University of Connecticut OpenCommons@UConn
Honors Scholar Theses Honors Scholar Program
Spring 5-1-2021
Sobolev Inequalities and Riemannian Manifolds
John Reever [email protected]
Follow this and additional works at: https://opencommons.uconn.edu/srhonors_theses
Part of the Analysis Commons, and the Geometry and Topology Commons
Recommended Citation Reever, John, "Sobolev Inequalities and Riemannian Manifolds" (2021). Honors Scholar Theses. 824. https://opencommons.uconn.edu/srhonors_theses/824 Sobolev Inequalities and Riemannian Manifolds
Jack Reever
A thesis presented to the Department of Mathematics in partial fulfillment of the requirements of a Bachelors of Science with Honors
Storrs, Connecticut Spring 2021 Abstract
Sobolev inequalities, named after Sergei Lvovich Sobolev, relate norms in Sobolev spaces and give insight to how Sobolev spaces are embedded within each other. This thesis begins with an overview of Lebesgue and Sobolev spaces, leading into an introduction to Sobolev inequalities. Soon thereafter, we consider the behavior of Sobolev inequalities on Riemannian manifolds. We discuss how Sobolev inequalities are used to construct isoperimetric inequal- ities and bound volume growth, and how Sobolev inequalities imply families of other Sobolev inequalities. We then delve into the usefulness of Sobolev inequalities in determining the geometry of a manifold, such as how they can be used to bound a manifold’s number of ends. We conclude with examples of real-world applications Sobolev inequalities.
1 Introduction to Sobolev Spaces
We will begin by introducing Sobolev spaces, and to do so we start with Lebesgue spaces. Let p ∈ [1, ∞). The space of functions which are Lebesgue integrable on a set Ω p-many times is denoted Z Lp(Ω) = {f : |f|p(x)dx < ∞}. Ω These Lebesgue spaces have the norm Z 1/p p ||f||Lp(Ω) = |f| (x)dx . Ω If p = ∞, the space L∞(Ω) is the space of all functions such that |f(x)| < ∞ “almost everywhere” on Ω, which is to say that |f(x)| can be infinite on only a set of measure zero. Here, the norm is the essential supremum of |f(x)| for x ∈ Ω [3]. The elements of a Lebesgue space are equivalence classes, and functions are considered equivalent if they differ only on a set of Lebesgue measure zero. Lebesgue spaces are complete normed spaces. Let k ∈ N ∪ {0} and p ∈ [1, ∞]. The Sobolev Space W k,p(Ω) is defined by
W k,p(Ω): = {u ∈ Lp(Ω) : Dαu ∈ Lp(Ω) ∀α with |α| ≤ k}.
1 This space has the norm X α ||u||W k,p(Ω) := ||D u||Lp(Ω) |α|≤k as discussed in [3]. For a function u to be included in a Sobolev space, all derivatives Dαu must exist and must be represented by an element of Lp(Ω). Note that if k = 0, the space W 0,p(Ω) = Lp(Ω) identically, as the restriction on the derivatives of each element is gone. Sobolev spaces are vector spaces, as elements can be added together and multiplied by scalars, with the resultant functions being elements of the Sobolev space. In addition, the k,p ∞ Sobolev space W0 (Ω) is the completion of C0 (Ω) (the space of infinitely differentiable functions with compact support) with respect to the norm of W k,p(Ω) [3]. Equipped with the definition of Sobolev spaces, we can intuitively embed Sobolev spaces in one another: let Ω ⊂ Rd be a domain with p ∈ [1, ∞) and k ≤ m. Then, W m,p(Ω) ⊂ W k,p(Ω). This is clear because if a function has m derivatives represented in Lp(Ω), then it must have k derivatives in Lp(Ω). Alternatively, for k ≥ 0 and p, q ∈ [1, ∞] with q > p it is clear that W k,q(Ω) ⊂ W k,p(Ω): if we can integrate q−many times, we can integrate p−many times. Now, we will consider the Sobolev Embedding Theorem. Consider W k,p(Rn) with k ∈ N+, p ∈ [1, ∞). If we have two real numbers l, q such that k > l, 1 ≤ p < q < ∞, and 1 k 1 l − = − , p n q n
then W k,p(Rn) ⊆ W l,q(Rn), and this embedding is continuous.
2 2 Introduction to Sobolev Inequalities
When we are working with smooth functions f : R → R with compact support, we have the inequality 1 Z ∞ |f(x)| ≤ |f 0(t)|dt. 2 −∞ This inequality bounds the value of |f(x)| using its derivative. As we move up to multidi- mensional spaces, the notions of absolute value and the derivative become more complicated, and we cannot use this inequality to bound our functions. This brings us to the topic of Sobolev inequalities: how can we construct an inequality that bounds the norm of a function using partial derivatives? While working in Rn, let p ∈ [1, n). The Sobolev conjugate of p is np p∗ := ∈ (p, ∞). n − p
Theorem 2.1: When n ≥ 2, the Sobolev inequality in Rn states Z 1/p∗ Z 1/p |f(x)|p∗ dx ≤ C(n, p) |∇f(x)|pdx (1) Rn Rn for all smooth functions f with compact support, for any p ∈ [1, n). Note that C(n, p) is a constant that depends on n and p [1].
