PI [gℓ ](ρ ) = PI [gℓ ] (ρ ), 0 ≤ q ≤ qm , ρ′ in a neighborhood of ρ.

Let 0 ≤ q ≤ qm . Since {αj (ℓ)}j is decreasing, it follows from Lemma 8.6 below that
d(ϕgqtℓm (ρ), ρ) ≤ δm,ℓ . (8.40)
Let ε0 be small enough that δm,ℓ ≤ δG for any choice of m, ℓ. Then, since ρ ∈ Γ, and δm,ℓ ≤ δG
by (6.8) and (6.9) there exists tq,m with such that
|tq,m − qtm | < CG δm,ℓ , ϕgtq,m
ℓ
(ρ) ∈ Γ̃, d(ϕgtq,m
ℓ
(ρ), ρ) < CG δm,ℓ . (8.41)
By Remark 6.10 there are p and Ĩ such that ρ ∈ D (p) [gℓ ],
Ĩ

P (p)
[gℓ ](ρ) = ϕgt1,m (ρ), T (p)
[gℓ ](ρ) = t1,m , max d(P (j) [gℓ ](ρ), Γ) < δG . (8.42)
Ĩ Ĩ 0≤j≤pqm Ĩ

Let c∗ be the constants given in Lemma 6.13. Now, by Lemma 6.13, we have BΓ̃ (ρ, cp∗ ) ⊂ D (p) [gℓ ].
Ĩ
Note that, since p ≤ Caℓ+1 , one can choose ε0 small so that CG δm,ℓ ≤ cp∗ for all m, ℓ and ε < ε0 .
Thus, [P (p) ]q (ρ) exists for 0 ≤ q ≤ qm .
Ĩ
58 YAIZA CANZANI AND JEFFREY GALKOWSKI

Next, we claim that for all 1 ≤ q ≤ qm

q
P (p) [gℓ ] (ρ) = ϕgtq,m
ℓ
(ρ). (8.43)
Ĩ

To see this, set t0,m = 0. Since the claim is true for q = 1, we may assume that for some
1 ≤ q < qm , and every 1 ≤ j ≤ q,

j−1 
T (p) [gℓ ] P (p) [gℓ ] (ρ) = tj,m − tj−1,m , (8.44)
Ĩ Ĩ

and hence that for 1 ≤ j ≤ q

j 
j−1 
P (p) [gℓ ] (ρ) = ϕgtj,m
ℓ
−tj−1,m P (p)
[gℓ ] (ρ) = ϕgtj,m
ℓ
(ρ). (8.45)
Ĩ Ĩ

Now, observe that there exists C depending only on G such that

|T (p) [gℓ ](ρ) − T (p) [gℓ ](ρ′ )| ≤ C p̥ d(ρ, ρ′ ).
Ĩ Ĩ

Hence, using (8.41), (8.42), and (8.45)


q 
t1,m − T (p) [gℓ ] P (p) [gℓ ] (ρ) ≤ C p̥ CG δm,ℓ .
Ĩ Ĩ

Thus, we have

q  
q 
ϕgtq+1,m
ℓ
−tq,m P (p) [gℓ ] (ρ) ∈ Γ̃, ϕgℓ 
q  P (p) [gℓ ] (ρ) ∈ Γ̃,
Ĩ (p) (p) Ĩ
T [gℓ ] P [gℓ ] (ρ)
Ĩ Ĩ

and, choosing ε0 small enough,


q 
tq+1,m − tq,m − T (p) [gℓ ] P (p) [gℓ ] (ρ) < (C p̥ + 3)CG δm,ℓ < cΓ ̥.
Ĩ Ĩ

In particular, this implies that


q 
tq+1,m − tq,m = T (p) [gℓ ] P (p) [gℓ ] (ρ) ,
Ĩ Ĩ

and hence (8.44) holds with q replaced by q + 1. This shows that (8.43) holds.
Let Ĩ := (ĩ0 , ĩ1 , . . . ). Using again that ε > 0 can be chosen small enough, this implies that,
with the chain
I := (ĩ0 , ĩ1 , . . . , ĩp , ĩ1 , . . . , ĩp , . . . ),
we have
(⌈qtm ⌉ − 1)̥ − CG δG ≤ TI(pq) [gℓ ](ρ) ≤ ⌈qtm ⌉̥ + CG δG , (8.46)
and hence
q
PI(qp) [gℓ ](ρ) = ϕgtq,m ℓ
(ρ) = P (p) [gℓ ] (ρ).
Ĩ
(p) (pq)
q
Moreover, by definition, using that the fact that DI [gℓ ] and DI [gℓ ] are open, we have PI(p) [gℓ ] (ρ′ ) =
PI(qp) [gℓ ](ρ′ ) for ρ′ in a neighborhood of ρ as claimed.

Step D. We now complete the proof of the lemma. It follows from (8.32) and (8.37) that
ρ ∈ Γ ∩ R̥ (km , CG δm,ℓ , gℓ ) ∩ N qm ,̥ (km , βm (ℓ), gℓ ), (8.47)
with δm,ℓ as in (8.38). Here, we have used that δm (ℓ) ≤ CG δm,ℓ since tm−1 ≤ ̥km ≤ The ̥aℓ+1 .
goal is then to use Lemma 6.18 with δ := CG δm,ℓ , and (β, q0 , k, m) := (βm (ℓ), qm , km , p). Let C3
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 59

be as in Lemma 6.18 and note that C0 > C3 by (8.27). Thus, to apply Lemma 6.18 we first need
to check that

CG δm,ℓ ≤ min δG , (2f(βm (ℓ), qm )C0km qm )−1 . (8.48)

To see this, note that by the definitions (8.38) and (6.14)

ℓ+1 +k
CG δm,ℓ 2f(βm (ℓ), qm )C0km qm = 2CG mC0̥a m qm
( 52 )2ℵ qm
4ℵ
αm−1 (ℓ)βm (ℓ)−2ℵ .

In addition, αm−1 (ℓ)βm (ℓ)−2ℵ ≤ γℓ−m+1 (ε) since βim ,jm ≥ βjm ,jm . Therefore, the bound in (8.48)
follows from (8.19) after noting that km qm ≤ aℓ+1 and (ℓ − m + 1)/b ≤ ℓ + 1, and choosing ε small
enough.
Combining (8.48), (8.47), and (8.39) we may apply Lemma 6.18 and obtain


ρ ∈ N ̥ n, (2f(βm (ℓ), qm )C0n )−1 , gℓ .
Here, we have used that, by (8.46) and (n − 1)̥ ≤ t0 ≤ n̥, we have

(n − 1)̥ − CG δG ≤ TI(qm p) [g](ρ) ≤ n̥ + CG δG .

Finally, using that βm (ℓ) ≥ βjm ,jm ≥ βℓ,ℓ one can check that C −k0 ≥ 2γℓ (5/2)2ℵ qm 4ℵ . In

particular, since qm ≤ aℓ+1 , b ∈ [0, 1], making ε0 < C0−1 would imply [2f(βm , qm )C0n ]−1 ≥ αℓ+1 ,
and the proof is complete. 
Lemma 8.6. Let ρ ∈ Γ satisfy (8.33) and (8.36). Then for all 0 ≤ j ≤ m − 1 and 0 ≤ q ≤ tj /tm ,
m−1
X Pm−1
tl
d(ϕgqtℓm (ρ), ρ) ≤ C0 l=i
αi (ℓ). (8.49)
i=j

Proof. First, note that for j = m − 1, the statement follows from (8.36). Let 0 ≤
Pjm≤ m − 2 and
assume (8.49) holds for 0 ≤ q ≤ tj+1 /tm . Next, let q < tj /tm and write qtm = i=j+1 qi ti with
qj < tj−1 /tj . Then, by (8.33) with i = j + 1, t = qj+1 tj+1 , and r = 0,
t
d(ϕgqj+1
ℓ j
tj+1 (ρ), ρ) ≤ C0 αj (ℓ).

Next, note that gℓ = Fℓ (g, σ) ∈ K̃ since g ∈ K. In particular, kϕg̥ℓ kC ν−2 ≤ Ckgℓ kC ν ≤ C0 for all
ℓ. It follows that
P Pm−1

tj + m
l=j+2 ql tl tl
d ϕgPℓ m qi ti (ρ), ϕgPℓ m q t (ρ) ≤ C0 αj (ℓ) ≤ C0 l=j αj (ℓ).
i=j+1 l=j+2 l l

Pm−1 Pm−1
l=i tl
By the induction hypothesis, d ϕgPℓ m qi ti
(ρ), ρ ≤ C
i=j+1 0 αi (ℓ). Thus,
i=j+2

Pm−1 m−1
X Pm−1 m−1
X Pm−1
tl tl tl
d(ϕgPℓ m+1 (ρ), ρ) ≤ C0 l=j
αj (ℓ) + C0 l=i
αi (ℓ) = C0 l=i
αi (ℓ)
i=j+1 qi ti
i=j+1 i=j

and the claim follows by induction. 

60 YAIZA CANZANI AND JEFFREY GALKOWSKI

8.2.3. Step 2: Preserving non-degeneracy under perturbation. In this section, we show that if the
non-degeneracy properties (8.25) and (8.26) listed in Proposition 8.3 hold for gℓ := Fℓ (g, σ), then
they also hold for the perturbed metric (gℓ )R,δ
σ := Q
R,δ (g , σ) for appropriate R, δ.
ℓ

Lemma 8.7. There is ε0 > 0 depending only on ν, N, y, b, a, ̥, Γ, G, K so that the following holds.
Suppose that the conclusions in Proposition 8.3 are valid up to the index ℓ and let 0 < ε < ε0 .
Then, for all (g, σ) ∈ K × Σℓ (Rε ) such that gℓ := Fℓ (g, σ) ∈ G ν \LT,ℓ the following holds. Let
(gℓ )R,δ
σ := Q
R,δ (g , σ), with R = R
ℓ ℓ+1 (ε) and δ = δℓ+1 (ε) as defined in (8.29). Then, with C0 as
in (8.27),
k(gℓ )R,δ
σ − gℓ kC ν ≤ C0 2
−ℓ−2
ε, (8.50)
and
• for 0 ≤ i ≤ ℓ, ai−1 < n ≤ ai ,
Γ ∩ R̥ (n, β̃i,ℓ+1 , (gℓ )R,δ R,δ
σ ) ⊂ N ̥ (n, β̃i,ℓ+1 , (gℓ )σ ). (8.51)
• for aℓ < n ≤ aℓ+1
Γ ∩ R̥ (n, 12 αℓ+1 , (gℓ )R,δ 1 R,δ 1 R,δ
σ ) \ S̥ (n, 2 αℓ+1 , (gℓ )σ ) ⊂ N ̥ (n, 2 αℓ+1 , (gℓ )σ ). (8.52)
• for b(ℓ + 1) ≤ j ≤ ℓ, 0 ≤ i ≤ j, ai−1 < n ≤ ai , and 1 ≤ q ≤ aj/b−i ,
Γ ∩ R̥ (n, βi,j,ℓ+1 q −3 , (gℓ )R,δ R,δ R,δ
σ ) ∩ S̥ (n, αj,ℓ+1 , (gℓ )σ ) ⊂ N q,̥ (n, βi,j,ℓ+1 , (gℓ )σ ). (8.53)

Proof. By Definition 7.2 and (8.27), the bound in (8.50) holds since
k(gℓ )R,δ
σ − gℓ kC ν = kQ
R,δ
(gℓ , σ) − QR,δ (gℓ , 0)kC ν ≤ 12 C0 δR−ϑν (8.54)
and, by definition,
δ ≤ εRϑν 2−(ℓ+1) . (8.55)
We next address (8.51). Let 0 ≤ i ≤ ℓ, ai−1 < n ≤ ai . We claim that
R̥ (n, β̃i,ℓ+1 , (gℓ )R,δ
σ ) ⊂ R̥ (n, β̃i,ℓ , gℓ ),
(8.56)
Γ ∩ R̥ (n, β̃i,ℓ , gℓ ) ∩ N ̥ (n, β̃i,ℓ , gℓ ) ⊂ N ̥ (n, β̃i,ℓ+1 , (gℓ )R,δ
σ ),

provided
i
δ ≤ C0−a ̥−1 Rϑ3 β̃i,ℓ (1 − sℓ ), 0 ≤ i ≤ ℓ. (8.57)
Once we prove (8.56), the claim in (8.51) will follow by applying the inductive assumption (8.25)
to gℓ . One can check that (8.57) is valid after asking that ε0 < C0−1 .
To see the first part of (8.56) suppose that ρ ∈ R̥ (n, β̃i,ℓ+1 , (gℓ )R,δ
σ ). Then, ρ ∈ R̥ (n, β̃i,ℓ , gℓ )
since Definition 7.2, (8.57), and (8.27) imply
(gℓ )R,δ
ϕgt ℓ − ϕt σ
C1
≤ C0t gℓ − (gℓ )R,δ
σ C3
≤ C0t+1 δR−ϑ3 < β̃i,ℓ (1 − sℓ ) = β̃i,ℓ − β̃i,ℓ+1 , (8.58)
provided (n − 1)̥ ≤ t ≤ n̥ and (8.57) holds.
For the second part of (8.56), note that with δG as in Lemma 6.12 and (6.11), if β̃i,ℓ < δG , we
may work in a single coordinate chart (U, ψ). Suppose ρ ∈ Γ ∩ R̥ (n, β̃i,ℓ , gℓ ) ∩ N ̥ (n, β̃i,ℓ , gℓ ).
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 61

By Definition 6.14 and Lemma 6.12 there exist I and m such that TI(m) [gℓ ](ρ) ∈ (n − 1)̥ −

CG β̃i,ℓ , n̥ + CG β̃i,ℓ and

−1 −1
sup I − d((PI(m) [gℓ ])ψ )(ρ) < β̃i,ℓ . (8.59)
{(ψ,U )∈A: B(ρ,CG δG )⊂U }

In addition, since otherwise we may modify the choice of chain, we may assume without loss of
generality that
d(PI(j) [gℓ ](ρ), Γ) < δG , 0 ≤ j ≤ m. (8.60)
Then, by Lemma 6.13 there is c > 0 depending only on G, Γ, ̥ such that
ρ ∈ DI(m) [(gℓ )R,δ
σ ],

as long as k(gℓ )R,δ m

σ − gℓ kC ν ≤ c . Here, we use that (8.54) holds with ν = 3 together with (8.29)
and ϑ5 + 1 − ϑ3 > 0 to justify
 that ε0 can be chosen small  enough that k(gℓ )R,δ m
σ − gℓ kC ν ≤ c .
Next, since TI(m) [gℓ ](ρ) ∈ (n − 1)̥ − CG β̃i,ℓ , n̥ + CG β̃i,ℓ , we have
TI(m) [(gℓ )R,δ n
σ ](ρ) ∈ [(n − 1)̥ − CG β̃i,ℓ − C0 δR
−ϑ3
, n̥ + CG β̃i,ℓ + C0n̥+2 δR−ϑ3 ]. (8.61)
Indeed, by (8.58), implicit function theorem, and that H|ξ|g is uniformly transverse to Γ,

|TI(m) [(gℓ )R,δ (m)

σ ](ρ) − TI [gℓ ](ρ)| ≤ CG C0n̥+1 δR−ϑ3
and (8.61) follows. Now, since for ε small enough,
CG C0n̥+1 δR−ϑ3 + CG β̃i,ℓ < CG β̃i,ℓ+1 < CG δG ,
we have
TI(m) [(gℓ )R,δ
σ ](ρ) ∈ [(n − 1)̥ − CG δG , n̥ + CG δG ]. (8.62)

Let (ψ, U ) ∈ A be such that B(ρ, CG δG ) ⊂ U . Abusing notation slightly, we work with PI(m) [gℓ ],
and PI(m) [(gℓ )R,δ (m)
σ ] for the maps induced in the (ψ, U ) coordinates. Let A = I − dPI [gℓ ](ρ) and
∆ = dPI(m) [gℓ ](ρ) − dPI(m) [(gℓ )R,δ
σ ](ρ), and observe that (8.58) yields k∆k ≤ β̃i,ℓ (1 − sℓ ). In
particular, by (8.59),
−1
−1 kA−1 k β̃i,ℓ −1 −1 −1
k(A + ∆) k≤ ≤ = β̃i,ℓ sℓ = β̃i,ℓ+1 .
1 − kA−1 kk∆k −1
1 − β̃i,ℓ β̃i,ℓ (1 − sℓ )

Combining this with (8.62) we conclude ρ ∈ N ̥ (n, β̃i,ℓ+1 , (gℓ )R,δ

σ ), and hence the second part
of (8.56). As noted earlier, this shows that (8.51) holds.
We next prove (8.52). Let aℓ < n ≤ aℓ+1 . We will show that
R̥ (n, 12 αℓ+1 , (gℓ )R,δ
σ ) ⊂ R̥ (n, αℓ+1 , gℓ ),
(8.63)
R̥ (n, αℓ+1 , gℓ ) ∩ N ̥ (n, αℓ+1 , gℓ ) ⊂ N ̥ (n, 21 αℓ+1 , (gℓ )R,δ
σ ),

S̥ (n, αℓ+1 , gℓ ) ⊂ S̥ (n, 12 αℓ+1 , (gℓ )R,δ

σ ), (8.64)
provided
ℓ+1 ̥−1
δ ≤ C0−a Rϑ3 21 αℓ+1 . (8.65)
62 YAIZA CANZANI AND JEFFREY GALKOWSKI

Note that combining (8.63) and (8.64) with Lemma 8.5 yields the claim in (8.52). Also, it is
immediate to check that (8.65) is valid after asking ε0 ≤ C0−1 .
The claims in (8.63) follow from the same proof that yielded (8.56). To see that (8.64)
holds, observe that if ρ ∈ / S̥ (n, 12 αℓ+1 , (gℓ )R,δ 1 1
σ )), then there is t ∈ ( 2 ̥, (n − 2 )̥) such that
(g )R,δ
R,δ
(g ) )
d ϕt ℓ σ (ρ), ρ < 12 α
ℓ+1 . Thus, by (8.58) and (8.65), we know kϕgt ℓ − ϕt ℓ σ kC 1 ≤ 12 αℓ+1 . In
particular, d ϕgt ℓ (ρ), ρ < αℓ+1 , and hence ρ ∈ / S̥ (n, αℓ+1 , κℓ ) as needed.
Finally, we address (8.53). Let 0 ≤ i ≤ j, ai−1 < n ≤ ai , b(ℓ + 1) ≤ j ≤ ℓ, and 1 ≤ q ≤ aj/b−i .
We claim that, provided (8.68) and (8.71) below hold,
R̥ (n, βi,j,ℓ+1 q −3 , (gℓ )R,δ −3
σ ) ⊂ R̥ (n, βi,j,ℓ q , gℓ ), S̥ (n, αj,ℓ+1 , (gℓ )R,δ
σ ) ⊂ S̥ (n, αj,ℓ , gℓ ) (8.66)

Γ ∩ R̥ (n, βi,j,ℓ q −3 , gℓ ) ∩ N q,̥ (n, βi,j,ℓ , gℓ ) ⊂ N q,̥ (n, βi,j,ℓ+1 , (gℓ )R,δ
σ ). (8.67)
Note that combining (8.66) and (8.67) with (8.26) yields the claim in (8.53).
For (8.66) we argue as in the proofs of the first part of (8.56) and (8.64), and use that αj,ℓ+1 ≥
1
2 αj,ℓ . We can do this provided
 ℓ ℓ

δ < min C0−a ̥−1 Rϑ3 (βi,j,ℓ − βi,j,ℓ+1 )(aj/b−i )−3 , C0−a ̥−1 Rϑ3 12 αj,ℓ . (8.68)

We note that (8.68) holds for ε0 small enough since {αj }j is decreasing and βℓ,ℓ ≤ βi,j .
The inclusion in (8.67) will follow from Lemma 8.8 below with β = βi,j,ℓ . Indeed, provided
−nq 2ℵ −4ℵ−1
kgℓ − (gℓ )R,δ
σ kC 3 ≤ min(C1 β q , δG − β), (8.69)
we have by Lemma 8.8
Γ ∩ R̥ (n, βi,j,ℓ q −3 , gℓ ) ∩ N q,̥ (n, βi,j,ℓ , gℓ )
⊂ Γ ∩ R̥ (n, βi,j,ℓ , gℓ ) ∩ N q,̥ (n, βi,j,ℓ , gℓ ) (8.70)
−2ℵ nq 4ℵ+1 1/2ℵ
⊂ N q,̥ (n, βi,j,ℓ (1 − kgℓ − (gℓ )R,δ
σ kC 3 βi,j,ℓ C1 q ) , (gℓ )R,δ
σ ).

