Download or view karatsuba.frink in plain text format
/** This library calculates nearest square roots of integers using the
Karatsuba square root recursive algorithm. It also contains
functions that build upon the integer routines to calculate floating-point
(and rational) square roots of integers and floating-point numbers to an
arbitrary number of digits.
For reference to the Karatsuba square root algorithm see:
https://hal.inria.fr/inria-00072854/PDF/RR-3805.pdf
*/
/** Perform Karatsuba square root, returning a floating-point square root of an
integer or calculating the floating-point square root of a floating-point
number (by first converting it to a giant integer.)
This works using a rather simple trick. If we need a lot of
digits in the result, we "shift left" the number by the number of bits
we need (by multiplying by 2^m bits, (see how m is calculated below))
and round it to an integer. We then do an integer square root of this big
integer, and divide the result by 2^(m/2) bits which turns the result into
the correct floating-point (or actually rational) square root of the
number.
(Note that this trick also elegantly works to *reduce* the number of digits
we have to calculate by "shifting right" when the requested number of
digits is small. For example, karatsubaFloat[10^1000, 3] would
normally take a 1000-digit number and use the full force of Frink to
calculate 3 digits of its square root.)
This algorithm is able to calculate a limited number of digits, which
makes it suitable for providing a good initial guess (with a large number
of correct digits) for an operation like Newton iteration for larger
numbers. Newton iteration approximately doubles the number of correct
digits per iteration if given a good guess, so an optimal algorithm *may*
use this algorithm and Newton iteration. Note that as of Java 1.8,
the integer operations from Frink were incorporated into Java, making
integer operations much faster than floating-point for large numbers.
If rational==false, (the default,) this returns a floating-point number.
If rational==true, this returns a rational number (which may be faster)
*/
karatsubaSqrt[n, digits=getPrecision[], rational=false] :=
{
oldPrec = getPrecision[]
try
{
m = ceil[digits log[10]/log[2] - approxLog2[n]]
setPrecision[digits+3]
// Make the number of bits even
if m mod 2 == 1
m = m + 1
shift = 2^m
[num, remainder] = sqrtRem[round[n * shift]]
denom = 2^(m/2)
// println["$num / $denom, remainder = $remainder"]
result = num/denom
/* This adjusts the result using the remainder and a Taylor series
approximation:
The Karatsuba square root algorithm only returns integers, so it
estimates the nearest integer to the result (called s in the paper)
as:
result = floor[sqrt[n]]
And the remainder (called r in the paper) is:
remainder = n - result^2
thus, the correct value of sqrt[n] is:
sqrt[n] = sqrt[result^2 + remainder]
However, this is not easily turned into a number without doing
another whole square root algorithm, so we correct it by calculating
a Taylor series for the corrections in the remainder. That is, the
Taylor series of the above: sqrt[result^2 + remainder]
Due to the choice of bits that we chose in the initial calculation,
1 term in the Taylor series seems to be just right at enough.
Without using the remainder, we have to calculate *twice* as many
digits in the recursive algorithm above to get correct answers. By
using the remainder, we can get all digits correct by calculating a
couple more digits than requested, rather than doubling the number of
digits.
The following terms are the first terms of the Taylor series.
First term: r / (2 s)
*/
radj = remainder / (2 result shift)
result = result + radj
//println["Remainder adjustment is $radj"]
/* Second-order Taylor series remainder adjustment. This is generally
not necessary and may only affect the last digit or so.
Second term: -r^2 (8 s^3)
*/
/* radj2 = - remainder^2 / (8 (result shift)^3)
// println["Remainder adjustment (second order) is $radj2"]
result = result + radj2 */
setPrecision[digits]
if rational == false
return 1. * result // Return floating-point number
else
return result // Return rational number
}
finally
{
setPrecision[oldPrec]
}
}
/* Karatsuba recursive square root.
For explication of algorithm, see:
https://hal.inria.fr/inria-00072854/PDF/RR-3805.pdf
Thanks to Jeremy Roach for much hard work at turning my broken
implementation into a working implementation.
*/
sqrtRem[n] :=
{
// Fallback case when numbers are smallish (TODO: tune this threshold)
if n < 2^256
{
s = intSqrt[n]
return [s, n - s^2]
}
[w, z] = divideAndRemainder[bitLength[n], 4]
w = ((4 w + z) + ((4-z) mod 4)) / 4
b = 2^w
if z == 1 || z == 2
{
fixup = true
n = 4 n
} else
{
fixup = false
}
a0 = bitAnd[n, b-1]
n = shiftRight[n, w]
a1 = bitAnd[n, b-1]
n = shiftRight[n, w]
a2 = bitAnd[n, b-1]
n = shiftRight[n, w]
a3 = bitAnd[n, b-1]
[s1, r1] = sqrtRem[a3*b + a2]
[q, u] = divideAndRemainder[r1*b + a1, 2*s1]
s = s1*b + q
r = u*b + a0 - q^2
if r < 0
{
r = r + 2*s - 1
s = s - 1
}
if fixup
return [s/2, r/4]
return [s, r]
}
/* Integer square root: returns the greatest integer less than or equal to the
square root of n. Only call this with integers.
This is Exercise 5.7 in Bressoud with my own modifications for better
initial guess.
*/
intSqrt[n] :=
{
a = 2^((bitLength[n]+1) div 2)
b = a - 1
while b<a
{
a = b
b = (a^2 + n) div (2*a)
}
return a
}
Download or view karatsuba.frink in plain text format
This is a program written in the programming language Frink.
For more information, view the Frink
Documentation or see More Sample Frink Programs.
Alan Eliasen was born 20217 days, 23 hours, 41 minutes ago.