# Linear algebra, reduced row echelon form

The code that I am sharing here for you to review today, is a segment of a JavaScript library that I am going to write as time goes by for fun. It is only the two functions in the following code:

/*jslint browser: true, indent: 8 */
/*global console */

/*
Sorts matrix like from something like this:
[
[0, 0, 0],
[0, 0, 0],
[0, 2, 1],
[0, 1, 3],
[1, 2, 3],
[0, 0, 3]
]
to this:
[
[1, 2, 3],
[0, 1, 3],
[0, 2, 1],
[0, 0, 0],
[0, 0, 0],
[0, 0, 3]
]

The reason why the [0, 0, 3] is last is because
the vector has the length of 3, so only
3 vectors are sorted and the rest (irrelevent vectors) are appended
later.
*/
function sort_reduced_matrix(matrix) {
'use strict';
var i, j, len, has_pivot, irrelevant, positions, new_matrix, count;

len       = {};
len.i     = matrix.length;    // matrix length (row)
len.j     = matrix.length; // vector length (column)
positions = [];

has_pivot = [];

// Find pivot positions
for (j = 0; j < len.j; j += 1) {
for (i = 0; i < len.i; i += 1) {
if (matrix[i][j] === 1 && has_pivot[i] !== i) {
has_pivot[i] = i;
positions[positions.length] = i;
break;
}
}
}

irrelevant = [];
count      = 0;

// Find irrelevant vectors positions
for (i = 0; i < len.i; i += 1) {
if (has_pivot[i] === undefined) {
irrelevant[count] = i;
count += 1;
}
}

new_matrix = [];
count      = 0;

// Sort positions
for (i = 0; i < len.i; i += 1) {
if (matrix[positions[i]] !== undefined) {
new_matrix[i] = matrix[positions[i]];
} else {
new_matrix[i] = matrix[irrelevant[count]];
count += 1;
}
}

return new_matrix;
}

function reduced_row_echolon_form(matrix) {
'use strict';
var i, p, tmp, len, mu, mv;

len = {}; // Length.
i   = {}; // Increment.
tmp = {}; // Temporary holder.
p   = {}; // Position.

len.r   = matrix.length;    // Row, length.
len.c   = matrix.length; // column, length.

i.r  = 0; // Row, increment.
i.r2 = 0; // Row2, increment.
i.c  = 0; // Column, increment.

tmp.v = []; // Vector, temporary holder.
tmp.p = 0;  // pivot value.

p.lp  = 0;  // Lead pivot, position.
p.rpd = []; // Reserved positions direct, position.

// Find lead pivots in matrix.
for (i.r = 0; i.r < len.r; i.r += 1) {

p.lp = null;
for (i.c = 0; i.c < len.c; i.c += 1) {
/* If position is not reserved nor is zero, then that is
if (matrix[i.r][i.c] !== 0 && p.rpd[i.c] === undefined) {
p.lp = i.c;
break;
}
}

if (p.lp !== null) {
p.rpd[p.lp] = p.lp;
// Reduce row such that the pivot is 1.
if (matrix[i.r][p.lp] !== 1) {
tmp.p = matrix[i.r][p.lp];
for (i.c = 0; i.c < len.c; i.c += 1) {
matrix[i.r][i.c] /= tmp.p;
}
}
/* Reduce other rows (i.r2) from row (i.r). */
for (i.r2 = 0; i.r2 < len.r; i.r2 += 1) {
/* Skip row (i.r) and don't reduce if desired
* value is already zero. */
if (i.r2 !== i.r && matrix[i.r2][p.lp] !== 0) {
/* Scale row (i.r) using pivot position
* from row (i.r2) as the multiplier. */
for (i.c = 0; i.c < len.c; i.c += 1) {
tmp.v[i.c] = matrix[i.r][i.c];
tmp.v[i.c] *= matrix[i.r2][p.lp];
}
// Row reduction.
for (i.c = 0; i.c < len.c; i.c += 1) {
matrix[i.r2][i.c] -= tmp.v[i.c];
}
}
}
}
}
// Finally, we sort our rows, keeping zeros at the bottom and return.
return sort_reduced_matrix(matrix);
}

// Compere this to wolframalpha.com answers.

var matrix = [
[5, -7, -8, -4],
[2, 8, -22, -55],
[-3, 0, -36, 12]
];

matrix = reduced_row_echolon_form(matrix);

console.log(matrix);

matrix = [
[5, -23, 2, 4, 5, 11],
[4, -3, 6, 4, 5, 2],
[3, 7, -18, 7, 9, -6],
[4, 87, -12, 7, 12, 6],
[5, 4, 7, 11, 7, -7]
];

matrix = reduced_row_echolon_form(matrix);
console.log(matrix);
console.log(matrix);
console.log(matrix);
console.log(matrix);
console.log(matrix);

matrix = [
[1, 2, 2, 2],
[1, 3, 3, 3],
[1, 4, 16, 5]
];

matrix = reduced_row_echolon_form(matrix);
console.log(matrix);

matrix = [
[0, 2, -1, -6],
[0, 3, -2, -16],
[0, 0, -3, 11]
];

matrix = reduced_row_echolon_form(matrix);
console.log(matrix);


Is there something...

