betaFeedback

WorksMathematics

Curvature estimates for shrinking gradient Ricci solitons

Abstract

Suppose (M5,g,f)(M^5,g,f) is a complete gradient shrinking Ricci soliton with bounded scalar curvature. We prove that if ν(g)>−log⁡2ν(g)>-\log 2, then its Riemannian curvature is bounded. The proof uses curvature blow-up, a local Ricci estimate, and the splitting given by the rescaled potential. An entropy estimate then allows us to apply the volume-ratio gap theorem for Ricci-flat four-manifolds.

Introduction

A gradient Ricci soliton is a complete Riemannian manifold (Mn,g)(M^n,g) with a smooth function ff satisfying

Ric⁡+∇2f=λg\operatorname{Ric}+\nabla^2 f=\lambda g

for a constant λ∈R\lambda\in\mathbb{R}. The soliton is shrinking, steady, or expanding according as λ>0\lambda>0, λ=0\lambda=0, or λ<0\lambda<0. In the shrinking case we rescale the metric so that

Ric⁡+∇2f=12g(1)\operatorname{Ric}+\nabla^2 f=\frac{1}{2}g \tag*{(1)}

Gradient Ricci solitons generate self-similar solutions to the Ricci flow by diffeomorphisms and scaling. They play a basic role in the analysis of singularities of the flow. In his study of ancient κ\kappa-solutions, Perelman [17, 18] established the structure underlying the three-dimensional singularity analysis. Together with the work of Ni–Wallach [16] and Cao–Chen–Zhu [4], this led to the classification of three-dimensional shrinking gradient Ricci solitons. In higher dimensions, classification results usually require additional curvature assumptions; see, for example, [2, 3, 5, 15, 19, 20, 24, 25].

Curvature boundedness is an important preliminary issue in these classification problems. In dimension four, Munteanu–Wang [14] proved that a scalar curvature upper bound implies an estimate ∣Rm⁡∣≤CR|\operatorname{Rm}|\leq CR, and hence bounded curvature. In dimensions five and higher, the scalar curvature does not give a direct bound for the curvature tensor. It is natural to seek conditions under which bounded scalar curvature still rules out curvature concentration at infinity.

In this paper, we prove the following entropy-gap criterion in dimension five. We use Perelman’s scale-invariant entropy ν(g)=inf⁡τ>0μ(g,τ)\nu(g)=\inf_{\tau>0}\mu(g,\tau), whose definition is recalled in § 2.

Theorem 1.1. Let (M5,g,f)(M^5,g,f) be a fixed complete gradient shrinking Ricci soliton satisfying (1.1). If

sup⁡MRg<∞,ν(g)>−log⁡2,(2)\sup_M R_g<\infty,\qquad\nu(g)>-\log2, \tag*{(2)}

then its Riemannian curvature is bounded:

sup⁡M∣Rm⁡g∣<∞.\sup_M|\operatorname{Rm}_g|<\infty.

We prove the theorem by contradiction. A point-selection argument, the noncollapsing theorem of Li–Wang [12], and their backward pseudolocality theorem [13] produce a complete smooth curvature blow-up limit. Bounded scalar curvature makes this limit Ricci flat. To prove that the limit splits, we use the local Ricci estimate of Chen–Wang [9 Theorem 1.2], in the higher-derivative form recorded by Bamler [1 Proposition 3.6], to obtain the bound

∣Ric⁡Qjg∣≤CQj−1/2\lvert\operatorname{Ric}_{Q_j g} \rvert\le C Q_j^{-1/2}

on each fixed pointed ball. Using this estimate, we show that the normalized potential converges to a function with parallel unit gradient. Thus the nonflat limit has a global splitting

Y5=R×X4,Ric⁡hX=0.Y^5 = \mathbb{R} \times X^4,\qquad\operatorname{Ric}_{h_X} = 0.

Wang’s continuity result for the local μ\mu-functional under pointed smooth Cheeger–Gromov convergence, together with domain monotonicity and exhaustion, provides one way to compare the entropies; see [23]. We give the needed one-sided inequality directly by transferring compactly supported test functions and using the scale invariance of ν\nu. This yields ν(g)≤ν(Y)\nu(g) \le\nu(Y). A Gaussian test-function argument then gives

ν(g)≤ν(Y)≤log⁡AVR⁡(Y),\nu(g) \le\nu(Y) \le\log\operatorname{AVR}(Y),

where AVR⁡\operatorname{AVR} denotes the Euclidean-normalized asymptotic volume ratio. The product splitting preserves this ratio. The entropy assumption therefore implies

AVR⁡(X4)=AVR⁡(Y5)>12.\operatorname{AVR}(X^4) = \operatorname{AVR}(Y^5) > \frac{1}{2}.

The sharp four-dimensional Ricci-flat gap theorem of Honda–Mondino [11 Proposition 3.7] forces XX to be Euclidean, contradicting the nonflatness of the normalized blow-up. The zeroth-order local Ricci estimate is sufficient for the Hessian estimate that produces the splitting.

The paper is organized as follows. Section 2 recalls the soliton identities, entropy, and local estimates used below. Section 3 constructs the Ricci-flat blow-up and proves its global real-line splitting under bounded scalar curvature. Section 4 establishes the entropy–volume estimate and invariance of the asymptotic volume ratio under Euclidean products. In Section 5, we prove Theorem 1.1.

Acknowledgement: The present work is a continuation of the research undertaken by the third author as a master’s student. The third author would like to thank Professor Meng Zhu and Wenqi Li for their constant encouragement and helpful discussions.

Disclosure on AI assistance. The research problem and the principal mathematical ideas were formulated by the authors. The authors used ChatGPT to assist with checking some computations and polishing the exposition. All mathematical claims, citations, and final arguments remain the responsibility of the authors.

Preliminaries

Throughout the paper, manifolds are smooth, connected, and without boundary. A shrinking gradient Ricci soliton is normalized by

Ric⁡g+∇g2f=12g.(3)\operatorname{Ric}_g+\nabla_g^2 f=\frac{1}{2}g. \tag*{(3)}

After adding a constant to ff, we also assume

Rg+∣∇f∣g2=f.(4)R_g+\lvert\nabla f\rvert_g^2=f. \tag*{(4)}

We write Δf=Δ−⟨∇f,∇⋅⟩\Delta_f=\Delta-\langle\nabla f,\nabla\cdot\rangle. We use the metric gg to take gradients, Hessians, traces, and norms unless another metric is written. The standard soliton identities are

dR=2Ric⁡(∇f,⋅),(5)dR=2\operatorname{Ric}(\nabla f,\cdot), \tag*{(5)}
Δf+R=n2,(6)\Delta_f+R=\frac{n}{2}, \tag*{(6)}
Δff=n2−f,(7)\Delta_f f=\frac{n}{2}-f, \tag*{(7)}
ΔfR=R−2∣Ric⁡∣2.(8)\Delta_f R=R-2\lvert\operatorname{Ric}\rvert^2. \tag*{(8)}

For completeness, these identities follow directly from (3). Taking its trace gives (6). The contracted Bianchi identity and the divergence formula for the Hessian give

0=12∇jR+∇jΔf+Ric⁡jℓ∇ℓf.0=\frac{1}{2}\nabla_jR+\nabla_j\Delta f+\operatorname{Ric}_{j\ell}\nabla_\ell f.

Since ∇jΔf=−∇jR\nabla_j\Delta f=-\nabla_jR, this is (5). It follows that

d(R+∣∇f∣2−f)=2Ric⁡(∇f,⋅)+2∇2f(∇f,⋅)−df=0.\begin{aligned} d\left(R+\lvert\nabla f\rvert^2-f\right)=2\operatorname{Ric}(\nabla f,\cdot)+2\nabla^2f(\nabla f,\cdot)-df \\ &=0. \end{aligned}

Thus adding a constant to ff indeed gives (4). Subtracting this identity from (6) proves (7). Finally, taking the divergence of ∇jR=2Ric⁡jℓ∇ℓf\nabla_jR=2\operatorname{Ric}_{j\ell}\nabla_\ell f gives

ΔR=2∇jRic⁡jℓ∇ℓf+2Ric⁡jℓ∇j∇ℓf=⟨∇R,∇f⟩+R−2∣Ric⁡∣2,\begin{aligned} \Delta R=2\nabla_j\operatorname{Ric}_{j\ell}\nabla_\ell f+2\operatorname{Ric}_{j\ell}\nabla_j\nabla_\ell f \\ &=\langle\nabla R,\nabla f\rangle+R-2\lvert\operatorname{Ric}\rvert^2, \end{aligned}

which proves (8). In particular, We also know the fact that every complete gradient shrinking Ricci soliton has nonnegative scalar curvature [8].