Since,
j/b −1
δ < Rϑ3 C1−a (aj/b−i )−4ℵ+1 βi,j,ℓ
2ℵ
(1 − s2ℵ
ℓ−j ), (8.71)
the bound on kgℓ − (gℓ )R,δ
σ kC 3 from (8.54) and (8.71) imply the first part of (8.69) and
−2ℵ nq 4ℵ+1 1/2ℵ
βi,j,ℓ+1 ≤ βi,j,ℓ (1 − kgℓ − (gℓ )R,δ
σ kC 3 βi,j,ℓ C q ) . (8.72)
δ δ
Furthermore, for ε < C0−1 , we have βi,j,ℓ ≤ 2G , and kgℓ − (gℓ )R,δ
σ kC 3 < 2 . Therefore, the second
G

part of (8.69) holds. Combining (8.70) with (8.72) then implies (8.67) as claimed. 
Lemma 8.8. Let K ⊂ G 3 bounded and C1 > 0 as above. Then for g1 , g2 ∈ K, β ∈ (0, δG ), q ∈ N,
and ε > 0 satisfying
kg1 − g2 kC 3 ≤ ε, ε ≤ min(C1−nq β 2ℵ q −4ℵ−1 , δG − β)
we have
Γ ∩ R(n, β, g1 ) ∩ N q,̥ (n, β, g1 ) ⊂ N q,̥ (n, β(1 − εβ −2ℵ C nq q 4ℵ+1 )1/2ℵ , g2 ).
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 63

Proof. Let ρ ∈ Γ ∩ R(n, β, g1 ) ∩ N q,̥ (n, β, g1 ) and (ψ, U ) ∈ A such that B(ρ, CG δG ) ⊂ U . Then,
 
there are I and m such that TI(m) [g1 ](ρ) ∈ (n − 1)̥ − CG β , n̥ + CG β ,
sup d(PI(m) [g1 ](ρ), Γ) < δG ,
m≤n

and
(I − [(d((PI(m) [g1 ])ψ )(ρ)]q )−1
(m)
≤ ( 25 q 2 β −1 )2ℵ . (8.73)
(1 + kd((PI [g1 ])ψ )(ρ)kq )2ℵ−1
Note that (1 + kd((PI(m) [g])ψ )(ρ′ )kq )2ℵ−1 ≤ C1nq for all g ∈ K and ρ′ ∈ DI(m) [g].
Now, by Lemma 6.13, ρ ∈ DI(m) [g2 ]. Therefore, defining
Aq = I − [(d((PI(m) [g1 ])ψ )(ρ)]q , ∆q = [(d((PI(m) [g1 ])ψ )(ρ)]q − [(d((PI(m) [g2 ])ψ )(ρ)]q ,
52ℵ nq −2ℵ 4ℵ
(8.73) yields kA−1
q k≤ C β
22ℵ 1
q . Next, note that
q
X
k∆q k = [(d((PI(m) [g1 ])ψ )(ρ)]q−j [d((PI(m) [g1 ])ψ )(ρ) − (d((PI(m) [g1 ])ψ )](ρ)[(d((PI(m) [g2 ])ψ )(ρ)]j
j=1
n(q+1)
≤ qC1nq kPI(m) [g1 ] − PI(m) [g2 ]kC 1 ≤ qC1 ε,
where we take the C 1 norm on the intersection of the two domains and the last inequality follows
from (6.11). In particular,
(I − [(d((PI(m) [g2 ])ψ )(ρ)]q )−1 kA−1
q k
(m)
≤ k(Aq + ∆q )−1 k ≤ n(q+1)
(1 + kd((PI [g2 ])ψ )(ρ)kq )2ℵ−1 1 − kA−1
q kqεC1
n(q+1) 1 22ℵ −nq 2ℵ −4ℵ
which yields the desired bound since qεC1 ≤ 2 52ℵ C1 β q ≤ 12 kA−1 −1
q k .
Finally, we also note that
(n − 1)̥ − CG δG ≤(n − 1)̥ − CG (β + ε) ≤ TI(m) [g2 ](ρ) ≤ n̥ + CG (β + ε)≤ n̥ + CG δG ,
and hence, ρ ∈ N q,̥ (n, β(1 − εβ −2ℵ C1nq q 4ℵ+1 )1/2ℵ , g2 ) as claimed. 

8.2.4. Step 3: Controlling the volume of bad perturbations. This step is dedicated to finishing the
induction argument that yields the proof of Proposition 8.3. We still assume the conclusions of
Proposition 8.3 hold up to some ℓ ≥ ⌈loga 2⌉ − 1, and will show they also hold up to ℓ + 1.
In what follows, we work with R := Rℓ+1 , δ := δℓ+1 , and for 0 ≤ i ≤ ℓ + 1 set
ℓ+1 ) dim sp(2ℵ)
βi,ℓ+1 = 1
2 min υn (ℓ), υn (ℓ) := 2βℓ+1,ℓ+1 ε(n−a . (8.74)
ai−1 <n≤ai

The first step is to show that for (g, σ) ∈ K × Σℓ (Rε ) such that Fℓ (g, σ) ∈ G ν \LT,ℓ ,
ℓ+1
n o a[
σℓ+1 ∈ Σ(Rℓ+1 ) : Fℓ+1 (g, (σ, σℓ+1 )) ∈ LT,ℓ+1 ⊂ SFR,δ
ℓ (g,σ̂ℓ )
(n, 12 αℓ+1 , υn (ℓ)) (8.75)
n=0

for SFR,δ
ℓ (g,σ̂ℓ )
as defined in Corollary 7.7.
64 YAIZA CANZANI AND JEFFREY GALKOWSKI

To prove (8.75), first note that if σℓ+1 is such that

gℓ+1 := Fℓ+1 (g, (σ, σℓ+1 )) = QR,δ (Fℓ (g, σ), σℓ+1 ) ∈ LT,ℓ+1 ,
by the definition of LT,ℓ+1 , then either (8.25) or (8.26) does not hold with (g, ℓ) = (gℓ+1 , ℓ + 1).
We claim that in this setting, one actually has that (8.26) does not hold with (g, ℓ, j) = (gℓ+1 , ℓ +
1, ℓ + 1). That is, there exist 0 ≤ i ≤ ℓ + 1, ai−1 < n ≤ ai , and 1 ≤ q ≤ a(ℓ+1)/b−i+1 , such that
Γ ∩ R̥ (n, βi,ℓ+1,ℓ+1 q −3 , gℓ+1 ) ∩ S̥ (n, αℓ+1,ℓ+1 , gℓ+1 ) * N q,̥ (n, βi,ℓ+1,ℓ+1 , gℓ+1 ). (8.76)

Indeed, suppose (8.25) does not hold with (g, ℓ) = (gℓ+1 , ℓ + 1). Then, since Fℓ (g, σ) ∈ G ν \LT,ℓ ,
by Lemma 8.7 equation (8.51), (8.25) can only fail when i = ℓ + 1. Thus, there are aℓ < n ≤ aℓ+1
and
ρ ∈ Γ ∩ R̥ (n, β̃ℓ+1,ℓ+1 , gℓ+1 )\N ̥ (n, β̃ℓ+1,ℓ+1 , gℓ+1 ).
Since β̃ℓ+1,ℓ+1 ≤ αℓ+1,ℓ+1 = 21 αℓ+1 , Lemma 8.7 equation (8.52) yields that ρ ∈ S̥ (n, αℓ+1,ℓ+1 , gℓ+1 ).
Furthermore, since β̃ℓ+1,ℓ+1 ≤ 2βℓ+1,ℓ+1 = βℓ+1,ℓ+1,ℓ+1 , we also have ρ ∈ Γ∩R̥ (n, βℓ+1,ℓ+1,ℓ+1 , gℓ+1 ).
Using that f(βℓ+1,ℓ+1,ℓ+1 , 1) ≤ 1/β̃ℓ+1,ℓ+1 , this implies (8.76) for q = 1 and i = ℓ + 1 since
N 1,̥ (n, βℓ+1,ℓ+1,ℓ+1 , gℓ+1 ) ⊂ N ̥ (n, β̃ℓ+1,ℓ+1 , gℓ+1 ).
For any σℓ+1 such that gℓ+1 = Fℓ+1 (g, (σ, σℓ+1 )) ∈ L̥,ℓ+1 , we have showed that (8.26) does not
hold with (g, ℓ) = (gℓ+1 , ℓ + 1). Since Lemma 8.7 equation (8.53) yields that (8.26) holds for all
j ≤ ℓ, we conclude the claim in (8.76). In particular, since βi,ℓ+1,ℓ+1 ≤ υn (ℓ) for all i ≤ ℓ + 1 and
ai−1 < n ≤ ai , we conclude there exists n ≤ aℓ+1 such that σℓ+1 ∈ SFR,δ ℓ (g,σ̂ℓ ),δ
(n, αℓ+1,ℓ+1 , υℓ (n)).
This yields the claim in (8.75).
Next, we apply Corollary 7.7 with R = Rℓ+1 (ε), δ = δℓ+1 (ε), α = 12 αℓ+1 , and β = υn (ℓ).
Note that since n ≤ aℓ+1 one may choose ε0 so that the conditions (7.19) on R, δ are satisfied.
Pℓ+1 ℓ+1−j aj/b ℓ+1
Furthermore, since βℓ+1,ℓ+1 = εd j=1 c c
β0,0 , one can check that there is d large enough
(depending only on a, b, ℵ) such that
∞
X
L−2ℵ−1 −L n(L−2ℵ−1)
C0 R δ C0 υn (ℓ) < ε2−ℓ−2 , (8.77)
n=0

and c ≥ 2ℵ(mν L + y′ (1 − dim sp(2ℵ))). To see this, we use estimate the left hand side of (8.77)
using the definitions of βℓ,ℓ (8.21), υn (8.74), Rℓ , and δℓ (8.29), αℓ (8.22) and γℓ (8.19) to obtain
that there is C > 0 such that
∞
X
L−2ℵ−1 −L n(L−2ℵ−1) c+2ℵ(dim Sp(2ℵ)−1)y′ −mν L (d−C)aℓ/b Cℓ Cℓ/b
C0 R δ C0 υn (ℓ) ≤ Cβℓ,ℓ ε 2 a .
n=0

Choosing d large enough the estimate (8.77) follows provided c ≥ 2ℵ(mν L+y′ (1−dim sp(2ℵ))).
y′
One can also check that β = υn (ℓ) ≤ cn Rℓ+1 provided c ≥ 2ℵy′ and hence, we may apply
Corollary 7.7. Optimizing in y′ ≥ y we obtain c = max( 2ℵm νL
L−2ℵ , 2ℵy).
Corollary 7.7 together with (8.77) and (8.75) then yield that (8.24) holds up to index ℓ + 1 in
place of ℓ.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 65

We next show that (8.23) is a consequence of (8.24). Given σ ∈ Σℓ+1 (R), if

∃j ≤ ℓ + 1 s.t. Fj (g, σ̂j ) ∈ L̥,j ,
then either there exists 1 ≤ j ≤ ℓ + 1 such that Fj (g, σ̂j ) ∈ L̥,j and Fj−1 (g, σ̂j−1 ) ∈ G ν \L̥,j−1
or F0 (g, σ̂0 ) ∈ L̥,0 . Therefore, by Fubini’s theorem and the induction hypothesis (8.24)
  Xℓ+1
mΣ (R)
σ ∈ Σ ℓ+1 (R) : ∃j ≤ ℓ + 1 s.t. Fj (g, b
σ j ) ∈ L ̥,j ≤ ε2−j−1 ≤ (1 − 2−(ℓ+2) )ε.
ℓ+1
j=0
Now, notice that
 
mΣℓ (R) σ ∈ Σℓ (R) : ∃j ≤ ℓ + 1 s.t. Fj (g, σ
bj ) ∈ L̥,j =
 
mΣ∞ (R) σ ∈ Σ∞ (R) : ∃j ≤ ℓ s.t. Fj (g, σj ) ∈ L̥,j
b
and hence (8.23) holds. This finishes the proof of Proposition 8.3.

8.3. Predominance of quantitative non-degeneracy. We now turn to the proof that (8.1)
holds for a predominant set of g. The next proposition is the key technical result of the article. We
Rε ,δε
will find sequences δε and Rε such that (8.1) holds for an F := {(F∞ , ∞)}ε predominant set
Rε ,δε
of metrics (with F∞ as in Section 8.1 built using any family perturbations QR,δ that are good
in the sense of Definition 7.2). More precisely, providedS we are able find a collection {(Γ, G)}G∈G
of (̥, b, y) admissible pairs for {QR,δ }R,δ such that G∈G G = G ν , for each ε > 0 and K ⊂ G ν
bounded we will find a C > 0 so that we can control the measure of the set of bad parameter
values σ ∈ Σ∞ (R) such that F∞ (g, σ) does not satisfy (8.1) with that C and any g ∈ K. This
result will be used in Section 11 to prove Theorem 1.3.
Proposition 8.9. Let ν ≥ 5, y ≥ 1, b > 0, N : (0, 1) → N, and {QR,δ }R,δ be a (ν, N )-good family
of perturbations. For 0 < ε < 1, there are δε = {δj (ε)}∞ ∞
j=0 , Rε := {Rj (ε)}j=0 , such that
∞
X
δj (ε)Rj (ε)−ϑν ≤ ε,
j=0

and, for all (Γ, G) an (̥, b, y)-admissible pair for {QR,δ }R,δ , K ⊂ G bounded, there is ε0 > 0 and
C > 0 such that, for all 0 < ε < ε0 and g ∈ K there exists Borel set Sg,ε ⊂ Σ∞ (Rε ) such that


sup mΣ∞ (R) Sg,ε ≤ ε,
g∈K
and for all g ∈ K 
σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ L∞ (ε) ⊂ Sg,ε
Rε ,δε
where, F∞ is defined in Lemma 8.1,
n o
L∞ (ε) = g ∈ G ν : ∃n such that Γ ∩ R̥ (n, βn (ε), g) * N ̥ (n, βn (ε), g) , (8.78)
and
γ γ −1 γ
βn (ε) := εCn C −n n−C log ε n , (8.79)
 2ℵ(2ℵ+3)

with γ := 1 + log2 max 2ℵ+1 mν , 2ℵy and mν = max(b, ϑν , ϑ5 + 1, 2ϑ̃4 , 1). In addition,
L∞ (ε) is Borel.
66 YAIZA CANZANI AND JEFFREY GALKOWSKI

Proof. From Proposition 8.3, for all 1 < a < 2 there are {δj (ε)}∞ ∞
j=0 and {Rj (ε)}j=0 such that
(8.5), (8.23) and (8.20) hold. Let (Γ, G) satisfying Definition 7.3 and K ⊂ G bounded. Suppose
that g ∈ K, and σ ∈ Σ∞ (R). Then, by Lemma 8.1, F∞ (g, σ) ∈ G ν is well defined. We claim
that the statement holds with
∞
[
Sg,ε := Sg,ε (ℓ), Sg,ε (ℓ) := {σ ∈ Σ∞ (R) : ∃j ≤ ℓ s.t. Fj (g, σ
bj ) ∈ L̥,j },
ℓ=1
with L̥,j as defined in Proposition 8.3. Note that by Proposition 8.3, there exists ε0 > 0 such that


for all 0 < ε < ε0 and g ∈ K, mΣ∞ (R) Sg,ε (ℓ) ≤ (1 − 2−ℓ−1 )ε, and so, since Sg,ε (ℓ + 1) ⊂ Sg,ε (ℓ),


for all 0 < ε < ε0 and g ∈ K, mΣ∞ (R) Sg,ε ≤ ε, as claimed.