∗ p n Note that this inequality can be written as ||f||Lp (Rn) ≤ C||∇f||L (R ), and moving
forward we will notate these norms as ||f||p∗ ≤ C||∇f||p. To prove Theorem 2.1, we will illustrate Gagliardo and Nirenberg’s proof of the Sobolev inequality in Rn for p = 1, and then expand upon that proof to show this inequality holds n for all real p ∈ [1, n). Note that µ = µn denotes the Lebesgue measure in R .
Proposition 2.2: When n ≥ 2, Z 1/1∗ Z 1/1 |f(x)|1∗ dx ≤ C(n, 1) |∇f(x)|1dx Rn Rn for all smooth functions f with compact support. Note that C(n, 1) is a constant that depends on n.
0 1 1 p p0 Proof. Let 1 ≤ p, p ≤ ∞ with 1 = p + p0 . For f ∈ L (µ), g ∈ L (µ), H¨older’sinequality states that for positive measure µ, Z
fgdµ ≤ ||f||p||g||p0 .
3 pi 1 1 1 By induction, for fi ∈ L , 1 ≤ i ≤ k, 1 ≤ pi ≤ ∞, + + ... + = 1, we see p1 p2 pk Z k Y f1f2...fkdµ ≤ ||fi||pi . (2)
i=1 ∞ n Fix f ∈ C0 (R ): a smooth function with compact support that. By the inequality 1 Z ∞ |f(x)| ≤ |f 0(t)|dt, 2 −∞ for any x = (x1, ..., xn) and any integer i ∈ [1, n] we observe 1 Z ∞ |f(x)| ≤ |∂xi f(x1, x2, ..., xi−1, t, xi+1, ..., xn)|dt. (3) 2 −∞ Fix Z ∞
Fi(x) = |∂xi f(x1, x2, ..., xi−1, t, xi+1, ..., xn)|dt, −∞ which depends on n − 1 coordinates of x with the integration variable replacing the ith coordinate. Additionally, fix
R ∞ R ∞ −∞ ... −∞ |∂xi f(x)|dx1...dxm when i ≤ m Fi,m(x) = R ∞ R ∞ −∞ ... −∞ Fi(x)dx1...dxm when i > m which depends on n − m or n − m − 1 variables, for all integers m ∈ [1, n]. Now, by (3) we have the inequality 1 |f|n ≤ (F ...F ). 2n 1 n Raising both sides of this equality to the power of 1/(n − 1), we see 1 |f|n/(n−1) ≤ (F ...F )1/(n−1). 2n/(n−1) 1 n 1 1 1 1 Setting k = n − 1, p1 = p2 = ... = pn−1 = n − 1, we see + + ... + = (n − 1) = 1. p1 p2 pk n−1 By (2) with induction on m ≤ n we see Z Z 1 ... |f(x)|n/(n−1)dx ...dx ≤ (F (x)...F (x))1/(n−1). 1 m 2n/(n−1) 1,m n,m When n = m, we see n 1/n Y ||f||n/(n−1) ≤ (1/2) ||∂xi f||1 . i=1 Note that both sides were raised to the power of (n − 1)/n when taking the norm in √ n/(n−1) n Qn 1/n 1 Pn Pn L (R ). Because ( 1 ai) ≤ n 1 ai for ai > 0 and n ∈ Z, and 1 |∂xi f| ≤ n|∇f|, this becomes n 1 X 1 ||f|| ≤ ||∂ f(x)|| dx ≤ √ ||∇f|| . n/(n−1) 2n xi 1 2 n 1 i=1 This proves Proposition 2.2, the Sobolev Inequality in Rn for p = 1. [1]
4 Next, we are going to build off of the above proof to show that Theorem 2.1 (1) holds for p ≥ 1.
Proposition 2.3: If Theorem 2.1, the Sobolev Inequality in Rn, holds for p = 1, then it holds for all p ∈ [1, n).
Proof. Knowing that (1) holds for p = 1, or that 1 ∀f ∈ C∞( n) ||f|| ≤ √ ||∇f|| , (4) 0 R n/(n−1) 2 n 1
∞ n α 1 fix a p > 1. For any α > 1 and function f ∈ C0 (R ), |f| is C0 and satisfies
|∇|f|α| = α|f|α−1|∇f|.