We first recall the entropy conventions. For a complete manifold (Nn,g)(N^n,g) put

A(N,g)={u∈Cc∞(N):u≥0, ∫Nu2 dVg=1}.\mathcal{A}(N,g)=\left\{u\in C^\infty_c(N):u\geq0,\ \int_Nu^2\,dV_g=1\right\}.

For u∈A(N,g)u\in\mathcal{A}(N,g) and τ>0\tau>0, define

W(g,u,τ)=∫N[τ(4∣∇u∣g2+Rgu2)−u2log⁡u2] dVg−n−n2log⁡(4πτ),(9)\mathcal{W}(g,u,\tau)=\int_N\left[\tau\left(4\lvert\nabla u\rvert_g^2+R_gu^2\right)-u^2\log u^2\right]\,dV_g-n-\frac{n}{2}\log(4\pi\tau), \tag*{(9)}

where u2log⁡u2u^2\log u^2 is interpreted as zero at u=0u=0. Perelman’s functionals [17] are

μ(g,τ)=inf⁡u∈A(N,g)W(g,u,τ),ν(g)=inf⁡τ>0μ(g,τ).(10)\mu(g,\tau)=\inf_{u\in\mathcal{A}(N,g)}\mathcal{W}(g,u,\tau),\qquad\nu(g)=\inf_{\tau>0}\mu(g,\tau). \tag*{(10)}

Constant metric rescaling satisfies

μ(ag,aτ)=μ(g,τ),ν(ag)=ν(g)(a>0).(11)\mu(ag,a\tau)=\mu(g,\tau),\qquad\nu(ag)=\nu(g)\quad(a>0). \tag*{(11)}

To check the scaling formula, let a>0a>0, put ga=agg_a=ag, and set ua=a−n/4uu_a=a^{-n/4}u. Then

dVga=an/2dVg,Rga=a−1Rg,∣∇ua∣ga2=a−n/2−1∣∇u∣g2.dV_{g_a}=a^{n/2}dV_g,\qquad R_{g_a}=a^{-1}R_g,\qquad\lvert\nabla u_a\rvert_{g_a}^2=a^{-n/2-1}\lvert\nabla u\rvert_g^2.

Consequently ua∈A(N,ga)u_a\in\mathcal{A}(N,g_a) if and only if u∈A(N,g)u\in\mathcal{A}(N,g). The energy term in W(ga,ua,aτ)\mathcal{W}(g_a,u_a,a\tau) is unchanged. The logarithmic term satisfies

−∫Nua2log⁡ua2 dVga=−∫Nu2log⁡u2 dVg+n2log⁡a.-\int_Nu_a^2\log u_a^2\,dV_{g_a}=-\int_Nu^2\log u^2\,dV_g+\frac{n}{2}\log a.

The last summand cancels the change in −n2log⁡(4πaτ)-\frac{n}{2}\log(4\pi a\tau). Hence W(ag,ua,aτ)=W(g,u,τ)\mathcal{W}(ag,u_a,a\tau)=\mathcal{W}(g,u,\tau). Taking the two infima proves (11). For a probability potential FF, with u2=(4πτ)−n/2e−Fu^2=(4\pi\tau)^{-n/2}e^{-F}, the equivalent expression is

W(g,F,τ)=∫M[τ(∣∇F∣g2+Rg)+F−n]e−F(4πτ)n/2 dVg.(12)\mathcal{W}(g,F,\tau)=\int_M\left[\tau\left(|\nabla F|_g^2+R_g\right)+F-n\right]\frac{e^{-F}}{(4\pi\tau)^{n/2}}\,dV_g. \tag*{(12)}

Indeed, on the set where u>0u>0,

4∣∇u∣g2=u2∣∇F∣g2,−log⁡u2=F+n2log⁡(4πτ).4|\nabla u|_g^2=u^2|\nabla F|_g^2,\qquad-\log u^2=F+\frac{n}{2}\log(4\pi\tau).

Substitution in (9) gives (12). We will use this expression for a noncompactly supported density only after showing that its value is the limit of values for compactly supported smooth test functions. This approximation is supplied in the proof of Lemma 4.1.

Proposition 2.1 ([12]). Let (Mn,g,f)(M^n,g,f) be a complete shrinker satisfying (3)–(4). Then

ν(g)=μ(g,1)=log⁡((4π)−n/2∫Me−f dVg)>−∞.(13)\nu(g)=\mu(g,1)=\log\left((4\pi)^{-n/2}\int_M e^{-f}\,dV_g\right)>-\infty. \tag*{(13)}

If F=f+ν(g)F=f+\nu(g), then (4π)−n/2∫Me−F dVg=1(4\pi)^{-n/2}\int_M e^{-F}\,dV_g=1 and W(g,F,1)=ν(g)\mathcal{W}(g,F,1)=\nu(g).

The constant added to obtain a probability potential should be kept separate from the normalization (4). If

I=(4π)−n/2∫Me−f dVg,I=(4\pi)^{-n/2}\int_M e^{-f}\,dV_g,

then Proposition 2.1 says that I=eν(g)I=e^{\nu(g)}. Thus F=f+log⁡IF=f+\log I satisfies

(4π)−n/2∫Me−F dVg=e−log⁡II=1.(4\pi)^{-n/2}\int_M e^{-F}\,dV_g=e^{-\log I}I=1.

Adding this constant changes neither ∇f\nabla f nor ∇2f\nabla^2 f. We continue to use ff with normalization (4) when we differentiate the potential.

Lemma 2.2. Let (Mn,g,f)(M^n,g,f) be a complete noncompact shrinker satisfying (3)–(4). For a fixed reference point oo, there is a constant CC such that

14(dg(o,x)−C)+2≤f(x)≤14(dg(o,x)+C)2.(14)\frac{1}{4}\left(d_g(o,x)-C\right)_+^2\leq f(x)\leq\frac{1}{4}\left(d_g(o,x)+C\right)^2. \tag*{(14)}

In particular, ff is proper and tends to infinity along every divergent sequence. Outside a compact set, the function ρf=2f\rho_f=2\sqrt{f} is smooth and satisfies

∣dρf∣g2=1−4Rgρf2,∇g2ρf=1ρf(g−2Ric⁡g−dρf⊗dρf).(15)|d\rho_f|_g^2=1-\frac{4R_g}{\rho_f^2},\qquad\nabla_g^2\rho_f=\frac{1}{\rho_f}\left(g-2\operatorname{Ric}_g-d\rho_f\otimes d\rho_f\right). \tag*{(15)}

Moreover, ∣dρf∣g≤1|d\rho_f|_g\leq1 wherever f>0f>0.

Proof. The growth estimate is the standard potential estimate of Cao–Zhou [6]; enlarging CC absorbs the choice of the reference point. The positive part is denoted by (a)+=max⁡{a,0}(a)_+=\max\{a,0\}. Differentiating ρf=2f\rho_f=2\sqrt{f} gives dρf=f−1/2dfd\rho_f=f^{-1/2}df. Equation (4) gives the gradient identity in (15). For the Hessian,

∇2ρf=f−1/2∇2f−12f−3/2df⊗df=1ρf(g−2Ric⁡−dρf⊗dρf),\begin{aligned} \nabla^2\rho_f=f^{-1/2}\nabla^2f-\frac{1}{2}f^{-3/2}df\otimes df \\ &=\frac{1}{\rho_f}\left(g-2\operatorname{Ric}-d\rho_f\otimes d\rho_f\right), \end{aligned}

where we used (3). The gradient bound follows from Rg≥0R_g\geq0. The lower growth bound implies that f>0f>0 outside a compact set, so all the differentiations above are valid there. It also implies that sublevel sets of ff are bounded. They are closed, and hence compact by completeness. This proves the properness assertion. □

We next state the local results used in the curvature blow-up. For a Ricci flow g(t)g(t), write

Pg(x0,t0;r,−T,0)=Bg(t0)(x0,r)×[t0−T,t0].\mathcal{P}_{g}(x_{0},t_{0};r,-T,0)=B_{g(t_{0})}(x_{0},r)\times[t_{0}-T,t_{0}].

Here the spatial ball is formed with the metric at the terminal time t0t_{0}. It is the same set of points at each time in the product. Curvature norms and scalar curvature at time tt are computed using g(t)g(t). All such neighborhoods are understood to lie in the time domain of the flow.

Theorem 2.3 (Scalar-curvature noncollapsing, [12]). For a fixed complete shrinker (Mn,g,f)(M^{n},g,f), there is a constant κ>0\kappa>0 such that

Rg≤r−2 on Bg(x,r)⟹Vol⁡gBg(x,r)≥κrn.R_{g}\leq r^{-2}\ \text{on }B_{g}(x,r)\quad\Longrightarrow\quad\operatorname{Vol}_{g}B_{g}(x,r)\geq\kappa r^{n}.

The statement is invariant under constant metric rescaling.

