Ann. I. H. Poincaré – AN 30 (2013) 251–274
www.elsevier.com/locate/anihpc
Conformal deformations of the Ebin metric and a generalized Calabi metric on the space of Riemannian metrics
Brian Clarke
∗, Yanir A. Rubinstein
Department of Mathematics, Stanford University, Stanford, CA 94305, USA Received 26 April 2011; received in revised form 3 July 2012; accepted 18 July 2012
Available online 21 July 2012
Abstract
We consider geometries on the space of Riemannian metrics conformally equivalent to the widely studied Ebin L2 metric.
Among these we characterize a distinguished metric that can be regarded as a generalization of Calabi’s metric on the space of Kähler metrics to the space of Riemannian metrics, and we study its geometry in detail. Unlike the Ebin metric, its geodesic equation involves non-local terms, and we solve it explicitly by using a constant of the motion. We then determine its completion, which gives the first example of a metric on the space of Riemannian metrics whose completion is strictly smaller than that of the Ebin metric.
©2012 Elsevier Masson SAS. All rights reserved.
1. Introduction
LetMbe ann-dimensional compact closed manifold, and consider the infinite-dimensional spaceMof all smooth Riemannian metrics on M. The spaceMis endowed with a naturalL2-type Riemannian structure, the Ebin met- ric[14],
gE(h, k)|g:=(h, k)E:=
M
tr
g−1hg−1k
dVg, (1)
whereg∈M,h, k∈TgM,TgMmay be identified with the spaceΓ (S2T∗M)of smooth symmetric(0,2)-tensor fields on M, andg−1h represents the(1,1)-tensor dual to hwith respect to g. This metric has received much at- tention since being introduced in the 1960s [14], see, e.g., [17,18,7,11], and has found various applications, for example in the Weil–Petersson geometry of moduli spaces of Riemann surfaces[16,25]and in the study of the moduli spaceM/Diff(M)of Riemannian structures (e.g.,[14,15,3]). A related pseudo-Riemannian metric, the DeWitt metric [13,23], has been used in the Hamiltonian formulation of general relativity.
Recently, the metric completion of ME of (M, gE) has been determined [9], and it was shown by means of examples that convergence inMEis too weak to control any geometric quantities or to imply geometric convergence of any sort (e.g., Gromov–Hausdorff convergence)[10]. Therefore, it seems natural to look for other metrics onM
* Corresponding author.
E-mail addresses:bfclarke@stanford.edu(B. Clarke),yanir@member.ams.org(Y.A. Rubinstein).
0294-1449/$ – see front matter ©2012 Elsevier Masson SAS. All rights reserved.
http://dx.doi.org/10.1016/j.anihpc.2012.07.003
with the property that their metric completions are strictly contained inME. In other words, metrics for which certain types of degenerations are excluded along convergent sequences. One purpose of this article is to take a first step in this direction by studying conformal deformations of the Ebin metric in the search for metrics with this and other distinguished properties.
Our first observation (Proposition3.1) is that there is a distinguished metric in the conformal class characterized by the property that the tautological vector fieldX|g=gonMis parallel. This metric, which we call thegeneralized Calabi metric(or sometimes thenormalized Ebin metric), is given by
gN:= 1 Vg
gE, g∈M,
whereVg :=Vol(M, g)is the volume function on M. We then restrict attention to conformal factors that depend on the volume, i.e., metrics on M of the form e2f (Vg)gE, with f a smooth function onR>0, and mostly to the metricsgp:=gE/Vp, which serve as the basic models within this family, as they capture the possible degenerations of manifolds in terms of either volume collapse or blow-up. By studying this family of metrics, we then show thatgN
has the smallest metric completion (Theorem5.3), and in particular one that is smaller than that of the Ebin metric.
This provides the first example of anL2-type metric onMwhose metric completion is strictly smaller than that of the Ebin metric.
An additional motivation for introducting gN comes from the study of the subspace of Kähler metrics H⊂M in a fixed Kähler class (whenM admits a Kähler structure). In our previous work[12], we studied the intrinsic and extrinsic geometry ofHinM. We observed that the Ebin metric induces the so-called Calabi geometry onH, and that this embedding is essentially as far from being totally geodesic as possible. It then seems natural to ask whether there exists a metric onMthat still induces the Calabi geometry onHbut with the property thatHis totally geodesic.
