I want to generate a sequence of strings with the following properties:
Lexically ordered
Theoretically infinite
Compact over a realistic range
Generated by a simple process of incrementation
Matches the regexp /\w+/
The obvious way to generate a lexically-ordered sequence is to choose a string length and pad the strings with a base value like this: 000000, 000001, etc. This approach poses a trade-off between the number of permutations and compactness: a string long enough to yield many permutations will be filled many zeros along the way. Plus, the length I choose sets an upper bound on the total number of permutations unless I have some mechanism for expanding the string when it maxes out.
So I came up with a sequence that works like this:
Each string consists of a "head", which is a base-36 number, followed by an underscore, and then the "tail", which is also a base-36 number padded by an increasing number of zeros
The first cycle goes from 0_0 to 0_z
The second cycle goes from 1_00 to 1_zz
The third cycle goes from 2_000 to 2_zzz, and so on
Once the head has reached z and the tail consists of 36 zs, the first "supercycle" has ended. Now the whole sequence starts over, except the z remains at the beginning, so the new cycle starts with z0_0, then continues to z1_00, and so on
The second supercycle goes zz0_0, zz1_00, and so on
Although the string of zs in the head could become unwieldy over the long run, a single supercycle contains over 10^56 permutations, which is far more than I ever expect to use. The sequence is theoretically infinite but very compact within a realistic range. For instance, the trillionth permutation is a succinct 7_bqd55h8s.
I can generate the sequence relatively simply with this javascript function:
function genStr (n) {
n = BigInt(n);
let prefix = "",
cycle = 0n,
max = 36n ** (cycle + 1n);
while (n >= max) {
n -= max;
if (cycle === 35n) {
prefix += "z";
cycle = 0n;
} else {
cycle++;
}
max = 36n ** (cycle + 1n);
}
return prefix
+ cycle.toString(36)
+ "_"
+ n.toString(36).padStart(Number(cycle) + 1, 0);
}
The n parameter is a number that I increment and pass to the function to get the next member of the sequence. All I need to keep track of is a simple integer, making the sequence very easy to use.
So obviously I spent a lot of time on this and I think it's pretty good, but I'm wondering if there is a better way. Is there a good algorithm for generating a sequence along the lines of the one I'm looking for?
A close idea to yours. (more rafined than my first edit...).
Let our alphabet be A = {0,1,2,3}.
Let |2| mean we iterate from 0 to 2 and |2|^2 mean we generate the cartesian product in a lexically sorted manner (00,01,10,11).
We start with
0 |3|
So we have a string of length 2. We "unshift" the digit 1 which "factorizes" since any 0|3|... is less than 1|3|^2.
1 |3|^2
Same idea: unshift 2, and make words of length 4.
2 |3|^3
Now we can continue and generate
3 |2| |3|^3
Notice |2| and not |3|. Now our maximum number becomes 32333. And as you did, we can now add the carry and start a new supercycle:
33 0|3|
This is a slight improvement, since _ can now be part of our alphabet: we don't need to reserve it as a token separator.
In our case we can represent in a supercycle:
n + n^2 + ... + n^(n-1) + (n-1) * n^(n-1)
\-----------------------/\--------------/
geometric special
In your case, the special part would be n^n (with the nuance that you have theorically one char less so replace n with n-1 everywhere)
The proposed supercycle is of length :
P = (n \sum_{k = 0}^{n-2} n^k) + (n-1) * n^(n-1)
P = (n \sum_{k = 0}^{n-3} n^k) + n^n
P = n(n^{n-2} - 1)/(n-1) + n^n
Here is an example diff with alphabet A={0,1,2}
my genStr(grandinero)
,00 0_0
,01 0_1
,02 0_2
,100 1_00
,101 1_01
,102 1_02
,110 1_10
,111 1_11
,112 1_12
,120 1_20
,121 1_21
,122 1_22
,2000 2_000
,2001 2_001
,2002 2_002
,2010 2_010
,2011 2_011
,2012 2_012
,2020 2_020
,2021 2_021
,2022 2_022
,2100 2_100
,2101 2_101
,2102 2_102
,2110 2_110
,2111 2_111
,2112 2_112
,2120 2_120
,2121 2_121
,2122 2_122
22,00 2_200 <-- end of my supercycle if no '_' allowed
22,01 2_201
22,02 2_202
22,100 2_210
22,101 2_211
22,102 2_212
22,110 2_220
22,111 2_221
22,112 2_222 <-- end of yours
22,120 z0_0
That said, for a given number x, we can can count how many supercycles (E(x / P)) there are, each supercycle making two leading e (e being the last char of A).
e.g: A = {0,1,2} and x = 43
e = 2
P = n(n^{n-2} - 1)/(n-1) + n^n = 3(3^1 -1)/2 + 27 = 30
// our supercycle is of length 30
E(43/30) = 1 // 43 makes one supercycle and a few more "strings"
r = x % P = 13 // this is also x - (E(43/30) * 30) (the rest of the euclidean division by P)
Then for the left over (r = x % P) two cases to consider:
either we fall in the geometric sequence
either we fall in the (n-1) * n^(n-1) part.
1. Adressing the geometric sequence with cumulative sums (x < S_w)
Let S_i be the cumsum of n, n^2,..
S_i = n\sum_{k = 0}^{i-1} n^k
S_i = n/(n-1)*(n^i - 1)
which gives S_0 = 0, S_1 = n, S_2 = n + n^2...
So basically, if x < S_1, we get 0(x), elif x < S_2, we get 1(x-S_1)
Let S_w = S_{n-1} the count of all the numbers we can represent.
If x <= S_w then we want the i such that
S_i < x <= S_{i+1} <=> n^i < (n-1)/n * x + 1 <= n^{i+1}
We can then apply some log flooring (base(n)) to get that i.
We can then associate the string: A[i] + base_n(x - S_i).
Illustration:
This time with A = {0,1,2,3}.
Let x be 17.
Our consecutive S_i are:
S_0 = 0
S_1 = 4
S_2 = S_1 + 4^2 = 20
S_3 = S_2 + 4^3 = 84
S_w = S_{4-1} = S_3 = 84
x=17 is indeed less than 84, we will be able to affect it to one of the S_i ranges.
In particular S_1==4 < x==17 <= S_2==20.
We remove the strings encoded by the leading 0(there are a number S_1 of those strings).
The position to encode with the leading 1 is
x - 4 = 13.
And we conclude the thirteen's string generated with a leading 1 is base_4(13) = '31' (idem string -> '131')
Should we have had x = 21, we would have removed the count of S_2 so 21-20 = 1, which in turn gives with a leading 2 the string '2001'.
2. Adressing x in the special part (x >= S_w)
Let's consider study case below:
with A = {0,1,2}
The special part is
2 |1| |2|^2
that is:
2 0 00
2 0 01
2 0 02
2 0 10
2 0 11
2 0 12
2 0 20
2 0 21
2 0 22
2 1 20
2 1 21
2 1 22
2 1 10
2 1 11
2 1 12
2 1 20
2 1 21
2 1 22
Each incremented number of the second column (here 0 to 1 (specified from |1|)) gives 3^2 combination.
This is similar to the geometric series except that here each range is constant. We want to find the range which means we know which string to prefix.
We can represent it as the matrix
20 (00,01,02,10,11,12,20,21,22)
21 (00,01,02,10,11,12,20,21,22)
The portion in parenthesis is our matrix.
Every item in a row is simply its position base_3 (left-padded with 0).
e.g: n=7 has base_3 value '21'. (7=2*3+1).
'21' does occur in position 7 in the row.
Assuming we get some x (relative to that special part).
E(x / 3^2) gives us the row number (here E(7/9) = 0 so prefix is '20')
x % 3^2 give us the position in the row (here base_3(7%9)='21' giving us the final string '2021')
If we want to observe it remember that we substracted S_w=12 before to get x = 7, so we would call myGen(7+12)
Some code
Notice the same output as long as we stand in the "geometric" range, without supercycle.
Obviously, when carry starts to appear, it depends on whether I can use '_' or not. If yes, my words get shorter otherwise longer.
// https://www.cs.sfu.ca/~ggbaker/zju/math/int-alg.html
// \w insensitive could give base64
// but also éè and other accents...
function base_n(x, n, A) {
const a = []
while (x !== 0n) {
a.push(A[Number(x % n)])
x = x / n // auto floor with bigInt
}
return a.reverse().join('')
}
function mygen (A) {
const n = A.length
const bn = BigInt(n)
const A_last = A[A.length-1]
const S = Array(n).fill(0).map((x, i) => bn * (bn ** BigInt(i) - 1n) / (bn - 1n))
const S_w = S[n-1]
const w = S_w + (bn - 1n) * bn ** (bn - 1n)
const w2 = bn ** (bn - 1n)
const flog_bn = x => {
// https://math.stackexchange.com/questions/1627914/smart-way-to-calculate-floorlogx
let L = 0
while (x >= bn) {
L++
x /= bn
}
return L
}
return function (x) {
x = BigInt(x)
let r = x % w
const q = (x - r) / w
let s
if (r < S_w) {
const i = flog_bn(r * (bn - 1n) / bn + 1n)
const r2 = r - S[i]
s = A[i] + base_n(r2, bn, A).padStart(i+1, '0')
} else {
const n2 = r - S_w
const r2 = n2 % w2
const q2 = (n2 - r2 ) / w2
s = A_last + A[q2] + base_n(r2, bn, A).padStart(n-1, '0')
}
// comma below __not__ necessary, just to ease seeing cycles
return A_last.repeat(2*Number(q)) +','+ s
}
}
function genStr (A) {
A = A.filter(x => x !== '_')
const bn_noUnderscore = BigInt(A.length)
return function (x) {
x = BigInt(x);
let prefix = "",
cycle = 0n,
max = bn_noUnderscore ** (cycle + 1n);
while (x >= max) {
x -= max;
if (cycle === bn_noUnderscore - 1n) {
prefix += "z";
cycle = 0n;
} else {
cycle++;
}
max = bn_noUnderscore ** (cycle + 1n);
}
return prefix
+ base_n(cycle, bn_noUnderscore, A)
+ "_"
+ base_n(x, bn_noUnderscore, A).padStart(Number(cycle) + 1, 0);
}
}
function test(a, b, x){
console.log(a(x), b(x))
}
{
console.log('---my supercycle is shorter if underscore not used. Plenty of room for grandinero')
const A = '0123456789abcdefghijklmnopqrstuvwxyz'.split('').sort((a,b)=>a.localeCompare(b))
let my = mygen(A)
const grandinero = genStr(A)
test(my, grandinero, 1e4)
test(my, grandinero, 1e12)
test(my, grandinero, 106471793335560744271846581685593263893929893610517909620n) // cycle ended for me (w variable value)
}
{
console.log('---\n my supercycle is greater if underscore is used in my alphabet (not grandinero since "forbidden')
// underscore used
const A = '0123456789abcdefghijklmnopqrstuvwxyz_'.split('').sort((a,b)=>a.localeCompare(b))
let my = mygen(A)
const grandinero = genStr(A)
test(my, grandinero, 1e12)
test(my, grandinero, 106471793335560744271846581685593263893929893610517909620n) // cycle ended for me (w variable value)
test(my, grandinero, 1e57) // still got some place in the supercycle
}
After considering the advice provided by #kaya3 and #grodzi and reviewing my original code, I have made some improvements. I realized a few things:
There was a bug in my original code. If one cycle ends at z_z (actually 36 z's after the underscore, but you get the idea) and the next one begins at z0_0, then lexical ordering is broken because _ comes after 0. The separator (or "neck") needs to be lower in lexical order than the lowest possible value of the head.
Though I was initially resistant to the idea of rolling a custom baseN generator so that more characters can be included, I have now come around to the idea.
I can squeeze more permutations out of a given string length by also incrementing the neck. For example, I can go from A00...A0z to A10...A1z, and so on, thus increasing the number of unique strings I can generate with A as the head before I move on to B.
With that in mind, I have revised my code:
// this is the alphabet used in standard baseN conversions:
let baseAlpha = "0123456789abcdefghijklmnopqrstuvwxyz";
// this is a factory for creating a new string generator:
function sequenceGenerator (config) {
let
// alphabets for the head, neck and body:
headAlpha = config.headAlpha,
neckAlpha = config.neckAlpha,
bodyAlpha = config.bodyAlpha,
// length of the body alphabet corresponds to the
// base of the numbering system:
base = BigInt(bodyAlpha.length),
// if bodyAlpha is identical to an alphabet that
// would be used for a standard baseN conversion,
// then use the built-in method, which should be
// much faster:
convertBody = baseAlpha.startsWith(bodyAlpha)
? (n) => n.toString(bodyAlpha.length)
// otherwise, roll a custom baseN generator:
: function (n) {
let s = "";
while (n > 0n) {
let i = n % base;
s = bodyAlpha[i] + s;
n = n / base;
}
return s;
},
// n is used to cache the last iteration and is
// incremented each time you call `getNext`
// it can optionally be initialized to a value other
// than 0:
n = BigInt(config.start || 0),
// see below:
headCycles = [0n],
cycleLength = 0n;
// the length of the body increases by 1 each time the
// head increments, meaning that the total number of
// permutations increases geometrically for each
// character in headAlpha
// here we cache the maximum number of permutations for
// each length of the body
// since we know these values ahead of time, calculating
// them in advance saves time when we generate a new
// string
// more importantly, it saves us from having to do a
// reverse calculation involving Math.log, which requires
// converting BigInts to Numbers, which breaks the
// program on larger numbers:
for (let i = 0; i < headAlpha.length; i++) {
// the maximum number of permutations depends on both
// the string length (i + 1) and the number of
// characters in neckAlpha, since the string length
// remains the same while the neck increments
cycleLength += BigInt(neckAlpha.length) * base ** BigInt(i + 1);
headCycles.push(cycleLength);
}
// given a number n, this function searches through
// headCycles to find where the total number of
// permutations exceeds n
// this is how we avoid the reverse calculation with
// Math.log to determine which head cycle we are on for
// a given permutation:
function getHeadCycle (n) {
for (let i = 0; i < headCycles.length; i++) {
if (headCycles[i] > n) return i;
}
}
return {
cycleLength: cycleLength,
getString: function (n) {
let cyclesDone = Number(n / cycleLength),
headLast = headAlpha[headAlpha.length - 1],
prefix = headLast.repeat(cyclesDone),
nn = n % cycleLength,
headCycle = getHeadCycle(nn),
head = headAlpha[headCycle - 1],
nnn = nn - headCycles[headCycle - 1],
neckCycleLength = BigInt(bodyAlpha.length) ** BigInt(headCycle),
neckCycle = nnn / neckCycleLength,
neck = neckAlpha[Number(neckCycle)],
body = convertBody(nnn % neckCycleLength);
body = body.padStart(headCycle , bodyAlpha[0]);
return prefix + head + neck + body;
},
getNext: function () { return this.getString(n++); }
};
}
let bodyAlpha = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ_abcdefghijklmnopqrstuvwxyz",
getStr = sequenceGenerator({
// achieve more permutations within a supercycle
// with a larger headAlpha:
headAlpha: "123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz",
// the highest value of neckAlpha must be lower than
// the lowest value of headAlpha:
neckAlpha: "0",
bodyAlpha: bodyAlpha
});
console.log("---supercycle length:");
console.log(Number(getStr.cycleLength));
console.log("---first two values:")
console.log(getStr.getNext());
console.log(getStr.getNext());
console.log("---arbitrary large value (1e57):");
console.log(getStr.getString(BigInt(1e57)));
console.log("");
// here we use a shorter headAlpha and longer neckAlpha
// to shorten the maximum length of the body, but this also
// decreases the number of permutations in the supercycle:
getStr = sequenceGenerator({
headAlpha: "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz",
neckAlpha: "0123456789",
bodyAlpha: bodyAlpha
});
console.log("---supercycle length:");
console.log(Number(getStr.cycleLength));
console.log("---first two values:");
console.log(getStr.getNext());
console.log(getStr.getNext());
console.log("---arbitrary large value (1e57):");
console.log(getStr.getString(BigInt(1e57)));
EDIT
After further discussion with #grodzi, I have made some more improvements:
I realized that the "neck" or separator wasn't providing much value, so I have gotten rid of it. Later edit: actually, the separator is necessary. I am not sure why I thought it wasn't. Without the separator, the beginning of each new supercycle will lexically precede the end of the previous supercycle. I haven't changed my code below, but anyone using this code should include a separator. I have also realized that I was wrong to use an underscore as the separator. The separator must be a character, such as the hyphen, which lexically precedes the lowest digit used in the sequence (0).