Only a scalar curvature bound is used in this noncollapsing statement. No bound for the full curvature tensor is required. For example, if Rg≤SR_{g}\leq S on all of MM, then every ball with r≤S−1/2r\leq S^{-1/2} satisfies its hypothesis. To check its scaling invariance, put ga=agg_{a}=ag. A ball of radius rr for gag_{a} is the ball of radius r/ar/\sqrt{a} for gg, and

Vol⁡gaBga(x,r)=an/2Vol⁡gBg(x,r/a).\operatorname{Vol}_{g_{a}}B_{g_{a}}(x,r)=a^{n/2}\operatorname{Vol}_{g}B_{g}(x,r/\sqrt{a}).

The scalar curvature assumption transforms into Rg≤(r/a)−2R_{g}\leq(r/\sqrt{a})^{-2}, and the same κ\kappa therefore works for gag_{a}. The assertion is also unchanged by a pullback under a diffeomorphism.

Theorem 2.4 (Backward pseudolocality, [13 Theorem 6.21]). For every 0<α<10<\alpha<1, there is an ε=ε(n,α)>0\varepsilon=\varepsilon(n,\alpha)>0 with the following property. Let g(t)g(t) be a Ricci flow induced by a complete shrinking gradient Ricci soliton. If

Vol⁡g(t0)Bg(t0)(x0,r)≥αrn,∣Rm⁡g(t0)∣≤(αr)−2 on Bg(t0)(x0,r),\operatorname{Vol}_{g(t_{0})}B_{g(t_{0})}(x_{0},r)\geq\alpha r^{n},\qquad\lvert\operatorname{Rm}_{g(t_{0})}\rvert\leq(\alpha r)^{-2}\ \text{on }B_{g(t_{0})}(x_{0},r),

then

∣Rm⁡g(t)∣≤(εr)−2 on Pg(x0,t0;(1−α)r,−(εr)2,0).\lvert\operatorname{Rm}_{g(t)}\rvert\leq(\varepsilon r)^{-2}\ \text{on }\mathcal{P}_{g}(x_{0},t_{0};(1-\alpha)r,-(\varepsilon r)^{2},0).

In the preceding theorem, both hypotheses concern only the terminal ball. The conclusion gives control on a smaller ball for a definite interval before t0t_{0}. The constants depend only on nn and α\alpha, so they do not change when we apply the theorem at different points of the same rescaled flow. In Section 3 we will apply it at every point of a fixed terminal ball. This gives the curvature bound on a whole spatial region, which is needed for the next estimate.

Proposition 2.5 (Local Ricci estimate, [1 Proposition 3.6]). Let g(t)g(t), t∈[−2,0]t\in[-2,0], be a Ricci flow on an nn-dimensional manifold, let (x0,t0)∈M×[−1,0](x_{0},t_{0})\in M\times[-1,0], and let 0<r<10<r<1. Suppose Bg(t0)(x0,r)B_{g(t_{0})}(x_{0},r) is relatively compact and

∣Rm⁡∣<r−2,∣R∣≤η<1on Pg(x0,t0;r,−r2,0),\lvert\operatorname{Rm}\rvert<r^{-2},\qquad\lvert R\rvert\leq\eta<1\quad\text{on }\mathcal{P}_{g}(x_{0},t_{0};r,-r^{2},0),

where η>0\eta>0. Then, for every integer ℓ≥0\ell\geq0,

∣∇ℓRic⁡∣(x0,t0)≤Cℓ(n)η1/2r−ℓ−1.(16)\lvert\nabla^{\ell}\operatorname{Ric}\rvert(x_{0},t_{0})\leq C_{\ell}(n)\eta^{1/2}r^{-\ell-1}. \tag*{(16)}

The zeroth-order estimate was proved earlier by Chen and Wang in On the conditions to extend Ricci flow (III) [9 Theorem 1.2]. Their theorem concerns a complete flow with

∣Rm⁡∣≤2on\lvert\operatorname{Rm}\rvert\leq2\quad\text{on}
Bg(0)(x0,1)×[−1/8,0].B_{g(0)}(x_{0},1)\times[-1/8,0].

and gives

sup⁡Bg(0)(x0,1/2)×[−1/16,0]∣Ric⁡∣≤A(n)(sup⁡Bg(0)(x0,1)×[−1/8,0]∣R∣)1/2\sup_{B_{g(0)}(x_0,1/2)\times[-1/16,0]} |\operatorname{Ric}| \le A(n)\left(\sup_{B_{g(0)}(x_0,1)\times[-1/8,0]} |R|\right)^{1/2}

After parabolic rescaling, this has the form ∣Ric⁡∣(x0,t0)≤C(n)r−1η1/2|\operatorname{Ric}|(x_0,t_0) \le C(n)r^{-1}\eta^{1/2}. Thus Chen–Wang’s theorem gives the case ℓ=0\ell=0 needed for the splitting argument below. The higher-derivative estimates in (16) are quoted from [1 Proposition 3.6].

The estimate (16) is local and does not require the ambient manifold to be compact. In its application below, the radius rr is fixed in the rescaled flow, while η\eta tends to zero. The square-root rate is what controls the rescaled Hessian of the potential. More precisely, ℓ=0\ell=0 gives ∣Ric⁡∣≤C0(n)r−1η1/2|\operatorname{Ric}| \le C_0(n)r^{-1}\eta^{1/2}. The constant is independent of η\eta, so this is a quantitative rate as the scalar curvature tends to zero. Smooth convergence alone would give only ∣Ric⁡∣→0|\operatorname{Ric}| \to0, without a specified rate.

Curvature blow-up and splitting

In this section we establish the blow-up and splitting statements used in the proof of the main theorem. These statements hold in arbitrary dimension. The scalar curvature bound will be used both to obtain a Ricci-flat limit and, through a local Ricci estimate, to produce a parallel direction in that limit.

Lemma 3.1 (Point selection). Let (M,g)(M,g) be a complete noncompact Riemannian manifold with sup⁡M∣Rm⁡g∣=∞\sup_M |\operatorname{Rm}_g|=\infty. There are points pj→∞p_j\to\infty and numbers Qj=∣Rm⁡g∣(pj)Q_j=|\operatorname{Rm}_g|(p_j) such that

Qj≥j4,∣Rm⁡g∣≤4Qjon Bg(pj,jQj).(17)Q_j \ge j^4,\qquad|\operatorname{Rm}_g| \le4Q_j \quad\text{on } B_g\left(p_j,\frac{j}{\sqrt{Q_j}}\right). \tag*{(17)}

Proof. Fix o∈Mo\in M. Curvature is bounded on each compact subset, so we can choose xjx_j with dg(o,xj)≥j+2d_g(o,x_j)\ge j+2 and ∣Rm⁡g∣(xj)≥j4|\operatorname{Rm}_g|(x_j)\ge j^4. Indeed, the closed ball Bg‾(o,j+2)\overline{B_g}(o,j+2) is compact by completeness and has bounded curvature. Unboundedness of curvature therefore allows us to choose xjx_j outside this ball with curvature as large as required.

Start with xjx_j as the current point. If the current point is xx and Q=∣Rm⁡g∣(x)Q=|\operatorname{Rm}_g|(x), stop if ∣Rm⁡g∣≤4Q|\operatorname{Rm}_g|\le4Q on Bg(x,j/Q)B_g(x,j/\sqrt{Q}). Otherwise choose a point in this ball where the curvature is greater than 4Q4Q, and repeat.

After kk moves the curvature is at least 4k∣Rm⁡g∣(xj)4^k|\operatorname{Rm}_g|(x_j). Consequently, counting the first move as k=0k=0, the distance moved at the kkth step is at most 2−kj/∣Rm⁡g∣(xj)2^{-k}j/\sqrt{|\operatorname{Rm}_g|(x_j)}. Thus the total distance from xjx_j is at most

2j∣Rm⁡g∣(xj)≤2j.\frac{2j}{\sqrt{|\operatorname{Rm}_g|(x_j)}}\le\frac{2}{j}.

The same geometric series bounds the distance between any two late points by a tail tending to zero. An infinite sequence of moves would therefore be Cauchy. By completeness it would converge to a point of MM, while its curvature tends to infinity. This contradicts continuity of curvature at the limit point. The process therefore terminates at a point pjp_j satisfying (3.1). Moreover, dg(o,pj)≥j+2−2/jd_g(o,p_j)\ge j+2-2/j, so pj→∞p_j\to\infty.

Lemma 3.2 (Parabolic control and a local Ricci estimate). Let (Mn,g,f)(M^n,g,f) be a fixed complete noncompact shrinker satisfying (3) and (4), with 0≤Rg≤S0\le R_g\le S. Let pj,Qjp_j,Q_j

satisfy (17). Write

gcan(s)=(−s)φs∗g,ddsφs=∇gf−s∘φs,φ−1=id⁡,(18)g_{\mathrm{can}}(s)=(-s)\varphi_s^*g,\qquad\frac{d}{ds}\varphi_s=\frac{\nabla_g f}{-s}\circ\varphi_s,\qquad\varphi_{-1}=\operatorname{id}, \tag*{(18)}
gj(t)=Qjgcan(−1+t/Qj).g_j(t)=Q_jg_{\mathrm{can}}(-1+t/Q_j).