As before, it is natural to restrict to conformal deformations depending on the volume, this time since the volume is an invariant of the Kähler class, and so any such metric will induce the Calabi geometry onH. We then show that to the extent possible,gN is the unique metric with the aforementioned property. In particular,His totally geodesic in the case thatMis a Riemann surface. In generalHis not totally geodesic, but by the Calabi–Yau Theorem it is isometric to the “Riemannian Kähler spaces”Pg∩Mv, consisting of metrics of fixed volume in a fixed conformal class, which are totally geodesic in(M, gN)(Corollary4.4).
One further possible application of the metricgNis to the Ricci flow. Recently, we showed that in the Kähler setting there is a connection between the existence of Einstein metrics, the smooth convergence of the normalized Ricci flow, and the metric geometry of(M, gE). Namely, a Kähler–Einstein metric exists on a Fano manifold if and only if the Kähler–Ricci flow converges in the metric completion of(M, gE), and in particular if and only if the flow path has finite length[12, Theorem 6.3, Corollary 6.9]. It would be very interesting to find analogous results for other classes of Riemannian manifolds, perhaps ones for which the singularities of the Ricci flow can be understood fairly well. In studying this problem, it might prove useful to use the metricgN, for which the submanifoldMv⊂Mof metrics of fixed volumev—which is preserved by the normalized Ricci flow—is also totally geodesic (Corollary3.3).
Motivated by these and possible other applications of the metricgNto geometric problems, we thus study the ge- ometry of(M, gN)in detail. Under the conformal change, the geodesic equation becomes substantially more difficult since it contains non-local terms involving integration over the whole manifold. The solution of the geodesic equation is obtained in several steps, building upon the work of Freed and Groisser forgE[17]. A key extra ingredient here is an invariant of thegN-geodesic flow, or a ‘constant of motion’ (Corollary3.2). The solution of the geodesic equation (Theorem 4.1) gives a precise sense to how geodesics in (M, gN)generalize those discovered by Calabi [4,5]for the subspace of Kähler metrics, which in turn bear several similarities with constrained geodesics of the Wasserstein metric in optimal transportation[6] (cf.[12]). We also compute the curvature of gN and compare it to that of the metricsgp(Section3.2).
Finally, it should be noted that “weighted”L2type metrics were also studied by several authors on the space of simple closed curves inR2(see[20,22,24]and references therein), and this can also be seen as another motivation for our study. Moreover, very recently, while the present article was being prepared, Bauer, Harms and Michor[2]have written down the geodesic equation for metrics conformal togEonM, as well as much more general Sobolev-type metrics and metrics weighted by the scalar curvature function. Their main result is that for some of these metrics the exponential mapping is a local diffeomorphism. In this article, we go into greater depth for a smaller class of
metrics by solving the geodesic equation, computing the curvature, estimating the distance function, and determining the metric completion.
The article is organized as follows. In Section2, we briefly review the relevant preliminaries aboutM. In Section3, we discuss general conformal changes, mostly focusing on those involving functions of the volume. For the model metrics gp we compute the curvature as well as find an invariant (or ‘constant of motion’) of the geodesic flow.
Section 4 contains the solution of the initial value problem for gN-geodesics, making use of the invariant of the geodesic flow. In Section5, we study the distance functions ofgpand determine their metric completions. Some of the technical facts needed in this analysis are proven in Appendix A. Section6concludes with some further remarks and a few open questions.
2. Preliminaries
Since the preliminaries relevant to our results are covered in detail in[14,17,18], we will simply briefly summarize what we need in this section.
The manifold of metrics,M, is easily seen to be an open cone in the Fréchet spaceΓ (S2T∗M)of smooth, sym- metric (0,2)-tensors on the finite-dimensional, compact manifoldM. As such, it is endowed with the structure of a Fréchet manifold (cf.[19]for background on Fréchet manifolds), and its tangent space atg∈Mis canonically identified withΓ (S2T∗M).
