The first and second derivertive of log moment generating function for $Z$, $\psi(\lambda)$, can be expressed with exponential tilting by $ \frac {e^{\lambda Z}} {\mathop{\mathbb{E}} \left \lbrack e^{\lambda Z} \right \rbrack } $ introduced in this post, and from this the convexity of $\psi(\lambda)$ is shown. Using the fact we see in this post, Hoeffding’s inequality is immediately derived, as we will see in the later post.

As before, $\psi(\lambda)$ denotes $ \log \mathbb{E} e^{ \lambda (Z - \mathbb{E} Z )} $ and by exponential tilting, the original measure $P$ is tilted into $ Q = P_{ \frac {e^{\lambda Z}} {\mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack } }.
$ Then, we have [ \psi’(\lambda) = \mathbb{E}_{Q} \left \lbrack Z - \mathbb{E}Z \right \rbrack, ] [ \psi’'(\lambda) = \text{Var}_{Q} \left \lbrack Z \right \rbrack, ] and [ \psi(\lambda) = \psi’’(\lambda) = 0, \ \ \psi’'(0) = \text{Var} \left \lbrack Z \right \rbrack, ] where $\mathbb{E}$ and $\text{Var}$ is the moment over $P$ as $\mathbb{E} = \mathbb{E}_{P}$ and $\text{Var} = \text{Var}_{P}$.

First derivertive

As $ \psi(\lambda) = \log \mathbb{E} e^{ \lambda Z } - \log \mathbb{E} e^{ \lambda \mathbb{E} Z } = \log \mathbb{E} e^{ \lambda Z } - \lambda \mathbb{E} Z$, we know [ \psi(0) = 0. ] The well known property of $\psi(\lambda)$ is that it is differentiable any number of times. The first derivertive is [ \psi’(\lambda) = \frac { \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack } { \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack } - \mathbb{E} Z = \mathbb{E}_{Q} Z - \mathbb{E} Z = \mathbb{E}_{Q} \left \lbrack Z - \mathbb{E} Z \right \rbrack. ] From the fact that $Q = P$ when $\lambda = 0$, we know [ \psi’(0) = 0. ]

Second derivertive

Next, since $\mathbb{E} Z$ is constant w.r.t. $\lambda$, we have [ \psi’'(\lambda) = \frac{d}{d \lambda} \left ( \frac { \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack } { \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack } \right ). ] As $ \left ( \frac{f}{g} \right ) = \frac{1}{g^{2}} \left ( f’g - fg’ \right ) $, the above is [ = \frac{1}{ \left ( \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack \right )^{2}} \left ( \frac{d}{d \lambda} \left ( \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack \right ) \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack - \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack \frac{d}{d \lambda} \left ( \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack \right ) \right ) ] [ = \frac{1}{ \left ( \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack \right )^{2}} \left ( \mathbb{E} \left \lbrack Z^{2} e^{\lambda Z} \right \rbrack \mathbb{E} \left \lbrack e^{\lambda Z} \right \rbrack - \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack \mathbb{E} \left \lbrack Z e^{\lambda Z} \right \rbrack \right ). ] By transforming the measure from $P$ to $Q$, the above is [ =\mathbb{E}_{Q} Z^{2} - \left ( \mathbb{E}_{Q} Z \right )^{2} = \text{Var}_{Q} \left \lbrack Z \right \rbrack \geq 0. ] As the second derivatibe is grater than $0$, it shows that $\psi(\lambda)$ is convex.

$psi$

Lastly, from the fact that $Q = P$ when $\lambda = 0$, we know [ \psi’'(0) = \text{Var}(Z). ]