I have taken #grodzi's suggestion to allow the length of the tail to continue growing indefinitely.
Here is the new code:
let baseAlpha = "0123456789abcdefghijklmnopqrstuvwxyz";
function sequenceGenerator (config) {
let headAlpha = config.headAlpha,
tailAlpha = config.tailAlpha,
base = BigInt(tailAlpha.length),
convertTail = baseAlpha.startsWith(tailAlpha)
? (n) => n.toString(tailAlpha.length)
: function (n) {
if (n === 0n) return "0";
let s = "";
while (n > 0n) {
let i = n % base;
s = tailAlpha[i] + s;
n = n / base;
}
return s;
},
n = BigInt(config.start || 0);
return {
getString: function (n) {
let cyclesDone = 0n,
headCycle = 0n,
initLength = 0n,
accum = 0n;
for (;; headCycle++) {
let _accum = accum + base ** (headCycle + 1n + initLength);
if (_accum > n) {
n -= accum;
break;
} else if (Number(headCycle) === headAlpha.length - 1) {
cyclesDone++;
initLength += BigInt(headAlpha.length);
headCycle = -1n;
}
accum = _accum;
}
let headLast = headAlpha[headAlpha.length - 1],
prefix = headLast.repeat(Number(cyclesDone)),
head = headAlpha[Number(headCycle)],
tail = convertTail(n),
tailLength = Number(headCycle + initLength);
tail = tail.padStart(tailLength, tailAlpha[0]);
return prefix + head + tail;
},
getNext: function () { return this.getString(n++); }
};
}
let alpha = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ_abcdefghijklmnopqrstuvwxyz",
genStr = sequenceGenerator({headAlpha: alpha, tailAlpha: alpha});
console.log("--- first string:");
console.log(genStr.getString(0n));
console.log("--- 1e+57");
console.log(genStr.getString(BigInt(1e+57)));
console.log("--- end of first supercycle:");
console.log(genStr.getString(63n*(1n-(63n**63n))/(1n-63n)-1n));
console.log("--- start of second supercycle:");
console.log(genStr.getString(63n*(1n-(63n**63n))/(1n-63n)));
It would be great if someone could point me towards an algorithm that would allow me to :
create a random square matrix, with entries 0 and 1, such that
every row and every column contain exactly two non-zero entries,
two non-zero entries cannot be adjacent,
all possible matrices are equiprobable.
Right now I manage to achieve points 1 and 2 doing the following : such a matrix can be transformed, using suitable permutations of rows and columns, into a diagonal block matrix with blocks of the form
1 1 0 0 ... 0
0 1 1 0 ... 0
0 0 1 1 ... 0
.............
1 0 0 0 ... 1
So I start from such a matrix using a partition of [0, ..., n-1] and scramble it by permuting rows and columns randomly. Unfortunately, I can't find a way to integrate the adjacency condition, and I am quite sure that my algorithm won't treat all the matrices equally.
Update
I have managed to achieve point 3. The answer was actually straight under my nose : the block matrix I am creating contains all the information needed to take into account the adjacency condition. First some properties and definitions:
a suitable matrix defines permutations of [1, ..., n] that can be build like so: select a 1 in row 1. The column containing this entry contains exactly one other entry equal to 1 on a row a different from 1. Again, row a contains another entry 1 in a column which contains a second entry 1 on a row b, and so on. This starts a permutation 1 -> a -> b ....
For instance, with the following matrix, starting with the marked entry
v
1 0 1 0 0 0 | 1
0 1 0 0 0 1 | 2
1 0 0 1 0 0 | 3
0 0 1 0 1 0 | 4
0 0 0 1 0 1 | 5
0 1 0 0 1 0 | 6
------------+--
1 2 3 4 5 6 |
we get permutation 1 -> 3 -> 5 -> 2 -> 6 -> 4 -> 1.
the cycles of such a permutation lead to the block matrix I mentioned earlier. I also mentioned scrambling the block matrix using arbitrary permutations on the rows and columns to rebuild a matrix compatible with the requirements.
But I was using any permutation, which led to some adjacent non-zero entries. To avoid that, I have to choose permutations that separate rows (and columns) that are adjacent in the block matrix. Actually, to be more precise, if two rows belong to a same block and are cyclically consecutive (the first and last rows of a block are considered consecutive too), then the permutation I want to apply has to move these rows into non-consecutive rows of the final matrix (I will call two rows incompatible in that case).
So the question becomes : How to build all such permutations ?
The simplest idea is to build a permutation progressively by randomly adding rows that are compatible with the previous one. As an example, consider the case n = 6 using partition 6 = 3 + 3 and the corresponding block matrix
1 1 0 0 0 0 | 1
0 1 1 0 0 0 | 2
1 0 1 0 0 0 | 3
0 0 0 1 1 0 | 4
0 0 0 0 1 1 | 5
0 0 0 1 0 1 | 6
------------+--
1 2 3 4 5 6 |
Here rows 1, 2 and 3 are mutually incompatible, as are 4, 5 and 6. Choose a random row, say 3.
We will write a permutation as an array: [2, 5, 6, 4, 3, 1] meaning 1 -> 2, 2 -> 5, 3 -> 6, ... This means that row 2 of the block matrix will become the first row of the final matrix, row 5 will become the second row, and so on.
Now let's build a suitable permutation by choosing randomly a row, say 3:
p = [3, ...]
The next row will then be chosen randomly among the remaining rows that are compatible with 3 : 4, 5and 6. Say we choose 4:
p = [3, 4, ...]
Next choice has to be made among 1 and 2, for instance 1:
p = [3, 4, 1, ...]
And so on: p = [3, 4, 1, 5, 2, 6].
Applying this permutation to the block matrix, we get:
1 0 1 0 0 0 | 3
0 0 0 1 1 0 | 4
1 1 0 0 0 0 | 1
0 0 0 0 1 1 | 5
0 1 1 0 0 0 | 2
0 0 0 1 0 1 | 6
------------+--
1 2 3 4 5 6 |
Doing so, we manage to vertically isolate all non-zero entries. Same has to be done with the columns, for instance by using permutation p' = [6, 3, 5, 1, 4, 2] to finally get
0 1 0 1 0 0 | 3
0 0 1 0 1 0 | 4
0 0 0 1 0 1 | 1
1 0 1 0 0 0 | 5
0 1 0 0 0 1 | 2
1 0 0 0 1 0 | 6
------------+--
6 3 5 1 4 2 |
So this seems to work quite efficiently, but building these permutations needs to be done with caution, because one can easily be stuck: for instance, with n=6 and partition 6 = 2 + 2 + 2, following the construction rules set up earlier can lead to p = [1, 3, 2, 4, ...]. Unfortunately, 5 and 6 are incompatible, so choosing one or the other makes the last choice impossible. I think I've found all situations that lead to a dead end. I will denote by r the set of remaining choices:
p = [..., x, ?], r = {y} with x and y incompatible
p = [..., x, ?, ?], r = {y, z} with y and z being both incompatible with x (no choice can be made)
p = [..., ?, ?], r = {x, y} with x and y incompatible (any choice would lead to situation 1)
p = [..., ?, ?, ?], r = {x, y, z} with x, y and z being cyclically consecutive (choosing x or z would lead to situation 2, choosing y to situation 3)
p = [..., w, ?, ?, ?], r = {x, y, z} with xwy being a 3-cycle (neither x nor y can be chosen, choosing z would lead to situation 3)
p = [..., ?, ?, ?, ?], r = {w, x, y, z} with wxyz being a 4-cycle (any choice would lead to situation 4)
p = [..., ?, ?, ?, ?], r = {w, x, y, z} with xyz being a 3-cycle (choosing w would lead to situation 4, choosing any other would lead to situation 4)
Now it seems that the following algorithm gives all suitable permutations:
As long as there are strictly more than 5 numbers to choose, choose randomly among the compatible ones.