The Ebin metric, defined in (1), is a smooth Riemannian metric. It is, however, a weak metric, meaning that the tangent spaces ofMare incomplete with respect to the scalar product induced on them by the Ebin metric. For weak Riemannian metrics, the existence of the Levi-Civita connection is not guaranteed by general results. Nevertheless, the Ebin metric has a Levi-Civita connection which can be directly computed. Geodesics and curvature may also be directly computed. The Riemannian curvature of(M, gE)is nonpositive, and the exponential mapping at any point g∈Mis a real-analytic diffeomorphism from an open neighborhood of zero inTgMto an open neighborhood ofg inM. (Both of these neighborhoods are taken in theC∞topology.)
With respect togE, we may orthogonally decompose the tangent spaceTgMinto the subspaces of traceless (sat- isfying tr(g−1h)=0) and pure-trace (satisfyingh=ρg for someρ∈C∞(M)) tensor fields. Corresponding to this decomposition is a product manifold structure forM. Denote byV the space of all smooth, positive volume forms onM; it is an open cone inΩn(M), the space of smooth top-degree forms. For anyg∈Mwe denote bydVg its induced volume form. Then for anyμ∈V, withMμ:= {g∈M: dVg=μ} ⊂M, there is a diffeomorphism
iμ:V×Mμ→M, iμ(ν, h)=(ν/μ)2/nh. (2)
That is,iμmaps(ν, h)to the unique metric conformal tohwith volume formν. ThusM∼=V×Mμ, and one sees that (iμ)∗(TV)is the subbundle ofTMconsisting of pure-trace tensor fields, while a tangent space to the submanifold Mμis identified with the subspace of traceless tensor fields.
An identity that will be repeatedly used below is that the differential of the mapg →dVgish → 12tr(g−1h) dVg. Therefore, if we denote byV =Vg:=
MdVg the volume function onM, then the differential ofg →Vg ish →
1 2(g, h)E.
For the remainder of the paper, all differentiability properties we refer to are implicitly meant to hold in the category of Fréchet manifolds; we again refer to[19]for background.
3. Conformal deformations of the Ebin metric
Letf:M→Rbe a twice continuously differentiable function, and consider the metric onM, gf(h, k)|g:=e2f (g)gE(h, k)|g=e2f (g)
M
tr
g−1hg−1k
dVg, h, k∈TgM, (3)
conformal to the Ebin metric. The purpose of this section is to characterize two metrics in the conformal class ofgE. One metric, the generalized Calabi metricgN=gE/V, is characterized by its Levi-Civita connection (Proposition3.1), and the other, the second Ebin metricg2=gE/V2, by its curvature tensor (Proposition3.6). We then restrict to the model metrics
gp:= 1 Vgp
gE, g∈M, (4)
for some integerp. We find invariants for their geodesic flows that will be important later in integrating the geodesic equation (Corollary3.2and Lemma3.4), compute their curvature (Section3.2), and describe a natural duality map on Mthat is a conformal isometry betweengp andg2−p (Proposition3.8) and that also conformally relates their curvature tensors.
3.1. Conformal deformations and the Levi-Civita connection
Our first observation is a characterization of thegeneralized Calabi metric gN:=g1=gE/Vg, g∈M.
Proposition 3.1.Letf :M→Rbe a differentiable function. Let∇f denote the Levi-Civita connection ofe2fgE, and suppose that∇fg=0, wheregdenotes the tautological vector fieldg →gonM. Thenf (g)= −12logVg+C for some constantC.
Proof. First, note that∇gEg=n4δ, whereδis the Kronecker tensor. This follows from the formula for the Levi-Civita connection ofgE[14, (4.1)],
∇hgEk|g=Dhk−1 2
hg−1k+kg−1h +1
4 tr
g−1k h+tr
g−1h k−tr
g−1hg−1k g
,
wherehandkare any vector fields onMandDhk|g=dtd|t=0k(g+t h).
Next, recall that[1, p. 58],
∇hfk= ∇hgEk+(∇hf )k+(∇kf )h−(h, k)E∇gEf, (5) so∇g=0 is equivalent to
0=n
4h+df (h)g+df (g)h−(h, g)E∇gEf, for allh.
Plugging in h=g shows that ∇gEf is proportional to g; and by inspecting the equation again then necessarily
−n4 =df (g)=Dgf = ∇ggEf anddF (h)g=(h, g)E∇gEf. Combining these two equations yields∇gEf = −4V1g and substituting this back into the second equation yieldsdf (h)= −14(g, h)N. Now consider a path{g(t)}. Then
d dtf
g(t)
= − 1 4V
M
tr
g(t)−1gt
dVg(t)= −1 2
d
dtlogVg(t), hencef (g)= −12logVg+C(asMis path connected), as desired. 2
Since, by the proof above, g is the gradient vector field of 2 logV with respect togN, we have the following corollary, which will prove crucial in integrating the geodesic equation forgNin Section4.
Corollary 3.2.The Hessian oflogV satisfies
∇gNdlogV=0.
In particular,logV is linear andV is either strictly monotone or constant alonggN-geodesics.
By the above corollary, ifg(t)is agN-geodesic and(Vg(t))t(0)=0, thenVg(t)is constant. This gives the following fact.
Corollary 3.3.For anyv∈R+, the submanifoldMv:= {g∈M: Vg=v}is totally geodesic in(M, gN).
As Corollary3.5below will imply, the above statement is true forgponly whenp=1. In particular, it is false for the Ebin metric.
By using the Koszul formula (or else by using (5) and the known expression for∇gE), one can directly compute the Levi-Civita connection ofgNfor constant vector fieldsh,kto be
∇hgNk|g=1
4gN(k, h)g−1 4tr
g−1hg−1k g−1
2hg−1k−1 2kg−1h +1
4tr g−1h
k+1 4tr
g−1k h−1
4gN(h, g)k−1
4gN(k, g)h. (6)
It is torsion free (symmetric inhandk), and one checks directly that it is metric compatible, hence it is the Levi-Civita connection.
It is well known that alonggE-geodesics the volume is quadratic[17]. This is explained by the following lemma and corollary, which are in a similar vein to Corollary3.2.
Lemma 3.4.We have
∇gpdV1−p=n
8(1−p)2gp.
Proof. By (5) and (6) for constant vector fieldsh,k,
∇hgpk|g=p
4(k, h)Ng−1 4tr
g−1hg−1k g−1
2hg−1k−1 2kg−1h +1
4tr g−1h
k+1 4tr
g−1k h−p
4(h, g)Nk−p2
4 (k, g)Nh.
Thus,
∇gpdV1−p(h, k)=1
2∇hgp(1−p)V−p(g, k)E−1
2(1−p)V−p
g,∇hgpk
E
= −p(1−p)
4Vp+1 (g, h)E(g, k)E−1−p
2Vp (h, k)E+1−p 4Vp
g,tr g−1h
k
E
−np(1−p)
8Vp (h, k)E+n(1−p)
8Vp (h, k)E+1−p 2Vp (h, k)E
−1−p 4Vp
g,tr g−1h
k
E+p(1−p)
4Vp+1 (g, h)E(g, k)E
=n
8(1−p)2gp(h, k). 2
Corollary 3.5.Along unit-speedgp-geodesicsg(t),V1−pgrows quadratically(p=1), V1−p(t)= n
16(1−p)2t2+1−p
2 a0t+V1−p(0), (7)
where
a0:= 1 Vgp0
M
f (0) dVg0, f (t ):=tr g−1gt
.
In particular,Vg(t)converges to0 (ifp <1)or to∞(ifp >1)in finite time precisely along constant conformal directions, i.e., ifgt(0)=λg(0)forλ∈Randλnegative(ifp <1)or positive(ifp >1). Also, along a unit-speed gp-geodesic(for anyp∈R),
d dt
1 Vp
M
f dVg
=n
4(1−p). (8)
Proof. Letg(t)be a unit-speedgp-geodesic. By the previous lemma, we have that d2
dt2V1−p(t)= d
dtdV1−p gt(t)
= ∇gt(t)dV1−p gt(t)
=n
8(1−p)2.
This proves thatV1−pgrows quadratically, and (7) follows from the observation that dtdV1−p(0)=1−2pa0.
