# concentration inequality for entropy from sample

Consider a measure $\mu$ on a finite set, and let $x_1, \ldots, x_n$ be i.i.d samples from $\mu$. Then the expression $S_n = -\frac{1}{n} \sum_{i=1}^n \log \mu(x_i)$ converges by a.s. to the entropy $H(\mu)$.

What concentration inequalities exist for finite $n$? In other words, what upper bounds are known for the expression $P(|S_n - H(\mu)|>t)$?

Of course, we can get some upper bounds using, for instance, Bernstein's inequality for bounded variables, with bound depending on the size of the smallest atom of $\mu$. However, using the sup norm bound seems very rough in this case, since small atoms inflate the norm but also have small contributions to the mean. On the other hand, using more general versions of Bernstein's inequality with moment bounds does not seem natural. ($\sum_{i \in X} \mu(i) \log^k \mu(i)$ does not seem like a natural quantity).

Since this is a classical situation, there probably exist sharper inequalities than sup norm, using some natural quantities. What are they?

• Not sure how to work out the precise connection, but this is relevant: "How much randomness can be extracted from memoryless Shannon entropy sources?" eprint.iacr.org/2015/591.pdf
– usul
Jan 21, 2016 at 8:13

Actually, Bernstein's inequality does not really require boundedness of the i.i.d. random summands; a finite exponential moment of the absolute value of a random summand will suffice. However, here we can just use Markov's inequality.

Let $X,X_1,\dots,X_n$ be independent identically distributed random variables (i.i.d. r.v.'s) such that $P(X=z)=\mu(z)\in(0,1)$ for all $z$ in a finite set $Z$, with $\sum_z\mu(z)=1$. Let $Y:=-\ln\mu(X)$, $Y_i:=-\ln\mu(X_i)$, and $S_n:=\frac1n\,\sum_1^n Y_i$. Then $$ES_n=EY=-\sum_z\mu(z)\ln\mu(z)=:H(\mu)>0$$ and $$Ee^{hY}=\sum_z\mu(z)^{1-h},\quad Ee^{hS_n}=(Ee^{hY})^n \tag{1}$$ for all real $h$. To avoid trivialities, assume that $\max Y:=-\min_z\ln\mu(z)>-\max_z\ln\mu(z)=:\min Y$. Then $$\min Y<EY=H(\mu)<\max Y.$$

Take now any real $t$ such that $H(\mu)=EY\le t<\max Y$. For all real $h\ge0$, by Markov's inequality, $$P(S_n\ge t)\le\exp\{-nht+\ln Ee^{nhS_n}\}=\exp\{-nht+n\ln Ee^{hY}\}. \tag{2}$$ The derivative of the exponent $-nht+n\ln Ee^{hY}$ in $h$ is $-nt+n\frac{EYe^{hY}}{Ee^{hY}}$, which strictly and continuously increases from $-nt+nEY\le 0$ to $-nt+n\max Y\ge0$ as $h$ increases from $0$ to $\infty$, and so, the upper bound $\exp\{-nht+n\ln Ee^{hY}\}$ on the right-tail probability $P(S_n\ge t)$ is minimized when $h=h_{t,+}=h_{t,\mu,+}$ is the only nonnegative root of the equation $$m(h):=m_\mu(h):=\frac{EYe^{hY}}{Ee^{hY}}=t. \tag{3}$$

Similarly, for any real $t$ such that $H(\mu)=EY\ge t>\min Y$, the best upper exponential bound on the left-tail probability $P(S_n\le t)$ is $\exp\{-nht+n\ln Ee^{hY/n}\}$, where now $h=h_{t,-}=h_{t,\mu,-}$ is the only non-positive root of the equation $(2)$.

Thus, $$P(S_n\ge t)\le e^{-na_+(t)},\quad\text{where a_+(t):=h_{t,+}t-\ln Ee^{h_{t,+}Y}>0} \tag{4}$$ if $H(\mu)=EY<t<\max Y$, $$P(S_n\le t)\le e^{-na_-(t)},\quad\text{where a_-(t):=h_{t,-}t-\ln Ee^{h_{t,-}Y}>0} \tag{5}$$ if $H(\mu)=EY>t>\min Y$ So, these bounds exponentially decrease in $n$ if $t$ is fixed. By formulas (3.7) and (3.8) in [Chernoff], bounds $(4)$ and $(5)$ cannot be improved by replacing $a_\pm(t)$ by greater values.

In view of $(1)$, equation $(3)$ can be rewritten as $$\sum_z(t+\ln\mu(z))\mu(z)^{1-h}=0, \tag{6}$$ and this equation can be easily solved for $h$ numerically if the set $Z$ is not too large.

All this is of course well known, even in the general case of i.i.d. random summands with finite exponential moment of the absolute value. Basically, in this particular situation I have just added the first equality in $(1)$ and rewrote $(3)$ as $(6)$.

