I. P RELIMINARIES Let Z1 , . . . , Zn be independent and identically distributed, real-valued, zero-mean random variables. Let m(τ ) = E e τ Z 1 be the moment-generating function of these random variables and ψ(τ ) = log m(τ ) be the cumulant- generating function. We use the notation m0 , m00 and m000 to denote the first three derivatives of m(τ ). Similarly, ψ 0 , ψ 00 , and ψ 000 denote the first three derivatives of ψ(τ ). A random variable Z is said to be lattice if it is supported on the points b, b ± h, b ± 2h, . . . for some b and h. A random variable that is not lattice will be referred to as nonlattice. Throughout, we will focus only on nonlattice random variables. Finally, we will denote by Q(·) the Gaussian Q function: Z∞ 2 1 u Q(x) = √ exp − du. 2π 2 x
II. T HE SADDLEPOINT EXPANSION
The goal is to estimate accurately the tail probability " n # 1X P Z` > γ n `=1
for some γ > 0, in the regime in which both the central-limit theorem and the large-deviation bound provided by Chernoff inequality provide loose estimates. The saddlepoint method [1] yields such an accurate estimate. The resulting expansion is given below. A self- contained proof (for a slightly more general setup) can be found in, e.g., [2, App. I.A]. Theorem 1 (saddlepoint approximation): Let the zero-mean i.i.d. random variables {Z` }n`=1 be nonlattice. Suppose that there exists a τ > 0 and a τ < 0 such that sup m000 (τ ) < ∞ τ ∈[τ ,τ ]
and inf ψ 00 (τ ) > 0. τ ∈[τ ,τ ]
Then, if γ ≥ 0 and the solution to the stationary equation ψ 0 (τ ) = γ gives a τ ∈ [0, τ ], we have " n # X K(τ, n) 1 n[ψ(τ )−τ ψ 0 (τ )] P Z` ≥ nγ = e Ψ(τ, n) + √ +o √ . (1) n n `=1
where u2 p ψ 00 (τ ) Ψ(τ, n) = en 2 Q u nψ 00 (τ )
ψ 000 (τ ) 1 u2 nψ 00 (τ ) 3 00 3/2 3/2 K(τ, n) = −√ + √ − u ψ (τ ) n Ψ(τ, n) 6ψ 00 (τ )3/2 2π 2π √ √ and o(1/ n) comprises terms that vanish faster than 1/ n and are uniform in τ , i.e., √ o(1/ n) lim sup √ = 0. n→∞ τ ∈[0,τ ) 1/ n 0 2
A. Remarks Here is the intuition behind the saddlepoint approximation. One performs the usual exponential tilting on the distribution of Z that is needed to prove the achievability of large-deviation exponent [ψ(τ ) − τ ψ 0 (τ )] (see e.g., [3, Ch. 5.11]). This allows one to pull out of the integral the exponential term. The pre-exponential factor is computed by approximating the distribution of the average of the tilted random variables by a Gaussian distribution using the central-limit theorem.
III. N UMERICAL EXAMPLE
Assume that the {Z` }n`=1 are independent Gamma(k, θ)-distributed random variables, i.e., their probability density function is given by: 1 fZ (x) = xk−1 e−x/θ . Γ(k)θk Then their sum is Gamma(nk, θ)-distributed. Hence, the tail probability " n # X P Z` > nγ `=1
can be easily evaluated numerically. We want to compare the approximation obtained using the saddlepoint method, as well as the normal approximation resulting from the central limit-theorem " n # X n(µ − γ) P Z` ≥ nγ ≈ Q √ (2) `=1 nσ 2 where µ = E [Z1 ] and σZ2 = Var[Z1 ], and the Chernoff bound " n # X 0 P Z` ≥ nγ ≤ en[ψ(τ )−τ ψ (τ )] (3) `=1
where τ is the solution of ψ 0 (τ ) = γ .
Throughout, we set k = 4, θ = 1, and n = 100. It then follows that µ = σ 2 = 4. Furthermore, ψ(τ ) = −4 log(1 − τ ). Hence, for all γ > 4, ψ 0 (τ ) = γ if τ = 1 − 4/γ . In the figure we compare the exact tail probability with the normal approximation√ (2), the Chernoff bound (3) and the saddle-point approximation in Theorem 1, obtained by neglecting the o(1/ n) term in (1). As expected, the normal approximation is accurate for values of γ close to the mean, and the Chernoff bound captures the correct slope of decay of the error probability. The saddlepoint expansion is accurate over the entire range of γ values considered in the figure.
R EFERENCES [1] J. L. Jensen, Saddlepoint approximations. Oxford, U.K.: Oxford Univ. Press, 1995. [2] A. Lancho, J. Ostman, G. Durisi, T. Koch, and G. Vazquez-Vilar, “Saddlepoint Approximations for Short-Packet Wireless Communications,” IEEE Trans. Wireless Commun., 2020. [3] G. R. Grimmett and D. R. Stirzaker, Probability and Random Processes, 3rd ed. Oxford, U.K.: Oxford Univ. Press, 2001. 3
γ Fig. 1. The probability P[ n `=1 Z` ≥ nγ] as a function of γ. The Z` are independent Gamma(4, 1) random variables. The exact tail P proability is compared with the normal approximation (2), the Chernoff bound (3), and the saddle-point approximation in Theorem 1.