By the quadratic formula applied to (7), the equation V1−p(t0)=0 has a real solution t0 if and only if a02− nV1−p(0)0. Furthermore, by the Cauchy–Schwarz inequality for the Hilbert–Schmidt scalar product on symmetric matrices, f (t )2=tr(g−1gt)2tr((g−1gt)2)tr(I )=ntr((g−1gt)2), with equality if and only if gt =λg for some λ∈R. Therefore, again using Cauchy–Schwarz,
a02= 1
Vgp0
M
f (0) dVg0
2
Vg10−p 1
Vgp0
M
f (0)2dVg0
Vg10−p n
Vgp0
M
tr
g(0)−1gt(0)2 dVg0
=nVg10−p,
where the last equality holds becausegp(gt(0), gt(0))=1 by assumption. Note that equality holds in the above if and only ifgt(0)is a constant scalar multiple ofg(0). Thus this is the only condition under whicha02−nV1−p(0)0, and the statement aboutVg(t)converging to 0 or∞in finite time follows.
Finally, to prove (8), note thatdtdV1−p(t)=1−2p·Vp1(t)
Mf (t ) dVg(t). The result then follows from differentiating this equation and applying the fact, shown above, that dtd22V1−p(t)=n8(1−p)2. 2
Eq. (8) in the corollary above provides an integral for the geodesic flow ofgp which allows solving the geodesic equation explicitly, in the spirit of the work of Freed–Groisser. This will be carried out forgN using Corollary3.2in Section4.
3.2. Conformal deformations and curvature
Our main purpose in this subsection is to study which conformal deformations ofgEstill have nonpositive curva- ture. The next result shows that there is precisely one metric of the formgE/Vp, besidesgEitself, whose curvature is nonpositive—it is also the unique such metric with curvature conformal to that ofgE—and this characterizes the second Ebin metric
g2=gE/V2,
among all conformal deformations that depend on the volume.
Proposition 3.6.The curvature ofgpis nonpositive if and only ifp=0or2. Moreover, the curvature ofg2=gE/V2 is conformal to the curvature ofgE,
Rg2=RgE/V2,
and this property characterizesg2, up to scaling, among all conformal deformationse2fgEwithf :M→Ra smooth function depending only onVg.
In the proof we make use of the following computation:
Proposition 3.7.The curvature tensor ofgpis given by Rgp= 1
VpRgE+2p−p2 16 gp∧
V2p−2gp⊗gp−n
2Vp−1gp
, (9)
wheregpis the1-form dual to the tautological vector fieldgwith respect togpand∧ denotes the Kulkarni–Nomizu product.
Leth,kbe tangent vectors that are orthonormal with respect togp. The sectional curvature of the planeR{h, k}is secgp(h, k)= 1
VpsecgE(h, k)−2p−p2 16
V2p−2(g, k)2p+V2p−2(g, h)2p−nVp−1
. (10)
Proof. The formula for the curvature under the conformal changegE →e2fgEis[1, p. 58], R= 1
V
RgE+gE∧
∇gEdf−df⊗df+1
2|df|2EgE
, (11)
and this applies in infinite dimensions as can be verified from its proof. (Note that our convention isR(X, Y )Z= (∇X∇Y − ∇Y∇X− ∇[X,Y])Z, the opposite of Besse’s.) Let f (g):= −12logVg andfp(g):=pf (g)= −p2logVg. Assume first thatp=1. We claim that
∇gEdf=1
8(g,·)N(g,·)N− n
16gN=2df⊗df− n
16gN. (12)
To see this, compute using the formula∇hgEg=n4h(cf. the proof of Proposition3.1) and the metric property of∇gE to deduce
∇gEdf (h, k)=
∇hgEdf
(k)= ∇hgE df (k)
−df
∇hgEk
= −1 4∇hgE
1 V(g, k)E
+1
4
g,∇hgEk
N
=1
8(g, h)N(g, k)N− n
16(h, k)N−1 4
g,∇hgEk
N+1 4
g,∇hgEk
N
=1
8(g, h)N(g, k)N− n
16(h, k)N. (13)
Second, note thatdf = −14g1,|df|2N= |∇f|2N=161|g|2N=16n, and|df|2NgN= |df|2EgE. Thus, RgN= 1
VRgE+ 1 16gN∧
g1⊗g1−n 2gN
. (14)
To conclude the proof, note now that∇gEdfp=p∇gEdf =2p df⊗df −pn16gN,dfp⊗dfp=p2df⊗df, and
1
2|dfp|2EgE=p22|df|2EgE=p162ngN. From (11) we thus obtain Rgp= 1
VpRgE+
2p−p2 gp∧
df⊗df− n
32Vp−1gp
.