If there are 5 numbers left to choose: if the remaining numbers contain a 3-cycle or a 4-cycle, break that cycle (i.e. choose a number belonging to that cycle).
If there are 4 numbers left to choose: if the remaining numbers contain three cyclically consecutive numbers, choose one of them.
If there are 3 numbers left to choose: if the remaining numbers contain two cyclically consecutive numbers, choose one of them.
I am quite sure that this allows me to generate all suitable permutations and, hence, all suitable matrices.
Unfortunately, every matrix will be obtained several times, depending on the partition that was chosen.
Intro
Here is some prototype-approach, trying to solve the more general task of
uniform combinatorial sampling, which for our approach here means: we can use this approach for everything which we can formulate as SAT-problem.
It's not exploiting your problem directly and takes a heavy detour. This detour to the SAT-problem can help in regards to theory (more powerful general theoretical results) and efficiency (SAT-solvers).
That being said, it's not an approach if you want to sample within seconds or less (in my experiments), at least while being concerned about uniformity.
Theory
The approach, based on results from complexity-theory, follows this work:
GOMES, Carla P.; SABHARWAL, Ashish; SELMAN, Bart. Near-uniform sampling of combinatorial spaces using XOR constraints. In: Advances In Neural Information Processing Systems. 2007. S. 481-488.
The basic idea:
formulate the problem as SAT-problem
add randomly generated xors to the problem (acting on the decision-variables only! that's important in practice)
this will reduce the number of solutions (some solutions will get impossible)
do that in a loop (with tuned parameters) until only one solution is left!
search for some solution is being done by SAT-solvers or #SAT-solvers (=model-counting)
if there is more than one solution: no xors will be added but a complete restart will be done: add random-xors to the start-problem!
The guarantees:
when tuning the parameters right, this approach achieves near-uniform sampling
this tuning can be costly, as it's based on approximating the number of possible solutions
empirically this can also be costly!
Ante's answer, mentioning the number sequence A001499 actually gives a nice upper bound on the solution-space (as it's just ignoring adjacency-constraints!)
The drawbacks:
inefficient for large problems (in general; not necessarily compared to the alternatives like MCMC and co.)
need to change / reduce parameters to produce samples
those reduced parameters lose the theoretical guarantees
but empirically: good results are still possible!
Parameters:
In practice, the parameters are:
N: number of xors added
L: minimum number of variables part of one xor-constraint
U: maximum number of variables part of one xor-constraint
N is important to reduce the number of possible solutions. Given N constant, the other variables of course also have some effect on that.
Theory says (if i interpret correctly), that we should use L = R = 0.5 * #dec-vars.
This is impossible in practice here, as xor-constraints hurt SAT-solvers a lot!
Here some more scientific slides about the impact of L and U.
They call xors of size 8-20 short-XORS, while we will need to use even shorter ones later!
Implementation
Final version
Here is a pretty hacky implementation in python, using the XorSample scripts from here.
The underlying SAT-solver in use is Cryptominisat.
The code basically boils down to:
Transform the problem to conjunctive normal-form
as DIMACS-CNF
Implement the sampling-approach:
Calls XorSample (pipe-based + file-based)
Call SAT-solver (file-based)
Add samples to some file for later analysis
Code: (i hope i did warn you already about the code-quality)
from itertools import count
from time import time
import subprocess
import numpy as np
import os
import shelve
import uuid
import pickle
from random import SystemRandom
cryptogen = SystemRandom()
""" Helper functions """
# K-ARY CONSTRAINT GENERATION
# ###########################
# SINZ, Carsten. Towards an optimal CNF encoding of boolean cardinality constraints.
# CP, 2005, 3709. Jg., S. 827-831.
def next_var_index(start):
next_var = start
while(True):
yield next_var
next_var += 1
class s_index():
def __init__(self, start_index):
self.firstEnvVar = start_index
def next(self,i,j,k):
return self.firstEnvVar + i*k +j
def gen_seq_circuit(k, input_indices, next_var_index_gen):
cnf_string = ''
s_index_gen = s_index(next_var_index_gen.next())
# write clauses of first partial sum (i.e. i=0)
cnf_string += (str(-input_indices[0]) + ' ' + str(s_index_gen.next(0,0,k)) + ' 0\n')
for i in range(1, k):
cnf_string += (str(-s_index_gen.next(0, i, k)) + ' 0\n')
# write clauses for general case (i.e. 0 < i < n-1)
for i in range(1, len(input_indices)-1):
cnf_string += (str(-input_indices[i]) + ' ' + str(s_index_gen.next(i, 0, k)) + ' 0\n')
cnf_string += (str(-s_index_gen.next(i-1, 0, k)) + ' ' + str(s_index_gen.next(i, 0, k)) + ' 0\n')
for u in range(1, k):
cnf_string += (str(-input_indices[i]) + ' ' + str(-s_index_gen.next(i-1, u-1, k)) + ' ' + str(s_index_gen.next(i, u, k)) + ' 0\n')
cnf_string += (str(-s_index_gen.next(i-1, u, k)) + ' ' + str(s_index_gen.next(i, u, k)) + ' 0\n')
cnf_string += (str(-input_indices[i]) + ' ' + str(-s_index_gen.next(i-1, k-1, k)) + ' 0\n')
# last clause for last variable
cnf_string += (str(-input_indices[-1]) + ' ' + str(-s_index_gen.next(len(input_indices)-2, k-1, k)) + ' 0\n')
return (cnf_string, (len(input_indices)-1)*k, 2*len(input_indices)*k + len(input_indices) - 3*k - 1)
# K=2 clause GENERATION
# #####################
def gen_at_most_2_constraints(vars, start_var):
constraint_string = ''
used_clauses = 0
used_vars = 0
index_gen = next_var_index(start_var)
circuit = gen_seq_circuit(2, vars, index_gen)
constraint_string += circuit[0]
used_clauses += circuit[2]
used_vars += circuit[1]
start_var += circuit[1]
return [constraint_string, used_clauses, used_vars, start_var]
def gen_at_least_2_constraints(vars, start_var):
k = len(vars) - 2
vars = [-var for var in vars]
constraint_string = ''
used_clauses = 0
used_vars = 0
index_gen = next_var_index(start_var)
circuit = gen_seq_circuit(k, vars, index_gen)
constraint_string += circuit[0]
used_clauses += circuit[2]
used_vars += circuit[1]
start_var += circuit[1]
return [constraint_string, used_clauses, used_vars, start_var]
# Adjacency conflicts
# ###################
def get_all_adjacency_conflicts_4_neighborhood(N, X):
conflicts = set()
for x in range(N):
for y in range(N):
if x < (N-1):
conflicts.add(((x,y),(x+1,y)))
if y < (N-1):
conflicts.add(((x,y),(x,y+1)))
cnf = '' # slow string appends
for (var_a, var_b) in conflicts:
var_a_ = X[var_a]
var_b_ = X[var_b]
cnf += '-' + var_a_ + ' ' + '-' + var_b_ + ' 0 \n'
return cnf, len(conflicts)
# Build SAT-CNF
#############
def build_cnf(N, verbose=False):
var_counter = count(1)
N_CLAUSES = 0
X = np.zeros((N, N), dtype=object)
for a in range(N):
for b in range(N):
X[a,b] = str(next(var_counter))
# Adjacency constraints
CNF, N_CLAUSES = get_all_adjacency_conflicts_4_neighborhood(N, X)
# k=2 constraints
NEXT_VAR = N*N+1
for row in range(N):
constraint_string, used_clauses, used_vars, NEXT_VAR = gen_at_most_2_constraints(X[row, :].astype(int).tolist(), NEXT_VAR)
N_CLAUSES += used_clauses
CNF += constraint_string
constraint_string, used_clauses, used_vars, NEXT_VAR = gen_at_least_2_constraints(X[row, :].astype(int).tolist(), NEXT_VAR)
N_CLAUSES += used_clauses
CNF += constraint_string
for col in range(N):
constraint_string, used_clauses, used_vars, NEXT_VAR = gen_at_most_2_constraints(X[:, col].astype(int).tolist(), NEXT_VAR)
N_CLAUSES += used_clauses
CNF += constraint_string
constraint_string, used_clauses, used_vars, NEXT_VAR = gen_at_least_2_constraints(X[:, col].astype(int).tolist(), NEXT_VAR)
N_CLAUSES += used_clauses
CNF += constraint_string
# build final cnf
CNF = 'p cnf ' + str(NEXT_VAR-1) + ' ' + str(N_CLAUSES) + '\n' + CNF
return X, CNF, NEXT_VAR-1
# External tools
# ##############
def get_random_xor_problem(CNF_IN_fp, N_DEC_VARS, N_ALL_VARS, s, min_l, max_l):