α The function |f| can be approximated by a sequence of smooth functions (fi) ∈ C0 such α α that ∇fi → ∇|f| . This allows us to substitute |f| into (4): 1 Z ||f||α ≤ √ α |f(x)α−1|∇f(x)|dx αn/(n−1) 2 n 1/p0 1/p 1 Z 0 Z ≤ √ α |f(x)|(α−1)p dx |∇f(x)|pdx 2 n
0 (n−1)p with 1/p + 1/p = 1. Setting α = (n−p) implies
1 (n − 1)p ||f||(n−1)p/(n−p) ≤ √ ||f||n(p−1)/(n−p)||∇f|| . np/(n−p) 2 n n − p np/(n−p) p
n(p−1)/(n−p) To simplify, divide both sides by ||f||np/(n−p) . We see that
(n − 1)p n(p − 1) pn − p − np + n n − p − = = = 1, n − p n − p n − p n − p and so 1 (n − 1)p ||f|| ≤ √ ||∇f|| . p∗ 2 n n − p p We have thereby proven a Sobolev inequality in Rn for p > 1: for any integer n ≥ 2 and p ∈ [1, n), then ∞ n np − p ∀f ∈ C ( ) ||f|| ∗ ≤ √ ||∇f|| . 0 R p 2(n − p) n p
5 We have shown the Sobolev Inequality in Rn holds for p ∈ [1, n), and thus have proven Theorem 2.1 [1]. Alternatively, in the case that n < p ≤ ∞, we have the H¨oldercontinuity estimate: |f(x) − f(y)| Z 1/p sup ≤ C |∇f(x)|pdx n 1−n/p x,y∈R |x − y| Rn for smooth functions f with compact support, for some constant C that depends only on n and p. We can generalize Sobolev inequalities to subsets of Rn: let’s consider the general n n Sobolev inequality for k < p . Let Ω ⊂ R be Lipschitz, with k ∈ {1, ..., n − 1} and n p ∈ [1, k ). Let 1 1 k np = − =⇒ q = > p. q p n n − kp Then, W k,p(Ω) ⊂ Lq(Ω) and
||u||Lq(Ω) ≤ C||u||W k,p(Ω) for some constant C that depends on k, p, n, and Ω. It is clear that Lq ⊂ Lp, but this inequality tells us that W k,p ⊂ Lq ⊂ Lp. The Gagliardo-Nirenberg-Sobolev inequality states for p ∈ [1, n),
p(n − 1) ||u|| p∗ n ≤ ||Du|| p n L (R ) n − p L (R )
1 n for u ∈ C0 (R ). Various Sobolev inequalities are related to one another; for example, there are equiva- lences between “weak” and “strong” Sobolev inequalities. Consider the Nash inequality on a Riemannian manifold M:
∞ 1+2/v 2/v ∀f ∈ C0 (M), ||f||2 ≤ C||∇f||2||f||1 this inequality is equivalent to the Sobolev inequality
∞ ∀f ∈ C0 (M), ||f||2v/(v−2) ≤ C||∇f||2.
We will examine more examples of inequality equivalences in further detail in Section 3.
6 3 Sobolev Inequalities on Manifolds
Sobolev inequalities were prevalent during the development of analysis on manifolds because Fourier analysis and other tools in Euclidean space are no longer available on manifolds. On manifolds, the original question of which inequalities are true or not is expanded into a search for necessary and sufficient conditions for Sobolev inequalities to hold true. Not every manifold admits global Sobolev inequalities; a workaround is to use families of local Sobolev inequalities. For example, we can construct a Sobolev inequality locally on a ball; for a ball B = B(x, r) on a complete Riemannian manifold, there is a constant C(B) such that for smooth functions f with compact support in B
Z 2/q 2 Z q C(B)r 2 −2 2 |f| dµ ≤ 2/v (|∇f| + r |f| )du B µ(B) B 1 1 1 for constants q, v > 2 such that q = 2 − v .
3.1 Isoperimetric Inequalities
Geometric properties of manifolds can be determined using isoperimetric inequalities, in- equalities that relate the measure of a set’s boundary to the set’s measure or volume. For some background, consider R2. A circle is the 2-dimensional or planar shape that maximizes area (A) in a fixed perimeter (circumference C). These two values are related by the equation
4πA = C2 derived from C = 2πr, A = πr2. Thus, the isoperimetric inequality in R2 tells us that for a closed curve of length L, with A the area of the region enclosed by L,
L2 ≥ 4πA.
Let Ω be a bounded subset of a Riemannian manifold M, with a smooth boundary ∂Ω. We want to find the maximal volume that can be enclosed in a hypersurface of a fixed (n − 1)-measure. Using the same intuition as the R2 example, in Euclidean space an n−dimensional sphere or ball maximizes enclosed volume with a fixed surface area. So, if M is n−dimensional Euclidean space, we can derive an isoperimetric inequality from the equation that defines the volume of a sphere in n dimensions. Now we will consider the isoperimetric inequality for a Riemannian manifold that may not be in Euclidean space. Let M be an n−dimensional Riemannian manifold. If M satisfies
1−1/v µn(Ω) ≤ C(M, v)µn−1(∂Ω) (5)
7 for any Ω ⊂ M with a smooth boundary for some v ≥ n, C(M, v) > 0, then rv V (x, r) ≥ (6) (vC(M, v))v
with V (x, r) denoting the volume of a ball with radius r [1]. This is because ∂rV (x, r) is the (n − 1)−dimensional Riemannian volume of the ball B(x, r)’s boundary: in general, the volume of a sphere in n dimensions is
bn/2c n V (x, r) = Cnπ r for some constant Cn [2]. Meanwhile, this sphere’s “surface area” can be found by the formula bn/2c n−1 An = Cnπ nr ,
which is equivalent to ∂r(V (x, r)) [2]. We see that
1−1/v V (x, r) ≤ C(M, v)∂rV (x, r) 1 =⇒ V (x, r)1−1/v ≤ C(M, v)V (x, r)n r 1 =⇒ V (x, r)−1/v ≤ C(M, v)n r r =⇒ V (x, r)1/v ≥ nC(M, v) rv rv =⇒ V (x, r) ≥ ≥ nvC(M, v)v vvC(M, v)v because v ≥ n.