There are c>0c>0 and K<∞K<\infty, independent of jj, such that, for each fixed L<∞L<\infty and all sufficiently large jj,

∣Rm⁡gj(t)∣≤Kon Bgj(0)(pj,L+2)×[−c,0].(19)|\operatorname{Rm}_{g_j(t)}|\le K\quad\text{on }B_{g_j(0)}(p_j,L+2)\times[-c,0]. \tag*{(19)}

For every nonnegative integer mm there is also a constant CmC_m, independent of jj and LL, such that, for all sufficiently large jj depending on LL,

sup⁡Bgj(0)(pj,L)∣∇mRic⁡gj(0)∣gj(0)≤CmQj−1/2.(20)\sup_{B_{g_j(0)}(p_j,L)}|\nabla^m\operatorname{Ric}_{g_j(0)}|_{g_j(0)}\le C_mQ_j^{-1/2}. \tag*{(20)}

The constants may depend on the fixed shrinker, nn, SS, and, in the last estimate, on mm.

Proof. We first note that the flow in (18) is well defined for t<Qjt<Q_j and satisfies gj(0)=Qjgg_j(0)=Q_jg. The potential growth estimate and ∣∇f∣2≤f|\nabla f|^2\le f show that ∇f\nabla f has at most linear growth. Its time-dependent multiple (−s)−1∇f(-s)^{-1}\nabla f thus has a complete flow on every compact subinterval of (−∞,0)(-\infty,0). Differentiating the first formula in (18) and using 2∇2f=g−2Ric⁡g2\nabla^2f=g-2\operatorname{Ric}_g gives

∂sgcan(s)=φs∗(−g+2∇2f)=−2Ric⁡gcan(s).\partial_sg_{\mathrm{can}}(s)=\varphi_s^*(-g+2\nabla^2f)=-2\operatorname{Ric}_{g_{\mathrm{can}}(s)}.

Parabolic rescaling therefore makes gj(t)g_j(t) a Ricci flow as well.

Theorem 2.3 gives κ>0\kappa>0, depending on the fixed shrinker, such that

Vol⁡gBg(q,r)≥κrnwhenever Rg≤r−2 on Bg(q,r).(21)\operatorname{Vol}_g B_g(q,r)\ge\kappa r^n\quad\text{whenever }R_g\le r^{-2}\text{ on }B_g(q,r). \tag*{(21)}

We apply Theorem 2.4, the backward part of Li–Wang’s pseudolocality theorem [13 Theorem 6.21]. This theorem does not require an a priori global curvature bound.

Choose

0<α<min⁡{κ,1/2},b=(2α)−1.0<\alpha<\min\{\kappa,1/2\},\qquad b=(2\alpha)^{-1}.

Fix LL. If j>L+2+bj>L+2+b and q∈Bgj(0)(pj,L+2)q\in B_{g_j(0)}(p_j,L+2), then (17) gives

∣Rm⁡gj(0)∣≤4=(αb)−2on Bgj(0)(q,b).|\operatorname{Rm}_{g_j(0)}|\le4=(\alpha b)^{-2}\quad\text{on }B_{g_j(0)}(q,b).

Here Bgj(0)(q,b)B_{g_j(0)}(q,b) lies in Bgj(0)(pj,j)B_{g_j(0)}(p_j,j), so the point-selected curvature bound applies to this entire ball. Since Rgj(0)=Qj−1Rg≤S/Qj≤b−2R_{g_j(0)}=Q_j^{-1}R_g\le S/Q_j\le b^{-2} for large jj, the rescaled form of (21) gives

Vol⁡gj(0)Bgj(0)(q,b)≥κbn≥αbn.\operatorname{Vol}_{g_j(0)}B_{g_j(0)}(q,b)\ge\kappa b^n\ge\alpha b^n.

Let ε=ε(n,α)>0\varepsilon=\varepsilon(n,\alpha)>0 be the pseudolocality constant. Backward pseudolocality gives a curvature bound (εb)−2(\varepsilon b)^{-2} on a backward neighborhood of (q,0)(q,0). In particular, this bound holds along the worldline with spatial coordinate qq for −(εb)2≤t≤0-(\varepsilon b)^2\le t\le0. We apply the theorem separately at every terminal point q∈Bgj(0)(pj,L+2)q\in B_{g_j(0)}(p_j,L+2). Thus the same bound holds on the whole fixed spatial region for this common time interval, proving (3.3). We can take K=(εb)−2K=(\varepsilon b)^{-2} and c≤(εb)2c\le(\varepsilon b)^2. The same constants work for every fixed LL; only the lower bound on jj changes. We may in particular assume c≤1c\le1.

The scalar curvature of a pullback metric is the pullback of the scalar curvature. Constant metric rescaling divides it by the scaling factor. With s=−1+t/Qjs=-1+t/Q_j, we have

gj(t)=(Qj−t)φs∗g,Rgj(t)(x)=Rg(φs(x))Qj−t.g_j(t)=(Q_j-t)\varphi_s^*g,\qquad R_{g_j(t)}(x)=\frac{R_g(\varphi_s(x))}{Q_j-t}.

The global scalar curvature bound on gg therefore implies, for t≤0t \le0,

0≤Rgj(t)≤SQj−t≤SQj.(22)0 \le R_{g_j}(t) \le\frac{S}{Q_j-t} \le\frac{S}{Q_j}. \tag*{(22)}

On the fixed spatial regions in (19), the curvature bound also bounds Ric⁡\operatorname{Ric} and hence controls the variation of the metric for −c≤t≤0-c \le t \le0. More precisely, since ∣Ric⁡∣≤C(n)K|\operatorname{Ric}| \le C(n)K there, the Ricci flow equation gives

e−2C(n)Kcgj(0)≤gj(t)≤e2C(n)Kcgj(0).e^{-2C(n)Kc}g_j(0) \le g_j(t) \le e^{2C(n)Kc}g_j(0).

Choose a>0a > 0 so small that a<1a < 1, a2<ca^2 < c, and K<a−2K < a^{-2}. For q∈Bgj(0)(pj,L)q \in B_{g_j(0)}(p_j,L), the terminal ball Bgj(0)(q,a)B_{g_j(0)}(q,a) is contained in Bgj(0)(pj,L+2)B_{g_j(0)}(p_j,L+2) and is relatively compact by completeness. Hence its backward parabolic neighborhood of radius aa is contained in the region where both (19) and (22) hold. The choice of aa depends only on cc and KK, and is independent of jj, LL, and qq.

We now apply Proposition 2.5, in the form of Bamler [1 Proposition 3.6], with r=ar=a and η=max⁡{S,1}/Qj\eta=\max\{S,1\}/Q_j, using (22). It gives

∣∇mRic⁡gj(0)∣(q)≤Am(max⁡{S,1}Qj)1/2a−m−1|\nabla^m\operatorname{Ric}_{g_j(0)}|(q) \le A_m\left(\frac{\max\{S,1\}}{Q_j}\right)^{1/2}a^{-m-1}

and proves (20). The time-domain assumption in that proposition causes no further restriction: gj(t)g_j(t) is defined on all of [−2,0][-2,0], whereas its curvature bound is needed only on the smaller parabolic neighborhood. Also, max⁡{S,1}/Qj<1\max\{S,1\}/Q_j < 1 for large jj, as required in that proposition. Absorbing the fixed factors a−m−1a^{-m-1} and max⁡{S,1}1/2\max\{S,1\}^{1/2} into CmC_m proves the claimed estimate uniformly at every point of the fixed ball. For m=0m=0, the same estimate follows directly from Chen–Wang [9 Theorem 1.2] after parabolic rescaling. □\square

Proposition 3.3 (A Ricci-flat blow-up limit). Let (Mn,g,f)(M^n,g,f) be a fixed complete noncompact shrinker satisfying (3) and (4). If 0≤Rg≤S0 \le R_g \le S and sup⁡M∣Rm⁡g∣=∞\sup_M|\operatorname{Rm}_g|=\infty, then, for some sequence pj→∞p_j \to\infty and Qj=∣Rm⁡g∣(pj)→∞Q_j=|\operatorname{Rm}_g|(p_j)\to\infty,

(M,Qjg,pj)⟶(Yn,h,p∞)(23)(M,Q_jg,p_j) \longrightarrow(Y^n,h,p_\infty) \tag*{(23)}

in the pointed smooth topology. The limit is complete and Ricci flat, and

∣Rm⁡h∣(p∞)=1,ν(Y,h)≥ν(M,g).(24)|\operatorname{Rm}_h|(p_\infty)=1,\qquad\nu(Y,h)\ge\nu(M,g). \tag*{(24)}

