The phenomenon of entropy concentration provides strong support for the maximum entropy method, MaxEnt, for inferring a probability vector from information in the form of constraints. Here we extend this phenomenon, in a discrete setting, to non-negative integral vectors not necessarily summing to 1. We show that linear constraints that simply bound the allowable sums suffice for concentration to occur even in this setting. This requires a new, `generalized' entropy measure in which the sum of the vector plays a role. We measure the concentration in terms of deviation from the maximum generalized entropy value, or in terms of the distance from the maximum generalized entropy vector. We provide non-asymptotic bounds on the concentration in terms of various parameters, including a tolerance on the constraints which ensures that they are always satisfied by an integral vector. Generalized entropy maximization is not only compatible with ordinary MaxEnt, but can also be considered an extension of it, as it allows us to address problems that cannot be formulated as MaxEnt problems.
翻译:映射浓度现象为最大引温方法( MaxEnt)提供了有力的支持, 以便从限制形式的信息中推断出一种概率矢量。 在这里, 我们将这种现象在离散的环境中扩大到非负性整体矢量, 而不一定是向1 。 我们显示, 仅仅约束允许量的线性限制足以使浓缩即使在这一背景下也发生。 这要求采用一种新的“ 通用” 的酶测量法, 使矢量的总和发挥作用。 我们用偏离最大通用的引温值或与最大通用向量的距离来衡量其浓度。 我们从各种参数的角度对集中度提供非被动的界限, 包括对各种限制的容忍度, 包括确保它们总是被一个整体矢量所满足的制约。 普通的 MaxEnt 最大化不仅与普通的 MaxEnt 兼容, 还可以被视为它的延伸, 因为它使我们能够解决无法被发展成为 MaxEnt 的问题。