Sincedf = −14(g,·)N= −14g1= −14Vp−1gp, (9) follows.
Next, recall that
G∧H (a, b, c, d)=G(a, c)H (b, d)+G(b, d)H (a, c)−G(a, d)H (b, c)−G(b, c)H (a, d).
So ifhandkaregp-orthonormal,gp∧gp(h, k, k, h)=2(h, k)2p−2|h|2p|k|2p= −2, and gp∧
gp⊗gp
(h, k, k, h)=2(h, k)p
gp⊗gp (h, k)
−(h, h)p
gp⊗gp
(k, k)−(k, k)p
gp⊗gp (h, h)
= −(g, k)2p−(g, h)2p.
By definition, secgp(h, k)=gp(Rgp(h, k)k, h), and so (10) follows. 2
Proof of Proposition 3.6. Let f =f (Vg) be a smooth function on M. Then df =fdV = 12f(g,·)E, or
∇gEf =12fg and 12|∇gEf|2EgE=n8(f)2V gE, whiledf ⊗df =14(f)2gE⊗gE. A computation similar to (13) gives∇gEdf=14fgE⊗gE+n8fgE. So
Rgf =e2fRgE+1
4e2fgE∧ f−
f2
gE⊗gE+n 2
f+V f2
gE
,
and analogously to the proof of (10), we may compute secgf(h, k)=e2fsecgE(h, k)−f−(f)2
4e4f
(g, k)2f+(g, h)2f
− n 4e2f
f+V f2
.
Suppose now that secgf =e2fsecgE. Then considering directionsh,ktangent to Mμgives that f+V (f)2=0, from which it follows that eitherf= −logV +C, i.e.,gf =e2CgE/V2, or elsef=0, i.e.,gf =e2CgE. 2 3.3. Conformal transformations and a duality map
By Proposition3.7,g2−p andgphave the same curvature tensor, up to a conformal factor. Here we observe that there is also a conformal diffeomorphismF :M→Mthat relates these two metrics, so they are in fact isometric, and in this senseg2does not provide a new geometry compared togE.
Consider the mapF:M→Mdefined byF (g):=Vqg. Leth∈TMbe a constant vector field. Then dF (h)= d
dt
t=0
(g+t h)Vgq+t h=Vqh+1
2q(g, h)EVq−1g.
Hence, a careful computation shows that gpdF (h), dF (k)
F (g)=gp+n
2q(p−1)(h, k)+ n
4q2+q
V(1−p)(1+n2q)−2(g, h)E(g, k)E.
To summarize, we have:
Proposition 3.8. The diffeomorphism F (g)=V−4ng of M is an isometry between the spaces (M, g2−p) and (M, gp), and we haveVF (g)=Vg−1andF−1=F. In particular,(M, g2)and(M, gE)are isometric.
It is interesting to note that using this result one obtains rather effortlessly the solution of the geodesic equation for g2, building on the much simpler one forgE([18, Thm. 3.2],[17, Thm. 2.3]). In fact, a direct solution of the g2-geodesic equation using the fact that the inverse of the volume is quadratic (Corollary3.5) is substantially more involved.
Remark 3.9.Ifφis a positive differentiable function, andF (g):=φ(Vg)g, then dF (h), dF (k)
p
F (g)=φn/2(VF (g))
Vp (h, k)E|g+VF (g)(g, h)E(g, k)E
φn/2φ φ
nφ 4φV +1
.
Hence, the only such mapF that is an isometry betweengEand anygpis given by Proposition3.8.
4. Geometry of the generalized Calabi metric
In this section, we study the geometry of the metricgNin more detail. In the first subsection we solve its geodesic equation for any given initial data. In the second subsection, we compute the sectional curvature, and examine the extrinsic geometry of certain submanifolds in the spirit of[12], showing that the Riemannian analogues of the space of Kähler metrics are totally geodesic. These spaces are naturally isometric (via the Calabi–Yau Theorem) to the usual spaces of Kähler metrics. These facts, together with the explicit formula for geodesics, give a precise meaning to the statement thatgNgeneralizes Calabi’s geometry on the space of Kähler metrics.
