Abstract.
The logarithmic Sobolev inequality is fundamental in mathematical physics. Associated stability estimates are equivalent to uncertainty principles. Via a second moment bound, estimates are obtained in one dimension and similar -quantitative estimates are investigated.
1. Introduction
The classical logarithmic Sobolev inequality LSI attributed to L. Gross states
(1.1) |
|
|
|
where , is normalized, , is the LSI deficit [6, 3].
Carlen [2] characterized the equality cases: equality holds in (1.1) if and only if .
Note that if is normalized and centered, equality is valid if and only if . Therefore, a natural problem is to identify a metric , , and such that
(1.2) |
|
|
|
with normalized and centered. In a recent paper, a variant of (1.2) with , a constant , and was established [4]. In [7], was proven to be sharp. Thus, the problem of identifying the strongest norm was a natural consequence which was addressed in [7].
Set
|
|
|
Then
|
|
|
|
|
|
Therefore (1.2) has an equivalent version
|
|
|
in the space of non-negative functions which satisfy
|
|
|
|
|
|
Suppose as ,
then [7] yields
|
|
|
in if and only if
|
|
|
Thus, some additional assumptions are necessary to prove
(1.3) |
|
|
|
with . Supposing the associated measures satisfy a PoincarΓ© inequality via , the stability with the sharp and also best constant was proven in [5].
It has already been underscored in [1] that a uniform upper bound on the second moment is necessary but not sufficient to prove (1.3) with the norm if . A sufficient condition was shown with some type of uniform exponential moment condition. Observe a fourth moment bound also yields (1.3) [7]. The problem of understanding the extra information that is needed with assuming a second moment bound was discussed on [1, p. 6]. The main result in my paper is that if , while the second moment control does not imply stability in , it indeed implies stability in :
Theorem 1.1.
(1) If is normalized and centered in
|
|
|
then there exists so that
|
|
|
(2) If is normalized and centered in
|
|
|
then there exists so that
|
|
|
Therefore a natural curiosity is the higher dimensional analog but that remains an open problem. However the following can be proven:
Corollary 1.2.
(1) If , is normalized and centered in
|
|
|
then
|
|
|
with as in Theorem 1.1.
(2) If , is normalized and centered in
|
|
|
then
|
|
|
with as in Theorem 1.1.
The Wasserstein distance between two probability measures with , , is
|
|
|
where the infimum is taken over all probability measures on with marginals and .
In particular, is called the KantorovichβRubinstein distance. The stability for has already appeared in [8]:
let be a centered probability measure, . There exists a constant such that
|
|
|
βstability is not true if one merely has finite second moments, therefore cannot be taken independent of [9, Theorem 1.2].
Examples in [9] allude to
|
|
|
with . Supposing a type of uniform exponential moment condition, this holds:
Theorem 1.3.
If is normalized and centered in , , ,
|
|
|
then there exists so that
|
|
|
Also, supposing combined with
|
|
|
the best possible necessarily has a lower bound:
|
|
|
When ,
|
|
|
In addition, the exponent on the deficit is sharp.