Proof. Choose pj,Qjp_j,Q_j by Lemma 3.1, and use the rescaled flows of Lemma 3.2. At the terminal time, (17) gives ∣Rm⁡Qjg∣≤4|\operatorname{Rm}_{Q_jg}| \le4 on BQjg(pj,j)B_{Q_jg}(p_j,j). For large jj, (21) also gives

Vol⁡QjgBQjg(pj,1)≥κ,\operatorname{Vol}_{Q_jg} B_{Q_jg}(p_j,1) \ge\kappa,

since RQjg≤S/Qj≤1R_{Q_jg} \le S/Q_j \le1. The curvature and volume bounds on the unit ball give, by the Cheeger–Gromov–Taylor estimate [7], a positive lower bound for the injectivity radius at pjp_j, independent of jj. The parabolic curvature bound (19) and local derivative estimates [22] give bounds for all curvature derivatives on every fixed smaller spatial region and on a common smaller backward time interval. Pointed Hamilton compactness gives a smooth limit Ricci flow on that interval. Here we take a convergent subsequence on a fixed pointed ball, then enlarge the radius through positive integers and take a diagonal subsequence. The resulting convergence is smooth on every compact subset, including at terminal time. The limit has complete spatial slices; the controlled terminal radii tend to infinity, and the metric comparison in the proof of Lemma 3.2 holds uniformly on every fixed ball on the common backward interval. By (22), the scalar curvature of the limiting flow is identically zero. Its scalar evolution equation

(∂t−Δ)R=2∣Ric⁡∣2(\partial_t-\Delta)R=2|\operatorname{Ric}|^2

shows that its Ricci tensor vanishes on the interior of the time interval, and hence also at terminal time by continuity. The Ricci flow equation now gives ∂th(t)=0\partial_t h(t)=0. Thus the limit is static; write its terminal slice as (Y,h,p∞)(Y,h,p_\infty). Smooth convergence at the basepoint gives ∣Rm⁡h∣(p∞)=1|\operatorname{Rm}_h|(p_\infty)=1.

It remains to verify the direction of the entropy inequality. Let u≥0u\geq0 be smooth and compactly supported on YY, with ∫Yu2 dVh=1\int_Y u^2\,dV_h=1, and let τ>0\tau>0. Transfer uu to MM by a pointed convergence map Ψj\Psi_j whose domain contains a neighborhood of supp⁡u\operatorname{supp}u. Define u~j=u∘Ψj−1\widetilde{u}_j=u\circ\Psi_j^{-1} on the image of this domain and extend it by zero on MM. The extension is smooth because uu vanishes near the boundary of the chosen domain. Put

cj2=∫Mu~j2 dVQjg,uj=cj−1u~j.c_j^2=\int_M\widetilde{u}_j^2\,dV_{Q_jg},\qquad u_j=c_j^{-1}\widetilde{u}_j.

Smooth convergence of the metrics on the fixed support gives cj→1c_j\to1. Thus uj∈A(M,Qjg)u_j\in\mathcal{A}(M,Q_jg) for all large jj. The volume forms, gradient terms, and scalar curvature terms converge smoothly on this support. The entropy terms also converge, since z2log⁡z2z^2\log z^2 extends continuously to zero. Consequently,

W(Qjg,uj,τ)⟶W(h,u,τ).\mathcal{W}(Q_jg,u_j,\tau)\longrightarrow\mathcal{W}(h,u,\tau).

Scale invariance of ν\nu yields

ν(M,g)=ν(M,Qjg)≤W(Qjg,uj,τ).\nu(M,g)=\nu(M,Q_jg)\leq\mathcal{W}(Q_jg,u_j,\tau).

Passing to the limit gives ν(M,g)≤W(h,u,τ)\nu(M,g)\leq\mathcal{W}(h,u,\tau). Since both uu and τ\tau were arbitrary, taking the infimum over this full class gives ν(M,g)≤ν(Y,h)\nu(M,g)\leq\nu(Y,h) and proves (24). □

Lemma 3.4 (A global parallel direction). The limit in Proposition 3.3 admits a globally defined smooth function v∞v_\infty satisfying

∣∇v∞∣h=1,∇h2v∞=0.(25)|\nabla v_\infty|_h=1,\qquad\nabla_h^2v_\infty=0. \tag*{(25)}

Consequently,

(Yn,h)≅(R×Xn−1,dt2+hX),(26)(Y^n,h)\cong(\mathbb{R}\times X^{n-1},dt^2+h_X), \tag*{(26)}

where XX is complete, Ricci flat, and nonflat.

Proof. Use the normalization (4) and set ρ=2f\rho=2\sqrt{f} outside a compact set. Lemma 2.2, based on [6], shows that ff is proper, ρ(pj)→∞\rho(p_j)\to\infty, and the identities (15) hold. In particular, ρ\rho is 1-Lipschitz. Since the selected ball has radius j/Qj≤1/jj/\sqrt{Q_j}\leq1/j in gg, it lies outside any fixed compact subset for all sufficiently large jj. The Lipschitz bound therefore gives

inf⁡Bg(pj,j/Qj)ρ≥ρ(pj)−1j⟶∞.\inf_{B_g(p_j,j/\sqrt{Q_j})}\rho\geq\rho(p_j)-\frac{1}{j}\longrightarrow\infty.

For each fixed LL and all j>Lj>L, the ball BQjg(pj,L)B_{Q_jg}(p_j,L) is contained in this selected ball. Thus ρ\rho tends uniformly to infinity on every fixed pointed ball in the rescaled metrics.

Put gj=Qjgg_j=Q_jg and define

vj=Qj(ρ−ρ(pj)).(27)v_j=\sqrt{Q_j}\bigl(\rho-\rho(p_j)\bigr). \tag*{(27)}

Then vj(pj)=0v_j(p_j)=0, and

∣dvj∣gj2=∣dρ∣g2=1−4Rgρ2⟶1(28)|dv_j|_{g_j}^2=|d\rho|_g^2=1-\frac{4R_g}{\rho^2}\longrightarrow1 \tag*{(28)}

locally uniformly. To estimate its Hessian, recall that constant metric rescaling leaves the Levi–Civita connection unchanged and divides the norm of a covariant two-tensor by QjQ_j. In particular, Ric⁡gj=Ric⁡g\operatorname{Ric}_{g_j}=\operatorname{Ric}_g as covariant two-tensors, while ∣Ric⁡g∣g=Qj∣Ric⁡gj∣gj\lvert\operatorname{Ric}_g\rvert_g=Q_j\lvert\operatorname{Ric}_{g_j}\rvert_{g_j}. Since ∣g∣g=n\lvert g\rvert_g=\sqrt{n} and ∣dρ⊗dρ∣g=∣dρ∣g2≤1\lvert d\rho\otimes d\rho\rvert_g=\lvert d\rho\rvert_g^2\le1, we obtain

∣∇gj2vj∣gj=Qj−1/2∣∇g2ρ∣g≤n+1+2∣Ric⁡g∣gQjρ.\lvert\nabla_{g_j}^2v_j\rvert_{g_j}=Q_j^{-1/2}\lvert\nabla_g^2\rho\rvert_g\le\frac{\sqrt{n}+1+2\lvert\operatorname{Ric}_g\rvert_g}{\sqrt{Q_j}\rho}.

Using the scaling of the Ricci norm, this becomes

∣∇gj2vj∣gj≤n+1Qjρ+2Qjρ∣Ric⁡gj∣gj≤n+1Qjρ+Cρ⟶0,(29)\begin{aligned} \lvert\nabla_{g_j}^2v_j\rvert_{g_j}&\le\frac{\sqrt{n}+1}{\sqrt{Q_j}\rho}+\frac{2\sqrt{Q_j}}{\rho}\lvert\operatorname{Ric}_{g_j}\rvert_{g_j}\\ &\le\frac{\sqrt{n}+1}{\sqrt{Q_j}\rho}+\frac{C}{\rho}\longrightarrow0, \tag*{(29)} \end{aligned}

where the second line uses (20) with m=0m=0. The convergence is uniform on every fixed pointed ball. In particular, no bound on Qj/f(pj)Q_j/f(p_j) is needed.

Pull back vjv_j by the convergence maps in (23). The estimate ∣dvj∣gj≤1\lvert dv_j\rvert_{g_j}\le1 and vj(pj)=0v_j(p_j)=0 bound vjv_j on each fixed pointed ball. In a fixed coordinate chart on YY, smooth convergence of the metrics also bounds the first coordinate derivatives of these pulled-back functions. Their second coordinate derivatives satisfy

∂a∂bvj=(∇gj2vj)ab+(Γgj)abd∂dvj.\partial_a\partial_bv_j=(\nabla_{g_j}^2v_j)_{ab}+(\Gamma_{g_j})^d_{ab}\partial_dv_j.