4.1. Geodesics
From (6) we obtain the geodesic equation for(M, gN), g−1gt
t=1 4tr
g−1gtg−1gt δ−1
2tr g−1gt
g−1gt+1
2(gt, g)Ng−1gt−1
4|gt|2Nδ, (15)
whereδ denotes the Kronecker tensor corresponding to the identity matrix. The last two terms are the new terms compared to the geodesic equation forgE. Since they are non-local, the solution of the equation becomes substantially more involved and requires making use the ‘constant of motion’ of the geodesic flow found in (8).
The solution of the initial value problem for the geodesic equation is given by the following theorem.
Theorem 4.1.Letg(0)∈M, and letμ0:=dVg(0). Then the geodesic in(M, gN)emanating fromg(0), with initial tangent vector(α, A)∈Tμ0Vol(M)×Tg(0)Mμ0, is given by the following.
Defineσ:= |(α, A)|Nand
a0:= 2 V (0)
M
α, b0:= nσ2−a20
4 , q:= α
μ0−a0
2, r:=
n 4tr
g(0)−1A2 .
First, ifb0=0, theng(t, x)=et σ√ng(0, x).
Ifb0=0, then for eachx∈M, g(t, x)=
1 2
1−q2+r2 b02
cos(b0t )+ q b0
sin(b0t )+1 2
1+q2+r2 b20
2
n
·ea0t /ng(0)exp 2
rtan−1
rsin(b0t )
b0+b0cos(b0t )+qsin(b0t )
g(0)−1A
. (16)
Here, we take the exponential term to be the identity ifA(x)=0.
IfA(x)=0, then in(16), arctangent takes values in[π k−π2, π k+π2]whenevert∈ [2π(k−b01)+θ,2π kb+θ
0 ]fork∈Z, where
θ (x):=
⎧⎪
⎨
⎪⎩
2π−cos−1(q(x)
2−b02
q(x)2+b02) ifq(x)0, cos−1(q(x)
2−b20
q(x)2+b20) ifq(x) <0.
(Here, arccosine takes values in[0, π].)
The domain of definition ofg(t)is[0,∞)ifb0=0. Ifb0=0, the domain of definition is[0, t0), wheret0 is the infimum ofθ (x)at points whereA(x)=0.(We take the infimum to be∞if there are no such points.)In the case where the geodesic exists for only finite time, it approaches a limit point on the boundary ofM⊂Γ (S2T∗M)ast→t0; i.e.,μg(t)(x)→0for at least one pointx∈M.
Remark 4.2.Theorem4.1gives a precise meaning to the statement thatgN generalizes Calabi’s geometry on the space of Kähler metrics. Indeed, on the level of volume forms, Calabi’s geodesics in the space of Kähler metrics (or, via the Calabi–Yau Theorem, on the space of volume forms with total volumev) are given by
dVg(t)=dVg
G√
vsin 1
2t /√ v
+cos
1 2t /√
v 2
,
where(dVg(0))t =G dVg(0) [12, Remark 5.7]. On the other hand, in proving (16) one shows that the volume forms alonggN-geodesics satisfy an equation of a similar form—see (28)—and the two equations can actually be shown to exactly coincide whenA≡0 anda0=0 by using trigonometric formulas and carefully identifying the integration constants.
Before we give the proof of this theorem, let us point out a contrast to the Ebin metric. Like the case ofgE(cf.[17,
§2]), geodesics in(M, gN)exist for all time ifA(x)=0 for allx∈M. However, the converse of this statement also holds—ifA(x)=0 for somex∈M, then the geodesic only exists for finite time—unlessb0=0. (In the case ofgE, this happens only when there is a point whereA(x)=0 and(α/μ0)(x) <0.)
Note also that, as in the case of the Ebin metric, any conformal class—a submanifold of the formPgwithg∈M—
is totally geodesic, as can be seen from Theorem4.1by puttingA≡0.
We will solve the geodesic equation in the following subsections, beginning with general considerations and then considering various special cases.
4.1.1. The general case
We letC:=g−1gt and decompose into pure trace and traceless parts:C=:E+fnI, with trE=0, i.e.,f =trC.
From (15), we obtain the pair of coupled equations Et= −1
2f E+ E 2V
M
f dVg, (17)
and
ft=n 4tr
E2
−f2 4 −nσ2
4 + f
2V
M
f dVg, (18)
whereσ= |gt|N, which is constant sincegis a geodesic. The last term in the first equation and the last two terms in the second equation are new compared to the unnormalized metric.
