Open chrisyeh96 opened 7 months ago
I'm sorry for the confusion. The key to this argument could, perhaps with some work, have been squeezed into a footnote (and I may try to do so now).
We begin with the expression on the LHS of Adler and Taylor's sufficient condition (their (1.4.4)), which, by no coincidence, was also discussed in #47:
E|ϕ - ϕ'|².
Let us assume that K is α-Hölder continuous and show that Adler and Taylor's condition is satisfied. Let's ignore the mean function for now and assume the process is centered, μ ≡ 0 [*].
By assumption, we have
E|ϕ - ϕ'|² = K(x, x) + K(x', x') - 2K(x, x') ≤ C|x - x'|ᵃ
for some C ≥ 0. Consider the function
h(d) = dᵃ |log d|¹⁺ᵃ
on the (compact) interval [0, 1]. Note that h(0) = h(1) = 0 on the boundary and that h > 0 on the interior (0, 1). Thus h achieves a maximum M > 0 (depending on alpha) somewhere in (0, 1). That implies
E|ϕ - ϕ'|² ≤ C|x - x'|ᵃ ≤ CM / |log |x - x'||¹⁺ᵃ
for |x - x'| < 1. That is sufficient for Adler and Taylor's condition.
[*] As before, to deal with nonzero μ just assume it's continuous and add it back in.
Did this help @chrisyeh96?
I'm trying to understand the Gaussian process sample continuity theorem stated on page 35. The text says:
Would it be possible to provide any additional guidance on how to derive the theorem stated on p.35 of the Bayes Opt book from Adler and Taylor's Theorem 1.4.1? I tried reading Adler and Taylor's Theorem, but I struggled to understand it.