The right-hand side is uniformly bounded on each compact subset. Thus the first derivatives are locally equicontinuous. Arzelà–Ascoli and a diagonal subsequence give local C1C^1 convergence to a function v∞v_\infty on all of YY, with v∞(p∞)=0v_\infty(p_\infty)=0.

Pass the Hessian equation against a compactly supported smooth function in such a chart and passing to the limit, We obtain, in the distributional sense,

∂a∂bv∞=Γabd(h)∂dv∞.\partial_a\partial_bv_\infty=\Gamma^d_{ab}(h)\partial_dv_\infty.

The right-hand side is continuous, so the first derivatives are C1C^1. Repeating this argument with the smooth coefficients of hh shows that v∞v_\infty is smooth. Hence ∇h2v∞=0\nabla_h^2v_\infty=0 classically. Local C1C^1 convergence and (28) also give ∣dv∞∣h=1\lvert dv_\infty\rvert_h=1, proving (25).

Let V=∇v∞V=\nabla v_\infty and X=v∞−1(0)X=v_\infty^{-1}(0). Since YY is complete and ∣V∣=1\lvert V\rvert=1, each integral curve of VV has speed one and cannot leave every compact subset in finite time. The flow Φt\Phi_t therefore exists for every t∈Rt\in\mathbb{R}. Moreover, LVh=2∇h2v∞=0\mathcal{L}_Vh=2\nabla_h^2v_\infty=0, so every Φt\Phi_t is an isometry. Along this flow,

ddtv∞(Φt(y))=∣V∣2=1.\frac{d}{dt}v_\infty(\Phi_t(y))=\lvert V\rvert^2=1.

Every flow line therefore meets XX exactly once. The map

R×X⟶Y,(t,x)⟼Φt(x)\mathbb{R}\times X\longrightarrow Y,\qquad(t,x)\longmapsto\Phi_t(x)

is a smooth bijection, with smooth inverse

y⟼(v∞(y),Φ−v∞(y)(y)).y\longmapsto\left(v_\infty(y),\Phi_{-v_\infty(y)}(y)\right).

Its differential sends the coordinate vector in the R\mathbb{R} direction to VV. Tangent vectors to XX are perpendicular to VV, and the flow preserves their inner products. Its pullback metric is therefore dt2+hXdt^2+h_X, proving the global isometry (26). The hypersurface XX is closed and totally geodesic, hence complete, and the product curvature formula gives Ric⁡hX=0\operatorname{Ric}_{h_X}=0. Finally, XX cannot be flat, because otherwise YY would be flat, contrary to ∣Rm⁡h∣(p∞)=1\lvert\operatorname{Rm}_h\rvert(p_\infty)=1. ▫

Entropy bounds and asymptotic volume growth on Ricci-flat manifolds

Throughout this section, ωn\omega_n denotes the volume of the unit ball in Euclidean Rn\mathbb{R}^n. For a complete connected Ricci-flat nn-manifold (N,h)(N,h) and o∈No \in N, define

AVR⁡(N,h):=lim⁡r→∞Vol⁡hBh(o,r)ωnrn.(30)\operatorname{AVR}(N,h) := \lim_{r\to\infty} \frac{\operatorname{Vol}_h B_h(o,r)}{\omega_n r^n}. \tag*{(30)}

Bishop–Gromov comparison shows that the ratio in (30) is nonincreasing and belongs to (0,1](0,1]. Its limit therefore exists in [0,1][0,1] and is independent of oo. Indeed, if o′∈No' \in N and D=dh(o,o′)D=d_h(o,o'), then, for r>Dr>D,

Bh(o,r−D)⊂Bh(o′,r)⊂Bh(o,r+D).B_h(o,r-D) \subset B_h(o',r) \subset B_h(o,r+D).

After division by ωnrn\omega_n r^n, both outer volume ratios tend to the same limit. Thus changing the centre does not change the asymptotic volume ratio. The entropy ν(N,h)\nu(N,h) is understood as the infimum of the W\mathcal{W}-functional over positive scales and nonnegative, compactly supported smooth test functions of unit L2L^2 norm. It is allowed to equal −∞-\infty.

Lemma 4.1 (Entropy and asymptotic volume ratio). Let (Nn,h)(N^n,h) be a complete connected Ricci-flat manifold. Then

ν(N,h)≤log⁡AVR⁡(N,h),(31)\nu(N,h) \le\log\operatorname{AVR}(N,h), \tag*{(31)}

where log⁡0=−∞\log0=-\infty. In particular, if ν(N,h)>−∞\nu(N,h)>-\infty, then

AVR⁡(N,h)≥eν(N,h)>0.(32)\operatorname{AVR}(N,h) \ge e^{\nu(N,h)} > 0. \tag*{(32)}

Proof. Fix o∈No \in N, write r(x)=dh(o,x)r(x)=d_h(o,x), and put

V(s)=Vol⁡hBh(o,s),θ(s)=V(s)ωnsn,a=AVR⁡(N,h).V(s)=\operatorname{Vol}_h B_h(o,s), \qquad\theta(s)=\frac{V(s)}{\omega_n s^n}, \qquad a=\operatorname{AVR}(N,h).

For τ>0\tau>0, define the Gaussian normalization

Iτ=(4πτ)−n/2∫Ne−r2/(4τ) dVh.(33)I_\tau=(4\pi\tau)^{-n/2}\int_N e^{-r^2/(4\tau)}\,dV_h. \tag*{(33)}

We first prove that

lim⁡τ→∞Iτ=a.(34)\lim_{\tau\to\infty} I_\tau=a. \tag*{(34)}

This limit also follows from [10 Lemma 8.10]; we give the details here. The bound V(s)≤ωnsnV(s)\le\omega_n s^n implies that IτI_\tau is finite and strictly positive. More generally, every Gaussian moment

∫Nrme−cr2 dVh\int_N r^m e^{-cr^2}\,dV_h

is finite for m≥0m\ge0 and c>0c>0. This follows, for example, by decomposing NN into the annuli Bh(o,ℓ+1)∖Bh(o,ℓ)B_h(o,\ell+1)\setminus B_h(o,\ell) and using the Euclidean upper bound for their volumes. These moment bounds will also justify the computations below. The pushforward of dVhdV_h under the distance function is the Stieltjes measure dV(s)dV(s). Radial integration and integration by parts therefore give

Iτ=(4πτ)−n/2∫0∞e−s2/(4τ) dV(s)=(4πτ)−n/2∫0∞s2τe−s2/(4τ)V(s) ds=ωn2(4π)n/2∫0∞θ(τu)un+1e−u2/4 du.(35)\begin{aligned} I_\tau &=(4\pi\tau)^{-n/2}\int_0^\infty e^{-s^2/(4\tau)}\,dV(s)\\ &=(4\pi\tau)^{-n/2}\int_0^\infty\frac{s}{2\tau}e^{-s^2/(4\tau)}V(s)\,ds\\ &=\frac{\omega_n}{2(4\pi)^{n/2}}\int_0^\infty\theta(\sqrt{\tau}u)u^{n+1}e^{-u^2/4}\,du. \tag*{(35)} \end{aligned}

The boundary terms vanish: V(s)→0V(s) \to0 as s↓0s \downarrow0, and the Gaussian dominates the Euclidean upper volume growth as s→∞s \to\infty. The measure

ωn2(4π)n/2un+1e−u2/4 du\frac{\omega_n}{2(4\pi)^{n/2}}u^{n+1}e^{-u^2/4}\,du

has total mass one. In fact,

∫0∞un+1e−u2/4 du=2n+1Γ(n2+1),ωn=πn/2Γ(n/2+1).\int_0^\infty u^{n+1}e^{-u^2/4}\,du=2^{n+1}\Gamma\left(\frac{n}{2}+1\right),\qquad\omega_n=\frac{\pi^{n/2}}{\Gamma(n/2+1)}.

The substitution s=τ us=\sqrt{\tau}\,u in (35) is useful because this probability measure no longer depends on τ\tau. Since 0≤θ≤10\leq\theta\leq1 and θ(τ u)→a\theta(\sqrt{\tau}\,u)\to a for every u>0u>0, dominated convergence yields (34). Moreover, the monotonicity of θ\theta shows directly from (35) that IτI_\tau is nonincreasing as a function of τ\tau.

Consider the normalized Gaussian test function

uτ=(4πτ)−n/4Iτ−1/2e−r2/(8τ),∫Nuτ2 dVh=1.(36)u_\tau=(4\pi\tau)^{-n/4}I_\tau^{-1/2}e^{-r^2/(8\tau)},\qquad\int_N u_\tau^2\,dV_h=1. \tag*{(36)}

Although uτu_\tau need not be smooth at the cut locus and does not have compact support, it is admissible by approximation. The distance function is locally Lipschitz, its weak gradient has length one almost everywhere away from oo, and the cut locus has measure zero. Thus the gradient of uτu_\tau is well defined almost everywhere. The Gaussian moment bounds give

∫N(uτ2+∣∇uτ∣h2+uτ2∣log⁡uτ2∣) dVh<∞.\int_N\left(u_\tau^2+|\nabla u_\tau|_h^2+u_\tau^2|\log u_\tau^2|\right)\,dV_h<\infty.