Theorem 3.1: A manifold M satisfies the conditions in (5) for some v, C(M, v) > 0 if and only if ∞ ∀f ∈ C0 (M), ||f||v/(v−1) ≤ C(M, v)||∇f||1. This is an (L1, v)-Sobolev inequality [1].
Definition 3.2: If we fix p, v such that 1 ≤ p < v, a Riemannian manifold M satisfies an (Lp, v)-Sobolev inequality if there exists a constant C(M, p, v) that satisfies
∞ ∀f ∈ C0 (M), ||f||pv/(p−v) ≤ C(M, p, v)||∇f||p.
Now, we have a definition of the (Lp, v)−Sobolev inequality. Let us consider the following theorem, which states that (Lp, v)−Sobolev inequalities imply other inequalities of this type.
8 Theorem 3.3: If M satisfies an (Lp, v)-Sobolev inequality, then M satisfies (Lq, v)-Sobolev inequalities for all q ∈ [p, v).
Proof. Let γ > 1 and use the (Lp, v) inequality on |f|γ:
Z 1/p γ p(γ−1) p ||f||γpv/(v−p) ≤ γC(M, p, v) |f| |∇f| dµ . M By the H¨olderinequality, we see Z p(γ−1) p p(γ−1) p |f| |∇f| ≤ |||f| ||q/(q−p)|||∇f| ||q/p
which implies Z 1/p Z 1/p−1/q Z 1/q |f|p(γ−1)|∇f|p dµ ≤ |f|pq(γ−1)/(q−p)dµ |∇f|qdµ . M M M Using these two inequalities, we see Z 1/p−1/q Z 1/q γ pq(γ−1)/(q−p) q ||f||γpv/(v−p) ≤ γC(M, p, v) |f| dµ |∇f| dµ M M
q(n−p) and if we select γ = p(n−q) we find γ − 1 = n(q − p)/p(n − q), implying q(n − p) ||f|| ∗ ≤ C(M, p, v)||∇f|| , q p(n − q) q an (Lq, v)-Sobolev inequality [1].
3.2 Bounding Volume Growth with Sobolev Inequalities
Let M be a manifold that satisfies an (Lp, v)-Sobolev Inequality:
∞ ∀f ∈ C0 (M), ||f||pv/(v−p) ≤ C(M, p, v)||∇f||p.
1 1 1 Let q = vp/(v − p) =⇒ q = p − v , and let r, s ∈ (0, ∞) and θ ∈ [0, 1] such that 1 1 1 1 − θ θ 1 − θ = θ − + = + . r p v s q s The H¨olderinequality tells us θ 1−θ ||f||r ≤ ||f||q||f||s and because q = vp/(v − p) we apply the (Lp, v)-Sobolev inequality above to find
θ 1−θ ||f||r ≤ (C(M, p, v)||∇f||p) ||f||s .
9 Thus, if M satisfies an (Lp, v)-Sobolev inequality, it also satisfies the above inequality for all 1 1 1 1−θ r, s ∈ (0, ∞) and θ ∈ [0, 1] such that r = θ p − v + s [1]. We want to use this result to generalize the isoperimetric inequality (6). In the introduc- tion to this section, we discussed that not all manifolds admit global Sobolev inequalities, and at times we have to use local inequalities based in balls. Now, we will use the above result to place a lower bound on the volume of a ball B(x, r). We will notate the volume growth function as V (x, r) = µ(B(x, r)), the measure of the ball of radius r centered at x ∈ M. The following result provides a lower bound for the volume growth V (x, t) in M, as seen in [1, Theorem 3.1.5].
Theorem 3.4: Assume that on a manifold M
∞ θ 1−θ ∀f ∈ C0 (M), ||f||r ≤ (C(M, p, v)||∇f||p) ||f||s
is satisfied for some r, s with 0 < s ≤ r ≤ ∞ and some θ ∈ (0, 1], such that θ 1 − θ 1 + − > 0. p s r Then, for v defined by θ θ 1 − θ 1 = + − v p s r we have tv t v V (x, t) ≥ = . 2v2/θr+v/θC(M, p, v)v 2v/θr+1/θC(M, p, v) Particularly, if a manifold M satisfies an (Lp, v)-Sobolev inequality with 1 ≤ p < v, then n ≥ v and inf {tvV (x, t)} > 0. x∈M t>0 Proof. For a fixed x ∈ M and t > 0, consider
f(y) = max{t − d(x, y), 0}.