Next, we claim that σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ L∞ (ε) ⊂ Sg,ε . For σ ∈ Σ∞ (R) \ Sg,ε we
/ L∞ (ε). Writing gℓ = Fℓ (g, σ̂ℓ ), we have that gℓ ∈ G ν \L̥,ℓ for all ℓ.
will show that F∞Rε ,δε (g, σ) ∈
In particular, for all ℓ, 0 ≤ i ≤ ℓ, and ai−1 < n ≤ ai ,
Γ ∩ R̥ (n, β̃i,ℓ , gℓ ) ⊂ N ̥ (n, β̃i,ℓ , gℓ ).
1+2−ℓ
By definition, β̃i,ℓ = 1+2−i
β̃i,i . Therefore, for ℓ, 0 ≤ i ≤ ℓ, and ai−1 < n ≤ ai , we have
Γ ∩ R̥ (n, 12 β̃i,i , gℓ ) ⊂ N ̥ (n, 21 β̃i,i , gℓ ).

Now, by (8.20), we have kgℓ − gℓ−1 kC ν ≤ C2−ℓ−1 ε. Therefore, gℓ → F∞ (g, σ) =: g∞ ∈ C ν with

kg∞ − gkC ν ≤ ε.
Fix ai−1 < n ≤ ai . We claim that
Γ ∩ R̥ (n, 41 β̃i,i , g∞ ) ⊂ N ̥ (n, 14 β̃i,i , g∞ ). (8.80)
To see this, suppose that ρ ∈ Γ ∩ R̥ (n, 14 β̃i,i , g∞ ). Then, for ℓ large enough,
ρ ∈ Γ ∩ R̥ (n, 21 β̃i,i , gℓ ) ⊂ N ̥ (n, 21 β̃i,i , gℓ ) ⊂ N ̥ (n, βn (ε), g∞ ).
In particular, ρ ∈ N ̥ (n, 14 β̃i,i , g∞ ).
Finally, we choose 1 < a < 2 (and hence find {Rε } and {δε }) such that (8.80) implies that for
all n ≥ 1,
Γ ∩ R̥ (n, βn (ε), g∞ ) ⊂ N ̥ (n, βn (ε), g∞ ), (8.81)
/ L∞ . To check (8.81), we need only show that for aℓ−1 < n ≤ aℓ , we have
and hence that g∞ ∈
1
4 β̃ℓ,ℓ ≥ βn (ε).
ℓ−1
Pℓ−1 1
−1 a b )j 1 log a
2
c εdcℓ
j=0 (c
− 1−log
By definition, βℓ,ℓ = β0,0 . In addition, lima→2− a− b = 2 2a = 0. There-
fore, we may choose a so that
ca−1/b = 1, (8.82)


where c = max 2ℵ(2ℵ+3)
2ℵ+1 mν , 2ℵy > 1 and mν = max(b, ϑν , ϑ5 + 1, 2ϑ̃4 , 1). Therefore,
ℓ ℓ−1 ℓ
β̃ℓ,ℓ = ε2ℵ−1 βℓ,ℓ
2ℵ 2ℵc 2ℵdc
= β0,0 ε ε2ℵ−1
ℓ
2ℵc 2ℵdc ℓ−1 ℓ+2ℵ−1
≥ β0,0 ε .
log2 c log2 c
Moreover, (8.82) implies log2 a = 1+log2 c and, in particular, log2 a = 1 + log2 c = γ.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 67

Notice that loga n < ℓ ≤ 1 + loga n since aℓ−1 < n ≤ aℓ and hence cℓ ≤ cnγ which, using that
β0,0 (ε) ≥ cεN for some c > 0 and N ≥ 0, yields, for C > 0 large enough (in the definition of
βn (ε)) depending on c, ℵ, and d, yields
log2 ε−1 γ
γ −2ℵdγ n γ +2ℵ−1
β̃ℓ,ℓ ≥ [β0,0 (ε)]2ℵcn n log 2 c
ε2ℵdn ≥ 4βn (ε).


9. Nice perturbations in the space of metrics

In this section, we define families of perturbations that satisfy Definition 7.2 and 7.3. The
basic perturbation takes place on a small ball through which a geodesic passes exactly once and is
inspired by those in [Ano82, Kli78]. We then build a large family of these perturbations in order
to be able to handle any possible geodesic.

9.1. The elementary family of perturbations. Let ν ≥ 0. We fix a reference metric, gf ∈

]
G ν+2 . Let ρ0 ∈ S ∗M and e , . . . , e ∈ T
1 d πM (ρ0 ) M an orthonormal set of vectors for gf .
For g⋆ ∈ G ν , consider the geodesic
γρg0⋆ (t) := πM (ϕgt ⋆ (ρ0 )).
We choose a basis E1g⋆ , . . . , Edg⋆ ∈ Tγρg⋆ (0) M , orthonormal with respect to g⋆ by applying Gram-
0
Schmidt process to (γ̇ρ0 (0), e1 , . . . , ed ) so that, in particular, Edg⋆ = γ̇ρ0 (0). (Here, we only use
ed if γ̇ρ0 (0), e1 , . . . , ed−1 are linearly dependent.) Let Eig⋆ (t) ∈ Tγρg⋆ (t) M be the parallel transport
0
(again with respect to g⋆ ) of Eig⋆ along γρg0⋆ . For εf > 0, define the map

g
X
d−1 
Φgρ⋆0 : BRd−1 (0, εf ) × R → M, Φgρ⋆0 (u, t) = expγfg⋆ (t) ui Eig⋆ (t) , (9.1)
ρ0
i=1
gf
where expγρ (t) denotes the exponential map for the metric gf with base point γρ0 (t). Then for
0
εf > 0 small enough depending only on gf , and any small enough (depending on g⋆ and gf )
interval I ⊂ R, Φgρ⋆0 |I×B(0,εf ) is a diffeomorphsim onto its image.
The reason for using gf above rather than g⋆ itself is that, in our perturbation argument, g⋆
′
will vary, remaining bounded in G ν , but not necessarily in G ν for ν ′ > ν. It will therefore be
g
important for us to control the C ν norms of Φρ⋆0 using only the G ν norms of g⋆ (See Lemma 9.4).
This is not possible if one replaces gf by g⋆ .
Remark 9.1. Although Φgρ⋆0 is not a global diffeomorphism, it is locally a diffeomorphism, and we
will often ignore this fact in order to simplify the presentation. When defining our perturbations
below, we will work in a (u, t) neighborhood where the map is a genuine diffeomorphism.
Lemma 9.2. In the (u, t) coordinates g⋆ satisfies
g⋆ij (0, t) = δij , g⋆dd (0, t) = 1, g⋆jd (0, t) = 0, ∂uj g⋆dd (0, t) = 0.
where, here and below, we write i, j, k, ℓ for indices in {1, . . . , d − 1}.
68 YAIZA CANZANI AND JEFFREY GALKOWSKI

Proof. The first three equalities follow from the fact that {Eig⋆ (t)}di=1 is an orthonormal frame
with respect to g⋆ . Since (0, t) is a geodesics for g⋆ , letting x(s) = (0, s) ∈ Rd−1
u × Rt , we have
1
∂s2 xβ + g⋆µβ (∂α g⋆,µν + ∂ν g⋆,µα − ∂µ g⋆,αν )∂s xα ∂s xν = 0.
2
Since xi = ∂s xi = ∂s2 xi = 0 for i = 1, . . . , d − 1, ∂s xd = 1, and ∂s2 xd = 0, we have that for
i = 1, . . . , d − 1
1 µi 1
g⋆ (∂d g⋆,µd + ∂d g⋆,µd − ∂µ g⋆,dd ) = − ∂ui g⋆,dd = 0.
2 2
dd
Since g⋆,ij (0, t) is the identity, ∂ui g⋆ = 0. 
Lemma 9.3. The map F : G ν → C ν−1 (Rd ; M ) given by F (g) := Φgρ0 is Frechet differentiable and,
moreover, for any bounded subset G ⊂ G ν there is C > 0 such that for g ∈ G,
kDF |g δg kC ν−1 ≤ Ckδg kC ν−1 .

Proof. To prove the lemma, we work locally in a coordinate chart so that we may identify vectors
along a curve with Rd .
Define the map E0 : G ν → (Rd )d by E0 (g) := (E1g (0), . . . , Edg (0)) with Eig defined as above.
Let E : G ν × C ν+1 ([0, 1]; Rd ) × (Rd )d → C ν ([0, 1]; (Rd )d ) be the map E(g, γ, (ei )di=1 ) := (egi )di=1 ,
where egi ∈ C ν ([0, 1]; Rd ) is a parallel frame along γ (with respect to g) with initial conditions
(egi (0))di=1 = (ei )di=1 .
g
In addition, define expgf : Rd × Rd → Rd by expgf (x, v) := expxf (v) and let Ψ : S ]∗M × G ν →
g g
C ν+1 ([0, 1]; Rd ) be the map Ψ(ρ, g) := γρ where γρ is the unit speed geodesic for the metric g
with initial conditions given by ρ. Finally, let Y : C ν+1 ([0, 1]; Rd ) × C ν ([0, 1]; (Rd )d ) → C ν ([0, 1] ×
Rd−1 ; Rd ),
  gf
X
d−1 
Y (γ, (Ei )di=1 ) (t, u) = expγ(t) ui Ei (t) .
i=1
We are interested in studying their composition


F (g) = Y Ψ(ρ0 , g) , E g, Ψ(ρ0 , g), E0 (g) = Φgρ0 .

To clarify the exposition, in what follows we write S ν for the space of symmetric tensors
endowed with the C ν topology.
Note that g 7→ Dg E0 is bounded and continuous in the G ν topology. By writing the ordinary
differential equation satisfied by a parallel vector field and using that γ̇ ∈ C ν−1 ([0, 1]; Rd ), ∂x g ∈
S ν−1 , one can check that Dg E : S ν−1 → C ν−1 ([0, 1]; (Rd )d ) and the map (g, γ) 7→ Dg E|(g,γ) is
continuous in the G ν × C ν topology with codomain the set of bounded linear maps from S ν−1 to
C ν−1 ([0, 1] × (Rd )d ).
Next, again using the parallel transport equation and γ̇ ∈ C ν−1 , g ∈ G ν , the map Dγ E :
S ν−1 → C ν−1 ([0, 1]; (Rd )d ) is bounded and the map (g, γ) 7→ Dγ E|(g,γ) continuous in the G ν × C ν
topology with codomain the set of bounded linear maps from S ν−1 to C ν−1 ([0, 1]; (Rd )d ).
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 69

Furthermore, it is easy to see that

D(γ,E) Y : C ν ([0, 1]; Rd ) × C ν−1 ([0, 1]; (Rd )d ) → C ν−1 ([0, 1] × Rd−1 ; Rd )

is bounded and (γ, E) 7→ D(γ,E) Y is continuous in the C ν ([0, 1]; Rd ) × C ν−1 ([0, 1]; (Rd )d ) topol-
ogy with codomain the set of bounded linear maps from C ν ([0, 1]; Rd ) × C ν−1 ([0, 1]; (Rd )d ) to
C ν−1 ([0, 1] × Rd−1 ; Rd ).
Finally, using the geodesic equation together with g ∈ G ν and that geodesics for a G ν metric lie
in C ν+1 , we have Dg Ψ : S ν−1 → C ν ([0, 1]; Rd ) is bounded and the map g0 7→ Dg Ψ|g0 is continuous
as a function of g0 in the G ν topology with with codomain the set of bounded linear maps from
S ν−1 to C ν ([0, 1]; Rd ).
With these estimates in place it is now easy to see that Dg F : S ν−1 → C ν−1 ([0, 1] × Rd−1 ; Rd )
is bounded and g0 7→ Dg F |g0 is continuous in the G ν topology with with codomain the set of
bounded linear maps from S ν−1 to C ν−1 ([0, 1] × Rd−1 ; Rd ). 
Lemma 9.4. Let ν ≥ 2 and K ⊂ G ν be bounded. Then there is C > 0 such that for all g⋆ ∈ K
and ρ0 ∈ Sg
∗M ,

kΦgρ⋆0 kC ν ≤ C.

Proof. First, recall that in any coordinate system x on M with dual coordinates ξ, if ρ0 =
(0, (0, . . . , 0, 1)), then γρ0 (s) is given by the solution x(s) to

α g⋆αβ ξβ ∂xα g⋆µβ ξµ ξβ

ẋ = , ξ̇α = − , (x(0), ξ(0)) = (0, (0, . . . , 0, 1)).
|ξ|g |ξ|g
Therefore, standard regularity theory for ODEs shows that γρ0 (t) ∈ C ν+1 (R; M ) and there exists
CK > 0 with kγρ0 kC ν+1 ≤ CK . In particular, kγ̇ρ0 kC ν ≤ CK . Next, recall that for each i the
equations for parallel transport take the form Ei (t) = eα (t)∂xα where eα (t) solves an equation
of the form ėα (t) = Γe(t)γ̇ρg0⋆ (t) with Γ representing the Christoffel symbol for g. In particular,
keα kC ν ≤ CK and hence kEi kC ν ≤ CK .
g
Now, since gf ∈ G ν+2 is independent of K, the exponential map (p, V ) 7→ exppf (V ) is in
C ν (T M ; M ) with bounds independent of K. Therefore, Φgρ⋆0 satisfies the desired estimates. 

For s ∈ R and ε > 0 small let

Zs := Φgρ⋆0 (B(0, εf ) × {s}), Hs := S^
∗ M ∩B
Zs ]
∗
(ϕgs⋆ (ρ0 ), ε). (9.2)
S M

]
Here, we use the metric gf to define the ball in S ∗M .

Note that Hs is a smooth hypersurface such that ρ0 ∈ H0 , and Hs is transverse to H|ξ|g⋆ ; in

particular, satisfying,
1
|H|ξ|g⋆ (ρ)f | ≥ |df |, for f defining Hs and all ρ ∈ Hs .
2
Despite the fact that we define Hs as a subset of S ]∗M , it will be convenient in the rest of this
]
section to identify S ∗
∗M with S M for each g i.e. the unit sphere bundle with respect to the metric
g
70 YAIZA CANZANI AND JEFFREY GALKOWSKI

]
g. To do this, we define the canonical isomorphism from S ∗M → S ∗ M to by (x, [ξ]) 7→ (x, ξ ).
g |ξ|g
 ξ 
] ∗
Here, we have used that [ξ] = |ξ|g in S M .
We now define a family of perturbations gσ ∈ G 4 of a metric g0 ∈ G 4 close to g⋆ as follows: Let
Σ1 := Rd−1 × Rd−1 , Σ2 := Sym(d − 1) × M(d − 1) × Sym(d − 1), Σ := Σ1 × Σ2
and denote the elements of Σ by


σ := (σ1 , σ2 ) = (A, B), (C, D, E) ∈ Σ. (9.3)
The closeness of g0 to g⋆ will be crucial in Lemma 9.12.
]
Then, for g0 ∈ G g , ρ0 ∈ S ∗M , σ ∈ Σ, R > 0, and t ∈ R, the family of perturbations
⋆

gσ = gσ (ρ0 , t⋆ , R, g0 , σ)
is defined in the (u, t) coordinates introduced in (9.1) as follows


gσdd (u, t) = g0dd (u, t)− hA, ui + 12 hCu, ui χR (u, t − t⋆ ),


gσjd (u, t) = g0jd (u, t) + 21 Bj + (Du)j χR (u, t − t⋆ ), (9.4)
gσij (u, t) = g0ij (u, t)
+ 21 Eij χR (u, t − t⋆ ),
√ √
χR (u, t) = R1 χ( R1 t)χ( R1 |u|) and χ ∈ Cc∞ ((− 2, 2); [0, 1]) even with χ ≡ 1 on [− 14 , 41 ] and χ = 1.
´

In order to identify gσ with a metric on M , we let K ⊂ G 3 be bounded and work with g⋆ ∈ K. In

addition, we require that 0 < R < R0 where R0 = R0 (K) is chosen such that for all 0 < R < R0 ,
g⋆ ∈ K, and t ∈ R, we have
 √ √ √

Φgρ⋆0 BRd−1 (0, 2R) × [t − 2R, t + 2R] ⊂ Bg⋆ (γρg0⋆ (t), 3R) (9.5)

with Φgρ⋆0 as in (9.1). This is possible since, by Lemma 9.2, g⋆ij (0, t) = δij , for all t ∈ R and
i, j = 1, . . . , d. In addition, will fix T0 > 0 below and require that
 
(Φgρ⋆0 )−1 (B(γρg0⋆ (t⋆ ), 3R)) ∩ BRd−1 (0, εf ) × [0, T0 ] ⊂ BRd−1 (0, εf ) × [t⋆ − 3R, T⋆ + 3R].

That is, the geodesic γρg0⋆ passes near γρg0⋆ (t⋆ ) only once between time 0 and T0 .
Remark 9.5. Observe that χR is chosen so that the integral along (0, t) is 1.