For fixed τ\tau, multiply uτu_\tau by a radial Lipschitz cutoff χL\chi_L that equals one on Bh(o,L)B_h(o,L), vanishes outside Bh(o,2L)B_h(o,2L), and satisfies 0≤χL≤10\leq\chi_L\leq1 and ∣dχL∣≤C/L|d\chi_L|\leq C/L. Put wL=χLuτw_L=\chi_Lu_\tau. Then

∥wL−uτ∥L2⟶0,∥∇wL−∇uτ∥L2⟶0.\|w_L-u_\tau\|_{L^2}\longrightarrow0,\qquad\|\nabla w_L-\nabla u_\tau\|_{L^2}\longrightarrow0.

For the second assertion, use

∇wL−∇uτ=(χL−1)∇uτ+uτ∇χL.\nabla w_L-\nabla u_\tau=(\chi_L-1)\nabla u_\tau+u_\tau\nabla\chi_L.

The first term tends to zero by integrability of ∣∇uτ∣2|\nabla u_\tau|^2, and the squared L2L^2 norm of the second is at most CL−2∫Nuτ2 dVhCL^{-2}\int_Nu_\tau^2\,dV_h. For the entropy term, write

wL2log⁡wL2=χL2uτ2log⁡uτ2+uτ2χL2log⁡χL2.w_L^2\log w_L^2=\chi_L^2u_\tau^2\log u_\tau^2+u_\tau^2\chi_L^2\log\chi_L^2.

The first term converges in L1L^1 by dominated convergence. The absolute integral of the second is bounded by e−1∫N∖Bh(o,L)uτ2 dVhe^{-1}\int_{N\setminus B_h(o,L)}u_\tau^2\,dV_h, which tends to zero. Finally, ∥wL∥L2→1\|w_L\|_{L^2}\to1, so replacing wLw_L by wL/∥wL∥L2w_L/\|w_L\|_{L^2} preserves all these limits. Each compactly supported Lipschitz approximation can in turn be approximated uniformly and in W1,2W^{1,2} by nonnegative compactly supported smooth functions and normalized again. On a fixed compact set, uniform approximation also gives convergence of the entropy integral, using the continuous extension t2log⁡t2=0t^2\log t^2=0 at t=0t=0. Consequently,

ν(N,h)≤W(h,uτ,τ).\nu(N,h)\leq\mathcal{W}(h,u_\tau,\tau).

Since ∣dr∣h=1|dr|_h=1 almost everywhere away from oo and Rh=0R_h=0, direct substitution into the W\mathcal{W}-functional gives

W(h,uτ,τ)=JτIτ+log⁡Iτ−n,(37)\mathcal{W}(h,u_\tau,\tau)=\frac{J_\tau}{I_\tau}+\log I_\tau-n, \tag*{(37)}
Jτ:=(4πτ)−n/2∫Nr22τe−r2/(4τ) dVh.J_\tau:=(4\pi\tau)^{-n/2}\int_N\frac{r^2}{2\tau}e^{-r^2/(4\tau)}\,dV_h.

To see the normalization explicitly, 4τ∣∇uτ∣2=(r2/(4τ))uτ24\tau|\nabla u_\tau|^2=(r^2/(4\tau))u_\tau^2, whereas

−log⁡uτ2=n2log⁡(4πτ)+log⁡Iτ+r24τ.-\log u_\tau^2=\frac{n}{2}\log(4\pi\tau)+\log I_\tau+\frac{r^2}{4\tau}.

The two quadratic terms therefore add to the moment in (37).

On each compact interval of positive τ\tau values, the differentiated integrand in (33) is bounded by a fixed polynomial in rr times a Gaussian. The moment bounds therefore justify differentiation under the integral sign:

Iτ′=−n2τIτ+(4πτ)−n/2∫Nr24τ2e−r2/(4τ) dVh.I_\tau'=-\frac{n}{2\tau}I_\tau+(4\pi\tau)^{-n/2}\int_N\frac{r^2}{4\tau^2}e^{-r^2/(4\tau)}\,dV_h.

Hence

2τIτ′=Jτ−nIτ,JτIτ=n+2τ(log⁡Iτ)′.(38)2\tau I_\tau'=J_\tau-nI_\tau,\qquad\frac{J_\tau}{I_\tau}=n+2\tau(\log I_\tau)'. \tag*{(38)}

Since IτI_\tau is nonincreasing, (37) gives

ν(N,h)≤W(h,uτ,τ)=log⁡Iτ+2τ(log⁡Iτ)′≤log⁡Iτ.(39)\nu(N,h)\leq\mathcal{W}(h,u_\tau,\tau)=\log I_\tau+2\tau(\log I_\tau)'\leq\log I_\tau. \tag*{(39)}

Letting τ→∞\tau\to\infty and using (34), we obtain

ν(N,h)≤log⁡a,\nu(N,h)\leq\log a,

with log⁡0=−∞\log0=-\infty. In particular, a=0a=0 implies ν(N,h)=−∞\nu(N,h)=-\infty. □

Proposition 4.2 (Euclidean product invariance of the volume ratio). Let (Xd,hX)(X^d,h_X) be a complete connected Ricci-flat manifold, let k≥1k\geq1, and give Y=Rk×XY=\mathbb{R}^k\times X the product metric h=gE+hXh=g_E+h_X. Then

AVR⁡(Y,h)=AVR⁡(X,hX).(40)\operatorname{AVR}(Y,h)=\operatorname{AVR}(X,h_X). \tag*{(40)}

Proof. Fix x0∈Xx_0\in X and write

θX(s)=Vol⁡hXBhX(x0,s)ωdsd.\theta_X(s)=\frac{\operatorname{Vol}_{h_X}B_{h_X}(x_0,s)}{\omega_d s^d}.

The product distance satisfies

dY((z,x),(0,x0))2=∣z∣2+dX(x,x0)2.d_Y\bigl((z,x),(0,x_0)\bigr)^2=|z|^2+d_X(x,x_0)^2.

Indeed, for a path parametrized on [0,1][0,1], the integral of its squared speed in the product is the sum of the corresponding integrals for its two components. Each component integral is bounded below by the square of the corresponding distance. Conversely, the straight segment in Rk\mathbb{R}^k and a minimizing geodesic in XX, parametrized on the same unit interval with constant speeds, realize this lower bound. Consequently, for fixed zz with ∣z∣<r|z|<r, the section of Bh((0,x0),r)B_h((0,x_0),r) over zz is exactly

{z}×BhX(x0,r2−∣z∣2).\{z\}\times B_{h_X}\left(x_0,\sqrt{r^2-|z|^2}\right).

The volume measure of the product is dz dVhXdz\,dV_{h_X}. Integrating the corresponding ball sections over the Euclidean factor by Fubini’s theorem gives the first equality below. For the second, set z=ruz=ru, so that dz=rkdudz=r^kdu, and use the definition of θX\theta_X. We obtain

Vol⁡hBh((0,x0),r)=∫∣z∣<rVol⁡hXBhX(x0,r2−∣z∣2) dz=ωdrd+k∫B1Rk(1−∣u∣2)d/2θX(r1−∣u∣2) du.(41)\begin{aligned} \operatorname{Vol}_h B_h((0,x_0),r) &=\int_{|z|<r}\operatorname{Vol}_{h_X}B_{h_X}\left(x_0,\sqrt{r^2-|z|^2}\right)\,dz\\ &=\omega_d r^{d+k}\int_{B_1^{\mathbb{R}^k}}(1-|u|^2)^{d/2}\theta_X\left(r\sqrt{1-|u|^2}\right)\,du. \tag*{(41)} \end{aligned}

For every ∣u∣<1|u| < 1, θX(r1−∣u∣2)→AVR⁡(X,hX)\theta_X\left(r\sqrt{1-|u|^2}\right) \to\operatorname{AVR}(X,h_X) as r→∞r \to\infty. Bishop–Gromov comparison gives 0≤θX≤10 \le\theta_X \le1, so the integrand is dominated by the integrable function (1−∣u∣2)d/2(1-|u|^2)^{d/2}.

Dividing (41) by ωd+krd+k\omega_{d+k}r^{d+k} and applying dominated convergence yields

AVR⁡(Y,h)=(ωdωd+k∫B1Rk(1−∣u∣2)d/2 du)AVR⁡(X,hX).\operatorname{AVR}(Y,h)=\left(\frac{\omega_d}{\omega_{d+k}}\int_{B_1^{\mathbb{R}^k}}(1-|u|^2)^{d/2}\,du\right)\operatorname{AVR}(X,h_X).