Then, t ||f|| ≥ V (x, t/2)1/r r 2 1/s ||f||s ≤ tV (x, t) 1/p ||∇f||p ≤ V (x, t) .
10 and thus 1 V (x, t)θ/p+(1−θ)/s ≥ (t/C)θV (x, t/2)1/r. 2 Here, define v by θ θ 1 − θ 1 = + − v p s r to find V (x, t) ≥ (2Cθ)−rv/(v+θr)tθrv/(v+θr)V (x, t/2)v/(v+θr).
For a = v/(v + θr), we find
V (x, t) ≥ (2Cθ)−ratθraV (x, t/2)a
θ −r Pi aj θr Pi aj −θr Pi (j−1)aj i ai ≥ (2C ) 1 t 1 2 1 V (x, t/2 ) .
Additionally,
∞ ∞ 2 X a v X a v2 aj = = , (j − 1)aj = = , 1 − a θr 1 − a θ2r2 j=1 j=1 so as we allow i → ∞ we find v 2 t V (x, t) ≥ (2Cθ)−v/θtv2−v /(θr) = . 2v/θ+v2/(θr)Cv
3.3 Weak and Strong Sobolev Inequalities
In order to prove that a certain manifold satisfies a Sobolev inequality, at times we must use the equivalences between weak and strong forms of Sobolev-type inequalities that were introduced in Section 2. As demonstrated in subsection 4.1, we know that if M satisfies an (Lp, v)-Sobolev inequality, then M satisfies (Lq, v)-Sobolev inequalities for all q ∈ [p, v). Any Sobolev inequality can be used in conjunction other inequalities, such as the H¨older inequality, to lead to “weaker” inequalities. Moving forward, we will discuss how “weaker” Sobolev inequalities imply stronger Sobolev inequalities. In the previous subsection, we saw the inequality
∞ 1/r θ 1/s 1−θ ∀f ∈ C0 (M), sup{tµ(|f| > t) } ≤ (C(M, p, v)||∇f||p) (||f||∞µ(supp(f)) ) . t>0
∗,θ This is called the (Sr,s) inequality, and is the weakest Sobolev inequality discussed so far. θ The (Sr,s) inequality is stronger and states
∞ θ 1−θ ∀f ∈ C0 (M), ||f||r ≤ (C(M, p, v)||∇f||p) ||f||s .
11 1 Here, 0 < r, s ≤ ∞ and θ ∈ (0, 1]. When θ = 1, we find (Sr,s) states
∞ ∀f ∈ C0 (M), ||f||r ≤ C(M, p, v)||∇f||p
∗,1 and (Sr,s ) states
∞ 1/r ∀f ∈ C0 (M), sup{tµ(|f| > t) } ≤ C(M, p, v)||∇f||p. t>0
∗,θ0 We will show how a weak inequality (Sr0,s0 ) implies a collection of stronger inequalities. First, consider the parameter q = q(r, s, θ) 6= 0 that satisfies
1 θ 1 − θ = + r q s
1 1 1 Recall that this q was introduced at the beginning of subsection 4.2 and satisfies q = p − v . We ∗,θ0 will prove that any weak inequality (Sr0,s0 ) for fixed r0, s0, θ0 implies all stronger inequalities θ ∗,θ0 (Sr,s) for all r, s ∈ (0, ∞] and θ ∈ (0, 1] with q(r, s, θ) = q(r0, s0, θ0): an inequality (Sr0,s0 ) θ for fixed r0, s0, θ0, and p0 implies all inequalities (Sr,s) for all p ≥ p0 with r, s, θ, p satisfying 1 θ 1−θ 1 1 1 r = q + s and q = p − v .
3.3.1 Equivalences when q ∈ (0, ∞)
∗,θ0 Begin by fixing p ∈ [1, ∞). Assume that, on a manifold M,(Sr0,s0 ) is satisfied for some 1 θ 1−θ 0 < r0, s0 ≤ ∞ and θ ∈ (0, 1]. Let q = q(r0, s0, θ0) satisfy r = q + s with p ≤ q < ∞. Let ∞ f ∈ C0 (M) be nonnegative and set
k k fρ,k = (f − ρ )+ ∧ ρ (ρ − 1)
for any ρ > 1 and k ∈ Z. Here, u+ = max{u, 0} and u ∧ v = min{u, v}: fρ,k reads as
min{(max{f − ρk, 0}), ρk(ρ − 1)}.
k k+1 We see that |∇fρ,k| ≤ |∇f|. Because {fρ,k ≥ (ρ − 1)ρ } = {f ≥ ρ }, and because M
∗,θ0 satisfies (Sr0,s0 ) we see
k k+1 1/r0 θ0 k k 1/s0 1−θ0 (ρ − 1)ρ µ({f ≥ ρ }) ≤ (C(M, p, v)||∇f||p) )((ρ − 1)ρ µ({f ≥ ρ }) ) .
