Version 4 of Incomplete Beta Function

Updated 2008-01-01 18:18:51 by EKB

EKB This is an implementation of the incomplete Beta function Ix(a, b), defined as:

                      / x
               1     |     a-1       b-1
  Ix(a,b) = -------  | dt t   (1 - t)
            B(a, b)  |
                    / 0

where B(a,b) is the Beta function. The incomplete Beta function is the cumulative probability function for the Beta distribution. This code has been tested as part of the tests run on the Beta distribution. Here are the results of the tests, where the expected results were calculated using R:

  Call: ::beta::cdf-beta 2.1 3.0 0.2
    Error: PASSED
    Result: 0.16220409270890568
    Expected: 0.16220409275804 ± 1.0e-9

  Call: ::beta::cdf-beta 4.2 17.3 0.5
    Error: PASSED
    Result: 0.998630771122991
    Expected: 0.998630771123192 ± 1.0e-9

  Call: ::beta::cdf-beta 500 375 0.7
    Error: PASSED
    Result: 1.0
    Expected: 1.0 ± 1.0e-9

  Call: ::beta::cdf-beta 250 760 0.2
    Error: PASSED
    Result: 0.0001252342758440994
    Expected: 0.000125234318666948 ± 1.0e-9

  Call: ::beta::cdf-beta 43.2 19.7 0.6
    Error: PASSED
    Result: 0.07288812943897244
    Expected: 0.0728881294218269 ± 1.0e-9

  Call: ::beta::cdf-beta 500 640 0.3
    Error: PASSED
    Result: 0.0
    Expected: 2.99872547567313e-23 ± 1.0e-9

  Call: ::beta::cdf-beta 400 640 0.3
    Error: PASSED
    Result: 3.040258600428558e-9
    Expected: 3.07056696205524e-09 ± 1.0e-9

  Call: ::beta::cdf-beta 0.1 30 0.1
    Error: PASSED
    Result: 0.9986410086716231
    Expected: 0.998641008671625 ± 1.0e-9

  Call: ::beta::cdf-beta 0.01 0.03 0.9
    Error: PASSED
    Result: 0.7658650057031781
    Expected: 0.765865005703006 ± 1.0e-9

  Call: ::beta::cdf-beta 2 3 0.9999
    Error: PASSED
    Result: 0.9999999999960003
    Expected: 0.999999999996 ± 1.0e-9

 Time for ::beta::cdf-beta 2 3 0.9999:
 98.56 microseconds per iteration

 Time for ::beta::cdf-beta 250 760 0.2:
 3343.65 microseconds per iteration

The time for "::beta::cdf-beta 250 760 0.2" is not good. it would take 3 seconds to run 1000 values, so performance at larger values of a & b does not appear to be very good. More efficient algorithms are available, such as the one used in DCDFLIB (http://people.scs.fsu.edu/~burkardt/f_src/dcdflib/dcdflib.html ).

Here's the code:

    package require math

    namespace import ::math::ln_Gamma
    namespace import ::math::Beta
    #
    # Implement the incomplete beta function Ix(a, b)
    #

    proc incompleteBeta {a b x {tol 1.0e-9}} {
        if {$x < 0.0 || $x > 1.0} {
            error "Value out of range in incomplete Beta function: x = $x, not in \[0, 1\]"
        }
        if {$a <= 0.0} {
            error "Value out of range in incomplete Beta function: a = $a, must be > 0"
        }
        if {$b <= 0.0} {
            error "Value out of range in incomplete Beta function: b = $b, must be > 0"
        }

        if {$x < $tol} {
            return 0.0
        }
        if {$x > 1.0 - $tol} {
            return 1.0
        }

        # Function will converge faster if x is smaller
        if {$a > $b} {
            return [incBeta_series $a $b $x $tol]
        } else {
            set z [incBeta_series $b $a [expr {1.0 - $x}] $tol]
            return [expr {1.0 - $z}]
        }
    }

    #####################################################
    #
    # Series expansion for Ix(a, b)
    #
    # Abramowitz & Stegun formula 26.5.4
    #
    # This series has terms
    #
    #   [B(a+1,n)/B(a+b,n)] * x^n
    #
    # This is guaranteed to converge only if
    # b < 1. Use the recurrence formula 26.5.15
    # from A&S to bring b below 1. Recurrence is:
    #
    # Ix(a,b) = (G(a+b)/(G(a+1)G(b))) * x^a * (1-x)^(b-1) +
    #               Ix(a+1, b-1)
    #
    # Also, use B(a,b) = G(a)G(b)/G(a+b) to
    # rewrite coeff as
    #
    #  Cn = [G(a+1)/G(a+b)] * G(a+b+n)/G(a+n+1)
    #
    #####################################################
    proc incBeta_series {a b x tol} {
        # a+b is invariant under recurrence formula
        set aplusb [expr {$a + $b}]
        set lnGapb [ln_Gamma $aplusb]

        # Calculate for convenience -- these don't change
        set lnx [expr {log($x)}]
        set ln1mx [expr {log(1.0 - $x)}]
        set z [expr {$x / (1.0 - $x)}]

        set retval 0.0
        # Pack everything in one big exp, since individual terms can over/underflow
        set factor [expr {exp($lnGapb - [ln_Gamma [expr {$a + 1}]] - [ln_Gamma $b] + \
                $a * $lnx + ($b - 1) * $ln1mx)}]
        while {$b > 1} {
            set retval [expr {$retval + $factor}]
            set a [expr {$a + 1}]
            set b [expr {$b - 1}]
            set factor [expr {$factor * $z * double($b)/double($a)}]
        }

        set pref_num [expr {pow($x, $a) * pow(1.0 - $x, $b)}]
        set pref_denom [expr {$a * [Beta $a $b]}]
        set pref [expr {$pref_num/$pref_denom}]

        set term 1.0
        set z 1.0
        set sum 1.0
        set n 0

        set adjtol [expr {$tol * $pref}]
        set factor 1.0

        while 1 {
            set z [expr {$z * $x}]

            set abn [expr {$a + $n + $b}]
            set an1 [expr {$a + $n + 1}]
            set factor [expr {$factor * double($abn)/double($an1)}]

            incr n

            set nexterm [expr {$z * $factor}]

            if {abs($nexterm - $term) < $adjtol} {
                break
            }
            #tk_messageBox -message "$x : $term : $nexterm"
            set term $nexterm
            set sum [expr {$sum + $term}]
        }

        # Add previous cumulative sum to the series expansion
        set retval [expr {$retval + $pref * $sum}]

        # Because of imprecision in underlying Tcl calculations, may fall out of bounds
        if {$retval < 0.0} {
            set retval 0.0
        } elseif {$retval > 1.0} {
            set retval 1.0
        }

        return $retval

    }