Finally, Fubini’s theorem applied to the Euclidean unit ball in Rd+k\mathbb{R}^{d+k} gives

ωd∫B1Rk(1−∣u∣2)d/2 du=ωd+k.(42)\omega_d\int_{B_1^{\mathbb{R}^k}}(1-|u|^2)^{d/2}\,du=\omega_{d+k}. \tag*{(42)}

Substitution proves (40), including the case in which both asymptotic volume ratios vanish. □

The proof of Theorem 1.1

We recall the four-dimensional Ricci-flat gap theorem used in the final step.

Proposition 5.1 ([11 Proposition 3.7]). Let (X4,hX)(X^4,h_X) be a complete Ricci-flat four-manifold. If

AVR⁡(X,hX)>12,\operatorname{AVR}(X,h_X)>\frac{1}{2},

then (X,hX)(X,h_X) is isometric to Euclidean R4\mathbb{R}^4.

Proof of Theorem 1.1. If MM is compact, its curvature is bounded by smoothness. Suppose that MM is noncompact and sup⁡M∣Rm⁡g∣=∞\sup_M|\operatorname{Rm}_g|=\infty. By Proposition 3.3 and Lemma 3.4, a pointed curvature blow-up produces a complete nonflat Ricci-flat manifold with a global product decomposition

(Y5,h)=(R×X4,dt2+hX).(Y^5,h)=(\mathbb{R}\times X^4,dt^2+h_X).

The factor XX is complete and Ricci flat. It is nonflat since the product curvature is that of XX and ∣Rm⁡h∣(p∞)=1|\operatorname{Rm}_h|(p_\infty)=1.

The compactly supported entropy transfer in Proposition 3.3 and the Gaussian estimate in Lemma 4.1 give

ν(M,g)≤ν(Y,h)≤log⁡AVR⁡(Y,h).(43)\nu(M,g)\leq\nu(Y,h)\leq\log\operatorname{AVR}(Y,h). \tag*{(43)}

By (2), the left-hand side is strictly greater than −log⁡2-\log2. In particular, ν(Y,h)>−∞\nu(Y,h)>-\infty, so Lemma 4.1 gives positive asymptotic volume ratio. Exponentiating (43), we obtain

AVR⁡(Y,h)≥eν(Y,h)≥eν(M,g)>12.\operatorname{AVR}(Y,h)\geq e^{\nu(Y,h)}\geq e^{\nu(M,g)}>\frac{1}{2}.

Proposition 4.2, with a one-dimensional Euclidean factor, yields

AVR⁡(X,hX)=AVR⁡(Y,h)>12.\operatorname{AVR}(X,h_X)=\operatorname{AVR}(Y,h)>\frac{1}{2}.

Proposition 5.1 now implies (X,hX)≅R4(X,h_X)\cong\mathbb{R}^4. Consequently YY is flat, contradicting ∣Rm⁡h∣(p∞)=1|\operatorname{Rm}_h|(p_\infty)=1. Thus the assumption sup⁡M∣Rm⁡g∣=∞\sup_M|\operatorname{Rm}_g|=\infty is impossible, and the theorem follows. □

Remark 5.2. The strict entropy inequality is used only in the final volume-ratio gap. The curvature blow-up and the real-line splitting in Section 3 require only the scalar curvature upper bound for the fixed shrinker. The local estimate in Lemma 3.2 provides the Ricci bound needed for the rescaled Hessian calculation.

References

Mathematical Science Research Center, Chongqing University of Technology, Chongqing 400054, China

Email address: [email protected]

School of Mathematical Sciences, Xiamen University, Xiamen 361005, China

Email address: [email protected]

School of Mathematical Sciences, Shanghai Key Laboratory of Pure Mathematics and Mathematical Practice, East China Normal University, Shanghai 200241, China

Email address: [email protected]

School of Science, Zhejiang Sci-Tech University, Hangzhou 310018, China

Email address: [email protected]

References

  1. [1]R. H. Bamber, Convergence of Ricci flows with bounded scalar curvature, Ann. of Math. (2) 188 (2018), no. 3, 753–831.
  2. [2]S. Brendle, Rotational symmetry of self-similar solutions to the Ricci flow, Invent. Math. 194 (2013), no. 3, 731–764.
  3. [3]S. Brendle, Rotational symmetry of Ricci solitons in higher dimensions, J. Differential Geom. 97 (2014), no. 2, 191–214.arxiv.org/abs/1203.0270
  4. [4]H.-D. Cao, B.-L. Chen, and X.-P. Zhu, Recent developments on Hamilton’s Ricci flow, Surveys in Differential Geometry 12 (2008), 47–112.DOI
  5. [5]H.-D. Cao and Q. Chen, On Bach-flat gradient shrinking Ricci solitons, Duke Math. J. 162 (2013), no. 6, 1149–1169.DOI
  6. [6]H.-D. Cao and D. Zhou, On complete gradient shrinking Ricci solitons, J. Differential Geom. 85 (2010), no. 2, 175–185.arxiv.org/abs/0903.3932
  7. [7]J. Cheeger, M. Gromov, and M. Taylor, Finite propagation speed, kernel estimates for functions of the Laplace operator, and the geometry of complete Riemannian manifolds, J. Differential Geom. 17 (1982), 15–53.DOI
  8. [8]B.-L. Chen, Strong uniqueness of the Ricci flow, J. Differential Geom. 82 (2009), 363–382.arxiv.org/abs/0706.3081
  9. [9]X. Chen and B. Wang, On the conditions to extend Ricci flow (III), Int. Math. Res. Not. IMRN 2013 (2013), no. 10, 2349–2367. doi:10.1093/imrn/rns117.DOI
  10. [10]B. Chow, S.-C. Chu, D. Glickenstein, C. Guenther, J. Isenberg, T. Ivey, D. Knopf, P. Lu, F. Luo, and L. Ni, The Ricci flow: techniques and applications. Part I. Geometric aspects, Mathematical Surveys and Monographs, vol. 135, American Mathematical Society, Providence, RI, 2007.
  11. [11]S. Honda and A. Mondino, Gap phenomena under curvature restrictions, Indag. Math. 37 (2026), no. 3, 744–763; arXiv:2410.04985.arxiv.org/abs/2410.04985
  12. [12]Y. Li and B. Wang, Heat kernel on Ricci shrinkers, Calc. Var. Partial Differential Equations 59 (2020), no. 6, Paper No. 194, 84 pp.
  13. [13]Y. Li and B. Wang, Heat kernel on Ricci shrinkers (II), Acta Math. Sci. Ser. B (Engl. Ed.) 44 (2024), no. 5, 1639–1695.arxiv.org/abs/2301.08430
  14. [14]O. Munteanu and J. P. Wang, Geometry of shrinking Ricci solitons, Compos. Math. 151 (2015), no. 12, 2273–2300.DOI
  15. [15]A. Naber, Noncompact shrinking four solitons with nonnegative curvature, J. Reine Angew. Math. 645 (2010), 125–153.DOI
  16. [16]L. Ni and N. Wallach, On a classification of gradient shrinking solitons, Math. Res. Lett. 15 (2008), no. 5, 941–955.DOI
  17. [17]G. Perelman, The entropy formula for the Ricci flow and its geometric applications, arXiv:math/0211159, 2002.
  18. [18]G. Perelman, Ricci flow with surgery on three-manifolds, arXiv:math/0303109, 2003.DOI
  19. [19]P. Petersen and W. Wylie, Rigidity of gradient Ricci solitons, Pacific J. Math. 241 (2009), no. 2, 329–345.arxiv.org/abs/0710.3174
  20. [20]P. Petersen and W. Wylie, On the classification of gradient Ricci solitons, Geom. Topol. 14 (2010), no. 4, 2277–2300.
  21. [22]W.-X. Shi, Deforming the metric on complete Riemannian manifolds, J. Differential Geom. 30 (1989), 223–301.DOI
  22. [23]B. Wang, The local entropy along Ricci flow—Part A: the no-local-collapsing theorems, Camb. J. Math. 6 (2018), no. 3, 267–346.arxiv.org/abs/1706.08485
  23. [24]J.-Y. Wu, P. Wu, and W. Wylie, Gradient shrinking Ricci solitons of half harmonic Weyl curvature, Calc. Var. Partial Differential Equations 57 (2018), no. 5, Art. 141, 15 pp.; correction, ibid. 60 (2021), Art. 76. doi:10.1007/s00526-020-01892-5.
  24. [25]Z. H. Zhang, Gradient shrinking solitons with vanishing Weyl tensor, Pacific J. Math. 242 (2009), no. 1, 189–200.arxiv.org/abs/0807.1582

Paper details

vlnc:W_uyd5xg9x
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 3

downloads · 0

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

  • Mathematics

ChatGPT

View statement
Looking for related papers…
Contents