Define the map ζ : BRd−1 (0, εf ) × BRd−1 (0, εf ) × G 4 × Σ → H0

ζ(z0 , g0 , σ) := (u0 , 0, ω0 , τ0 (z0 , g0 , σ)), z0 := (u0 , ω0 ) (9.6)
where τ0 > 0 is defined so that |(ω0 , τ0 )|gσ (u0 ,0) = 1 and ζ(z0 , g0 , σ) ∈ H0 . In particular,
ρ0 = ζ(0, g0 , 0) = (0, 0, 0, 1) ∈ H0 . (9.7)
Let u, t, ω, τ be functions defined on R × BRd−1 (0, εf ) × BRd−1 (0, εf ) × G4 × Σ such that


ϕgsσ (ζ(z0 , g0 , σ)) = u, t, ω, τ (s, z0 , g0 , σ). (9.8)
Note that by the definition of ζ, for all s and σ
|(ω, τ )|gσ (u,t) = 1.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 71

Also, t(s, 0, g⋆ , σ) = s. Next, define the matrices

 
∂u0 u(s, z0 , g0 , σ) ∂ω0 u(s, z0 , g0 , σ)
∆σ (s, z0 , g0 ) := ,
∂u0 ω(s, z0 , g0 , σ) ∂ω0 ω(s, z0 , g0 , σ)
  (9.9)
−1 ∂s u


˜
∆σ := ∆σ − [∂s t] ∂u0 t ∂ω0 t .
∂s ω
We note that ∆˜ σ ∈ Sp(2(d − 1)). Indeed, for each fixed s, observe that ∆
˜ σ is the derivative of the
following Poincaré map (a symplectomorphism)
Pσ : H0 → Ht(s,z0 ,g0,σ) , Pσ (z0 ) = (u(S, z0 , g0 , σ), ω(S, z0 , g0 , σ)),
where S = S(s, z0 , g0 , σ) is defined such that ϕgSσ (ζ(z0 , σ)) ∈ Ht(s,z0,g0 ,σ) . Here, we are identifying
H0 with B(0, εf ) × B(0, εf ) via z0 = (u0 , ω0 ) 7→ ζ(z0 , σ). Note also that the symplectic form
induced on Hs from T ∗ M is the standard symplectic form in (u, ω) coordinates.
Let T0 ∈ R, g0 ∈ G 4 , and z0 ∈ BRd−1 (0, εf ) × BRd−1 (0, εf ). Define the map Ψgz 0 : Σ →
0
Rd−1 × Rd−1 × Sp(2(d − 1)) by
Ψzg00 ,T0 (σ) := (Ξ1 (z0 , g0 , σ), Ξ2 (z0 , g0 , σ)) (9.10)


Ξ1 (z0 , g0 , σ) := u(T, z0 , g0 , σ), ω(T, z0 , g0 , σ) , ˜ σ∆
Ξ2 (z0 , g0 , σ) := ∆ ˜ −1 (T, z0 , g0 ),
0
where T = T (z0 , g0 , σ, T0 ) is defined by
ϕgTσ(z (ζ(z0 , g0 , σ)) ∈ HT0 .
0 ,g0 ,σ,T0 )

Note that T (0, g⋆ , 0, T0 ) = T0 , i.e., ϕgT⋆ (ρ0 ) ∈ HT0 .

0

Remark 9.6. The choice of ∆ ˜ σ∆ ˜ −1 as the last entry of Ψgz00 (σ) is motivated by the fact that
0
we intend to write an ODE (in the s variable) which ∂δ ∆ ˜ δσ |δ=0 ∆
˜ −1 (s) solves. Because ∆
˜ δσ (s) ∈
0
˜
Sp(2d − 1), ∂δ ∆δσ (s)|δ=0 ∈ T∆ ˜ ˜ −1
˜ 0 (s) Sp(2d − 1), while ∂δ ∆δσ (s)|δ=0 ∆0 ∈ TI Sp(2d − 1) = sp(2d − 1).
Because this is a linear subspace of M(2(d − 1)) which is independent of s, ODEs posed in this
space are much simpler to work with.

For the purposes of the next calculations, we take g0 = g⋆ . In Lemma 9.10 we show how to
handle g0 6= g⋆ . Note that, with g0 = g⋆ , we have (u, t, ω, τ )|σ=0 = (0, s, 0, 1) when z0 = (0, 0).
Let
uσ (s) := u(s, 0, g⋆ , σ), ωσ (s) = ω(s, 0, g⋆ , σ), tσ (s) = t(s, 0, g⋆ , σ). (9.11)


We continue to write σ = A, B, C, D, E and define
 
∂uk gσdj gσjk
Lσ := ∈ sp(2(d − 1)).
− 21 ∂u2j uk gσdd −∂uj gσdk

Next, we calculate how derivatives of ∆δσ and Ξ1 (z0 , g0 , δσ) with respect to δ behave.
Lemma 9.7. Let K ⊂ G 3 bounded. Then there is R0 > 0 such that for all g⋆ ∈ K the following
holds. Let ρ0 ∈ Sg
∗M , T ∈ R, 0 < R < R and t ∈ [3R, T − 3R] be such that the set
0 0 ⋆ 0

{t ∈ [0, T0 ] : γρg0⋆ (t) ∩ B(γρg0⋆ (t⋆ ), 3R)} (9.12)

is connected.
72 YAIZA CANZANI AND JEFFREY GALKOWSKI

HT0

H0

ϕgTσ(z0 ) (ζ)

ϕgT⋆0 (ρ0 )
ζ
ρ0

Figure 4. The figure shows the setup for the definition of the functions t and T .
Here, we abbreviate ζ = ζ(z0 , g0 , σ) and T (z0 ) = T (z0 , g0 , σ). The hypersurface
H0 parametrized by z0 = (u0 , ω0 ) via z0 7→ ζ(z0 , σ), while the corresponding points
in HT0 are (u, ω) = Ξ1 (z0 , g0 , σ).

Let g0 = g⋆ , σ = A, B, C, D, E ∈ Σ and δ > 0. Then, with uσ and ωσ as defined in (9.11),
for s ∈ [0, T0 ]      
∂δ uδσ |δ=0 ∂δ uδσ |δ=0 1 B
∂s = L0 +2 χR (0, s − t⋆ ), (9.13)
∂δ ωδσ |δ=0 ∂δ ωδσ |δ=0 A
and
 
−1

1 Dj E jk
∂s ∂δ ∆δσ δ=0 ∆0 = 2 k χR (0, s − t⋆ ) + [L0 , ∂δ ∆δσ δ=0 ∆−1
0 ] + FR (A, B, s), (9.14)
Cjk −Djk
where Lσ is given by (9.18), FR (A, B, s) ∈ sp(2(d − 1)) and
|FR (A, B, s)| ≤ CeC|s−t⋆ | k(A, B)k
with C = C(kg⋆ kC 3 ).

Proof. There is R0 = R0 (K) > 0 such that (9.5) holds and we can work with coordinates (u, t)
on the ball B(γρg0⋆ (t⋆ ), 3R0 ). Moreover, by the connectedness of the set in (9.12)
 
(Φgρ⋆0 )−1 (B(γρg0⋆ (t⋆ ), 3R0 )) ∩ BRd−1 (0, εf ) × [0, T0 ] ⊂ BRd−1 (0, εf ) × [t⋆ − 3R0 , T⋆ + 3R0 ].

Since the support of the perturbation is inside B(γρg0⋆ (t⋆ ), 3R0 ), this allows us to identify the
perturbation gσ with a genuine metric perturbation on M and to treat (u, t) as though they were
global coordinates for the purposes of the calculations in this lemma. By (9.8), the Hamiltonian
flow, ϕgsσ , for |ξ|gσ is defined by
∂s uj = gσij ωi + gσdj τ , ∂s ωj = − 21 ∂uj gσiℓ ωi ωℓ + ∂uj gσdℓ τ ωℓ + 21 ∂uj gσdd τ 2 ,
(9.15)
∂s t = gσdj ωj + gσdd τ , ∂s τ = − 12 ∂t gσiℓ ωi ωℓ + ∂t gσdℓ τ ωℓ + 21 ∂t gσdd τ 2 ,
where ξ = (ω, τ ).
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 73

Combining the Hamilton equations (9.15) together with (9.4) and Lemma 9.2 yields (9.13).
Here, we are using that ∂t g⋆jd (0, t) = g⋆jd (0, t) = ∂t ∂uj g⋆dd = ∂uj g⋆dd = 0, ∂δ uδσ |s=0 ≡ 0, and
∂δ ωδσ |s=0 ≡ 0. The last two equalities follow from the facts that ∂δ uδ |s=0 = ∂δ (uδ |s=0 ) and
uδ |s=0 ≡ u0 , and the analogue for ωδ .
The same computation yields that at z0 = (0, 0),
∂s ∆0 = L0 ∆0 , ∆0 |s=0 = I. (9.16)
Differentiating in (u0 , ω0 ), then δ, using that ∆σ |s=0 = I, and that ∂δ uδσ |δ=0 and ∂δ ωδσ |δ=0
depend only on (A, B) and satisfy
k(∂δ uδσ |δ=0 (s), ∂δ ωδσ |δ=0 (s)k ≤ CeC|s−t⋆ | , (9.17)
we find there is FR (A, B, s) such that at z0 = (0, 0),
∂s ∂δ ∆δσ |δ=0 = L0 ∂δ ∆δσ |δ=0 + ∂δ Lδσ |δ=0 ∆0 + FR (A, B, s)∆0 , ∂δ ∆δσ |s=0 = 0.
Note that  
1 Dkj E jk
∂δ Lδσ |δ=0 = χR (0, s − t⋆ ). (9.18)
2 Cjk −Djk
Therefore, at z0 = (0, 0), by (9.16) we have (∂δ ∆δσ |δ=0 ∆−1
0 )|s=0 = 0 and conclude (9.14). We
claim that (9.14) implies
FR (A, B, s) ∈ sp(2(d − 1)).
Indeed, this follows from the fact that 1) ∆0 ∈ Sp(2(d − 1)) and 2) ∂δ ∆δσ |δ=0 ∆−1 0 ∈ TI Sp(2(d −
1)) = sp(2(d − 1)). Claim 1) follows from the fact that ∂s (u, ω)|(u0 ,ω0 )=0 = 0 and hence that ∆0 =
∆˜ 0 ∈ Sp(2(d−1)). Claim 2) The claim is not obvious since we do not know that ∆σ ∈ Sp(2(d−1))
for all σ 6= 0. However, using that ∂u0 t|δ=0 = ∂ω0 t|δ=0 = 0 and ∂s u|δ=0 = ∂s ω|δ=0 = 0, we have
˜ δσ |δ=0 , and hence, since ∆
∂δ ∆δσ |δ=0 = ∂δ ∆ ˜ σ ∈ Sp(2(d − 1)), ∂δ (∆T J∆δσ )|δ=0 = 0. In particular,
δσ
this implies that
(∂δ ∆δσ |δ=0 ∆−1 T −1 T −1 T −1
0 ) J∂δ ∆δσ |δ=0 ∆0 = (∆0 ) (∂δ (∆δσ J∆δσ )|δ=0 )∆0 = 0,

and hence that ∂δ ∆δσ |δ=0 ∆−1

0 ∈ sp(2(d − 1)).
The estimate on FR (A, B, s), now follows from (9.17). 

In what follows we prove that ∂σ Ψg0 ⋆ |σ=0 is bijective and bound its inverse. We write σ =
(σ1 , σ2 ) as introduced in (9.3). We note that Ψzg00 ,T0 (σ) is defined using the Hamiltonian flow for
gσ = gσ (ρ0 , t⋆ , R, g0 , σ), where gσ is a perturbation of the metric g0 done in a neighborhood of the
point γρg0⋆ (t⋆ ). Indeed,
√ in the (u, t) coordinates,
√ the perturbation is supported where χR (u, t − t⋆ )
is, i.e. for |u| ≤ 2R and |t − t⋆ | ≤ 2R.
Lemma 9.8. Let K ⊂ G 3 bounded. Then there is C > 0 and R0 > 0 suc that for all g⋆ ∈ K the
following holds. Let ρ0 ∈ Sg
∗M , T ∈ R, 0 < R < R and t ∈ [3R, T − 3R] be such that the set
0 0 ⋆ 0

{t ∈ [0, T0 ] : γρg0⋆ (t) ∩ B(γρg0⋆ (t⋆ ), 3R)} (9.19)

is connected. For 0 < R < R0 and j = 1, 2, let Ξj (σ) := Ξj (0, g⋆ , σ) be the maps Ξj introduced in
(9.10), which are defined using the perturbations gσ = gσ (ρ0 , t⋆ , R, g⋆ , σ) introduced in (9.4).
74 YAIZA CANZANI AND JEFFREY GALKOWSKI

Then, the map ∂σ1 Ξ1 (0)|Σ1 : Σ1 → R2d−2 is bijective with inverse bounded by CeC|T0 −t⋆ | and
∂σ2 Ξ1 (0)|Σ1 = 0. Also, for every fixed σ1 ∈ Σ1 , the map ∂σ2 Ξ2 (σ1 , 0)|Σ2 : Σ2 → sp(2(d − 1))
is bijective with inverse bounded by CeC|T0 −t⋆ | , and the map ∂σ1 Ξ2 (0)|Σ1 : Σ1 → sp(2(d − 1)) is
bounded by CeC|T0 −t⋆ | . In particular,
∂σ Ψg0 ⋆ ,T0 |σ=0 is bijective with inverse bounded by CeC|T0 −t⋆ | .

Proof. First, note that ∂δ Ξ1 (δσ)|δ=0 = (∂δ uδσ (T0 ), ∂δ ωδσ (T0 )) δ=0 since ∂s (u, ω)(T0 , z0 , g⋆ , 0) = 0.
Fix v ∈ R2d−2 . By the connectedness of the set (9.19), and the fact that t⋆ ∈ [3R, T0 − 3R], the
right hand side of the equation (9.13) is the same as that in Lemma A.2. In particular, taking R0
small enough, Lemma A.2 together with (9.13), there are R0 , C1 > 0 depending on kgkC 2 such
that for all R < R0 , there are σ1 = (A, B) ∈ R2d−2 satisfying
(∂δ uδ(σ1 ,0) (T0 ), ∂δ ωδ(σ1 ,0) (T0 )) = v, kσ1 k ≤ C1 eC1 |T0 −t⋆ | kvk.
Thus, the map σ1 7→ (∂δ uδ(σ1 ,0) (T0 ), ∂δ ωδ(σ1 ,0) (T0 )), which equals ∂δ Ξ1 (0)|Σ1 , is bijective with
inverse bounded by C1 eC1 |T0 −t⋆ | .
Checking that ∂δ Ξ1 (δσ)|δ=0 = 0 when σ1 = 0 is straightforward. It follows that ∂σ2 Ξ1 (0)|Σ1 = 0.
Next, since Ξ2 (δσ) = [∆˜ δσ ∆
˜ −1 ](T (z0 , g⋆ , δσ, T0 ), z0 , g⋆ ), by (9.9)
0
 h  
−1 ∂s uδσ

i
∂δ Ξ2 (δσ) = ∂δ ∆δσ − ∂δ [∂s tδσ ] ∂u0 tδσ ∂ω0 tδσ ∆−1 (T0 )
s=T0 0
δ=0 ∂s ωδσ
δ=0
 
−1
= ∂δ ∆δσ δ=0 ∆0 . (9.20)
s=T0

Where to obtain the final equality, we observe that

 

∂s u
∂u0 t ∂ω0 t s=T0 = 0, and = 0.
z0 =0
∂s ω s=T0
z0 =0

We now write Q0 for the solution of

∂s Q0 = [L0 , Q0 ] + FR (A, B, s), Q0 (0) = 0. (9.21)
Note that since L0 (s), FR (A, B, s) ∈ sp(2(d − 1)) we have Q0 (s) ∈ sp(2(d − 1)). Observe that,
working with σ = (σ1 , 0), we have ∂δ Ξ2 (δσ) = ∂σ1 Ξ2 (0)σ1 and so (9.14) yields
δ=0

∂σ1 Ξ2 (0)σ1 = kQ0 (T0 )k ≤ CeC|T0 −t⋆ | kσ1 k,

where C = C(kgkC 3 ), and hence k∂σ1 Ξ2 (0)|Σ1 k ≤ CeC|T0 −t⋆ | , as claimed.
Next, we show that ∂σ2 Ξ2 (0)|Σ2 is invertible. Fix L ∈ sp(2(d − 1)). Then, we use Lemma A.2
to find σ2 = (C, D, E) such that the solution Q1 to
 
1 Dkj E jk
∂s Q1 = [L0 , Q1 ] + χR (0, s − t⋆ ), Q1 (0) = 0 (9.22)
2 Cjk −Djk
satisfies Q1 (T0 ) = L and such that
∂σ2 Ξ2 (0)σ2 = Q1 (T0 ) = L, kσ2 k ≤ CeC|T0 −t⋆ | kLk,
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 75

and hence ∂σ2 Ξ2 (0) is invertible as claimed.

The application Lemma A.2 requires a brief explanation. Observe that, again using the con-
nectedness of (9.19), (9.22) is a differential equation on sp(2(d − 1)) of the form
X
R
∂s Q1 = RQ1 + ai fi,t⋆
i

where R(s) is a linear map on sp(2(d − 1)) with kRkL∞ ≤ CkgkC 2 and the ai are a basis for
sp(2(d − 1)).
Finally, we show that ∂σ Ψg0 ⋆ ,T0 |σ=0 has the required properties. Notice that, as a map from
R2(d−1)× sp(2(d − 1)) to itself,
   
g⋆ ,T0 ∂σ1 Ξ1 ∂σ2 Ξ1 ∂σ1 Ξ1 0
∂σ Ψ0 |σ=0 = = .
∂σ1 Ξ2 ∂σ2 Ξ2 ∂σ1 Ξ2 ∂σ2 Ξ2

Remark 9.9. Here it is crucial that ∂σ1 Ξ2 : R2d−2 → sp(2(d − 1)) so that we may think of the
map ∂σ Ψg0 ⋆ ,T0 |σ=0 acting on R2d−2 × sp(2(d − 1)). This follows from the fact that FR (A, B, s) ∈
sp(2(d − 1)).

In particular, we have
 
g⋆ ,T0 −1 (∂σ1 Ξ1 )−1 0
[∂σ Ψ0 |σ=0 ] = ,
−[∂σ2 Ξ2 ]−1 ∂σ1 Ξ2 (∂σ1 Ξ1 )−1 [∂σ2 Ξ2 ]−1
from which the estimates on [∂σ Ψg0 ⋆ ,T0 |σ=0 ]−1 easily follow. 