# .cnf not part of arg!
p = subprocess.Popen(['./gen-wff', CNF_IN_fp,
str(N_DEC_VARS), str(N_ALL_VARS),
str(s), str(min_l), str(max_l), 'xored'],
stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
result = p.communicate()
os.remove(CNF_IN_fp + '-str-xored.xor') # file not needed
return CNF_IN_fp + '-str-xored.cnf'
def solve(CNF_IN_fp, N_DEC_VARS):
seed = cryptogen.randint(0, 2147483647) # actually no reason to do it; but can't hurt either
p = subprocess.Popen(["./cryptominisat5", '-t', '4', '-r', str(seed), CNF_IN_fp], stdin=subprocess.PIPE, stdout=subprocess.PIPE)
result = p.communicate()[0]
sat_line = result.find('s SATISFIABLE')
if sat_line != -1:
# solution found!
vars = parse_solution(result)[:N_DEC_VARS]
# forbid solution (DeMorgan)
negated_vars = list(map(lambda x: x*(-1), vars))
with open(CNF_IN_fp, 'a') as f:
f.write( (str(negated_vars)[1:-1] + ' 0\n').replace(',', ''))
# assume solve is treating last constraint despite not changing header!
# solve again
seed = cryptogen.randint(0, 2147483647)
p = subprocess.Popen(["./cryptominisat5", '-t', '4', '-r', str(seed), CNF_IN_fp], stdin=subprocess.PIPE, stdout=subprocess.PIPE)
result = p.communicate()[0]
sat_line = result.find('s SATISFIABLE')
if sat_line != -1:
os.remove(CNF_IN_fp) # not needed anymore
return True, False, None
else:
return True, True, vars
else:
return False, False, None
def parse_solution(output):
# assumes there is one
vars = []
for line in output.split("\n"):
if line:
if line[0] == 'v':
line_vars = list(map(lambda x: int(x), line.split()[1:]))
vars.extend(line_vars)
return vars
# Core-algorithm
# ##############
def xorsample(X, CNF_IN_fp, N_DEC_VARS, N_VARS, s, min_l, max_l):
start_time = time()
while True:
# add s random XOR constraints to F
xored_cnf_fp = get_random_xor_problem(CNF_IN_fp, N_DEC_VARS, N_VARS, s, min_l, max_l)
state_lvl1, state_lvl2, var_sol = solve(xored_cnf_fp, N_DEC_VARS)
print('------------')
if state_lvl1 and state_lvl2:
print('FOUND')
d = shelve.open('N_15_70_4_6_TO_PLOT')
d[str(uuid.uuid4())] = (pickle.dumps(var_sol), time() - start_time)
d.close()
return True
else:
if state_lvl1:
print('sol not unique')
else:
print('no sol found')
print('------------')
""" Run """
N = 15
N_DEC_VARS = N*N
X, CNF, N_VARS = build_cnf(N)
with open('my_problem.cnf', 'w') as f:
f.write(CNF)
counter = 0
while True:
print('sample: ', counter)
xorsample(X, 'my_problem', N_DEC_VARS, N_VARS, 70, 4, 6)
counter += 1
Output will look like (removed some warnings):
------------
no sol found
------------
------------
no sol found
------------
------------
no sol found
------------
------------
sol not unique
------------
------------
FOUND
Core: CNF-formulation
We introduce one variable for every cell of the matrix. N=20 means 400 binary-variables.
Adjancency:
Precalculate all symmetry-reduced conflicts and add conflict-clauses.
Basic theory:
a -> !b
<->
!a v !b (propositional logic)
Row/Col-wise Cardinality:
This is tough to express in CNF and naive approaches need an exponential number
of constraints.
We use some adder-circuit based encoding (SINZ, Carsten. Towards an optimal CNF encoding of boolean cardinality constraints) which introduces new auxiliary-variables.
Remark:
sum(var_set) <= k
<->
sum(negated(var_set)) >= len(var_set) - k
These SAT-encodings can be put into exact model-counters (for small N; e.g. < 9). The number of solutions equals Ante's results, which is a strong indication for a correct transformation!
There are also interesting approximate model-counters (also heavily based on xor-constraints) like approxMC which shows one more thing we can do with the SAT-formulation. But in practice i have not been able to use these (approxMC = autoconf; no comment).
Other experiments
I did also build a version using pblib, to use more powerful cardinality-formulations
for the SAT-CNF formulation. I did not try to use the C++-based API, but only the reduced pbencoder, which automatically selects some best encoding, which was way worse than my encoding used above (which is best is still a research-problem; often even redundant-constraints can help).
Empirical analysis
For the sake of obtaining some sample-size (given my patience), i only computed samples for N=15. In this case we used:
N=70 xors
L,U = 4,6
I also computed some samples for N=20 with (100,3,6), but this takes a few mins and we reduced the lower bound!
Visualization
Here some animation (strengthening my love-hate relationship with matplotlib):
Edit: And a (reduced) comparison to brute-force uniform-sampling with N=5 (NXOR,L,U = 4, 10, 30):
(I have not yet decided on the addition of the plotting-code. It's as ugly as the above one and people might look too much into my statistical shambles; normalizations and co.)
Theory
Statistical analysis is probably hard to do as the underlying problem is of such combinatoric nature. It's even not entirely obvious how that final cell-PDF should look like. In the case of N=odd, it's probably non-uniform and looks like a chess-board (i did brute-force check N=5 to observe this).
One thing we can be sure about (imho): symmetry!
Given a cell-PDF matrix, we should expect, that the matrix is symmetric (A = A.T).
This is checked in the visualization and the euclidean-norm of differences over time is plotted.
We can do the same on some other observation: observed pairings.
For N=3, we can observe the following pairs:
0,1
0,2
1,2
Now we can do this per-row and per-column and should expect symmetry too!
Sadly, it's probably not easy to say something about the variance and therefore the needed samples to speak about confidence!
Observation
According to my simplified perception, current-samples and the cell-PDF look good, although convergence is not achieved yet (or we are far away from uniformity).
The more important aspect are probably the two norms, nicely decreasing towards 0.
(yes; one could tune some algorithm for that by transposing with prob=0.5; but this is not done here as it would defeat it's purpose).
Potential next steps
Tune parameters
Check out the approach using #SAT-solvers / Model-counters instead of SAT-solvers
Try different CNF-formulations, especially in regards to cardinality-encodings and xor-encodings
XorSample is by default using tseitin-like encoding to get around exponentially grow
for smaller xors (as used) it might be a good idea to use naive encoding (which propagates faster)
XorSample supports that in theory; but the script's work differently in practice
Cryptominisat is known for dedicated XOR-handling (as it was build for analyzing cryptography including many xors) and might gain something by naive encoding (as inferring xors from blown-up CNFs is much harder)
More statistical-analysis
Get rid of XorSample scripts (shell + perl...)
Summary
The approach is very general
This code produces feasible samples
It should be not hard to prove, that every feasible solution can be sampled
Others have proven theoretical guarantees for uniformity for some params
does not hold for our params
Others have empirically / theoretically analyzed smaller parameters (in use here)
(Updated test results, example run-through and code snippets below.)
You can use dynamic programming to calculate the number of solutions resulting from every state (in a much more efficient way than a brute-force algorithm), and use those (pre-calculated) values to create equiprobable random solutions.
