mathjs/src/function/algebra/sparse/csSqr.js

Math.js is an extensive math library for JavaScript and Node.js. It features a flexible expression parser with support for symbolic computation, comes with a large set of built-in functions and constants, and offers an integrated solution to work with dif

import { csPermute } from './csPermute'
import { csPost } from './csPost'
import { csEtree } from './csEtree'
import { createCsAmd } from './csAmd'
import { createCsCounts } from './csCounts'
import { factory } from '../../../utils/factory'

const name = 'csSqr'
const dependencies = [
  'add',
  'multiply',
  'transpose'
]

export const createCsSqr = /* #__PURE__ */ factory(name, dependencies, ({ add, multiply, transpose }) => {
  const csAmd = createCsAmd({ add, multiply, transpose })
  const csCounts = createCsCounts({ transpose })

  /**
   * Symbolic ordering and analysis for QR and LU decompositions.
   *
   * @param {Number}  order           The ordering strategy (see csAmd for more details)
   * @param {Matrix}  a               The A matrix
   * @param {boolean} qr              Symbolic ordering and analysis for QR decomposition (true) or
   *                                  symbolic ordering and analysis for LU decomposition (false)
   *
   * @return {Object}                 The Symbolic ordering and analysis for matrix A
   *
   * Reference: http://faculty.cse.tamu.edu/davis/publications.html
   */
  return function csSqr (order, a, qr) {
    // a arrays
    const aptr = a._ptr
    const asize = a._size
    // columns
    const n = asize[1]
    // vars
    let k
    // symbolic analysis result
    const s = {}
    // fill-reducing ordering
    s.q = csAmd(order, a)
    // validate results
    if (order && !s.q) { return null }
    // QR symbolic analysis
    if (qr) {
      // apply permutations if needed
      const c = order ? csPermute(a, null, s.q, 0) : a
      // etree of C'*C, where C=A(:,q)
      s.parent = csEtree(c, 1)
      // post order elimination tree
      const post = csPost(s.parent, n)
      // col counts chol(C'*C)
      s.cp = csCounts(c, s.parent, post, 1)
      // check we have everything needed to calculate number of nonzero elements
      if (c && s.parent && s.cp && _vcount(c, s)) {
        // calculate number of nonzero elements
        for (s.unz = 0, k = 0; k < n; k++) { s.unz += s.cp[k] }
      }
    } else {
      // for LU factorization only, guess nnz(L) and nnz(U)
      s.unz = 4 * (aptr[n]) + n
      s.lnz = s.unz
    }
    // return result S
    return s
  }

  /**
   * Compute nnz(V) = s.lnz, s.pinv, s.leftmost, s.m2 from A and s.parent
   */
  function _vcount (a, s) {
    // a arrays
    const aptr = a._ptr
    const aindex = a._index
    const asize = a._size
    // rows & columns
    const m = asize[0]
    const n = asize[1]
    // initialize s arrays
    s.pinv = [] // (m + n)
    s.leftmost = [] // (m)
    // vars
    const parent = s.parent
    const pinv = s.pinv
    const leftmost = s.leftmost
    // workspace, next: first m entries, head: next n entries, tail: next n entries, nque: next n entries
    const w = [] // (m + 3 * n)
    const next = 0
    const head = m
    const tail = m + n
    const nque = m + 2 * n
    // vars
    let i, k, p, p0, p1
    // initialize w
    for (k = 0; k < n; k++) {
      // queue k is empty
      w[head + k] = -1
      w[tail + k] = -1
      w[nque + k] = 0
    }
    // initialize row arrays
    for (i = 0; i < m; i++) { leftmost[i] = -1 }
    // loop columns backwards
    for (k = n - 1; k >= 0; k--) {
      // values & index for column k
      for (p0 = aptr[k], p1 = aptr[k + 1], p = p0; p < p1; p++) {
        // leftmost[i] = min(find(A(i,:)))
        leftmost[aindex[p]] = k
      }
    }
    // scan rows in reverse order
    for (i = m - 1; i >= 0; i--) {
      // row i is not yet ordered
      pinv[i] = -1
      k = leftmost[i]
      // check row i is empty
      if (k === -1) { continue }
      // first row in queue k
      if (w[nque + k]++ === 0) { w[tail + k] = i }
      // put i at head of queue k
      w[next + i] = w[head + k]
      w[head + k] = i
    }
    s.lnz = 0
    s.m2 = m
    // find row permutation and nnz(V)
    for (k = 0; k < n; k++) {
      // remove row i from queue k
      i = w[head + k]
      // count V(k,k) as nonzero
      s.lnz++
      // add a fictitious row
      if (i < 0) { i = s.m2++ }
      // associate row i with V(:,k)
      pinv[i] = k
      // skip if V(k+1:m,k) is empty
      if (--nque[k] <= 0) { continue }
      // nque[k] is nnz (V(k+1:m,k))
      s.lnz += w[nque + k]
      // move all rows to parent of k
      const pa = parent[k]
      if (pa !== -1) {
        if (w[nque + pa] === 0) { w[tail + pa] = w[tail + k] }
        w[next + w[tail + k]] = w[head + pa]
        w[head + pa] = w[next + i]
        w[nque + pa] += w[nque + k]
      }
    }
    for (i = 0; i < m; i++) {
      if (pinv[i] < 0) { pinv[i] = k++ }
    }
    return true
  }
})