Before we prove that the estimates in Lemma 9.8 are stable under small changes of the metric
or initial position, we need to control how much these changes affect ∂σ Ψg0 ⋆ ,T0 |σ=0 . This is done
in our next lemma.
Lemma 9.10. Given K ⊂ G 4 bounded, there is C > 0 such that the following holds.
Let g⋆ ∈ K and R0 be as in Lemma 9.8. Let ρ0 ∈ Sg ∗M , T ∈ R, 0 < R < R , and t ∈
0 0 ⋆
g⋆
[3R, T0 − 3R] be such that the set {t ∈ [0, T0 ] : γρ0 (t) ∩ B(γρg0⋆ (t⋆ ), 3R)} is connected. Then, for
all g0 ∈ G 4 with kg0 − g⋆ kC 4 ≤ 1, z0 ∈ Rd−1 (0, εf ) × BRd−1 (0, εf ), and all R > 0,


kdσ (Ψgz00 ,T0 − Ψg0⋆ ,T0 )|σ=0 kC 0 ≤ C(1 + R−2 ) kg0 − g⋆ kC 1 + |z0 | eC|T0 | , (9.23)
where Ψgz00 is the map introduced in (9.10), defined using the perturbations gσ = gσ (ρ0 , t⋆ , R, g0 , σ)
introduced in (9.4).
Remark 9.11. It is possible to replace R−2 in (9.23) by R−1 through a more careful study of the
terms produced by perturbing g⋆ and z0 . In particular, using the fact that the L1 norm of the
worst terms produced gains a power of R. This would, in turn, lead to a slightly smaller power
Ων in our main theorems: Theorems 1.1, 1.2, and 1.3. However, the form of the power would still
be the same so we do not pursue this here.

Proof. First, observe that kH|ξ|g − H|ξ|g⋆ kC k−1 ≤ kg0 − g⋆ kC k . In particular, defining ρ0 (t) and
0
ρ⋆ (t) such that ∂t ρ0 (t) = H|ξ|g (ρ0 (t)), ∂t ρ⋆ (t) = H|ξ|g⋆ (ρ⋆ (t)), and ρ0 (0) = ρ⋆ (0), we obtain for
0
76 YAIZA CANZANI AND JEFFREY GALKOWSKI

all t ∈ R
ˆ t
|ρ0 (t) − ρ⋆ (t)| ≤ |H|ξ|g (ρ0 (s)) − H|ξ|g⋆ (ρ⋆ (s))|ds
0
0
ˆ t
≤ (kg0 kC 2 + kg⋆ kC 2 )|ρ0 (s) − ρ⋆ (s)|ds + tkg0 − g⋆ kC 1 .
0
Therefore, using Gronwall’s inequality, there is C = C(kg⋆ kC 2 , kg0 kC 2 ) such that
d(ϕgt 0 (ρ), ϕgt ⋆ (ρ)) ≤ Ckg0 − g⋆ kC 1 eC|t| , for all ρ ∈ T ∗ M \ {0}.
In addition, recall that for every metric g there is C = C(kgkC 2 ) such that for all ρ1 , ρ2 ∈ T ∗ M \{0}
d(ϕgt (ρ1 ), ϕgt (ρ2 )) ≤ Cd(ρ1 , ρ2 )eC|t| .
In particular, there is C = C(kg⋆ kC 2 ) such that for kg⋆ − g0 kC 2 ≤ 1, and ρ0 , ρ⋆ ∈ T ∗ M \ {0}, t ∈ R,
d(ϕgt 0 (ρ0 ), ϕgt ⋆ (ρ⋆ )) ≤ C(kg⋆ − g0 kC 1 + d(ρ0 , ρ⋆ ))eC|t| . (9.24)
Next, in analogy with (9.11), for σ ∈ Σ let
uσ (s) := u(s, z0 , g0 , σ), ωσ (s) = ω(s, z0 , g0 , σ), tσ (s) = t(s, z0 , g0 , σ).
and differentiate (9.15) to get
     
∂δ uδσ |δ=0 ∂δ uδσ |δ=0 1 B
∂s = L0 +2 χR (0, s − t⋆ ) + E1 , ∂s ∂δ tδσ |δ=0 = Ẽ1 , (9.25)
∂δ ωδσ |δ=0 ∂δ ωδσ |δ=0 A
where  
kẼ1 k + kE1 k ≤ C (1 + R−2 )(kg0 − g⋆ kC 1 + |z0 |)eC|s| kσk. (9.26)
Note that (9.25) reduces to (9.13) when (z0 , g0 ) = (0, g⋆ ). To obtain (9.25), we differentiate (9.15)
and use (9.24) together with Lipschitz bounds on g0 , g⋆ , and χR to estimate the change from (9.13).
The bound on (9.26) follows from combining (9.24) together with |∂δ u| + |∂δ ω| + |∂δ τ | + |∂δ t| ≤
CeC|s| kσk where C = C(kg0 kC 2 , kg⋆ kC 3 ).
Similarly, differentiating again as to obtain (9.14),
 
−1

1 D j E jk
∂s (∂δ ∆δσ |δ=0 )∆0 = 2 k χR (0, s − t⋆ ) + [L0 , (∂δ ∆δσ |δ=0 )∆−1
0 ] + FR (A, B, s) + E2 ,
Cjk −Djk
∂s ∂δ ∂(u0 ,ω0 ) t|δ=0 = Ẽ2
where  
kẼ2 k + kE2 k ≤ C(kg⋆ kC 4 ) (1 + R−2 )(kg0 − g⋆ kC 1 + |z0 |)eCs kσk.
By (9.13) and (9.14), the bounds on E1 and E2 directly imply that for any s > 0,


kdσ (u, ω, ∆σ ∆−1 −1
0 )(s, z0 , g0 , σ) − (u, ω, ∆σ ∆0 )(s, 0, g⋆ , σ) |σ=0 k
≤ C(kg⋆ kC 4 )(1 + R−2 )(kg0 − g⋆ kC 1 + |z0 |)eCs .
In addition, computing as in (9.20), we have
˜ σ∆
kdσ ∆ ˜ −1 (T0 , z0 , g0 , σ) − dσ ∆σ ∆−1 (T0 , z0 , g0 , σ)k ≤ C(kg⋆ kC 4 )(1 + R−2 )(kg0 − g⋆ kC 1 + |z0 |)eCT0
0 0
and hence, we obtain (9.23). 
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 77

Vi0 supp(g0 − gσ ) HT̃0 Wi1

H0

γρ̃g0
ρ̃ ρ̃1
ρ ϕgT⋆0 (ρ)
ρ0 γρg⋆

Figure 5. The figures shows the geometric setup used in Lemma 9.12 to pass
from the special hypersurfaces defined using the coordinates Φgρ⋆ to a given well
separated set (Γ, Γ̃).

In what follows we use the notion of a well-separated set introduced in Definition 6.5.
Lemma 9.12. Let G ⊂ G 4 bounded, ̥ > 0, and {(Wi , Vi )}N i=1 be ̥-well separated for G. There
exist δ > 0, R0 > 0, C > 0, and C1 > 0, such that the following holds. Let g⋆ ∈ G, T0 > 0,
t⋆ ∈ [3R0 , T0 − 3R0 ], ρ ∈ Vi0 , and i1 ∈ {1, . . . , N } such that
ϕgT⋆0 (ρ) ∈ Wi1 , d(ϕgT⋆0 (ρ), Vi1 ) < δ,
and the set
{t ∈ [0, T0 ] : γρg⋆ (t) ∩ B(γρg⋆ (t⋆ ), 3R0 ) 6= ∅} (9.27)
is connected. Then for all
0 < R < R0 , 0 < ε ≤ R2 e−C1 |T0 | /C1 ,
g0 ∈ G 4 and ρ̃ ∈ Vi0 with
kg⋆ − g0 kG 4 ≤ ε, kd(ρ̃, ρ)k ≤ ε,
the map
dσ (Pσ |q=ρ̃ , dq Pσ |q=ρ̃ )|σ=0 : Σ → TP0 (ρ̃) Wi0 × Tdq P0 |q=ρ̃ Sp(Tρ̃ Wi0 ; TP0 (ρ̃) Wi1 ),
is bijective with inverse bounded by CeC|T0 | , where Pσ : BVi0 (ρ̃, e−C1 |T0 | /C1 ) ⊂ Vi0 → Wi1 is the
map
Pσ (q) := ϕgTσgσ (q) (q), T g⋆ (ρ̃) = T0 , ϕgTσgσ (q) (q) ∈ Wi1 ,
with T gσ (q) smooth in σ and q, gσ = gσ (ρ, t⋆ , R, g0 , σ) is the perturbed metric defined in (9.4),
and Sp(Tρ̃ Wi0 ; TP 0 (ρ̃) Wi1 ) denotes the symplectic linear maps from Tρ̃ Wi to TP0 (ρ̃) Wi1 .

Proof. Let (u0 , t0 ) be such that Φgρ⋆ (u0 , t0 ) = πM (ρ̃). Next, let ρ0 := γρg⋆ (t0 ). From now on we
work with the system of coordinates (u, t) induced by (u, t) 7→ Φgρ⋆0 (u, t) and with Hs as defined
in (9.2). Note that ρ̃ ∈ Vi0 ∩ H0 .
Let ρ̃1 := P0 (ρ̃) = ϕgT0g0 (ρ̃) (ρ̃) ∈ Wi1 , and T̃0 be such that ρ̃1 ∈ Wi1 ∩ HT̃ Define the maps
0

Q0,σ : Wi0 → H0 , Q1,σ : HT̃ → Wi1 ,

0
78 YAIZA CANZANI AND JEFFREY GALKOWSKI

by
Q0,σ (q) = ϕgSσ (q) (q), Q1,σ (q) = ϕgSσ (q) (q),
0 1

where S0 (q) is the real number with smallest absolute value such that ϕgSσ(q) (q) ∈ Ht0 , and S1 (q)
0
is the real number with smallest absolute value such that ϕgSσ(q) (q) ∈ Wi1 .
1
√
Note that, by construction, the perturbation is supported on (u, t) such that |t − t⋆ | ≤ 2R.
Hence, since t⋆ > 3R0 and R < R0 , the perturbation is supported away from Wi0 and Wi1 .
Therefore, choosing R0 small enough, Q0,σ (q) = Q0,0 (q) for all q with d(q, ρ) ≪ R0 and Q1,σ (q) =
Q1,0 (q) for all q with d(q, ρ1 ) ≪ R0 .
In the (u, t) coordinate system, with ω dual to u, we write Z0,σ and Z1,σ for the (u, ω) repre-
sentation of Q0,σ and Q1,σ respectively. Next, consider the map

Ψzg0 ,T̃0 (σ) := (Ξ1 (z0 , g0 , σ), Ξ2 (z0 , g0 , σ))

0

defined in (9.10) with T̃0 in place of T0 . Using that the perturbation is supported away from Wi0
and Wi1 , we have Zj,σ (z0 ) = Zj,0 (z0 ) for |z0 | ≪ R0 .
˜ 0 (T, z0 , g0 ), with
Now, abbreviating Ξj,σ (z0 ) := Ξj (z0 , g0 , σ), and putting Ξ̃2,σ (z0 ) := Ξ2,σ (z0 )∆
˜ 0 as in (9.9) and T = T (z0 , g0 , σ, T̃0 ) as in (9.10) with T0 replaced by T̃0 , we have (in the (u0 , ω0 )
∆
coordinates),
Pσ = Z1,σ ◦ Ξ1,σ ◦ Z0,σ , dρ Pσ = dρ Z1,σ ◦ Ξ̃2,σ ◦ dρ Z0,σ . (9.28)
Thus, since for |z0 | < ε with C1 to be determined later, and kσk ≤ 1, we have |Ξ1,σ (Z0,σ (z0 ))| ≪ R0
and |z0 | ≪ R0 , and hence we may replace Zj,σ by Zj,0 in (9.28) to obtain
dσ ((Pσ (z0 ), dρ Pσ (z0 )))|σ=0

= dρ Z1,0 |Ξ1,0 (Z0,0 (z0 )) dσ Ξ1,σ (Z0,0 (z0 ))|σ=0 , d2ρ Z1,0 |Ξ1,0 (Z0,0 (z0 )) Ξ̃2,0 dρ Z0,0

+ dρ Z1,0 |Ξ1 (Z0,0 (z0 ) dσ Ξ̃2,σ (Z0,0 (z0 ))|σ=0 dρ Z0,0 |z0 .

Next, observe that by Lemmas 9.8 and 9.10

kdσ2 Ξ1,σ (Z0,0 (z0 ))|σ=0 k ≤ C(1 + R−2 )εeC|T̃0 | ,

kdσ1 Ξ1,σ (Z0,0 (z0 ))|σ=0 k + kdσ1 Ξ2,σ (Z0,0 (z0 ))|σ=0 k ≤ CeC|T̃0 | ,
and, provided that C1 is chosen large enough,

k(dσ1 Ξ1,σ (Z0,0 (z0 ))|σ=0 )−1 k + k(dσ2 Ξ2,σ (Z0,0 (z0 ))|σ=0 )−1 k ≤ CeC|T̃1 | .
Together with the fact that

kd2ρ Zk,0 k + kdρ Zk,0 k + k(dρ Zk,0 )−1 k ≤ C, k∆0 k ≤ CeC|T̃0 | ,

this implies the lemma. To see this, observe that, taking
C1 large enough, the estimates above
A R , where kBA−1 Rk < 1 kD −1 k.
imply that dσ ((Pσ (q), dρ Pσ (q)))|σ=0 takes the form B 
D 2
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 79

9.2. Probing families of metrics. Before defining our family of probing metrics, we need one
more auxiliary lemma which will allow us to control the size of the perturbation, gσ − g0 , in the
C ν norm on symmetric tensors. Here, we recall that, to make sense of C ν norms, we use the fixed
norm as in Remark 1.5. Note that, in the (u, t) coordinates from (9.4), controlling the size of
the perturbation is trivial. However, we need to estimate how the coordinate change, Φgρ⋆ affects
these norms. This is the purpose of our next lemma.
Lemma 9.13. Let ν ≥ 2 and K ⊂ G ν be bounded. Then there are Cν > 0 and R0 > 0 such that
for all g⋆ ∈ K, t0 ∈ R, 0 < R < R0 and v ∈ C ν (Rn ) supported in the ball B((0, t0 ), 2R), we have
k[(Φgρ⋆ )−1 ]∗ vkC ν ≤ Cν kvkC ν ,

Proof. Once we understand the C ν norms of (Φgρ⋆ )−1 , it will remain to apply the Faà di Bruno
formula. To do this, we first choose R0 ≪ 1 small enough such that for all g⋆ in K, Φgρ⋆ is a
diffeomorphism on B((0, t0 ), 2R) for any t0 and R < R0 .
To estimate the C ν norms of (Φgρ⋆ )−1 , it is enough to estimate (dΦgρ⋆ |(0,t) )−1 from below and
kΦgρ⋆ kC ν from above. To see that estimating the inverse at u = 0 is sufficient, observe that
kdΦgρ⋆ |(0,t) − dΦgρ⋆ |(ũ,t) k ≤ kΦgρ⋆ kC 2 |ũ| ≤ R0 kΦgρ⋆ kC 2 ,
for (ũ, t) ∈ supp v. Therefore, provided k(dΦgρ⋆ |(0,t) )−1 k < 1
2R0 kΦgρ⋆ kC 2
, we have dΦgρ⋆ is also
invertible in the support of v.
Observe that there is C > 0 such that for all g ∈ K and V ∈ T M
C −1 |V |g ≤ |V | ]
∗
≤ C|V |g . (9.29)
S M

Therefore, since
dΦgρ⋆ |(0,t0 ) (E i (0)δui + E d (0)δt ) = E d (t0 )δt + E i (t0 )δui ,
we have
C|dΦgρ⋆ |(0,t0 ) (E i (0)δui + E d (0)δt )|2]
∗
≥ |dΦgρ⋆ |(0,t0 ) (E i (0)δui + E d (0)δt )|2g⋆
S M
= |(E i (0)δui + E d (0)δt |2g⋆ ≥ C −1 |E d (0)δt + E i (0)δui )|2]
∗
.
S M
and hence k(dΦgρ⋆ |(0,t) )−1 k ≤ C 2.
Next, Lemma 9.4 implies that there is C, depending only on K, such that kΦgρ⋆ kC ν ≤ C. In
particular, we obtain that there is c0 > 0 depending on K such that for R ≤ c0 there is Cν
depending on K and ν such that k(Φgρ⋆ )−1 kC ν ≤ Cν on Φgρ⋆ (supp v). Using this in the Faà di
Bruno formula then completes the proof of the lemma. 
N (r) ]
Let r > 0 and {ρi (r)}i=1 be a maximal r separated set in S ∗M so that

N (r)
[
]
S ∗M ⊂ B(ρi , r). (9.30)
i=1

We associate to each ρi (r), a family of perturbations defined by iterative application of the

construction (9.4). In particular, let

N (r)
Σ(r) := BRL (0, 1)
80 YAIZA CANZANI AND JEFFREY GALKOWSKI

We define for δ, R > 0, σ ∈ Σ(r), and g⋆ ∈ G ν ,

N (r)
X
r,R,δ
gσ,g ⋆
= g⋆ + (gδσj (ρj , 0, R, g⋆ , δσj ) − g⋆ ),
j=1

where gσ = gσ (ρj , 0, R, g⋆ , σ) is the metric perturbation of g⋆ defined along γρgj⋆ (t) in R-neighborhood
of ρj = γρgj⋆ (0) (see (9.4)).
In what follows, we write S ν for the space of C ν symmetric tensors on M .
For R, r, δ > 0 define the map Q : G ν × Σ(R) → S ν by
Qr,R,δ (g, σ) := gσ,g
r,R,δ
. (9.31)
Lemma 9.14. Let ι : G ν → G ν−1 be the natural embedding. Then, for all K ⊂ G ν bounded, there
is ε0 > 0 such that if
δR−1 r −2d+1 max(r, R)d < ε0 ,
then
Qr,R,δ : K × Σ(R) → G ν . (9.32)
Moreover, with Q̃r,R,δ := ι ◦ Qr,R,δ : G ν × Σ(R) → G ν is Frechet differentiable, the map g →
D(g,σ) Q̃r,R,δ is continuous, and
Dg Q̃r,R,δ : S ν−1 → S ν−1
is bijective, where we have extended Dg Q̃r,R,δ to S ν−1 by density. Moreover, for all K ⊂ G ν
bounded, there is C > 0 such that for all δ, r, R > 0, g ∈ K, and σ ∈ Σ(R),
kDσ Qr,R,δ kℓ∞ →C ν ≤ CδR−1−ν r −2d+1 max(r, R)d , (9.33)
r,R,δ −ν −2d+1 d
kDg Q (g, σ) − IkC ν−1 →C ν−1 ≤ C1 δR r max(r, R) , (9.34)
k∂σj Qr,R,δ (g, σ)kB (0,1)→C ν ≤ CδR−1−ν , (9.35)
RL