• Thanks! But can we say something about the roots of equation (3) and (2) as function of $\mu$? I mean something simpler than that $h$ is determined by the roots of equation that is determined by $\mu$. I'm interested in: "what is the right and simple parameter that controls the deviations of S_n". Jan 21, 2016 at 5:18
• If by "right" you mean "accurate and general enough", then the simplest parameter that controls the large deviations of $S_n$ (say, with $t$ fixed) is the function $m=m_\mu$ defined in $(3)$. You cannot make it simpler: I have now added the corresponding reference to Chernoff. Jan 21, 2016 at 16:26
• If the deviations are only moderately large, that is, $|t|<<n^{-1/3}$, then, by the central limit theorem and Cramer's theorem for large deviations, the tail probabilities $P(S_n\ge|t|)$ and $P(S_n\le-|t|)$ are asymptotic to the corresponding normal ones; so, that normal asymptotic depends only on $n$, $t$, $EY$, and $EY^2$. Jan 21, 2016 at 16:38
• Thanks again for the detailed answer and especially for the Chernoff reference. I agree that the "probabilistic" part of the question is closed. However, in this form, (4) and (5) are difficult to apply. For instance, one could be interested in the question: for general $\mu$, for which $t$ we have $a_{+}(t) > \frac{1}{2}$? Is there a simpler (perhaps not quite tight) form for this $t$? Or alternatively, suppose we set $t = (1+\epsilon) H(\mu)$. Is there a simpler form for $a_{+}(t)$? Jan 21, 2016 at 23:02
• Unfortunately, I don't know good answers to the questions in your latest comment. My suggestion at this point would be to play numerically with various $\mu$'s, and then perhaps you could come up with a plausible conjecture. Jan 21, 2016 at 23:40

Here's a step that seems nice enough to point out. It still leaves a parameter to pick, and I'm not sure it's ever better than applying Bernstein, but it does something different.

We can get a probability bound in terms of how much $S_n$ exceeds the Renyi entropy $H_{\alpha}$ of $\mu$ (equivalently, worded in terms of the $\ell_{\alpha}$ norm of $\mu$), for any $0 < \alpha < 1$. The unresolved question is if we can to pick $\alpha$ to get a nice closed form of some kind. Maybe someone more clever than I can speak to that.

Claim. Let $X_1,\dots,X_n$ be i.i.d. according to $\mu$ and $Y_i = \log(1/\mu(X_i))$; let $S_n = \frac{1}{n} \sum_{i=1}^n Y_i$. Then for any $0 < \alpha < 1$, \begin{align} \Pr[ S_n \geq t ] &\leq 2^{-n (1-\alpha) \left( t - H_{\alpha}(\mu) \right) } \\ &= 2^{-n \left( (1-\alpha)t - \alpha \log \| \mu \|_{\alpha} \right) } . \end{align} Here I'm writing $\mu = (\mu_1,\dots,\mu_m)$ as a vector of probabilities. Note that $H_{\alpha}$ is decreasing in $\alpha$ and $H_1 = H$, Shannon entropy. So as $n \to \infty$, we can pick $\alpha \to 1$ and get tail bounds for $t \to H(\mu)$.

Proof. Using the general Chernoff method, \begin{align} \Pr[S_n \geq t] &= \Pr\left[ 2^{\lambda S_n} \geq 2^{\lambda t}\right] & (\forall \lambda \geq 0) \\ &\leq \frac{\mathbb{E} 2^{\lambda S_n} }{2^{\lambda t}} & (\text{Markov's}). \end{align} We have \begin{align} \mathbb{E} 2^{\lambda S_n} &= \left( \mathbb{E} 2^{\frac{\lambda}{n} Y_1} \right)^n \\ &= \left( \mathbb{E} \mu(X_1)^{-\lambda/n} \right)^n \\ &= \left( \sum_{j=1}^m \mu_j^{1-\lambda/n} \right)^n . \end{align} Hence \begin{align} \Pr[S_n \geq t] \leq 2^{-n \left(\frac{\lambda}{n} t - \log \sum_j \mu_j^{1-\lambda/n} \right)} . \end{align} Pick $\lambda$ such that $1-\lambda/n = \alpha$, for a chosen $\alpha \in [0,1]$. In other words, $\frac{\lambda}{n} = 1-\alpha$, and factoring this out and substituting, \begin{align} \Pr[S_n \geq t] \leq 2^{-n (1-\alpha) \left(t - \frac{1}{1-\alpha} \log \sum_j \mu_j^{\alpha} \right)} . \end{align}

• Your bound is actually better than Bernstein's, but is it different from "mine", except in form? Jan 21, 2016 at 14:02
• Also, there seems to be a typo in your definition of $Y_i$: you probably meant $\mu(X_i)$ there instead of $X_i$. Jan 21, 2016 at 14:24
• @IosifPinelis thanks. I expect you're right that there's no real difference between the bounds.
– usul
Jan 21, 2016 at 15:01
• Renyi entropy is indeed a nice interpretation. The bound itself is identical to (2) in @IosifPinelis answer, except to the base of the logarithm. Jan 21, 2016 at 22:16