12 Set N(f) = sup{ρkµ({f ≥ ρk})1/q}. We then find k
θ0 C||∇f||p µ({f ≥ ρk+1})1/r0 ≤ ρ−k(θ0+(1−θ0)q/s0) N(f)q(1−θ0)/s0 ρ − 1 θ0 C||∇f||p ≤ ρ−kq/r0 N(f)q(1−θ0)/s0 ρ − 1 θ0 C||∇f||p =⇒ N(f)q/r0 ≤ ρq/r0 N(f)q(1−θ0)/s0 ρ − 1 ρq/r0 =⇒ N(f) ≤ C||∇f|| ρ − 1 p by the definition of q. This tells us
1+q/r0 1/q ρ sup{tµ({f ≥ t}) } ≤ ρN(f) ≤ C||∇f||p. t>0 ρ − 1
When we set ρ = 1 + r0θ0/q this gives 1/q q sup{tµ({f ≥ t}) } ≤ e 1 + C||∇f||p. (7) t>0 r0θ0
p This is the “weak” form of the (L , v)-Sobolev inequality ||f||pv/(v−p) ≤ AC||∇f||p. We
∗,θ0 θ want to show that if M satisfies (Sr0,s0 ), then all inequalities (Sr,s) with 0 < r, s ≤ ∞, θ ∈ (0, 1] with q(r, s, θ) = q are satisfied. Particularly, we want to show that there is a constant A such that
∞ ∀f ∈ C0 (M), ||f||q ≤ AC||∇f||p, (8)
p pv an (L , v)-Sobolev inequality (recall q = p−v ). We move forward using the fact that for 1 ≤ p < q < ∞, if
∞ 1/q ∀f ∈ C0 (M), sup{tµ({f ≥ t}) } ≤ C1||∇f||p t>0 then ∞ ∀f ∈ C0 (M), ||f||q ≤ 2(1 + q)C1||∇f||p. Using this with (7) we find the desired result (8), with A = 2e(1 + q) 1 + q and C r0θ0 ∗,θ0 the constant from (Sr0,s0 ).
3.3.2 Equivalences when q = ∞
1 θ 1−θ s Again, let us fix p ∈ [1, ∞). When q = ∞, the relationship r = q + s becomes r = 1−θ . s ∗,1−s0/r0 Thus, θ = 1 − r . We want to show that if (Sr0,s0 ) is satisfied on a manifold M for some
13 1−s/r 0 < s0 < r0 < ∞, then all inequalities (Sr,s ) with 0 < s < r < ∞ are satisfied on M. ∞ Following the same approach as the previous sub-subsection, if f ∈ C0 (M) is nonnegative and we consider fρ,k, we find
r0−s0 C||∇f||p ρkr0 µ({f ≥ ρk+1}) ≤ ρks0 µ({f ≥ ρk}). ρ − 1
t Set a real number t > 0 and set r = r0 + t, s = s0 + t. Multiplying both sides by ρ brings us to C||∇f|| r0−s0 ρkrµ({f ≥ ρk+1}) ≤ p ρksµ({f ≥ ρk}). ρ − 1
Note that r0 − s0 = r − s because r0 = r − t, s0 = s − t. If we sum over all k, we see r−s X C||∇f||p X ρkrµ({f ≥ ρk+1}) ≤ ρksµ({f ≥ ρk}). ρ − 1 k k Building off of these sums, we can find Z ρk+2 1 r 1 X s−1 kr k+1 2r r ||f||r ≤ 2r r r t µ({f ≥ t})dt ≤ ρ µ({f ≥ ρ }) ρ − ρ ρ − ρ k+1 k ρ and X ρs ρksµ({f ≥ ρk}) ≤ ||f||s. ρs − 1 s k
Putting this together, we find for all t > 1, r = r0 + t, s = s0 + t ρr+s(ρr − 1) ||f||r ≤ (C||∇f|| )r−s||f||s r (ρs − 1)(ρ − 1)r−s p s
ρr+s(ρr − 1) 1/r =⇒ ||f|| ≤ (C||∇f|| )1−s/r||f||s/r r (ρs − 1)(ρ − 1)r−s p s 1−s/r which is the definition of (Sr,s ), with a constant depending on the fixed ρ > 1. Thus, all
1−s/r ∗,1−s0/r0 inequalities (Sr,s ) with s0 ≤ s < r < ∞, r0 ≤ r are satisfied on M if M satisfies (Sr0,s0 ) θ for some 0 < s0 < r0 < ∞. Meanwhile, the H¨olderinequality shows that if r ≥ s,(Sr,s) θ0 0 0 0 0 0 ∗,1−s0/r0 implies (Sr0,s0 ) for all r , s such that s ≤ r ≤ r and s ≤ s. Thus, if M satisfies (Sr0,s0 ), 1−s/r M satisfies all (Sr,s ) with 0 < s < r < ∞.
