For any number , we define the positive part and negative part of x as, respectively:
(Some authors use .)
The following identities hold:
For any , we have:
If , then
Let with pdf f and cdf F. Let and be the pdf and cdf, respectively, of the standard normal distribution.
We define
for . Moreover,
Throughout, we use and to refer to the first‐ and second‐order loss functions, and and to refer to the corresponding complementary loss functions. 1 It would be equally appropriate to use for the first‐order loss function, but we drop the superscript for notational simplicity, and often omit the phrase “first‐order” when describing this function and its complement. For the standard normal distribution, we replace n with in these functions.
Let X be a continuous random variable with pdf f and cdf F. Let be the complementary cdf. The loss function and complementary loss function are given by
The loss function and its complement are related as follows:
The derivatives of the loss function and its complement are given by
The loss function and its complement are therefore both convex.
The second‐order loss function and its complement are given by
The second‐order loss function and its complement are related as follows:
The derivatives of the second‐order loss function and its complement are given by
Let , with pdf , cdf , and complementary cdf . The standard normal loss function, its complement, and their derivatives are given by
Also:
(The second equality follows from the fact that .)
The second‐order standard normal loss function, its complement, and their derivatives are given by
Let with pdf f, cdf F, and complementary cdf . The normal loss function can be computed using the standard normal loss function as follows:
where . (In many instances, we assume so that the probability that is small; in these cases, we often replace the lower limit of the integral in (C.32) with 0.) The derivatives of and are given by (C.15)–(C.16).
The second‐order normal loss function and its complement are given by
Let X be a discrete random variable with pmf f and cdf F. Let be the complementary cdf. The loss function and complementary loss function are given by
The loss function and its complement are related as follows:
The second‐order loss function and its complement are given by
The second‐order loss function and its complement are related as follows:
If X is nonnegative, then equations C.37) and (C.40) can facilitate the calculation of and , since and contain infinite sums, but and contain finite ones.
Let with pmf f, cdf F, and complementary cdf . The Poisson loss function and complementary loss function are given by
The second‐order Poisson loss function and its complement are given by
Equation (C.49) is known as Leibniz's rule.
If , then:
Microsoft Excel and MATLAB have several built‐in functions for computing normal distributions. Let with pdf f and cdf F and with pdf and cdf . Then, in Excel:
And, in MATLAB:
The following formulas computes partial expectations of a random variable with pdf f and cdf F. (If and , these each equal the true mean.)
Discrete versions are also available:
For a continuous random variable X and constants a and b, the identities above can be used to prove: