0
$\begingroup$

In the paper Wasserstein GANs are Minimax Optimal Distribution Estimators [1], the authors state within Lemma 3.3, p.12, that Besov spaces [2] of generalized smoothness (i.e., with the added parameter $b$ as compared to [2], defined in section 3.2.1 of [1]) are compactly embedded in Hölder spaces [3] of the same integer regularity, namely $$ \mathcal{B}^{\alpha,1+\epsilon}_{\infty,\infty}(\mathbb{R}^p,\mathbb{R})\hookrightarrow\mathcal{H}^{\alpha}(\mathbb{R}^p,\mathbb{R}),\quad \alpha\in\mathbb{N},\epsilon>0 \label{1}\tag{1} $$

They cite [4] (Proposition 4.3.23) and [5] (4.63) for Lemma 3.3, but the former doesn't treat the specific embedding I'm looking for and the latter doesn't include anything numbered 4.63, at least in the 2007 print [5] I have. And I can't seem to find the result in there myself, likely because I lack background to see it.

I am looking to verify the specific embedding \eqref{1} and would be very thankful for any reference on this.

[1] https://arxiv.org/abs/2311.18613
[2] https://en.wikipedia.org/wiki/Besov_space
[3] https://en.wikipedia.org/wiki/H%C3%B6lder_condition
[4] https://www.cambridge.org/core/books/mathematical-foundations-of-infinitedimensional-statistical-models/C9731BF27A4CDBDB297404EBF1B7820E
[5] https://www.taylorfrancis.com/books/mono/10.1201/9781584887515/envelopes-sharp-embeddings-function-spaces-dorothee-haroske

This question is cross-posted in https://math.stackexchange.com/questions/5091931/reference-request-besov-spaces-are-compactly-embedded-in-h%c3%b6lder-spaces.

Edit: A proof for a continuous embedding has been given in the cross-post on stackexchange.

$\endgroup$
4
  • $\begingroup$ Good point -- could it be a continuous embedding, though? $\endgroup$ Commented Aug 25 at 12:56
  • $\begingroup$ It will be good if you can write down the necessary definitions for your notations and Lemmas, and recall what happen there using your own language. It's very inconvenient to jump between the references. $\endgroup$ Commented Aug 26 at 5:28
  • $\begingroup$ I don't have access to Haroske's book, but perhaps the (4.63) is a typo? Have you checkked chapter 7 in that book? Or at least in part 2? Based on the web preview I see, Besov-type spaces are not treated until part 2 of the book, which starts on Chapter 7. $\endgroup$ Commented Aug 26 at 12:08
  • $\begingroup$ In fact, (7.63) states a condition on $b$ for $B^{\frac{n}{p},b}_{p,q}\hookrightarrow \mathcal{C}$, with $\mathcal{C}$ the bounded, uniformly continuous functions. It seems close, but at least I am unable to see how this would also imply the seemingly stronger embedding into $\mathcal{H}^{\alpha}$. $\endgroup$ Commented Aug 26 at 16:01

1 Answer 1

4
$\begingroup$

Lemma 3.3 cannot be true as stated.

The classical Holder spaces (which the authors denote by $\mathcal{H}^\eta$), for $\eta$ non-integral and positive, is exactly equal to the Besov space $B^{\eta}_{\infty,\infty}$, and so the embedding from Holder into Besov of the same regularity cannot be compact. (It is the identity, and so is continuous.)

For the first embedding, compactness is also false: let $w^{(i)}$ be an enumeration of points in $\mathbb{Z}^p$, and set $f_i = \psi_{0w^{(i)}}$. Then you can check that by their definition the $B^{s,b}_{\infty,\infty}$ norm of $(f_i - f_j)$ is exactly $2$ when $i\neq j$, for any $b$. So this is a bounded sequence in $B^{\alpha,1+\epsilon}_{\infty,\infty}$ that has no convergent subsequence in $B^{\alpha,0}_{\infty,\infty}$.

(More generally, any of the spaces in the $B^{\alpha,\epsilon}_{p,q}$ scale is translation invariant, and so on a non-compact domain you can generate bounded non-convergent sequences by translations, which will prevent compactness of embedding between them.)


As the OP mentioned, in the MSE version of the question, it was shown that $B^{k,1+\epsilon}_{\infty,\infty}$ embeds into $B^{k}_{\infty,1}$ which then embeds into $\mathcal{H}^k$ for $k$ a non-negative ineger. The original article in question also mentioned compactness, so I wish to add one more remark:

Observe that given the smoothness and compact support of the wavelet functions, you have that

$$ |\psi_{j\ell w}(x) - \psi_{j\ell w}(y)| \leq 2^{jp/2} M \min(1, 2^j|x-y|) $$

(The first uses triangle inequality, the second uses the fundamental theorem of calculus, and so $M$ depends on the uniform bound on the wavelet and its first derivative.)

Additionally, the compact support of the wavelet functions means that there is a universal constant $c$ such that at every point $x$ and at every scale $j$ there is at most $c$ of the points $w$ such that $\psi_{j\ell w}(x) \neq 0$.

So using the wavelet expansion, you can estimate

$$ |f(x) - f(y)| \leq 2c M |x-y| \Big[ \sup_w |\alpha_f(w)| + \sum_{j \leq j_*} \sum_\ell \sup_w |\alpha_f(j,\ell,w)| 2^{j(p/2+1)} \Big] \\+ 2c M \sum_{j > j_*} \sum_\ell \sup_w|\alpha_f(j,\ell,w)|2^{jp/2}$$

where $j_*$ is the largest value satisfying $2^{j_*}|x-y| \leq 1$.

The sum inside the brackets can be estimated by

$$ \sum_{j\leq j_*} \sum_\ell \sup_w |\alpha_f(j,\ell,w)| 2^{j(p/2) + j)} \leq \Big[ \sup_j \Big(2^{jp/2} (1+j)^b \sum_\ell \sup_w |\alpha_f(j,\ell,w)| \Big) \Big]\cdot \sum_{j \leq j_*} \frac{2^j}{(1+j)^b} $$

Similarly, the second sum outisde brackets can be estimated by

$$ \sum_{j> j_*} \sum_\ell \sup_w |\alpha(j,\ell,w)|2^{jp/2} \leq \sum_{j > j_*} \frac{1}{(1+j)^b} \cdot \Big[ \sup_j \Big(2^{jp/2} (1+j)^b \sum_\ell \sup_w |\alpha_f(j,\ell,w)| \Big) \Big] $$

So we find that

$$ |f(x) - f(y)| \leq 2cM \|f\|_{B^{0,b}_{\infty,\infty}} \cdot \Big(\sum_{j \leq j_*} \frac{2^{j-j_*}}{(1+j)^b} + \sum_{j > j_*} \frac{1}{(1+j)^b} \Big) $$

when $b > 1$, the final sum is absolutely summable, and is bounded by $(1+j_*)^{1-b}$. The first sum can be estimated in different ways, a very cheap and inefficient way is to split the sum into $j < j_*/2$ and $j \in [j_*/2,j_*]$. We get $$ \sum_{j \leq j_*/2} \frac{2^{j-j_*}}{(1+j)^b} \leq \sum_{j \leq j_*/2} \frac{2^{-j_*/2}}{1} \leq j_* 2^{-j_*/2 - 1} $$ and $$ \sum_{j = j_*/2}^{j_*} \frac{2^{j-j_*}}{(1+j)^b} \leq \sum_{j = j_*/2}^{j_*} \frac{1}{(1+j_*/2)^b} \leq \frac{j_*/2}{(1+j_*/2)^b} \leq (1+j_*/2)^{1-b} $$

To summarize, we find, when $b > 1$, that for $|x-y| < \frac12$ $$ |f(x) - f(y)| \leq \|f\|_{B^{0,b}_{\infty,\infty}} \cdot O( |\log(|x-y|)|^{1-b}) $$ This allows us to draw two conclusions:

  1. The continuous embedding of $B^{0,b}_{\infty,\infty}$ into $\mathcal{H}^0$.
  2. That the unit ball in $B^{0,b}_{\infty,\infty}$ is in fact uniformly equicontinuous, and so restricting to functions with support within a fixed compact set $K$, the embedding is in fact compact.
$\endgroup$
5
  • $\begingroup$ Thank you very much for the detailed answer. If you don't mind, I would appreciate it a lot if you could leave me a pointer to what makes continuity of the first embedding obvious, I struggle to see an immediate way to bound the Hölder norm. $\endgroup$ Commented Aug 26 at 15:45
  • $\begingroup$ @feltshire: you need to take for granted the fact that the Holder norm is equivalent to the $B^{s,0}_{\infty,\infty}$ norm. If you do that, then clearly reducing the $b$ exponent makes the norm smaller. $\endgroup$ Commented Aug 26 at 18:23
  • $\begingroup$ Oh, yes, for non-integer $\eta$, I can see it. I thought you were referring to the integer case. Since the $\mathcal{H}^{\eta}$ and $\mathcal{B}^{\eta,0}_{\infty,\infty}$ norms are not equivalent there (if I am not mistaken), I would expect it to be less immediate. I can follow the proof in the math-se sister-thread, but that would not have been obvious to me, so I thought there might be some property of the wavelet characterisation you were thinking of. $\endgroup$ Commented Aug 26 at 20:34
  • $\begingroup$ @feltshire: I misread your question (and the paper). I thought you were looking at $\alpha \not\in \mathbb{N}$; in the case where $\alpha\in \mathbb{N}$ then you need to do as the MSE answer does. I deleted the irrelevant portions. $\endgroup$ Commented Aug 26 at 21:09
  • $\begingroup$ @feltshire: I added a slightly more involved proof of the continuity of embedding, which has the benefit of showing equicontinuity and hence shows that the embedding is in fact compact with the spatial domain is restricted. $\endgroup$ Commented Aug 27 at 14:16

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.