∗,θ0 For all bounded sets Ω ⊂ M and all r ≥ 1, if M satisfies (Sr0,s0 ) for some 0 < s0 < r0 < ∞
then there exists a constant A1 such that
∞ 1/r ∀f ∈ C0 (Ω), ||f||r ≤ A1(1 + r)µ(Ω) ||∇f||p.
In addition, there exist constants α > 0 and A2 such that Z ∞ α|f|/||∇f||p ∀f ∈ C0 (Ω), e dµ ≤ A2µ(Ω) Ω
14 3.3.3 Equivalences when q ∈ (−∞, 0)
∗,θ0 Begin by fixing p ∈ [1, ∞). Assume that on a manifold M,(Sr0,s0 ) is satisfied for some
0 < s0 < r0 ≤ ∞ and θ0 ∈ (0, 1]. Again, take q = q(r0, s0, θ0) as before. With s0 < r0, we θ have q ∈ (−∞, 0). We want to show that all inequalities (Sr,s) with 0 < s < r ≤ ∞, θ ∈ (0, 1]
and q(r, s, θ) = q(r0, s0, θ0) are satisfied: particularly, we want to show that there is a constant A ∈ R such that
∞ 1/(1−s/q) 1/(1−q/s) ∀f ∈ C0 (M), ||f||∞ ≤ A(C||∇f||p) ||f||s for all s ∈ (0, ∞), with C the constant from (S∗,θ0 ). Set a nonnegative function f ∈ C∞(M) f0,s0 0 with ||f||∞ 6= 0. In adition, fix ε > 0 and ρ > 1, and define k(f) to be the largest integer k k such that ρ < ||f||∞. Then, set
k k−1 fρ,k = min{(max{f − ||f||∞ + ε + ρ , 0}), ρ (ρ − 1)} k k−1 = (f − (||f||∞ − ε − ρ ))+ ∧ ρ (ρ − 1)
k for all k ≤ k(f). Fix λk = ||f||∞ − ε − ρ . Note that fρ,k has support in {f ≥ λk} and
k−1 {fρ,k ≥ ρ (ρ − 1)} = {f ≥ λk−1}.
∗,θ0 Because M satisfies (Sr0,s0 ), apply this inequality to fρ,k to find
θ0 C||∇f||p ρk−1µ({f ≥ λ })1/r0 ≤ ρ(k−1)(1−θ0)µ({f ≥ λ })(1−θ0)/s0 k−1 ρ − 1 k θ0 C||∇f||p =⇒ µ({f ≥ λ })1/r0 ≤ ρ(1−θ0)µ({f ≥ λ })(1−θ0)/s0 k−1 ρ − 1 k
k−1 qk δ(k−1) by dividing by ρ . Now, set ak = ρ µ({f ≥ λk}) and multiply by ρ with r0(1+δ) = q to find θ0 C||∇f||p a1/r0 ≤ ρ−q(1−θ0)/s0 a(1−θ0)/s0 k−1 ρ − 1 k holds for all k ≤ k(f). Note that ak > 0 for all k ≤ k(f), and lim ak = ∞. Set k→−∞ a = inf ak > 0 to find k≤k(f)
θ0 C||∇f||p a1/r0 ≤ ρ−q(1−θ0)/s0 a(1−θ0)/s0 ρ − 1 which implies θ0 2 C||∇f||p a ≥ ρ−q (1−θ0)/(s0θ0) ρ − 1
15 s s Because λ µ({f ≥ λ}) ≤ ||f||s, we find
θ0 s C||∇f||p λ−sρpk||f||s ≥ ρ−q (1−θ0)/s0θ0 k s ρ − 1
1 Letting ε = 0 and setting k = k(f) − 1, ρ = 1 + 1+|q| we find
(1−θ0)/s0θ0 1/(1−s/q) 1/(1−q/s) ||f||∞ ≤ 4(1 + |q|)(e C||∇f||p) ||f||s .
θ Thus, all inequalities (Sr,s) with 0 < s < r ≤ ∞, θ ∈ (0, 1] with q(r, s, θ) = q(r0, s0, θ0) are satisfied on M.
∗,θ0 In addition, if M satisfies (Sr0,s0 ) with q ∈ (−∞, 0) then there exists a constant A such that ∞ −1/q ∀f ∈ C0 (Ω), ||f||∞ ≤ ACµ(Ω) ||∇f||p
∗,θ0 for bounded domains Ω ∈ M, with C the constant from (Sr0,s0 ).
3.3.4 Equivalences under different values of p
In the previous sections when we examined equivalences between weak- and strong-Sobolev ∗,θ inequalities, we fixed p and found equivalences under the same value of q. Allow (Sr,s (p)) θ ∗,θ θ and (Sr,s(p)) refer to the inequalities (Sr,s ) and (Sr,s) as we will now vary p ∈ [1, ∞).