∂σαj Qr,R,δ (g, σ) = 0, |α| ≥ 2. (9.36)

Proof. First observe that (9.33) implies (9.32), so we only check the estimates (9.33) through (9.36).
N (r)
Let ρ ∈ {ρi }i=1 and σ ∈ BRL (0, 1), and set gσ := gδσ (ρ, 0, R, g, σ). Observe that (Φgρ )∗ gδσ =
(Φgρ )∗ g + δhσ where hσ does not depend on g and
|∂ α hσ | ≤ Cα R−|α| kσk, |∂ α Dσ hσ | ≤ Cα R−|α| , ∂σα hσ = 0, |α| ≥ 2
and hσ is supported in a ball of radius 2R around (Φgρ )−1 πM (ρ). Therefore, using Lemma 9.13
kDσ gδσ kC ν = kδDσ ((Φgρ )−1 )∗ hσ kC ν ≤ Cν R−1−ν δ, Dσα gδσ = 0, |α| ≥ 2.
This proves (9.35) and (9.36).
Next, let gi ∈ G ν , i = 0, 1 and note
 
(g0 )δσ − g0 − [(g1 )δσ − g1 ] = ([Φgρ0 ]−1 )∗ − ([Φgρ1 ]−1 )∗ δhσ .
Therefore, by Lemmas 9.3 and 9.13,
kDg [(g)δσ − g]kC ν−1 →C ν−1 ≤ Cν δkσkR−ν , (9.37)
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 81

the map (g, σ) 7→ Dg [gδσ − g] is continuous in the G ν × BRL (0, 1) topology, and the range of the
derivative is supported in a ball of radius CR around πM (ρ) where C > 0 depends only on K.
Using these estimates together with the definition of Qr,R,δ , the lemma follows after a counting
argument. In particular, it is enough to bound
sup #{i ∈ {1, . . . , N (r)} : B(πM (ρj ), 2R) ∩ B(πM (ρi ), 2R) 6= ∅}.
j

To do this, note that since ρj are a maximal r separated set, there is a constant D depending
only on d and C such that there are {Jℓ }D
ℓ=1 , Jℓ ⊂ {1, . . . N (r)} such that
[
{1, . . . , N (r)} = Jℓ , B(ρj , Cr) ∩ B(ρi , Cr) = ∅, i 6= j, i, j ∈ Jℓ .
ℓ
In particular, for each fixed j,
#{i ∈ {1, . . . , N (r)} : B(πM (ρj ), 2R) ∩ B(πM (ρi ), 2R) 6= ∅}
≤ #{i ∈ {1, . . . , N (r)} : πM (ρi ) ∈ BM (πM (ρj ), 4R)}
≤ #{i ∈ {1, . . . , N (r)} : πM (B ] (ρi , 2r)) ⊂ BM (πM (ρj ), 4R + 2r)}
S ∗M
 
≤ Cd D vol ρ ∈ S ] ∗M : π (ρ) ∈ B (π (ρ ), 2r + 4R) r −2d+1
M M M j

≤ Cd D̃(max(r, R))d r −2d+1

Combining this bound with (9.35) implies (9.33), and with (9.37) implies (9.34). 

10. Construction of well separated sets

In this section, we construct a well-separated set for a small neighborhood of a metric g† ∈ G 3 .

We then show that these well-separated sets are admissible in the sense of Definition 7.3 for the
perturbation from Section 9.

10.1. Construction of the well-separated set. Define for g ∈ G 3

]
Ψg : R × S ]
∗M → S ∗M , Ψg (t, q) = ϕgt (q),
]
and for T > 0, V ⊂ S ∗M , consider the geodesic tube


TVg (T ) := Ψg (−T, T ) × V .
Lemma 10.1. Let g† ∈ G 3 . There exist N > 0, δ > 0, hypsersurfaces Zi ⊂ M , i = 1, . . . , N ,
and open subsets Vi ⋐ Wi ⊂ S^
∗
Zi M such that H|ξ|g is transverse to Wi and there is a bounded
†
neighborhood G ⊂ G 3 of g† such that for all g ∈ G,
N
[
TWg i (δ) ∩ TWg j (δ) = ∅, i 6= j, and S]
∗M ⊂ TVgi ( 13 T† ), (10.1)
i=1
T T
where T† = injg (M ). Moreover, Ψg : (− 3† , 3† ) × Wi → Sg
∗M is a C 1 -diffeomorphism onto its
†
image.
82 YAIZA CANZANI AND JEFFREY GALKOWSKI

Proof. For each q ∈ S ]∗M , let Z ⊂ M be a hypersurface with q ∈ N ∗ Z and such that we can work
q q
with Fermi normal coordinates associated to Zq . That is, we assume that there is εq > 0 such
that (x1 , x′ ) are well defined coordinates for |(x1 , x′ )| < 2εq with the property that (0, 0) = πM (q)
and Zq = {(x1 , x′ ) : |x′ | < εq , x1 = 0}. Let
ε ]
Hqq := {(0, x′ , ξ1 , ξ ′ ) ∈ S ∗M : |ξ | > 1 |ξ| , |x′ | < ε },
1 g† q
2

where we write (ξ1 , ξ ′ ) for the dual coordinates so that we have q = (0, 0, 1, 0). Shrinking εq if
3ε ]
necessary, we may further assume that Ψg† : (− 12 T† , 12 T† ) × Hq q → S ∗M is a C 1 -diffeomorphism

onto its image.

Next, for each q ∈ S ] ∗M and x ∈ (−ε , ε ) let Z (x ) := {(x , x′ ) : |x′ | < 3ε }. Observe that
1 q q q 1 1 q
2ε
there exist 0 < δq < εq and Cq > 0 such that for x1 ∈ (0, δq ) and ρ ∈ Hq q ,
n g
o
Tx1 (ρ) := inf t > 0 : πM (ϕt † (ρ)) ∈ Zq (x1 ) < Cq |x1 |,
g 2εq
and the map ρ 7→ ϕT†x (ρ) is a C 1 -diffeomorphism from Hq onto its image. In particular,
1 (ρ)
2ε
choosing δq small enough, we have for x1 ∈ (0, δq ), V ⊂ Hq q open, we have
g g† g
TV † ( 15 T† ) ⊂ TV [x1 ] ( 41 T† ), V [x1 ] := TV † ( 12 T† ) ∩ S ∗^
Zq (x1 ), (10.2)
and V [x1 ] is C 1 -diffeomorphic to V . Therefore, we can shift slightly the initial position of our
tubes without changing covering properties. Moreover, after possibly shrinking δq , we may assume
that

sup diam V [x1 ] ≤ 4 diam(V ). (10.3)
x1 ∈[0,δq )
SL
By compactness, there are {qi }L ]∗ ]∗ 1
i=1 ⊂ S M such that S M ⊂ i=1 THεqi ( 5 T† ). Fix r > 0. For
qi
i N (r) 2εqi
each i = 1, . . . , L, we let {ρij }j=1 be a maximal r/2 separated set on Hqi so that
Ni (r)
εq [
Hq i i ⊂ B 2εq (ρij , r/2),
Hq i
j=1 i

and there is D > 0, independent of r > 0, and a partition {1, . . . , Ni (r)} = ∪D

k=1 Jk such that for
any k ∈ 1, . . . , D and any j1 6= j2 , j1 , j2 ∈ Jk
B 2εq (ρij1 , 4r) ∩ B 2εq (ρij2 , 4r) = ∅.
Hq i Hq i
i i

In particular, since L is independent of r, there is D̃ > 0 independent of r such that the tubes
Tij (4r, 12 T† ) := Ψg† ((− 12 T† , 21 T† ) × B 2εq (ρij , 4r)),
Hq i
i

can be divided into D̃ collections of disjoint tubes, and

L
[ L N[
[ i (r)
]
S ∗M ⊂ T εq ( 15 T† ) ⊂ Tij ( 12 r, 51 T† ). (10.4)
Hq i
i
i=1 i=1 j=1
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 83

δ0
We now fix 0 < r < 100D̃
, where δ0 = mini δqi . By (10.3), for |x1 | < δ0 ,


max diam B 2εq (ρij , 4r) [x1 ] ≤ 32r.
i,j Hq i
i

Therefore, we may find |x1 (i, j)| < δ0 for all i, j such that for (i, j) 6= (k, ℓ),


Wij ∩ Wkl = ∅, Wij := B 2εq (ρi,j , 4r) [x1 (i, j)]. (10.5)
Hq i
i

(This is possible, for instance, using a greedy algorithm where we simply select x1 (i, j) iteratively
for each intersecting tube.)
Thus, by (10.2) and (10.4),
L N[
[ i (r)
g

]
S ∗M ⊂ TVij† ( 14 T† ), Vij := B 2εq (ρij , 21 r) [x1 (i, j)].
Hq i
i=1 j=1 i

Moreover, by (10.5) there is δ > 0 such that for (i, j) 6= (k, ℓ),
g g
TW†ij (2δ) ∩ TW†kl (2δ) = ∅.

Finally, since the map G 3 ∋ g 7→ Ψg ∈ C 1 is continuous, Ψg† : (− 21 T† , 12 T† ) × Hq

3εq ]
→S ∗M is a

C 1 -diffeomorphism, which this proves the claim. 

Lemma 10.2. Let g† ∈ G 3 and {(Wi , Vi )}N
i=1 as in Lemma 10.1. Let Vi = Vi , Vi ⋐ Wi ⋐ Wi .
Then, there is a bounded neighborhood G ⊂ G 3 of g† and ̥ > 0 such that {(Wi , Vi )}N
i=1 is a
̥-well separated set for G.

Proof. Let G, δ, and T† as in Lemma 10.1. Shrinking G if necessary, we may assume that H|ξ|g is
uniformly transverse to Γ̃ for g ∈ G.
We need to check the conditions (6.3). Let ̥ = 12 δ. Then, by the second condition in (10.1),
4T†
the first condition in (6.3) follows with CΓ = 3δ .
Next, by the first condition in (10.1), the second condition in (6.3) holds with cΓ = 2. Finally,
T T
]
since for all g ∈ G, and i, Ψg (− 3† , 3† ) × Wi → S ∗M is a C 1 -diffeomorphism onto its image, and

W i ⋐ Wi ,
2T 4T
inf inf{t > 0 : ϕgt (ρ) ∈ W i } ≥ † = † ̥ ≥ CΓ ̥,
ρ∈Wi 3 3δ
and hence, the last condition in (6.3) holds. 

10.2. Admissibility of the family of perturbations from Section 9. In order to show that
the well separated-sets constructed above are admissible for the perturbations from Section 9, we
study these perturbations near simple points.
We start by showing that one can find a ball through which the geodesic emanating from a
simple point passes only once.
84 YAIZA CANZANI AND JEFFREY GALKOWSKI

Lemma 10.3. Let K ⊂ G 3 bounded, ̥ < inf g∈K injg (M ). Then there are c0 > 0 and c1 > 0
such that if 0 < α < c0 , n ∈ N, and ρ ∈ S̥ (n, α, g), then for every R > 0 and any maximal R
J(R)
separated set {xj }j=1 on M , and any g ∈ K
n2 ̥2
#{j : Ig (xj , R) 6= ∅ and Ig (xj , 4R) is disconnected} ≤ ,
αcn1
where Ig (xj , R) := {t ∈ [0, (n − 12 )̥] : πM (ϕgt (ρ)) ∩ B(xj , R) 6= ∅}.

Proof. We claim that since ρ ∈ S̥ (n, α, g), there is ε > 0, 0 < c1 < 1 such that if g ∈ K, α < ε ,
0 ≤ t, s ≤ (n − 12 )̥, and |t − s| > c1n−1 α, then
d(ϕgs (ρ), ϕgt (ρ)) ≥ cn1 α. (10.6)

To prove the claim in (10.6), recall that there is C > 0 such that d(ϕgt (ρ1 ), ϕgt (ρ2 )) ≤ C |t| d(ρ1 , ρ2 ),
]
for g ∈ K and all ρ1 , ρ2 ∈ S ∗M . In addition, by Lemma 6.6, for all g ∈ K and ρ ∈ S ]∗M we have

d(ϕgt (ρ), ρ) ≥ |t|/C, |t| ≤ ̥. (10.7)

By contradiction, suppose there are 0 ≤ t ≤ s−c1n−1 α and s ≤ (n− 21 )̥ such that d(ϕgs (ρ), ϕgt (ρ)) ≤
cn1 α. First, by (10.7), if |t − s| ≤ ̥, we have
C −1 |s − t| ≤ d(ϕgs−t (ϕgt (ρ)), ϕgt (ρ)) ≤ cn1 α.
This contradicts t ≤ s − c1n−1 α for c1 chosen small enough. In particular, we have |s − t| ≥ ̥.
Next, applying the flow for time −t,
d(ϕgs−t (ρ), ρ) ≤ cn1 C t α ≤ cn1 C n̥ α.
Since |s − t| ≥ 21 ̥, this contradicts the fact that ρ ∈ S̥ (n, α, g) and hence proves the claim in
(10.6).
We next claim that there is c0 > 0 such that if g ∈ K, ρ ∈ S̥ (n, α, g), t0 , s0 ∈ [0, (n − 12 )̥],
|s0 − t0 | ≥ cn1 α, and
d(πM (ϕgt0 (ρ)), πM (ϕgs0 (ρ))) < 4R,
then for |t − t0 | < c0 and |s − s0 | < c0 ,
d(πM (ϕgt (ρ)), πM (ϕgs (ρ)) > c0 (|t − t0 | + |s − s0 |)cn1 α − 8R. (10.8)

We now prove (10.8). Without loss of generality, we may assume cn1 α ≥ 8R since other-
wise (10.8) is trivial. Since we are working locally near ϕgs0 (ρ), we may work in geodesic normal
coordinates such that ϕgs0 (ρ) = (0, e1 ) where e1 := (1, 0, . . . , 0) and ϕgt0 (ρ) = (x1 , ξ1 ). Observe
that by (10.6),
cn1 α ≤ |(x1 , ξ1 ) − (0, e1 )| ≤ 4R + |ξ1 − e1 |.
Thus, |ξ1 − e1 | ≥ 12 cn1 α. Then, define
(x(t, (x0 , ξ0 )), ξ(t, (x0 , ξ0 ))) := ϕgt (x0 , ξ0 ).
We first claim that
|ξ(t0 − s0 , (0, e1 )) + e1 | > 12 cn1 α. (10.9)
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 85

Indeed, suppose not. Then,

|x(t0 − s0 , (0, e1 ))| + |ξ(t0 − s0 , (0, e1 )) + e1 | ≤ 4R + 12 cn1 α < cn1 α.
In particular, applying the backward flow, we have



d x(t0 −s0 −t, (0, e1 )), ξ(t0 −s0 −t, (0, e1 )) , x(−t, (0, −e1 )), ξ(−t, (0, −e1 )) < C |t| cn1 α. (10.10)

Next, observe that

(x(−t, (0, −e1 )), ξ(−t, (0, −e1 ))) = (x(t, (0, e1 )), −ξ(t, (0, e1 )).
Therefore, putting t = t0 −s 2
0
in (10.10), we have



d x( t0 −s
2
0
, (0, e1 )), ξ( t0 −s0
2 , (0, e1 )) , x( t0 −s0
2 , (0, e1 )), −ξ( t0 −s0
2 , (0, e1 ) < C n̥ cn1 α.

Choosing c1 small enough (uniformly in K), this is a contradiction since |ξ( t0 −s

2 , (0, e1 ))|g = 1.
0

In particular, (10.9) holds.

Now, we prove (10.8). Since we work in geodesic normal coordinates at x(s0 , (0, e1 )), we have
|x(s − s0 , (0, e1 )) − x(t − t0 , (x1 , ξ1 ))|
≥ |x(s − s0 , (0, e1 )) − x(t − t0 , (0, ξ1 ))| − |x(t − t0 , (0, ξ1 )) − x(t − t0 , (x1 , ξ1 ))|
≥ |(s − s0 )e1 − (t − t0 )ξ1 | − (1 + C|t − t0 |)|x1 |
≥ c0 cn1 α(|s − s0 | + |t − t0 |) − 8R,
where in the last line we have used that |ξ1 |g(x1 ) = |e1 | = 1, min(|ξ1 − e1 |, |ξ1 + e1 |) > c0 cn1 α, and
that C|t − t0 | < Cc0 ≤ 1. This proves the claim in (10.8).
We now use (10.8) to control the length of time for which the (projection
 S to the base of the) seg- 
g
ment of the geodesic near t0 can be close to that near s0 . Let Ω := B |s−s0 |<c0 πM (ϕs (ρ)), 8R .
We claim that
n o CR
t : |t − t0 | < c0 , B(πM (ϕgt (ρ)), 4R) ∩ Ω 6= ∅ ≤ , (10.11)
c0 cn1 α
where C depends only on K. To see (10.11), note that if t lies in the left hand side, then there is
s such that |s − s0 | < c0 and, by (10.8), we have
12R > d(πM (ϕt (ρ)), πM (ϕs (ρ))) ≥ c0 cn1 α|t − t0 | − 8R.
In particular, (10.11) holds.
̥(n− 1 )−2c
Now, let Iℓ := [ℓc0 , (ℓ + 2)c0 ), ℓ = 1, . . . , L − 1, L = ⌊ 2
c0
0
⌋ + 1, IL := [Lc0 , ̥(n − 12 )], and
[ 
Ωℓ := B πM (ϕgt (ρ)), 8R .
t∈Iℓ

Since for |ℓ − k| > 1, inf t∈Iℓ ,s∈Ik |t − s| ≥ 2c0 , using (10.11) for each pair ℓ, k with |ℓ − k| > 1,
CR n2 ̥2
{t : ∃k, ℓ s.t. |k − ℓ| > 1, t ∈ Iℓ , B(πM (ϕgt (ρ)), 4R) ∩ Ωk 6= ∅} ≤ .
c0 cn1 α c20
86 YAIZA CANZANI AND JEFFREY GALKOWSKI

Next, note that there exists c > 0 such that the following holds. Let g ∈ K and j such that
Ig (xj , R) 6= ∅ and Ig (xj , 4R) is disconnected there is an interval of length ≥ cR inside the set
{t : ∃k, ℓ s.t. |ℓ − k| > 1, t ∈ Iℓ , B(πM (ϕgt (ρ)), 4R) ∩ Ωk 6= ∅},
and therefore the lemma follows.
Indeed, to see the last claim, shrinking c0 so that 4c0 < inf g∈K injg (M ) if necessary, we can find
t1 ∈ Ik1 , t2 ∈ Ik2 such that πM (ϕgt1 (ρ)) ∈ B(xj , R) and πM (ϕgt2 (ρ)) ∈ B(xj , 4R), with |t1 −t2 | > 4c0
and hence |k1 − k2 | > 1. Notice that B(xj , 4R) ⊂ Ωk2 and hence, since πM (ϕgt1 (ρ)) ∈ B(xj , R),
we have πM (ϕgt (ρ)) ∈ Ωk2 for |t − t1 | ≤ cR (here, again c > 0 is a constant depending only on
K). 

