McDiarmid不等式

X_{1},X_{2},...,X_{n}为一组独立随机变量,假设存在常数c_{1},c_{2},...,c_{n},对任意1\leqslant i\leq n,映射函数f满足:

\underset{x_{1},...,x_{n},x_{i}^{'}}{\sup }\left | f(x_{1},...,x_{n}) -f(x_{1},...,x_{i-1},x_{i}^{'},x_{i+1},...,x_{n})\right |\leq c_{i}

也就说,改变第i个随机变量的值,函数f的值最大变化c_{i}

那么对任何\epsilon > 0,有

P((f(X_{1},...,X_{n})-E\left [ f(X_{1},...,X_{n}) \right ])\geq \epsilon )\leq exp(-\frac{2\epsilon ^{2}}{\sum_{i=1}^{n}c_{i}^{2}})

P((f(X_{1},...,X_{n})-E\left [ f(X_{1},...,X_{n}) \right ]) \leq -\epsilon )\leq exp(-\frac{2\epsilon ^{2}}{\sum_{i=1}^{n}c_{i}^{2}})

P(\left | f(X_{1},...,X_{n})-E\left [ f(X_{1},...,X_{n}) \right ] \right |\geq \epsilon )\leq 2exp(-\frac{2\epsilon ^{2}}{\sum_{i=1}^{n}c_{i}^{2}})

参考资料:

McDiarmid's inequality | encyclopedia article by TheFreeDictionary

你可能感兴趣的:(概率论)