The following relations hold betweenE,f, and data related to the splittingM∼=V×Mμ0, whereμ0:=dVg0 (cf. Section2). We writeg=(μμg
0)2/nh, whereμg=dVg andh∈Mμ0, i.e.,his the unique metric conformal tog withdVh=μ0. Theng−1gt=h−1ht+2n(μμgg)tI, implying thatE=h−1ht andf=2(μμg)t
g . We define
φ:=f− 1 V
M
f dVg.
Note thatV−1
Mf dVg=2dtd(logVg(t)). Hence, by Corollary3.2, this quantity is constant alongg(t). So defining a0:= 1
V (0)
M
f (0) dVg(0), we haveφ=f −a0andφt=ft.
Now, note that
−f2 4 + f
2V
M
f= −1 4φ2+1
4 1
V
M
f 2
. (19)
Using this, together with the considerations of the previous paragraph, we can rewrite (17)–(18) in terms ofφ, Et= −φ
2E, (20)
φt=n 4tr
E2
−nσ2 4 −φ2
4 +a02
4 . (21)
Note that tr
E2
t =2 tr(EtE)=
V−1
M
f dVg−f
tr E2
= −φtr E2 (so tr(E2)=exp(t
0(V−1
Mf dVg−f ) ds)tr(E2(0))). Hence, differentiating (21) yields φt t= −n
4φtr E2
−1 2φφt,
and substituting for tr(E2)using (21) we obtain 4φt t+6φφt+φ3=φ
a02−nσ2
. (22)
We now let p:=μg
μ0e−a0t /2. (23)
It follows that 2pt/p=2(μg)t/μ0−a0=φ. Thus, the left-hand side of (22) equals 8pt t t/p, hencepsatisfies 4pt t t−
a02−nσ2
pt=0. (24)
Let
b0:= nσ2−a02
4 .
Note thatb0is real-valued and positive, sinceσ= |gt(0)|Nand so σ2= 1
V (0)
M
tr C(0)2
dVg(0) 1 nV (0)
M
f (0)2dVg(0)
1
n 1
V (0)
M
f (0) dVg(0)
2
=a20
n , (25)
where the second inequality is Cauchy–Schwarz. Note that the first inequality is an equality if and only ifE(0)≡0, and the second inequality is an equality if and only iff is constant. Therefore,b0=0 if and only ifgt(0)=λg(0)for someλ∈R.
Now, integrating (24), we have
pt t+b20p=C, (26)
for someC∈R. It follows that p(t )=
C1cos(b0t )+C2sin(b0t )+C3, ifb0=0,
C1t2+C2t+C3, ifb0=0. (27)
By (23), then, μg μ0 =
(C1cos(b0t )+C2sin(b0t )+C3)ea0t /2, ifb0=0,
(C1t2+C2t+C3)ea0t /2, ifb0=0. (28)
We now consider the initial value data needed to determine the constants of integration. Note that (23) implies that p(0)=1 andpt(0)=α/μ0−a0/2.
To determinept t(0), we first use thatφ=2pt/pto see that on the one hand, φt(0)=2
pt
p
t
(0)=2pt t(0)p(0)−(pt(0))2
p(0)2 =2pt t(0)−2 α
μ0−a0
2 2
.
On the other hand, we see by (21) and the fact thatf (0)=2α/μ0that pt t(0)=n
4tr E(0)2
−nσ2 4 −φ2
4 +a02 4 +2
α μ0−a0
2 2
=n 4tr
g(0)−1A2
−1 4
2 α μ0 −a0
2
−b20+2 α
μ0−a0
2 2
=1 2
q2+r2−b20 ,
whereq:=α/μ0−a0/2 andr:=
n
4tr((g(0)−1A)2).
This gives all the information needed to solve forμg/μ0in the individual cases. To solve forh, we must use (20) and the fact thatφ=2pt/pto see thatEt= −(logp)tE, implyingE=E(0)/p=g(0)−1A/p. SinceE=h−1ht, this gives
h−1ht(t)=g(0)−1A/p(t ). (29)
We now give the solution of the geodesic equation for each special case.