Consider the example of a 7x7 matrix; at the start, the state is:
0,0,0,0,0,0,0
meaning that there are seven adjacent unused columns. After adding two ones to the first row, the state could be e.g.:
0,1,0,0,1,0,0
with two columns that now have a one in them. After adding ones to the second row, the state could be e.g.:
0,1,1,0,1,0,1
After three rows are filled, there is a possibility that a column will have its maximum of two ones; this effectively splits the matrix into two independent zones:
1,1,1,0,2,0,1 -> 1,1,1,0 + 0,1
These zones are independent in the sense that the no-adjacent-ones rule has no effect when adding ones to different zones, and the order of the zones has no effect on the number of solutions.
In order to use these states as signatures for types of solutions, we have to transform them into a canonical notation. First, we have to take into account the fact that columns with only 1 one in them may be unusable in the next row, because they contain a one in the current row. So instead of a binary notation, we have to use a ternary notation, e.g.:
2,1,1,0 + 0,1
where the 2 means that this column was used in the current row (and not that there are 2 ones in the column). At the next step, we should then convert the twos back into ones.
Additionally, we can also mirror the seperate groups to put them into their lexicographically smallest notation:
2,1,1,0 + 0,1 -> 0,1,1,2 + 0,1
Lastly, we sort the seperate groups from small to large, and then lexicographically, so that a state in a larger matrix may be e.g.:
0,0 + 0,1 + 0,0,2 + 0,1,0 + 0,1,0,1
Then, when calculating the number of solutions resulting from each state, we can use memoization using the canonical notation of each state as a key.
Creating a dictionary of the states and the number of solutions for each of them only needs to be done once, and a table for larger matrices can probably be used for smaller matrices too.
Practically, you'd generate a random number between 0 and the total number of solutions, and then for every row, you'd look at the different states you could create from the current state, look at the number of unique solutions each one would generate, and see which option leads to the solution that corresponds with your randomly generated number.
Note that every state and the corresponding key can only occur in a particular row, so you can store the keys in seperate dictionaries per row.
TEST RESULTS
A first test using unoptimized JavaScript gave very promising results. With dynamic programming, calculating the number of solutions for a 10x10 matrix now takes a second, where a brute-force algorithm took several hours (and this is the part of the algorithm that only needs to be done once). The size of the dictionary with the signatures and numbers of solutions grows with a diminishing factor approaching 2.5 for each step in size; the time to generate it grows with a factor of around 3.
These are the number of solutions, states, signatures (total size of the dictionaries), and maximum number of signatures per row (largest dictionary per row) that are created:
size unique solutions states signatures max/row
4x4 2 9 6 2
5x5 16 73 26 8
6x6 722 514 107 40
7x7 33,988 2,870 411 152
8x8 2,215,764 13,485 1,411 596
9x9 179,431,924 56,375 4,510 1,983
10x10 17,849,077,140 218,038 13,453 5,672
11x11 2,138,979,146,276 801,266 38,314 14,491
12x12 304,243,884,374,412 2,847,885 104,764 35,803
13x13 50,702,643,217,809,908 9,901,431 278,561 96,414
14x14 9,789,567,606,147,948,364 33,911,578 723,306 238,359
15x15 2,168,538,331,223,656,364,084 114,897,838 1,845,861 548,409
16x16 546,386,962,452,256,865,969,596 ... 4,952,501 1,444,487
17x17 155,420,047,516,794,379,573,558,433 12,837,870 3,754,040
18x18 48,614,566,676,379,251,956,711,945,475 31,452,747 8,992,972
19x19 17,139,174,923,928,277,182,879,888,254,495 74,818,773 20,929,008
20x20 6,688,262,914,418,168,812,086,412,204,858,650 175,678,000 50,094,203
(Additional results obtained with C++, using a simple 128-bit integer implementation. To count the states, the code had to be run using each state as a seperate signature, which I was unable to do for the largest sizes. )
EXAMPLE
The dictionary for a 5x5 matrix looks like this:
row 0: 00000 -> 16 row 3: 101 -> 0
1112 -> 1
row 1: 20002 -> 2 1121 -> 1
00202 -> 4 1+01 -> 0
02002 -> 2 11+12 -> 2
02020 -> 2 1+121 -> 1
0+1+1 -> 0
row 2: 10212 -> 1 1+112 -> 1
12012 -> 1
12021 -> 2 row 4: 0 -> 0
12102 -> 1 11 -> 0
21012 -> 0 12 -> 0
02121 -> 3 1+1 -> 1
01212 -> 1 1+2 -> 0
The total number of solutions is 16; if we randomly pick a number from 0 to 15, e.g. 13, we can find the corresponding (i.e. the 14th) solution like this:
state: 00000
options: 10100 10010 10001 01010 01001 00101
signature: 00202 02002 20002 02020 02002 00202
solutions: 4 2 2 2 2 4
This tells us that the 14th solution is the 2nd solution of option 00101. The next step is:
state: 00101
options: 10010 01010
signature: 12102 02121
solutions: 1 3
This tells us that the 2nd solution is the 1st solution of option 01010. The next step is:
state: 01111
options: 10100 10001 00101
signature: 11+12 1112 1+01
solutions: 2 1 0
This tells us that the 1st solution is the 1st solution of option 10100. The next step is:
state: 11211
options: 01010 01001
signature: 1+1 1+1
solutions: 1 1
This tells us that the 1st solutions is the 1st solution of option 01010. The last step is:
state: 12221
options: 10001
And the 5x5 matrix corresponding to randomly chosen number 13 is:
0 0 1 0 1
0 1 0 1 0
1 0 1 0 0
0 1 0 1 0
1 0 0 0 1
And here's a quick'n'dirty code example; run the snippet to generate the signature and solution count dictionary, and generate a random 10x10 matrix (it takes a second to generate the dictionary; once that is done, it generates random solutions in half a millisecond):
function signature(state, prev) {
var zones = [], zone = [];
for (var i = 0; i < state.length; i++) {
if (state[i] == 2) {
if (zone.length) zones.push(mirror(zone));
zone = [];
}
else if (prev[i]) zone.push(3);
else zone.push(state[i]);
}
if (zone.length) zones.push(mirror(zone));
zones.sort(function(a,b) {return a.length - b.length || a - b;});
return zones.length ? zones.join("2") : "2";
function mirror(zone) {
var ltr = zone.join('');
zone.reverse();
var rtl = zone.join('');
return (ltr < rtl) ? ltr : rtl;
}
}
function memoize(n) {
var memo = [], empty = [];
for (var i = 0; i <= n; i++) memo[i] = [];
for (var i = 0; i < n; i++) empty[i] = 0;
memo[0][signature(empty, empty)] = next_row(empty, empty, 1);
return memo;
function next_row(state, prev, row) {
if (row > n) return 1;
var solutions = 0;
for (var i = 0; i < n - 2; i++) {
if (state[i] == 2 || prev[i] == 1) continue;
for (var j = i + 2; j < n; j++) {
if (state[j] == 2 || prev[j] == 1) continue;
var s = state.slice(), p = empty.slice();
++s[i]; ++s[j]; ++p[i]; ++p[j];
var sig = signature(s, p);
var sol = memo[row][sig];
if (sol == undefined)
memo[row][sig] = sol = next_row(s, p, row + 1);
solutions += sol;
}
}
return solutions;
}
}
function random_matrix(n, memo) {
var matrix = [], empty = [], state = [], prev = [];
for (var i = 0; i < n; i++) empty[i] = state[i] = prev[i] = 0;
var total = memo[0][signature(empty, empty)];
var pick = Math.floor(Math.random() * total);
document.write("solution " + pick.toLocaleString('en-US') +
" from a total of " + total.toLocaleString('en-US') + "<br>");
for (var row = 1; row <= n; row++) {
var options = find_options(state, prev);
for (var i in options) {
var state_copy = state.slice();
for (var j in state_copy) state_copy[j] += options[i][j];
var sig = signature(state_copy, options[i]);
var solutions = memo[row][sig];
if (pick < solutions) {
matrix.push(options[i].slice());
prev = options[i].slice();
state = state_copy.slice();
break;
}
else pick -= solutions;
}
}
return matrix;
function find_options(state, prev) {
var options = [];
for (var i = 0; i < n - 2; i++) {
if (state[i] == 2 || prev[i] == 1) continue;
for (var j = i + 2; j < n; j++) {
if (state[j] == 2 || prev[j] == 1) continue;
var option = empty.slice();
++option[i]; ++option[j];
options.push(option);
}
}
return options;
}
}
var size = 10;
var memo = memoize(size);
var matrix = random_matrix(size, memo);
for (var row in matrix) document.write(matrix[row] + "<br>");
The code snippet below shows the dictionary of signatures and solution counts for a matrix of size 10x10. I've used a slightly different signature format from the explanation above: the zones are delimited by a '2' instead of a plus sign, and a column which has a one in the previous row is marked with a '3' instead of a '2'. This shows how the keys could be stored in a file as integers with 2×N bits (padded with 2's).
