Collatz Conjecture

For me the most fasinating thing about the Collatz Conjecture is how it so blantantly points to a deficiency in mathematics. It seems incredibly obvious that it is true. The problem lies in a way to say that no loops are created excluding one - a limited understanding of process dynamics.

Recently, I've changed the problem to eliminate the split condition - there is no reason to divide even numbers by two. Without the split we have a simplified algorithm, but it must work with larger numbers. This doesn't seem like an intuitive step.


1. y <= 1+Shr[Xor[x,x-1],1]

2. if x=y then step 5

3. x <= 3x + y

4. goto step 1

5. number is a power of two



The first step calculates the value of least set bit. Instead of dividing by two, one is scaled up to match the bit which would make (x) odd if it were down scaled. Only when a single bit is set does the algorithm end (step 2). (y) can be tracked progressively instead of looking at the complete representation of (x) as step 3 always clears the least set bit (or more). Of course, I've coded this in x86:

; check Collatz Conjecture on multi-percision number
;
; carry flag set on array overflow
; carry flag clear on pass
; (otherwise it loops forever: not likely, but hasn't been proven)
;
; NOTE: array should be twice the size of number to test - too large
; is not a problem as only active bits are operated on.
CollatzArrayQ:
        label .arr at esp+4+32 ; BigInt
        label .len at esp+8+32 ; length (dwords)

        pushad
        mov ebx,[.arr]
        mov eax,[.len]
        mov edx,1
        mov esi,ebx
        lea ebp,[ebx+eax*4]
; address past number in EDI (active end marker)
; (need at least one zero dword on end!)
.z:     dec eax
        test dword [ebx+eax*4],-1
        je .z
        lea edi,[ebx+eax*4+4]
        cmp edi,ebp
        je .over

; advance to lowest set bit
.0:     test [esi],edx
        jne .1
.0a:    rol edx,1
        jnc .0 ; HINT: taken
        add esi,4
        cmp esi,ebp     ; end
        cmovnc esi,ebx  ; reset
        jmp .0
; ESI[ECX] is first set bit, multiply by three and add ESI[ECX]
; all (virtual) lower bits are zero.
.1:
        ; test if buffer is == ESI[ECX] then done!
        cmp [esi],edx
        jne .kx ; HINT: taken
        ; if EDI==ESI+4 or (ESI+4==EBP && EDI==EBX)
        ; (faster than testing all of buffer)
        mov ecx,edi
        lea eax,[esi+4]
        cmp edi,ebx
        cmove ecx,ebp
        cmp eax,ecx
        je .x
.kx:
        push edx
        push esi
.2:
        mov eax,3
        mov ecx,edx
        mul dword [esi]
        add eax,ecx
        adc edx,0
        mov [esi],eax

        add esi,4
        cmp esi,ebp     ; end
        cmovnc esi,ebx  ; reset

        cmp esi,edi
        jne .2

        test edx,edx
        je .5

        cmp edi,[esp]
        je .over
        mov [edi],edx
        add edi,4
        cmp edi,ebp     ; end
        cmovnc edi,ebx  ; reset

.5:     pop esi
        pop edx
        jmp .0a ; we just cleared that bit

.x:     popad
        clc
        retn 8

.over:  popad
        stc
        retn 8

Try a really big number - like 2^131072 + 3

TestBuffer rb 32*1024 ; zero bytes

mov dword [TestBuffer],3
mov dword [TestBuffer+16*1024],1

push 32*1024/4 ; L1 cache size (c:
push TestBuffer1
call CollatzArrayQ
; 3829ms on my machine

Why is this important? The algorithm is no longer concerned with value of (x) except for the exit condition - the same operation takes place each loop. It is sufficient to show that (y) -> (x) for all (x), and the conjecture is true. The multiplication by three is irrelavant because the addition propells the least set bit at a rate greater than the constant multiple!

(c:

Update:

Status of the 3x+1 class record progress as of: Dec 13, 2007. Intervals are
indicated in blocks of 20,000,000,000,000 (20 E+12 or 20 trillion numbers).
Calculation of a single block takes approximately 20 days of (idle) time on a
Pentium 1000 MHz.

Basically, a trillion per day, per Ghz.

Naturally, I'm currious how my routine compares:
Currently, they are working on 7 * 2^53 block of numbers. I ran a test of my algorithm for 2^24 numbers (in that block range) which took 38781ms. Which means it is roughly 25 times slower. Their algorithm is tuned for this size and uses large data structures (to avoid processing numbers). I might be able to double the speed of my algorithm, but confirming ranges of numbers is not what it's designed for.