https://blog.paulhankin.net/fibonacci/
This code, somewhat surprisingly, generates Fibonacci numbers.
def fib(n):
return (4 << n(3+n)) // ((4 << 2n) - (2 << n) - 1) & ((2 << n) - 1)
In this blog post, I'll explain where it comes from and how it works.
Before getting to explaining, I'll give a whirlwind background overview of Fibonacci numbers and how to compute them. If you're already a maths whiz, you can skip most of the introduction, quickly skim the section "Generating functions" and then read "An integer formula".
Overview
The Fibonacci numbers are a well-known sequence of numbers:
1,1,2,3,5,8,13,21,34,55,…
The nth number in the sequence is defined to be the sum of the previous two, or formally by this recurrence relation:
Fib(0)Fib(1)Fib(n)===11Fib(n−1)+Fib(n−2)
I've chosen to start the sequence at index 0 rather than the more usual 1.
Computing Fibonacci numbers
There's a few different reasonably well-known ways of computing the sequence. The obvious recursive implementation is slow:
def fib_recursive(n):
if n < 2: return 1
return fib_recursive(n - 1) + fib_recursive(n - 2)
An iterative implementation works in O(n) operations:
def fib_iter(n):
a, b = 1, 1
for _ in xrange(n):
a, b = a + b, a
return b
And a slightly less well-known matrix power implementation works in O(log n) operations.
def fib_matpow(n):
m = numpy.matrix('1 1 ; 1 0') ** n
return m.item(0)
The last method works by considering the a and b in fib_iter as sequences, and noting that
(an+1bn+1)=(1110)(anbn)
From which follows
(anbn)=(1110)n(11)
and so if m=(1110)n then bn=m11 (noting that unlike Python, matrix indexes are usually 1-based).
It's O(log n) based on the assumption that numpy's matrix power does something like exponentation by squaring.
Another method is to find a closed form for the solution of the recurrence relation. This leads to the real-valued formula: Fib(n)=(ϕn+1−ψn+1)/5‾√) where ϕ=(1+5‾√)/2 and ψ=(1−5‾√)/2. The practical flaw in this method is that it requires arbitrary precision real-valued arithmetic, but it works for small n.
def fib_phi(n):
phi = (1 + math.sqrt(5)) / 2.0
psi = (1 - math.sqrt(5)) / 2.0
return int((phi ** (n+1) - psi ** (n+1)) / math.sqrt(5))
Generating Functions
A generating function for an arbitrary sequence an is the infinite sum Σnanxn. In the specific case of the Fibonacci numbers, that means ΣnFib(n)xn. In words, it's an infinite power series, with the coefficient of xn being the nth Fibonacci number.
Now,
Fib(n+2)=Fib(n+1)+Fib(n)
Multiplying by xn+2 and summing over all n, we get:
ΣnFib(n+2)xn+2=ΣnFib(n+1)xn+2+ΣnFib(n)xn+2
If we let F(x) to be the generating function of Fib, which is defined to be ΣnFib(n)xn then this equation can be simplified:
F(x)−x−1=x(F(x)−1)+x2F(x)
and simplifying,
F(x)=xF(x)+x2F(x)+1
We can solve this equation for F to get
F(x)=11−x−x2
It's surprising that we've managed to find a small and simple formula which captures all of the Fibonacci numbers, but it's not yet obvious how we can use it. We'll get to that in the next section.
A technical aside is that we're going to want to evaluate F at some values of x, and we'd like the power series to converge. We know the Fibonacci numbers grow like ϕn and that geometric series Σnan converge if |a|<1, so we know that if |x|<1/ϕ≃0.618 then the power series converges.
An integer formula
Now we're ready to start understanding the Python code.
To get the intuition behind the formula, we'll evaluate the generating function F at 10−3.
F(x)=11−x−x2F(10−3)=11−10−3−10−6=1.001002003005008013021034055089144233377610988599588187…
Interestingly, we can see some Fibonacci numbers in this decimal expansion: 1,1,2,3,5,8,13,21,34,55,89. That seems magical and surprising, but it's because F(10−3)=Fib(0)+Fib(1)/103+Fib(2)/106+Fib(3)/109+….
In this example, the Fibonacci numbers are spaced out at multiples of 1/1000, which means once they start getting bigger that 1000 they'll start interfering with their neighbours. We can see that starting at 988 in the computation of F(10−3) above: the correct Fibonacci number is 987, but there's a 1 overflowed from the next number in the sequence causing an off-by-one error. This breaks the pattern from then on.
But, for any value of n, we can arrange for the negative power of 10 to be large enough that overflows don't disturb the nth Fibonacci. For now, we'll just assume that there's some k which makes 10−k sufficient, and we'll come back to picking it later.
Also, since we'd like to use integer maths (because it's easier to code), let's multiply by 10kn, which also puts the nth Fibonacci number just to the left of the decimal point, and simplify the equation.
10knF(10−k)=10kn1−10−k−10−2k=10kn+2k102k−10k−1
If we take this result modulo 10k, we'll get the nth Fibonacci number (again, assuming we've picked k large enough).
Before proceeding, let's switch to base 2 rather than base 10, which changes nothing but will make it easier to program.
2knF(2−k)=2k(n+2)22k−2k−1
Now all that's left is to pick a value of k large enough so that Fib(n+1)<2k. We know that the Fibonacci numbers grow like ϕn, and ϕ<2, so k=n+1 is safe.
So! Putting that together:
Fib(n)≡2(n+1)nF(2−(n+1)) mod 2n+1≡2(n+1)(n+2)(2n+1)2−2n+1−1 mod 2n+1≡2(n+1)(n+2)22n+2−2n+1−1 mod 2n+1
If we use left-shift notation that's available in python, where a<<k=a⋅2k then we can write this as:
Fib(n)≡4<<n(3+n)(4<<2n)−(2<<n)−1mod (2<<n)
Observing that mod (2<<n) can be expressed as the bitwise and (&) of (2<<n)−1, we reconstruct our original Python program:
def fib(n):
return (4 << n(3+n)) // ((4 << 2n) - (2 << n) - 1) & ((2 << n) - 1)
Although it's curious to find a non-iterative, closed-form solution, this isn't a practical method at all. We're doing integer arithmetic with integers of size O(n2) bits, and in fact, before performing the final bit-wise and, we've got an integer that is the first n Fibonacci numbers concatenated together!