Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • Edge Worth

    Uploaded by

    Eduardo Fe
    2 views3 pages

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Report this Document

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    2 views3 pages

    Edge Worth

    Uploaded by

    Eduardo Fe

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Flag for inappropriate content
    Download as pdf or txt
    of 3

    Edgeworth expansions

    1 Four preliminary facts


    n
    1. You already know that (1 + a/n) → ea . But how good is this approxi-
    mation? The binomial theorem shows (after quite a bit of algebra) that
    for a fixed nonnegative integer k,
       
     a n−k a(a + k) 1
    1+ = ea 1 − +o (1)
    n 2n n
    as n → ∞.

    2. Hermite polynomials: If φ(x) denotes the standard normal density


    function, then we define the Hermite polynomials Hk (x) by the equation

    dk
    (−1)k φ(x) = Hk (x)φ(x). (2)
    dxk
    Thus, we obtain H1 (x) = x, H2 (x) = x2 − 1, H3 (x) = x3 − 3x, and so on.
    By differentiating (2), we obtain

    d
    [Hk (x)φ(x)] = −Hk+1 (x)φ(x). (3)
    dx

    3. An inversion formula for characteristic functions: R Suppose X ∼


    ∞
    G(x) and ψX (t) denotes the characteristic function of X. If −∞ |ψX (t)| dt <
    0
    ∞, then g(x) = G (x) exists and
    Z ∞
    1
    g(x) = e−itx ψX (t) dt. (4)
    2π −∞

    We won’t prove this fact here, but its proof can be found in most books
    on theoretical probability.
    4. An identity involving φ(x): For any positive integer k,
    ∞ ∞
    (−1)k dk
    Z Z
    1 −itx −t2 /2 2
    e e k
    (it) dt = e−itx e−t /2
    dt
    2π −∞ 2π dxk −∞
    dk
    = (−1)k
    φ(x) (5)
    dxk
    = Hk (x)φ(x), (6)
    2
    where (5) follows from (4) since e−t /2 is the characteristic function for a
    standard normal distribution, and (6) follows from (2).

    1
    2 The setup
    Let X1 , . . . , Xn be a simple random sample from F (x). Without loss of general-
    ity, suppose that E X√1 = 0 and Var X1 = 1; for otherwise, we can replace each
    Xj by (Xj − E X1 )/ Var X1 without changing anything that follows. Let

    γ = E Xj3 and τ = E Xj4

    and suppose that τ < ∞. We wish to study the distribution of the standardized
    sum
    n
    1 X
    Sn = √ Xj .
    n j=1

    The central limit theorem tells us that for every x, P (Sn ≤ x) → Φ(x), where
    Φ(x) denotes the standard normal distribution function. But we would like a
    better approximation to P (Sn ≤ x) than Φ(x), and we begin by constructing
    the characteristic function of Sn :
    √ X  √ n
    ψSn (t) = E exp{(it/ n) Xj } = ψX (t/ n) .
    j

    √
    We next use a Taylor expansion of exp{itX/ n}: As n → ∞,

    (it)2 X 2 (it)3 X 3 (it)4 X 4


         
    t itX 1
    ψX √ = E 1+ √ + + √ + +o
    n n 2n 6n n 24n2 n2
    t2 (it)3 γ (it4 )τ
       
    1
    = 1− + √ + +o .
    2n 6n n 24n2 n2

    If we raise this tetranomial to the nth power, most terms are o(1/n):
    n " n  n−1 
    t2 t2 (it)3 γ (it)4 τ
      
    t
    ψX √ = 1− + 1− √ +
    n 2n 2n 6 n 24n
    n−2
    #
    t2 (n − 1)(it)6 γ 2
       
    1
    + 1− 2
    +o . (7)
    2n 72n n

    By equations (1) and (7) we conclude that

    t4 (it)3 γ (it)4 τ (it)6 γ 2


       
    −t2 /2 1
    ψSn (t) = e 1− + √ + + +o
    8n 6 n 24n 72n n
    (it)3 γ (it)4 (τ − 3) (it)6 γ 2
       
    −t2 /2 1
    = e 1+ √ + + +o . (8)
    6 n 24n 72n n

    2
    3 The conclusion
    Putting (8) together with (4), we obtain the following density function as an
    approximation to the distribution of Sn :
    Z ∞ Z ∞
    1 2 γ 2
    g(x) = e−itx e−t /2
    dt + √ e−itx e−t /2
    (it)3 dt
    2π −∞ 6 n −∞
    !
    ∞ ∞
    τ −3 γ2
    Z Z
    2 2
    + e−itx e−t /2
    (it)4 dt + e−itx e−t /2
    (it)6 dt . (9)
    24n −∞ 72n −∞

    Next, combine (9) with (6) to yield

    γH3 (x) (τ − 3)H4 (x) γ 2 H6 (x)


     
    g(x) = φ(x) 1 + √ + + . (10)
    6 n 24n 72n

    By (3), the antiderivative of g(x) equals

    γH2 (x) (τ − 3)H3 (x) γ 2 H5 (x)


     
    G(x) = Φ(x) − φ(x) √ + +
    6 n 24n 72n
    γ(x − 1) (τ − 3)(x − 3x) γ 2 (x5 − 10x3 + 15x)
    2 3
     
    = Φ(x) − φ(x) √ + + .
    6 n 24n 72n

    The expression above is called the second-order Edgeworth expansion. By car-


    rying out the expansion in (8) to more terms, we may obtain higher-order Edge-
    worth expansions. The first-order Edgeworth expansion is

    γ(x2 − 1)
     
    G(x) = Φ(x) − φ(x) √ .
    6 n

    Thus, for a symmetric distribution F (x), γ = 0 and the usual (zero-order)


    central limit theorem approximation Φ(x) is already first-order accurate.
    Incidentally, the second-order Edgeworth expansion explains why the defi-
    nition of kurtosis of a distribution with mean 0 and variance 1 is τ − 3.

    You might also like