1In this work, we do not consider the important problem of getting sharp optimal constants.
Statement
1In this work, we do not consider the important problem of getting sharp optimal constants. Universal property of the empirical mean. Let us first recall the universal estima- tion property satisfied by the empirical mean in the Gaussian model. Let S be a symetric set of Rd and denote by k\cdotkS : x \in Rd \to supv\inS v, x the associated pseudo-norm. Let G1 , . . . , GN be N i.i.d. N (µ∗ , Σ) Gaussian vectors in PN Rd (with mean µ and covariance matrix Σ) and denote by ḠN = (1/N ) i=1 Gi their empirical mean. It follows from the Borell-TIS inequality (Theorem 7.1 in [2] or pages 56-57 in [3]) that for all 0 < \delta < 1, with probability at least 1 − \delta, p ḠN − µ∗ S = sup v, ḠN − µ∗ \leq E sup v, ḠN − µ∗ + \sigmaS 2 log(1/\delta) v\inS v\inS q 2 where \sigmaS = supv\inS E v, ḠN − µ∗ is called the weak variance. It follows that with probability at least 1 − \delta, p ∗ ℓ∗ (Σ1/2 S) supv\inS Σ1/2 v 2 log(1/\delta) (1) ḠN − µ S \leq \sqrt{ } + \sqrt{ } N N ∗ 1/2 1/2 1/2 where ℓ (Σ S) = sup G, x : x \in Σ S = E Σ G S , for G \sim N (0, Id ), is the Gaussian mean width of the set Σ1/2 S. In particular, in the case where S = B2d , we recover in (1) the classical subgaussian rate for the mean estimation w.r.t. the ℓd2 -norm r s Tr(Σ) kΣkop log(1/\delta) (2) + . N N Moreover, the rate achieved by the empirical mean for the mean estimation prob- lem
Record
Comments
Nothing has been published against this problem yet, and nothing has been said about it.
Sign in with an institutional address to take part in the discussion. Reading every thread stays open to everyone.
Sign inSolve with an agent
Open the statement in a chat, with the problem and the ground rules already written into the prompt.