function signature(state, prev) {
var zones = [], zone = [];
for (var i = 0; i < state.length; i++) {
if (state[i] == 2) {
if (zone.length) zones.push(mirror(zone));
zone = [];
}
else if (prev[i]) zone.push(3);
else zone.push(state[i]);
}
if (zone.length) zones.push(mirror(zone));
zones.sort(function(a,b) {return a.length - b.length || a - b;});
return zones.length ? zones.join("2") : "2";
function mirror(zone) {
var ltr = zone.join('');
zone.reverse();
var rtl = zone.join('');
return (ltr < rtl) ? ltr : rtl;
}
}
function memoize(n) {
var memo = [], empty = [];
for (var i = 0; i <= n; i++) memo[i] = [];
for (var i = 0; i < n; i++) empty[i] = 0;
memo[0][signature(empty, empty)] = next_row(empty, empty, 1);
return memo;
function next_row(state, prev, row) {
if (row > n) return 1;
var solutions = 0;
for (var i = 0; i < n - 2; i++) {
if (state[i] == 2 || prev[i] == 1) continue;
for (var j = i + 2; j < n; j++) {
if (state[j] == 2 || prev[j] == 1) continue;
var s = state.slice(), p = empty.slice();
++s[i]; ++s[j]; ++p[i]; ++p[j];
var sig = signature(s, p);
var sol = memo[row][sig];
if (sol == undefined)
memo[row][sig] = sol = next_row(s, p, row + 1);
solutions += sol;
}
}
return solutions;
}
}
var memo = memoize(10);
for (var i in memo) {
document.write("row " + i + ":<br>");
for (var j in memo[i]) {
document.write(""" + j + "": " + memo[i][j] + "<br>");
}
}
Just few thoughts. Number of matrices satisfying conditions for n <= 10:
3 0
4 2
5 16
6 722
7 33988
8 2215764
9 179431924
10 17849077140
Unfortunatelly there is no sequence with these numbers in OEIS.
There is one similar (A001499), without condition for neighbouring one's. Number of nxn matrices in this case is 'of order' as A001499's number of (n-1)x(n-1) matrices. That is to be expected since number
of ways to fill one row in this case, position 2 one's in n places with at least one zero between them is ((n-1) choose 2). Same as to position 2 one's in (n-1) places without the restriction.
I don't think there is an easy connection between these matrix of order n and A001499 matrix of order n-1, meaning that if we have A001499 matrix than we can construct some of these matrices.
With this, for n=20, number of matrices is >10^30. Quite a lot :-/
This solution use recursion in order to set the cell of the matrix one by one. If the random walk finish with an impossible solution then we rollback one step in the tree and we continue the random walk.
The algorithm is efficient and i think that the generated data are highly equiprobable.
package rndsqmatrix;
import java.util.ArrayList;
import java.util.Collections;
import java.util.List;
import java.util.stream.IntStream;
public class RndSqMatrix {
/**
* Generate a random matrix
* #param size the size of the matrix
* #return the matrix encoded in 1d array i=(x+y*size)
*/
public static int[] generate(final int size) {
return generate(size, new int[size * size], new int[size],
new int[size]);
}
/**
* Build a matrix recursivly with a random walk
* #param size the size of the matrix
* #param matrix the matrix encoded in 1d array i=(x+y*size)
* #param rowSum
* #param colSum
* #return
*/
private static int[] generate(final int size, final int[] matrix,
final int[] rowSum, final int[] colSum) {
// generate list of valid positions
final List<Integer> positions = new ArrayList();
for (int y = 0; y < size; y++) {
if (rowSum[y] < 2) {
for (int x = 0; x < size; x++) {
if (colSum[x] < 2) {
final int p = x + y * size;
if (matrix[p] == 0
&& (x == 0 || matrix[p - 1] == 0)
&& (x == size - 1 || matrix[p + 1] == 0)
&& (y == 0 || matrix[p - size] == 0)
&& (y == size - 1 || matrix[p + size] == 0)) {
positions.add(p);
}
}
}
}
}
// no valid positions ?
if (positions.isEmpty()) {
// if the matrix is incomplete => return null
for (int i = 0; i < size; i++) {
if (rowSum[i] != 2 || colSum[i] != 2) {
return null;
}
}
// the matrix is complete => return it
return matrix;
}
// random walk
Collections.shuffle(positions);
for (int p : positions) {
// set '1' and continue recursivly the exploration
matrix[p] = 1;
rowSum[p / size]++;
colSum[p % size]++;
final int[] solMatrix = generate(size, matrix, rowSum, colSum);
if (solMatrix != null) {
return solMatrix;
}
// rollback
matrix[p] = 0;
rowSum[p / size]--;
colSum[p % size]--;
}
// we can't find a valid matrix from here => return null
return null;
}
public static void printMatrix(final int size, final int[] matrix) {
for (int y = 0; y < size; y++) {
for (int x = 0; x < size; x++) {
System.out.print(matrix[x + y * size]);
System.out.print(" ");
}
System.out.println();
}
}
public static void printStatistics(final int size, final int count) {
final int sumMatrix[] = new int[size * size];
for (int i = 0; i < count; i++) {
final int[] matrix = generate(size);
for (int j = 0; j < sumMatrix.length; j++) {
sumMatrix[j] += matrix[j];
}
}
printMatrix(size, sumMatrix);
}
public static void checkAlgorithm() {
final int size = 8;
final int count = 2215764;
final int divisor = 122;
final int sumMatrix[] = new int[size * size];
for (int i = 0; i < count/divisor ; i++) {
final int[] matrix = generate(size);
for (int j = 0; j < sumMatrix.length; j++) {
sumMatrix[j] += matrix[j];
}
}
int total = 0;
for(int i=0; i < sumMatrix.length; i++) {
total += sumMatrix[i];
}
final double factor = (double)total / (count/divisor);
System.out.println("Factor=" + factor + " (theory=16.0)");
}
public static void benchmark(final int size, final int count,
final boolean parallel) {
final long begin = System.currentTimeMillis();
if (!parallel) {
for (int i = 0; i < count; i++) {
generate(size);
}
} else {
IntStream.range(0, count).parallel().forEach(i -> generate(size));
}
final long end = System.currentTimeMillis();
System.out.println("rate="
+ (double) (end - begin) / count + "ms/matrix");
}
public static void main(String[] args) {
checkAlgorithm();
benchmark(8, 10000, true);
//printStatistics(8, 2215764/36);
printStatistics(8, 2215764);
}
}
The output is:
Factor=16.0 (theory=16.0)
rate=0.2835ms/matrix
552969 554643 552895 554632 555680 552753 554567 553389
554071 554847 553441 553315 553425 553883 554485 554061
554272 552633 555130 553699 553604 554298 553864 554028
554118 554299 553565 552986 553786 554473 553530 554771
554474 553604 554473 554231 553617 553556 553581 553992
554960 554572 552861 552732 553782 554039 553921 554661
553578 553253 555721 554235 554107 553676 553776 553182
553086 553677 553442 555698 553527 554850 553804 553444
Here is a very fast approach of generating the matrix row by row, written in Java:
public static void main(String[] args) throws Exception {
int n = 100;
Random rnd = new Random();
byte[] mat = new byte[n*n];
byte[] colCount = new byte[n];
//generate row by row
for (int x = 0; x < n; x++) {
//generate a random first bit
int b1 = rnd.nextInt(n);
while ( (x > 0 && mat[(x-1)*n + b1] == 1) || //not adjacent to the one above
(colCount[b1] == 2) //not in a column which has 2
) b1 = rnd.nextInt(n);
//generate a second bit, not equal to the first one
int b2 = rnd.nextInt(n);
while ( (b2 == b1) || //not the same as bit 1
(x > 0 && mat[(x-1)*n + b2] == 1) || //not adjacent to the one above
(colCount[b2] == 2) || //not in a column which has 2
(b2 == b1 - 1) || //not adjacent to b1
(b2 == b1 + 1)
) b2 = rnd.nextInt(n);
//fill the matrix values and increment column counts
mat[x*n + b1] = 1;
mat[x*n + b2] = 1;
colCount[b1]++;
colCount[b2]++;
}
String arr = Arrays.toString(mat).substring(1, n*n*3 - 1);
System.out.println(arr.replaceAll("(.{" + n*3 + "})", "$1\n"));
}
It essentially generates each a random row at a time. If the row will violate any of the conditions, it is generated again (again randomly). I believe this will satisfy condition 4 as well.