∗,θ0 Assume that on a manifold M, the inequality (Sr0,s0 (p0)) is satisfied for some p0 ∈
[1, ∞), 0 < r0, s0 ≤ ∞, 0 ∈ (0, 1]. Assume that q0 = q(r0, s0, θ0) satisfies 1/q0 < 1/p0, and let v = p0q0 1 = 1 − 1 . By the results of the previous sub-subsections, we know the q0−p0 q0 p0 v σ0 inequalities (Sp0,u0 (p0)) are satisfied for all u0, σ0 such that q(p0, u0, σ0) = q0. ∞ σ0 γ Let f ∈ C0 (M) and γ ≥ 1. Apply some (Sp0,u0 (p0)) to |f| to find
Z σ0/p0 γ (γ−1)p0 p0 γ(1−σ0) ||f||p0γ ≤ C γ |f| |∇f| dµ ||f||u0γ . M The H¨olderinequality states
Z 1/p0 (γ−1)p0 p0 γ−1 f g dµ ≤ ||f||p0γ ||g||p0γ,
which brings us to
γ σ0/p0 (γ−1)σ0 σ0 γ(1−σ0) ||f||p0γ ≤ Cγ ||f||p0γ ||∇f||p0γ||f||u0γ
σ0 pu0 σ If we set p = γp0, σ = , u = we see that the inequality (S (p))is satisfied. In σ0+γ(1−σ0) p0 p,u addition, we have 1 − σ 1 − σ = γ 0 , σ σ0
16 σ0 σ and that the parameter v has not changed from (Sp0,u0 (p0)) to (Sp,u(p)); 1 1 1 1 1 = − = − . v p0 q0 p q(p, u, σ)
σ0 σ Thus, if M satisfies (Sp0,u0 (p0)), M satisfies all (Sp,u(p)) such that 1 1 1 = − . q(p, u, σ) p v This allows us to increase p and construct a strong Sobolev inequality. In addition, the results from the previous three sections show that even under the new value of p, all inequalities θ 1 1 1 (Sr,s(p)) with 0 < r, s ≤ ∞, θ ∈ (0, 1] such that q(r,s,θ) = p − v are satisfied.
3.4 Pseudo-Poincar´eInequalities
A pseudo-Poincar´einequality is tool that can be used to prove a manifold satisfies a Sobolev inequality. Here, we will provide a brief overview of what a psuedo-Poincar´einequality is, and demonstrate its ability to prove what Sobolev inequalities a manifold satisfies. In the next section, a connection between a manifold’s geometry and a psuedo-Poincar´einequality is illustrated. ∞ For any f ∈ C0 (M), set 1 Z ft(x) = f(y)dµ(y). V (x, t) B(x,t) It is said that M satisfies a psuedo-Poincar´einequality in Lp if there exists a constant C such that ∞ ∀f ∈ C0 (M), ∀t > 0, ||f − ft||p ≤ Ct||∇f||p.
||f||1 Note that with this definition of ft, we have ||f||∞ ≤ V (x,t) . The following result pro- vides a connection between psuedo-Poincar´einequalities and Sobolev inequalities of the θ (Sr,s(p))−type, as seen in [1, Theorem 3.3.1].
Theorem 3.5: Assume that M satisfies
∞ ∀f ∈ C0 (M), ∀t > 0, ||f − ft||p ≤ Ct||∇f||p. for some p0 ∈ [1, ∞) and that inf {t−vV (x, t)} > 0 x∈M t>0 θ for some v > 0. Then the inequalities (Sr,s(p)) are satisfied on M for all p ≥ p0, and all 0 < r, s ≤ ∞, θ ∈ (0, 1] such that q(r, s, θ) satisfies 1/q = 1/p − 1/v.
17 Proof. By the result of sub-subsection 3.3.4 Equivalences under different values of p, we only
have to treat the case p = p0. Let M be a manifold that satisfies
∞ ∀f ∈ C0 (M), ∀t > 0, ||f − ft||p ≤ Ct||∇f||p
v for some p0 ∈ [1, ∞). There exists a constant c > 0 such that V (x, t) ≥ (ct) for all ∞ x ∈ M, t > 0. Set a nonnegative f ∈ C0 (M) and a constant λ > 0. For any t > 0, write
µ({f ≥ λ}) ≤ µ({|f − ft| ≥ λ/2}) + µ({ft ≥ λ/2})
||f||1 and set t such that (ct)v = λ/4. Then, µ({ft ≥ λ/2}) = 0 and
µ({f ≥ λ}) ≤ µ({|f − ft| ≥ λ/2}) p p ≤ (2/λ) ||f − ft||p p ≤ (2Ct||∇f||p/λ) 1+2/v 1/v −1−1/vp = (2 C/c)||f||1 ||∇f||pλ .
and so 21+2/vC p λ(vp+p)/vµ({f ≥ λ}) ≤ ||f||1/v||∇f|| . c 1 p v Raising this to the power p+pv we find
v 1/(1+v) p+pv v/(1+v) v/(1+v) λµ({f ≥ λ}) ≤ 4(C/c) ||f||1 ||∇f||p ,