Talagrand's concentration inequality

From HandWiki

In the probability theory field of mathematics, Talagrand's concentration inequality is an isoperimetric-type inequality for product probability spaces.[1][2] It was first proved by the French mathematician Michel Talagrand.[3] The inequality is one of the manifestations of the concentration of measure phenomenon.[2]

Statement

The inequality states that if [math]\displaystyle{ \Omega = \Omega_1 \times \Omega_2 \times \cdots \times \Omega_n }[/math] is a product space endowed with a product probability measure and [math]\displaystyle{ A }[/math] is a subset in this space, then for any [math]\displaystyle{ t \ge 0 }[/math]

[math]\displaystyle{ \Pr[A] \cdot \Pr\left[{A^c_t}\right] \le e^{-t^2/4} \, , }[/math]

where [math]\displaystyle{ {A^c_t} }[/math] is the complement of [math]\displaystyle{ A_{t} }[/math] where this is defined by

[math]\displaystyle{ A_t = \{ x \in \Omega ~:~ \rho(A,x) \le t \} }[/math]

and where [math]\displaystyle{ \rho }[/math] is Talagrand's convex distance defined as

[math]\displaystyle{ \rho(A,x) = \max_{\alpha, \|\alpha\|_2 \le 1} \ \min_{y \in A} \ \sum_{i~:~x_i \neq y_i} \alpha_i }[/math]

where [math]\displaystyle{ \alpha \in \mathbf{R}^n }[/math], [math]\displaystyle{ x,y \in \Omega }[/math] are [math]\displaystyle{ n }[/math]-dimensional vectors with entries [math]\displaystyle{ \alpha_i, x_i, y_i }[/math] respectively and [math]\displaystyle{ \|\cdot\|_2 }[/math] is the [math]\displaystyle{ \ell^2 }[/math]-norm. That is,

[math]\displaystyle{ \|\alpha\|_2=\left(\sum_i\alpha_i^2\right)^{1/2} }[/math]

References

  1. Alon, Noga; Spencer, Joel H. (2000). The Probabilistic Method (2nd ed.). John Wiley & Sons, Inc.. ISBN 0-471-37046-0. 
  2. 2.0 2.1 Ledoux, Michel (2001). The Concentration of Measure Phenomenon. American Mathematical Society. ISBN 0-8218-2864-9. 
  3. Talagrand, Michel (1995). "Concentration of measure and isoperimetric inequalities in product spaces". Publications Mathématiques de l'IHÉS (Springer-Verlag) 81: 73–205. doi:10.1007/BF02699376. ISSN 0073-8301.