Adding a quick note that it will spin forever for N-s where there is no solutions (like N=3).
Imagine you have 3 buckets, but each of them has a hole in it. I'm trying to fill a bath tub. The bath tub has a minimum level of water it needs and a maximum level of water it can contain. By the time you reach the tub with the bucket it is not clear how much water will be in the bucket, but you have a range of possible values.
Is it possible to adequately fill the tub with water?
Pretty much you have 3 ranges (min,max), is there some sum of them that will fall within a 4th range?
For example:
Bucket 1 : 5-10L
Bucket 2 : 15-25L
Bucket 3 : 10-50L
Bathtub 100-150L
Is there some guaranteed combination of 1 2 and 3 that will fill the bathtub within the requisite range? Multiples of each bucket can be used.
EDIT: Now imagine there are 50 different buckets?
If the capacity of the tub is not very large ( not greater than 10^6 for an example), we can solve it using dynamic programming.
Approach:
Initialization: memo[X][Y] is an array to memorize the result. X = number of buckets, Y = maximum capacity of the tub. Initialize memo[][] with -1.
Code:
bool dp(int bucketNum, int curVolume){
if(curVolume > maxCap)return false; // pruning extra branches
if(curVolume>=minCap && curVolume<=maxCap){ // base case on success
return true;
}
int &ret = memo[bucketNum][curVolume];
if(ret != -1){ // this state has been visited earlier
return false;
}
ret = false;
for(int i = minC[bucketNum]; i < = maxC[bucketNum]; i++){
int newVolume = curVolume + i;
for(int j = bucketNum; j <= 3; j++){
ret|=dp(j,newVolume);
if(ret == true)return ret;
}
}
return ret;
}
Warning: Code not tested
Here's a naïve recursive solution in python that works just fine (although it doesn't find an optimal solution):
def match_helper(lower, upper, units, least_difference, fail = dict()):
if upper < lower + least_difference:
return None
if fail.get((lower,upper)):
return None
exact_match = [ u for u in units if u['lower'] >= lower and u['upper'] <= upper ]
if exact_match:
return [ exact_match[0] ]
for unit in units:
if unit['upper'] > upper:
continue
recursive_match = match_helper(lower - unit['lower'], upper - unit['upper'], units, least_difference)
if recursive_match:
return [unit] + recursive_match
else:
fail[(lower,upper)] = 1
return None
def match(lower, upper):
units = [
{ 'name': 'Bucket 1', 'lower': 5, 'upper': 10 },
{ 'name': 'Bucket 2', 'lower': 15, 'upper': 25 },
{ 'name': 'Bucket 3', 'lower': 10, 'upper': 50 }
]
least_difference = min([ u['upper'] - u['lower'] for u in units ])
return match_helper(
lower = lower,
upper = upper,
units = sorted(units, key = lambda u: u['upper']),
least_difference = min([ u['upper'] - u['lower'] for u in units ]),
)
result = match(100, 175)
if result:
lower = sum([ u['lower'] for u in result ])
upper = sum([ u['upper'] for u in result ])
names = [ u['name'] for u in result ]
print lower, "-", upper
print names
else:
print "No solution"
It prints "No solution" for 100-150, but for 100-175 it comes up with a solution of 5x bucket 1, 5x bucket 2.
Assuming you are saying that the "range" for each bucket is the amount of water that it may have when it reaches the tub, and all you care about is if they could possibly fill the tub...
Just take the "max" of each bucket and sum them. If that is in the range of what you consider the tub to be "filled" then it can.
Updated:
Given that buckets can be used multiple times, this seems to me like we're looking for solutions to a pair of equations.
Given buckets x, y and z we want to find a, b and c:
a*x.min + b*y.min + c*z.min >= bathtub.min
and
a*x.max + b*y.max + c*z.max <= bathtub.max
Re: http://en.wikipedia.org/wiki/Diophantine_equation
If bathtub.min and bathtub.max are both multiples of the greatest common divisor of a,b and c, then there are infinitely many solutions (i.e. we can fill the tub), otherwise there are no solutions (i.e. we can never fill the tub).
This can be solved with multiple applications of the change making problem.
Each Bucket.Min value is a currency denomination, and Bathtub.Min is the target value.
When you find a solution via a change-making algorithm, then apply one more constraint:
sum(each Bucket.Max in your solution) <= Bathtub.max
If this constraint is not met, throw out this solution and look for another. This will probably require a change to a standard change-making algorithm that allows you to try other solutions when one is found to not be suitable.
Initially, your target range is Bathtub.Range.
Each time you add an instance of a bucket to the solution, you reduce the target range for the remaining buckets.
For example, using your example buckets and tub:
Target Range = 100..150
Let's say we want to add a Bucket1 to the candidate solution. That then gives us
Target Range = 95..140
because if the rest of the buckets in the solution total < 95, then this Bucket1 might not be sufficient to fill the tub to 100, and if the rest of the buckets in the solution total > 140, then this Bucket1 might fill the tub over 150.
So, this gives you a quick way to check if a candidate solution is valid:
TargetRange = Bathtub.Range
foreach Bucket in CandidateSolution
TargetRange.Min -= Bucket.Min
TargetRange.Max -= Bucket.Max
if TargetRange.Min == 0 AND TargetRange.Max >= 0 then solution found
if TargetRange.Min < 0 or TargetRange.Max < 0 then solution is invalid
This still leaves the question - How do you come up with the set of candidate solutions?
Brute force would try all possible combinations of buckets.
Here is my solution for finding the optimal solution (least number of buckets). It compares the ratio of the maximums to the ratio of the minimums, to figure out the optimal number of buckets to fill the tub.
private static void BucketProblem()
{
Range bathTub = new Range(100, 175);
List<Range> buckets = new List<Range> {new Range(5, 10), new Range(15, 25), new Range(10, 50)};
Dictionary<Range, int> result;
bool canBeFilled = SolveBuckets(bathTub, buckets, out result);
}
private static bool BucketHelper(Range tub, List<Range> buckets, Dictionary<Range, int> results)
{
Range bucket;
int startBucket = -1;
int fills = -1;
for (int i = buckets.Count - 1; i >=0 ; i--)
{
bucket = buckets[i];
double maxRatio = (double)tub.Maximum / bucket.Maximum;
double minRatio = (double)tub.Minimum / bucket.Minimum;
if (maxRatio >= minRatio)
{
startBucket = i;
if (maxRatio - minRatio > 1)
fills = (int) minRatio + 1;
else
fills = (int) maxRatio;
break;
}
}
if (startBucket < 0)
return false;
bucket = buckets[startBucket];
tub.Maximum -= bucket.Maximum * fills;
tub.Minimum -= bucket.Minimum * fills;
results.Add(bucket, fills);
return tub.Maximum == 0 || tub.Minimum <= 0 || startBucket == 0 || BucketHelper(tub, buckets.GetRange(0, startBucket), results);
}
public static bool SolveBuckets(Range tub, List<Range> buckets, out Dictionary<Range, int> results)
{
results = new Dictionary<Range, int>();
buckets = buckets.OrderBy(b => b.Minimum).ToList();
return BucketHelper(new Range(tub.Minimum, tub.Maximum), buckets, results);
}