Let Qr,R,δ be the map that defines the perturbation of the metric as introduced in (9.31). We
now check that for any ν0 and y > 2, there is b such that (Γ, G) is a (̥, b, y) admissible pair for
y
{QR ,R,δ }R,δ . (see Definition 7.3).
Lemma 10.4. Let ν0 ≥ 5 and g† ∈ G 3 . Let ̥ > 0, G ⊂ G 3 be the bounded neighborhood G ⊂ G 3
FN
of g† , and {(Wi , Vi )}N
i=1 be the ̥-well separated set for G given by Lemma 10.2. Let Γ := i=1 Vi .
Then, for y > 2,
(Γ, G) is a (̥, b, y) − admissible pair for {QR,δ }R,δ ,
y ,R,δ
with b := 7 − y + d(2y − 1) and QR,δ := QR as defined in (9.31).

Proof. Let K ⊂ G bounded in G ν0 . We will show that there are c > 0, ε > 0 such that for all
δ > 0, g ∈ K, α ∈ (0, c), n ∈ N, 0 < R < cn α, 0 < δ ≤ Rb cn+1 , and
ρ0 ∈ Γ ∩ S̥ (n, α, g) ∩ R̥ (n, ε, g),

then there are i0 ∈ {1, . . . , N (R)}, I0 , and m0 such that BΓ (ρ0 , Ry ) ⊂ DI(m0 ) [QR,δ (g, σ)] for all
  0
σ ∈ ΣN(R) , TI(m0 ) [g](ρ) ∈ (n−1)̥−CG δG , n̥+CG δG , and for all ρ ∈ BΓ (ρ0 , Ry ) and σ ∈ ΣN(R) ,
0
the lower bound in (7.6) holds.
By Lemma 6.12, for c, R, ε small enough there are I and m such that ρ0 ∈ DI(m) [g], TI(m) [g](ρ) ∈
 
(n − 1)̥ − CG ε , n̥ + CG ε , and d(PI(m) [g](ρ0 ), ρ0 ) < CG ε holds.
Let δ0 be as in Lemma 6.13. Next, since
sup sup inf{t > 0 : ϕgt (ρ) ∈ Γ} < CΓ ̥,
g∈G ρ∈Γ̃

by removing elements from I, there is C > 0 depending only on G such that for any δ0 > 0, we
find (m0 , I0 ) such that TI(m0 ) [g](ρ0 ) ∈ (n − 1)̥ − CG ε , n̥ + CG ε and
0

sup d(PI(m) [g](ρ0 ), Γ) ≤ max(CG ε, δ0 ).

0
m≤m0

In particular, for c small enough (depending only on G) and

δ< R1−y+3+d(2y−1) cn+1 , 0 < R̃ < cn+1 , (10.12)

 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 87

(m)
we have by Lemma 6.13 that B(ρ0 , R̃) ⊂ DI0 [QR,δ (g, σ)], for all σ ∈ ΣN(R) and m ≤ m0 . Indeed,
since
CΓ ̥m0 ≥ TI(m0 ) [g](ρ0 ) ≥ cΓ ̥m0 > ̥m0 ,
0

we have n ≤ CΓ m0 + 2 and m0 ≤ (n + 1). Thus, R̃ < cn ≤ cm0 and the condition on δ yields
kQR,δ (g, σ) − gkC 3 < cn+1 ≤ cm0 as needed, after applying Lemma 9.14 with (Ry , R, δ, 3) in place
of (r, R, δ, ν).
We now study the action of the perturbation QR,δ on ΨR,δ
g,I0 as defined in (7.7). To do this, we
will use Lemma 9.12 with g⋆ = g, and some appropriate choice of ρ, g0 , and T0 .
J(R)
Note that for any maximal R separated set {xj }j=1 ⊂ M , since ̥ < injg (M ),

# j : there exists t ∈ [̥/5, ̥/4] satisfying πM (ϕgt (ρ0 )) ∈ B(xj , R) ≥ c̥R−1 .
Therefore, by Lemma 10.3 if
n2 ̥2
α < c0 and < c̥R−1 , (10.13)
αcn1
then there is an index j0 such that
{t ∈ [0, (n − 12 )̥] : πM (ϕgt (ρ0 )) ∩ B(xj0 , 4R) 6= ∅} is connected (10.14)
and there is t ∈ [̥/5, ̥/4] satisfying πM (ϕgt (ρ0 )) ∈ B(xj0 , R).
N (R)
Moreover, there is c = c(̥) > 0 such that, with {ρi }i=1 as in (9.30), if
y
R ≤ cR̃, (10.15)
then for some i0 = 1, . . . , N (R) satisfying πM (ρi0 ) ∈ B(xj0 , R + Ry ) there are ρ⋆ ∈ B(ρ0 , R̃) and
t⋆ ∈ (1/6̥, ̥/3) such that ϕgt⋆ (ρ⋆ ) = ρi0 . In addition, if
R̃ < cn R, (10.16)
then (10.14) implies
{t ∈ [0, (n − 12 )̥] : πM (ϕgt (ρ⋆ )) ∩ B(xj0 , 27 R) 6= ∅} is connected. (10.17)
To obtain the lower bound in (7.6), we aim to apply Lemma  9.12. To do this, let T0 =
TI(m0 ) [g](ρ⋆ ), since TI(m0 ) [g](ρ0 ) ∈ (n − 1)̥ − CG ε , n̥ + CG ε ,
0 0

(n − 1)̥ − CG ε − C n R̃ ≤ T0 ≤ n̥ + CG ε + C n R̃.
Now, we claim that
{t ∈ [0, T0 ] : πM (ϕgt (ρ⋆ )) ∩ B(πM (ϕgt (ρ⋆ )), 72 R) 6= ∅} is connected. (10.18)
Indeed, if T0 ≤ (n − 12 )̥, then this is true by (10.17). Suppose instead that T0 > (n − 21 )̥.
We need to show that there is no t ∈ [(n − 12 )̥, T0 ] such that πM (ϕgt (ρ⋆ )) ∈ B(xj , 72 R). To do
this, recall that ρ⋆ ∈ Wl ⊂ Γ̃, Wl ⊂ S ^∗ M for some hypersurface Z ⊂ M , H is transverse
Z |ξ|g
ψ
to Wl and ̥ is chosen small enough that there are local coordinates Ω ∋ (y1 , y ′ ) 7→ V ⊂ M
with Z = ψ({y1 = 0} ∩ Ω), πM (ϕt (Wl )) ⊂ V for |t| ≤ ̥, and H|ξ|g (ψ −1 )∗ y1 > c > 0 on ϕt (Wl )
for |t| ≤ ̥. Therefore, y1 (πM (ϕt (ρ⋆ ))) ≥ ct. In particular, since πM (ϕgt⋆ (ρ⋆ )) ∈ B(xj0 , R + Ry )
88 YAIZA CANZANI AND JEFFREY GALKOWSKI

and t⋆ ∈ (1/6̥, ̥/3), this implies for R small enough (depending only on G and ̥), we have
B(xj0 , 27 R) ⊂ {y1 > 0}.
We claim that ϕgT0 (ρ⋆ ) ∈ Wl . To see this, note that if R̃ is small enough, then ρ0 ∈ Wl .
This implies that PI(m0 ) [g](ρ0 ) ∈ Wl , and hence, since B(ρ0 , R̃) ∈ DI(m0 ) [g], PI(m0 ) [g](ρ⋆ ) ∈ Wl .
0 0 0
Therefore, the claim holds. Now, since ϕgT0 (ρ⋆ ) ∈ Wl , we have y1 (πM (ϕgT0 −s (ρ⋆ )) ≤ −cs for
0 ≤ s < ̥. In particular, since T0 ≤ n̥+CG ε+C n R̃ < (n+ 12 )̥ (for ε and R̃ chosen small enough
as above), for t ∈ [(n − 21 )̥, T0 ], πM (ϕt (ρ⋆ )) ∈ {y1 ≤ 0}, and hence πM (ϕgt (ρ⋆ )) ∈
/ B(xj0 , 72 R), as
claimed. This proves (10.18).
Shrinking δ0 if necessary and collecting (10.15), (10.12),(10.13),(10.16), and that, by (9.33),
kQR,δ (g, σ) − gkG 4 < CδR−5−y(2d−1)+d < R2 δR−7−d(2y−1)+y cn+1 < cn+1 < R2 e−C1 (n+1)̥ /C1 ,
(10.19)
b
for c chosen small enough, provided δ < R cn+1 with b = 7−y+d(2y−1). we conclude that (10.18)
yields that the hypotheses of Lemma 9.12 hold with ρ = ρ⋆ , g⋆ = g, ε = max(Ry , kRy − g0 kG 4 ),
g0 = QR,δ (g, σ), R0 = 76 R, provided
Ry ≤ cR̃ ≤ min(cn+1 , cn R2 ), R ≤ cn α, δ < R7−y+d(2y−1) cn+1 .
Next, noticing that, since ϕgt⋆⋆ (ρ⋆ ) = ρi0 ,
QR,δ (g, σ + δσi0 ) = gδσi0 (ρi0 , 0, R, g0 , δσi0 ) = gδσi0 (ρ⋆ , t⋆ , R, g0 , δσi0 ),
with Pσ : BΓ (ρ⋆ , Ry ) ⊂ Wi0 → Wim0 as defined in Lemma 9.12 with T0 = TI(m0 ) [g](ρ⋆ ) we have
0


R,δ (m0 ) (m0 )
Ψg,I (m0 , σ + δσi0 , ρ) = PI [gδσi0 ](ρ) , dρ PI [gδσi0 ] (ρ)
0 0 0

= (Pδσi0 (ρ), dρ Pδσi0 (ρ)).

Since y > 2
Ry < R2 Ry−2 < R2 cn(y−2) < R2 e−C1 (n+1)̥ /C1 , (10.20)
for c chosen small enough and C1 as in Lemma 9.12. Note that we have used y > 2 and d ≥ 2
to obtain the next to last inequality. In particular, (10.20) and (10.19) show that Lemma 9.12
implies (7.6): that for all ρ ∈ BΓ (ρ⋆ , Ry ),
|dσi0 ΨR,δ n
g,I (m0 , σ, ρ)v| ≥ δc |v|, v ∈ Tσi0 BRL (0,1) .
0
y ,R,δ
From Definition 7.3 it follows that (Γ, G) is (̥, b, y) admissible for QR with b = 7 − y + d(2y −
1). 

11. Proof of Theorem 1.3

We now use Proposition 8.9 to prove Theorem 1.3. Let ν ≥ 5 and Ω > Ων .
y
Let y > 2 to be chosen close to 2 later, and Q̃R,δ (g, σ) := QR ,R,δ (g, σ) with Q as in (9.31). Let
Rε ,δε
F := {(F∞ , ∞)}ε be the G ν−1 family of probing maps for G ν constructed as in Proposition 8.9
(see also Lemma 8.1).
S
Let Kn ⊂ G ν bounded such that Kn ⊂ Kn+1 and n Kn = G ν . Then for each g ∈ Kn , by
Lemmas 10.2 and 10.4, there are rg > 0, ̥g > 0, and a symplectic manifold Γg such that (Γg , Gg )
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 89

is ̥g well separated with Gg := BG 3 (g, 2rg ) ∩ G ν . Moreover, (Γg , Gg ) is (̥g , b, y) admissible for
Q̃R,δ with b = 7 − y + d(2y − 1).
SNn
Since ν > 3, Kn is compact in G 3 , there are {gi,n }N
i=1 such that Kn ⊂
n
i=1 BG 3 (gi,n , rgi,n ). We
∞ Nn ν
relabel {gi }i=1 = ∪n ∪i=1 {gi,n }. Next, fix K ⊂ G bounded and let NK be such that
NK
[
K⊂ BG 3 (gi , rgi ).
i=1

To each i ∈ {1, . . . , NK } apply Proposition 8.9 to (Γgi , Ggi ∩ K) in place of (Γ, K). Let εi , Ci be
the constants ε0 , C given by the proposition. Let εK,1 = min1≤i≤NK εi .
Next, let
Rε ,δε
εK,2 = sup{ε > 0 : kF∞ (g, σ) − gkG ν < min rgi for all g ∈ K, σ ∈ Σ∞ },
1≤i≤NK

and set εK := min(εK,1 , εK,2 ) > 0. By Proposition 8.9, for each i ∈ {1, . . . , NK } and 0 < ε < εK
there exists Sg,ε (gi ) ⊂ Σ∞ (Rε ) an mΣ∞ (R) measurable set such that for all g ∈ K ∩ Ggi


σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ L∞ (ε, gi ) ⊂ Sg,ε (gi ), sup mΣ∞ (R) Sg,ε (gi ) ≤ ε, (11.1)
g∈K∩Ggi
n o
where L∞ (ε, gi ) := g ∈ G ν : ∃n such that Γgi ∩ R̥g (n, βn (ε, gi ), g) * N ̥g (n, βn (ε, gi ), g) ,
i i
with
γ γ −1 nγ
βn (ε, gi ) := εCi n Ci−n n−Ci log ε ,
where, as in (8.79), γ = γν (y) is given by


γν (y) := 1 + log2 max 2ℵ(2ℵ+3)
2ℵ+1 mν (y), 2ℵy , mν (y) := 1 + ν + (2d − 1)y − d.
Here, we have used mν (y) = 1 + max(ν, 6) + (2d − 1)y − d = max(b, ϑν ) and that ϑν = 1 + ν +
(2d − 1)y − d by (9.33) with r = Ry .
Then, let
 there exists C > 0 such that for all n > 0 
G̃i := g ∈ Ggi γ γ . (11.2)
Γgi ∩ R̥g (n, (Cn)−Cn , g) ⊂ N ̥g (n, (Cn)−Cn , g)
i i

S
We claim that G̃ := i G̃i is F -predominant. To see this, fix ε > 0. We claim that for all
g ∈ K, there is an mΣ∞ (R) -measurable set Sg,ε such that

σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ G̃c } ⊂ Sg,ε , mΣ∞ (R) (Sg,ε ) ≤ ε. (11.3)

To obtain (11.3), fix g ∈ K. Then there is i such that g ∈ K∩GG 3 (gi , rgi ) and F∞Rε ,δε (g, σ) ∈ Ggi
for all σ ∈ Σ∞ (Rε ) (by our choice of εK,2 ). Now, since for all i and ε > 0 there is C > 0 such
γ
that (Cn)−Cn ≤ βn (ε, gi ) for all n, we have
\
G̃c ∩ Ggi ⊂ L∞ (s, gi ) ∩ Ggi .
s>0
90 YAIZA CANZANI AND JEFFREY GALKOWSKI

Therefore,
  \
σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ G̃c } ⊂ σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ L∞ (s, gi ) ∩ Ggi
s>0

⊂ σ ∈ Σ∞ (Rε ) : F∞Rε ,δε (g, σ) ∈ L∞ (ε, gi ) ∩ Ggi ⊂ Sg,ε (gi ).
In particular, (11.3) follows from (11.1).
We next show that if Ω > Ων , we may choose y > 2 such that if g ∈ G̃, and W = {W Ui }i is a
family of transition maps for g, then there is CΩ > 0 such that for all t > ̥gi /2
 −CΩ (t+1)Ω −1
  −CΩ (t+1)Ω −1

R t, CΩ , g ⊂ N t, CΩ , (g, W) . (11.4)

To see (11.4), let g ∈ G̃ and i such that g ∈ G̃i . Let Cg > 0 be such that for all n > 0
γ
γ

Γgi ∩ R̥g n, (Cg n)−Cg n , g ⊂ N ̥g n, (Cg n)−Cg n , g . (11.5)
i i

γ
Suppose that t > ̥gi /2 and ρ ∈ R(t, (Λ(t + 1))−Λ(t+1) , g) for some Λ > 0. Define

s(ρ) := inf{s ≥ 0 : ϕgs (ρ) ∈ Γgi }, s̃± (ρ) := inf{s ≥ 0 : ϕg±s (ρ) ∈ Γ̃gi }
where Γ̃gi is defined as in (6.2). In what follows, C, c are two positive constants that depend only
on K. Note also that supρ∈S ∗M s(ρ) < C injgi (M ). Define ρ+ := ϕgs(ρ) (ρ) and observe that there
is a choice of ± such that
γ
d(ρ+ , ρt ) < C(Λ(t + 1))−Λ(t+1) , ρt := ϕ±s̃± (ϕgt (ρ)) (ϕgt (ρ)),
γ γ
|s(ρ) − s̃± (ϕgt (ρ))| < C(Λ(t + 1))−Λ(t+1) , or |s(ρ) + s̃± (ϕgt (ρ))| < C(Λ(t + 1))−Λ(t+1) .
In particular, by Remark 6.10, there are I and m such that
γ
PI(m) [g](ρ+ ) = ρt , |TI(m) [g](ρ+ ) − t| < C(Λ(t + 1))−Λ(t+1) ,
1 t 1 t
and so, choosing Λ > 0 large enough, there is n with 4 < ̥gi − 4 < n ≤ ̥gi + 43 , such that
γ
ρ+ ∈ Γgi ∩ R̥g (n, (cΛn)−cΛn , g). Furthermore, choosing Λ large enough, we have cΛ > Cg and
i
hence, by (11.5),
γ
ρ+ ∈ N ̥g (n, (Cg n)−Cg n , g).
i