1. That I am doing in these two functions that you would consider as a bad practice and why?
2. That would explain why the code is slower than it needs to be?
3. That is just bad in some other way?

I am not working as a programmer and I don't know anyone who makes a living as a programmer, so any hint or tips would be welcome.

Partial answer. Wait around and I'm sure someone will do the actual reduction function justice. It's a bit much for me to digest right now.

From a cursory glance, though, I'd suggest breaking things out into functions where possible. But again, I haven't delved too deep.

Overall stuff:

• You misspelled echelon in the function name ("echolon")
• Conventionally, names in JavaScript are camelCased, not under_scored
• You have a lot of really short variable names. Yes, there are comments explaining them where they're declared, but that doesn't help much, when you're in the middle of a dense line later, having to refer back to the comments again and again. Yes, lines will get longer with longer variable names, but they'll also become a lot more readable!
• There's no need to first assign an empty object to a variable, and then add properties to that object. It'd be clearer to define the object's initial state in one go:

len = {
i: matrix.length,    // matrix length (rows)
j: matrix.length  // vector length (columns)
};


Looking at the above code, the object should perhaps be called simply size, and i and j could well be called rows and columns - it's what the comments say, so you might as well just call them that.

The above things are fixable without changed to the actual logic. What's more troublesome to me is that your reduced_row_echelon_form function produces side effects: It alters the matrix you pass it, rather than returning a new one. After calling it, you have the answer, but you've lost the question.

I'd suggest making a lot more use of map(). Again, I haven't had the courage to dive into the reduced_row_echelon_form function, but - for instance - the sorting function can be cleaned up rather nicely using it:

function sortRows(matrix) {
// map rows to arrays with the structure
// [<column index>, <pivot coefficient>, <row index>]
// which makes them easily sortable
var pivots = matrix.map(function (row, r) {
for(var c = 0, l = row.length ; c < l ; c++) {
if(row[c]) {
return [c, row[c], r]; // found a non-zero coefficient at matrix[r][c]
}
}
return [Infinity, null, r]; // row is all-zero
});

// sort the pivots array, and use it to (re)build
// the matrix with the rows in the correct order
return pivots.sort().map(function (pivot) {
return matrix[pivot];
});
}


Which will do something like (in some sort of ascii-notation)

[ 0, 0, 0 ]                   [ 1, 2, 3 ]
[ 0, 0, 0 ]                   [ 0, 1, 3 ]
[ 0, 2, 1 ]         =>        [ 0, 2, 1 ]
[ 0, 1, 3 ]                   [ 0, 0, 3 ]
[ 1, 2, 3 ]                   [ 0, 0, 0 ]
[ 0, 0, 3 ]                   [ 0, 0, 0 ]

• @user1235831 No problem. I may still take a crack at reviewing the main function, but I haven't found the time. Generally, though, using Array's map, reduce and their friends should help (also things like findIndex - see link in answer - which aren't available everywhere, but easy to write yourself). Factoring out some commonly used functions like scaling a vector/row, transposing a matrix, etc., would also help clean things up, I'd bet. Lastly, if you've got a normal for loop, consider using continue to skip an iteration instead of wrapping everything in an if block Jun 18, 2014 at 21:25
• Don't worry about the main function; there is usable information in the fact that you read it, but found it too time consuming to digest it. You literally told me the simple and valid reason: too short variable names. (They are also just bad) I was mostly asking about general review about my style, and you delivered nicely. The code itself is too terse and that is a big problem. I will add the revision in another answer later, it might be the next weekend when I have more time. I will let you know here in the comment section when it is ready. Again, thanks. Jun 18, 2014 at 22:42
• @user1235831 Awesome, glad you found it useful! And yes, do feel free to add your own answer, or (better yet, if you feel like it) to post a new question with the revised code. Let more people look it over. I should add that I also got bogged down simply because linear algebra's never been my strong suit and it's been a long time since I last dealt with it :) Jun 18, 2014 at 23:14
• I have made the revision. Here is the link: codereview.stackexchange.com/questions/54667/… I made this into a single function and added your method within it as a choice. I am new to this so let me know if I am giving credits incorrectly. Jun 18, 2014 at 23:23
• Linear algebra is also my weakness... That is why I am challenging it ;) Jun 18, 2014 at 23:30