EXP(x) approximation in old 1980's computer ROM

337 Views Asked by At

The presentation :

Old 1980’s ROM (Apple 2e, Commodore 64, ...) uses a Taylor’s series-like to evaluate the exponential function EXP(x) :

EXP(x) = 1 + x + 1/2! x^2 + ... + 1/7! x^7

x is in [-1, 1], thanks to a preliminary work. Another preliminary work is to multiply x by a correction factor 1/ln(2) (better for the algorithm I suppose). So the series must take into account this factor and modify the coefficients :

EXP(x) = 1 + ln(2) x + ln(2)^2 1/2! x^2 + ... + ln(2)^7 1/7! x^7
EXP(x) = a0 + a1 x + a2 x^2 + ... a7 x^7

a0 = 1.000000000
a1 = 0.693147181 ln(2)
a2 = 0.240226507
a3 = 0.055504109
a4 = 0.009618129
a5 = 0.001333356
a6 = 0.000154035
a7 = 0.000015253

To reduce the Taylor error, especially near -1 and 1, the old 1980’s algorithm modify the coefficients with the Chebyshev polynomials approximation. As the comment says :

“TCHEBYSHEV MODIFIED TAYLOR SERIES COEFFICIENTS FOR E**X”

b0 = 1.000000000
b1 = 0.693147186
b2 = 0.240226385
b3 = 0.055505127
b4 = 0.009614017
b5 = 0.001342263
b6 = 0.000143523
b7 = 0.000021499

(see https://github.com/wsoltys/multicomp/blob/master/ROMS/6809/basic.asm line 3852)

Now, the problem :

Where these Chebyshev values come from ?

I did a lot of calculations to retreive these values. I used the normal method : compute the Chebyshev coefficients for EXP(x) up to 8 terms, then replace the Ti(x) by its x polynomial. Or by the economization method : compute Taylor EXP(x) up to x^9, and remove a9*T9(x)/2^8 and a8*T8(x)/2^7 (normally equivalent to the first method). No chance !

The computed Chebyshev coefficients I obtain, in accordance with the literature, are :

c0 = 0.999999801   delta   b0 - c0 =  0.000000199
c1 = 0.693147113           b1 - c1 =  0.000000073
c2 = 0.240229556           b2 - c2 = -0.000003171
c3 = 0.055504542           b3 - c3 =  0.000000585
c4 = 0.009610825           b4 - c4 =  0.000003192
c5 = 0.001332605           b5 - c5 =  0.000009658
c6 = 0.000159622           b6 - c6 = -0.000016099
c7 = 0.000015734           b7 - c7 =  0.000005765

Have you an idea ? Is the ROM coefficients really Chebyshev coefficients ? Is this another approximation method ? Am I wrong ?

Thank you for your help.

BR,

see https://retrocomputing.stackexchange.com/questions/7065/1980s-rom-which-expn-algorithm-used in retrocomputing forum