Now, let v ∈ Tρ (S ∗M )g . Then, there is w ∈ RH|ξ|g (ρ) such that

dϕs(ρ) ṽ ∈ Tρ+ Γgi , ṽ := w + v.
Therefore, choosing coordinates on Γgi near ρ+ to identify Tρ+ Γgi and Tρt Γgi , and writing the
identification of tangent spaces as W̃ρt ,ρ+ : Tρ+ Γgi → Tρt Γgi ,
γ
k(dPI(m) [g] − W̃ρt ,ρ+ )dϕgs(ρ) ṽk ≥ (Cg n)−Cg n kṽk.
In particular,
γ
k(dϕg−s(ρ) dPI(m) [g]dϕgs(ρ) − dϕg−s(ρ) W̃ρt ,ρ+ dϕgs(ρ) )ṽk ≥ c(Cg n)−Cg n kṽk.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 91

γ
Now, there is t̃ with |t̃ − t| < C(Λ(t + 1))−Λ(t+1) , such that
dϕg−s(ρ) dPI [g](m) dϕgs(ρ) = dϕgt̃ ,
and hence
γ
k(dϕgt̃ − dϕg−s(ρ) W̃ρt ,ρ+ dϕgs(ρ) )ṽk ≥ c(Cg n)−Cg n kṽk.
Applying dϕgt−t̃ on the left, we have
γ
k(dϕgt − dϕgt−t̃ dϕg−s(ρ) W̃ρt ,ρ+ dϕgs(ρ) )ṽk ≥ c(Cg n)−Cg n kṽk. (11.6)

Now, the map dϕgt−t̃ dϕg−s(ρ) W̃ρt ,ρ+ dϕgs(ρ) identifies Tρ S ∗M/RHp with Tϕgt (ρ) S ∗M/RHp and has
uniformly bounded derivatives in t. Suppose W := W Uk is a transition map such that (ρ, ϕgt (ρ)) ∈
Uk . Then, there is CW > 0
γ
kWϕgt (ρ),ρ − dϕgt−t̃ dϕg−s(ρ) W̃ρt ,ρ+ dϕgs(ρ) k ≤ CW d(ϕgt (ρ), ρ) ≤ CW (Λ(t + 1))−Λ(t+1) . (11.7)
Hence, letting Λ be large enough, there is CW,g > 0, depending on g and W, such that
γ
k(dϕgt − Wϕgt (ρ),ρ )ṽk ≥ (CW,g (t + 1))−CW,g (t+1) kṽk. (11.8)
Next, we claim that
γ
k(dϕgt − Wϕgt (ρ),ρ )v + RHp k ≥ (CW,g (t + 1))−CW,g (t+1) kv + RHp k. (11.9)
γ
Note that then (11.9) implies ρ ∈ N (t, (CW,g (t + 1))−CW,g (t+1) , g). Thus, increasing Λ again if
necessary, we would conclude that there is CW,g = CW,g (g, W) > 0 such that, for t > ̥gi /2,
γ γ
R(t, (CW,g (t + 1))−CW,g (t+1) , g) ⊂ N (t, (CW,g (t + 1))−CW,g (t+1) , g). (11.10)

To prove (11.9) recall the smooth decomposition Tρ S ] ∗M = H (ρ)⊕RH (ρ) and the fact that, by
p
Definition (1.3), W preserves this splitting. Therefore, letting v = vH + tHp (ρ) with vH ∈ H (ρ),
(dϕgt − Wϕgt (ρ),ρ )v = (dϕgt − Wϕgt (ρ),ρ )ṽ = (dϕgt − Wϕgt (ρ),ρ )vH ∈ H (ϕgt (ρ)).
In particular, since H (ρ) is uniformly transverse to Hp (ρ), there is c > 0 such that
k(dϕgt − Wϕgt (ρ),ρ )v + RHp (ϕgt (ρ))k = inf k(dϕgt − Wϕgt (ρ),ρ )v + sHp (ϕgt (ρ))k
s∈R
(11.11)
≥ ck(dϕgt − Wϕgt (ρ),ρ ṽk.
Next, since v = ṽ + w, with w ∈ RHp , we have
kṽk ≥ inf kv + sHp k = kv + RHp k. (11.12)
s∈R

Combining (11.8), (11.11), and (11.12) implies (11.9).

Finally, to finish the proof of Theorem 1.3, we observe that y 7→ γ(y) is an increasing function
and Ων = γν (2). Therefore, for every Ω > Ων there is y > 2 such that γ(y) < Ω. In particular,
there is CΩ > 0 such that for t > ̥gi /2,
Ω −1 γ
CΩ−CΩ (t+1) < (C(t + 1))−C(t+1) ,
92 YAIZA CANZANI AND JEFFREY GALKOWSKI

and hence (11.10) implies that (11.4) holds for g ∈ G̃i and t > ̥gi /2. Thus, since ∪i G̃i is
predominant, Theorem 1.3 follows after recalling that Lemma 6.6 implies that for all i there is
c > 0 such that for g ∈ Gi ,
d(ϕgt (ρ), ρ) ≥ c|t|, |t| < ̥gi .


Appendix A. Elementary Control theory for ODEs

Let ℵ ∈ N, R > 0, t⋆ ∈ R, and for 1 ≤ i ≤ ℵ and t ∈ R define

R
fi,t (t) := 1 1
R χ( R (t − t⋆ ))ei , ei := (0, . . . , 0, 1, 0, . . . , 0)t ,
⋆
| {z } | {z }
i−1 ℵ−i

where χ ∈ Cc∞ ((−2, 2); [0, 1]) with

´
χ = 1.
Lemma A.1 (Observation). Let L ∈ L∞ (R; M(ℵ)) and u0 ∈ Rℵ and suppose that u solves
u̇ + L∗ u = 0, u(0) = u0 .
Then, there are C, R0 > 0 depending on kLkL∞ such that for any t0 ∈ R, 0 < R < R0 , and T ∈ R
ℵ
X
ku(T )k2 ≤ CeC(|T −t⋆ |) R
|hu, fi,t i 2 |2 .
⋆ Lt
i=1

Proof. Note that

¨ t
R R
hu, fi,t⋆
i − u (t
i 0 ) = hu(t 0 ), f i,t⋆ i − u (t
i 0 ) − L∗ (s)u(s)fi,t
R (t)dsdt
⋆
0
¨ t
=− L∗ (s)u(s)fi,t R (t)dsdt.
⋆
0
Then, since there is C > 0 depending only on kL∗ kL∞ such that ku(s)k ≤ Cku(t0 )k for |s−t0 | ≤ 1,
we obtain for R < 21 ,
¨ t ˆ 2R
∗ R
L (s)u(s)fi,t⋆ (t)dsdt ≤ Cku(t0 )k kL∗ (s)kds ≤ CRku(t0 )kkL∗ kL∞ .
0 −2R

R i − u (t ) ≤ CRkL∗ k ∞ ku(t )k for R < 1 . Therefore,

In particular, hu, fi,t⋆ i 0 L 0 2
ℵ
X
R 2 1
hu, fi,t⋆
i ≥ ku(t0 )k2 ( − CR2 kL∗ k2L∞ ).
2
i=1
1
In particular, taking R < 2CkL∗ kL∞ , we have
ℵ
X
R 2 1
hu, fi,t⋆
i ≥ ku(t0 )k2 .
4
i=1
The claim now follows from standard estimates on first order ODE’s. 
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 93

Lemma A.2 (Control). Let L ∈ L∞ (R; M(ℵ)). There are C, R0 > 0, depending only on kLkL∞ ,
such that for all 0 < R < R0 , v ∈ Rℵ , t⋆ > 2R and T > t⋆ + 2R, there is a ∈ Rℵ such that the
solution, w to
ℵ
X
R
ẇ − Lw = ai fi,t ⋆
, w(0) = 0 (A.1)
i=1

satisfies w(T ) = v and, moreover, kak ≤ CeC|T −t⋆ | kvk.

Proof. Define a map K : Rℵ → Rℵ by

Xˆ
R
K(u0 ) = hu, fi,t⋆
i(s)dsei
i

where u solves
u̇ + L∗ u = 0, u(0) = u0 . (A.2)
Next, fix v ∈ Rℵ and define the linear map ℓv : K(Rℵ ) → R by

ℓv (K(u0 )) = hu(T ), vi

where again u is as in (A.2). Then, by Lemma A.1, there are R0 , C > 0 such that for 0 < R < R0 ,

ku(T )k ≤ CeC|T −t⋆ | kK(u0 )k.

In particular, ℓv is a bounded linear function on the subspace K(Rℵ ) with kℓv k ≤ CeC|T −t⋆ | kvk.
Thus, by the Hahn–Banach theorem, we can extend ℓ to a linear functional on Rℵ with the same
norm and there is a ∈ Rℵ with
kak ≤ CeC|T −t⋆ | kvk
such that ℓ(b) = hb, ai, for all b ∈ Rℵ . Now, suppose that w solves
X
R
ẇ − Lw = ai fi,t ⋆
, w(T) = v.
i

Then, integrating by parts, we have for u solving (A.2),

ˆ T X ˆ T
R
−h ai fi,t⋆ , ui(s)ds = hw(s), u̇ + L∗ ui(s)ds + hw(0), u(0)i − hw(T ), u(T )i
0 i 0

= hw(0), u0 i − hv, u(T )i.

But, by construction of a, and that T > t⋆ + 2R > 4R,

DXˆ E
R
hu, fi,t⋆
i(s)ds e i , a = hu(T ), vi.
i

In particular, hw(0), u0 i = 0 for all u0 ∈ Rℵ and hence w(0) = 0 and the lemma is proved. 
94 YAIZA CANZANI AND JEFFREY GALKOWSKI

References
[Abr70] R. Abraham. Bumpy metrics. In Global Analysis (Proc. Sympos. Pure Math., Vol. XIV, Berkeley, Calif.,
1968), pages 1–3. Amer. Math. Soc., Providence, R.I., 1970.
[Ano82] D. V. Anosov. Generic properties of closed geodesics. Izv. Akad. Nauk SSSR Ser. Mat., 46(4):675–709,
896, 1982.
[Ava56] Vojislav G. Avakumović. Über die Eigenfunktionen auf geschlossenen Riemannschen Mannigfaltigkeiten.
Math. Z., 65:327–344, 1956.
[Bér77] Pierre H. Bérard. On the wave equation on a compact Riemannian manifold without conjugate points.
Math. Z., 155(3):249–276, 1977.
[Bon17] Yannick Bonthonneau. The Θ function and the Weyl law on manifolds without conjugate points. Doc.
Math., 22:1275–1283, 2017.
[Bow72] Rufus Bowen. Periodic orbits for hyperbolic flows. Amer. J. Math., 94:1–30, 1972.
[BW17] B. Bourgain and N. Watt. Mean square of zeta function, circle problem and divisor problem revisited.
arXiv:1709.04340, 2017.
[CG20] Yaiza Canzani and Jeffrey Galkowski. Weyl remainders: an application of geodesic beams.
arXiv:2010.03969, 2020.
[Con10] Gonzalo Contreras. Geodesic flows with positive topological entropy, twist maps and hyperbolicity. Ann.
of Math. (2), 172(2):761–808, 2010.
[DG75] J. J. Duistermaat and V. W. Guillemin. The spectrum of positive elliptic operators and periodic bichar-
acteristics. Invent. Math., 29(1):39–79, 1975.
[dlCS21] Ralph J. de la Cruz and Philip Saltenberger. On the density of semisimple matrices in indefinite scalar
product spaces. Electron. J. Linear Algebra, 37:387–401, 2021.
[DZ16] Semyon Dyatlov and Maciej Zworski. Dynamical zeta functions for Anosov flows via microlocal analysis.
Ann. Sci. Éc. Norm. Supér. (4), 49(3):543–577, 2016.
[Hal15] Brian Hall. Lie groups, Lie algebras, and representations, volume 222 of Graduate Texts in Mathematics.
Springer, Cham, second edition, 2015. An elementary introduction.
[Hin84] N. Hingston. Equivariant Morse theory and closed geodesics. J. Differential Geom., 19(1):85–116, 1984.
[HN12] Joachim Hilgert and Karl-Hermann Neeb. Structure and geometry of Lie groups. Springer Monographs in
Mathematics. Springer, New York, 2012.
[Hör68] Lars Hörmander. The spectral function of an elliptic operator. Acta Math., 121:193–218, 1968.
[HSY92] Brian R. Hunt, Tim Sauer, and James A. Yorke. Prevalence: a translation-invariant “almost every” on
infinite-dimensional spaces. Bull. Amer. Math. Soc. (N.S.), 27(2):217–238, 1992.
[Hux03] M. N. Huxley. Exponential sums and lattice points. III. Proc. London Math. Soc. (3), 87(3):591–609,
2003.
[Ivr80] V. Ja. Ivriı̆. The second term of the spectral asymptotics for a Laplace-Beltrami operator on manifolds
with boundary. Funktsional. Anal. i Prilozhen., 14(2):25–34, 1980.
[Ivr16] Victor Ivrii. 100 years of Weyl’s law. Bull. Math. Sci., 6(3):379–452, 2016.
[IW21] A. Iosevich and E. Wyman. Weyl law improvement for products of spheres. Analysis Mathematica,
47(3):593–612, 2021.
[Kal97] V. Yu. Kaloshin. Prevalence in spaces of finitely smooth mappings. Funktsional. Anal. i Prilozhen.,
31(2):27–33, 95, 1997.
[Kal00] Vadim Yu. Kaloshin. Generic diffeomorphisms with superexponential growth of number of periodic orbits.
Comm. Math. Phys., 211(1):253–271, 2000.
[KH07] Vadim Yu. Kaloshin and Brian R. Hunt. Stretched exponential estimates on growth of the number of
periodic points for prevalent diffeomorphisms. I. Ann. of Math. (2), 165(1):89–170, 2007.
[Kli78] Wilhelm Klingenberg. Lectures on closed geodesics. Grundlehren der Mathematischen Wissenschaften,
Vol. 230. Springer-Verlag, Berlin-New York, 1978.
[Kni98] Gerhard Knieper. The uniqueness of the measure of maximal entropy for geodesic flows on rank 1 mani-
folds. Ann. of Math. (2), 148(1):291–314, 1998.
[KS06] Vadim Kaloshin and Maria Saprykina. Generic 3-dimensional volume-preserving diffeomorphisms with
superexponential growth of number of periodic orbits. Discrete Contin. Dyn. Syst., 15(2):611–640, 2006.
 PREDOMINANT WEYL IMPROVEMENTS AND BOUNDS ON CLOSED GEODESICS 95

[KT72] Wilhelm Klingenberg and Floris Takens. Generic properties of geodesic flows. Math. Ann., 197:323–334,
1972.
[Lev52] Boris M. Levitan. On the asymptotic behavior of the spectral function of a self-adjoint differential equation
of the second order. Izvestiya Akad. Nauk SSSR. Ser. Mat., 16:325–352, 1952.
[Mar69] G. A. Margulis. Certain applications of ergodic theory to the investigation of manifolds of negative cur-
vature. Funkcional. Anal. i Priložen., 3(4):89–90, 1969.
[Mil97] John Milnor. Fubini foiled: Katok’s paradoxical example in measure theory. Math. Intelligencer, 19(2):30–
32, 1997.
[OCV11] Kevin C. O’Meara, John Clark, and Charles I. Vinsonhaler. Advanced topics in linear algebra. Oxford
University Press, Oxford, 2011. Weaving matrix problems through the Weyr form.
[Pat99] Gabriel P. Paternain. Geodesic flows, volume 180 of Progress in Mathematics. Birkhäuser Boston, Inc.,
Boston, MA, 1999.
[PS87a] Vesselin Petkov and Luchezar Stojanov. Periods of multiple reflecting geodesics and inverse spectral
results. Amer. J. Math., 109(4):619–668, 1987.
[PS87b] Vesselin Petkov and Luchezar Stojanov. Spectrum of the Poincaré map for periodic reflecting rays in
generic domains. Math. Z., 194(4):505–518, 1987.
[Sto87] Luchezar Stojanov. Generic properties of periodic reflecting rays. Ergodic Theory Dynam. Systems,
7(4):597–609, 1987.
[SZ02] Christopher D. Sogge and Steve Zelditch. Riemannian manifolds with maximal eigenfunction growth.
Duke Math. J., 114(3):387–437, 2002.
[Vol90a] A. V. Volovoy. Improved two-term asymptotics for the eigenvalue distribution function of an elliptic
operator on a compact manifold. Comm. Partial Differential Equations, 15(11):1509–1563, 1990.
[Vol90b] A. V. Volovoy. Verification of the Hamilton flow conditions associated with Weyl’s conjecture. Ann. Global
Anal. Geom., 8(2):127–136, 1990.
[Wey12] Hermann Weyl. Das asymptotische Verteilungsgesetz der Eigenwerte linearer partieller Differentialgle-
ichungen (mit einer Anwendung auf die Theorie der Hohlraumstrahlung). Math. Ann., 71(4):441–479,
1912.
[YC04] Yosef Yomdin and Georges Comte. Tame geometry with application in smooth analysis, volume 1834 of
Lecture Notes in Mathematics. Springer-Verlag, Berlin, 2004.
[Yom85] Y. Yomdin. A quantitative version of the Kupka-Smale theorem. Ergodic Theory Dynam. Systems,
5(3):449–472, 1985.

Department of Mathematics, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA
Email address: [email protected]

Department of Mathematics, University College London, London, UK

Email address: [email protected]