science.stats.js

(function(){science.stats = {};
// Bandwidth selectors for Gaussian kernels.
// Based on R's implementations in `stats.bw`.
science.stats.bandwidth = {

  // Silverman, B. W. (1986) Density Estimation. London: Chapman and Hall.
  nrd0: function(x) {
    var hi = Math.sqrt(science.stats.variance(x));
    if (!(lo = Math.min(hi, science.stats.iqr(x) / 1.34)))
      (lo = hi) || (lo = Math.abs(x[1])) || (lo = 1);
    return .9 * lo * Math.pow(x.length, -.2);
  },

  // Scott, D. W. (1992) Multivariate Density Estimation: Theory, Practice, and
  // Visualization. Wiley.
  nrd: function(x) {
    var h = science.stats.iqr(x) / 1.34;
    return 1.06 * Math.min(Math.sqrt(science.stats.variance(x)), h)
      * Math.pow(x.length, -1/5);
  }
};
science.stats.distance = {
  euclidean: function(a, b) {
    var n = a.length,
        i = -1,
        s = 0,
        x;
    while (++i < n) {
      x = a[i] - b[i];
      s += x * x;
    }
    return Math.sqrt(s);
  },
  manhattan: function(a, b) {
    var n = a.length,
        i = -1,
        s = 0;
    while (++i < n) s += Math.abs(a[i] - b[i]);
    return s;
  },
  minkowski: function(p) {
    return function(a, b) {
      var n = a.length,
          i = -1,
          s = 0;
      while (++i < n) s += Math.pow(Math.abs(a[i] - b[i]), p);
      return Math.pow(s, 1 / p);
    };
  },
  chebyshev: function(a, b) {
    var n = a.length,
        i = -1,
        max = 0,
        x;
    while (++i < n) {
      x = Math.abs(a[i] - b[i]);
      if (x > max) max = x;
    }
    return max;
  },
  hamming: function(a, b) {
    var n = a.length,
        i = -1,
        d = 0;
    while (++i < n) if (a[i] !== b[i]) d++;
    return d;
  },
  jaccard: function(a, b) {
    var n = a.length,
        i = -1,
        s = 0;
    while (++i < n) if (a[i] === b[i]) s++;
    return s / n;
  },
  braycurtis: function(a, b) {
    var n = a.length,
        i = -1,
        s0 = 0,
        s1 = 0,
        ai,
        bi;
    while (++i < n) {
      ai = a[i];
      bi = b[i];
      s0 += Math.abs(ai - bi);
      s1 += Math.abs(ai + bi);
    }
    return s0 / s1;
  }
};
// Based on implementation in http://picomath.org/.
science.stats.erf = function(x) {
  var a1 =  0.254829592,
      a2 = -0.284496736,
      a3 =  1.421413741,
      a4 = -1.453152027,
      a5 =  1.061405429,
      p  =  0.3275911;

  // Save the sign of x
  var sign = x < 0 ? -1 : 1;
  if (x < 0) {
    sign = -1;
    x = -x;
  }

  // A&S formula 7.1.26
  var t = 1 / (1 + p * x);
  return sign * (
    1 - (((((a5 * t + a4) * t) + a3) * t + a2) * t + a1)
    * t * Math.exp(-x * x));
};
science.stats.phi = function(x) {
  return .5 * (1 + science.stats.erf(x / Math.SQRT2));
};
// See <http://en.wikipedia.org/wiki/Kernel_(statistics)>.
science.stats.kernel = {
  uniform: function(u) {
    if (u <= 1 && u >= -1) return .5;
    return 0;
  },
  triangular: function(u) {
    if (u <= 1 && u >= -1) return 1 - Math.abs(u);
    return 0;
  },
  epanechnikov: function(u) {
    if (u <= 1 && u >= -1) return .75 * (1 - u * u);
    return 0;
  },
  quartic: function(u) {
    if (u <= 1 && u >= -1) {
      var tmp = 1 - u * u;
      return (15 / 16) * tmp * tmp;
    }
    return 0;
  },
  triweight: function(u) {
    if (u <= 1 && u >= -1) {
      var tmp = 1 - u * u;
      return (35 / 32) * tmp * tmp * tmp;
    }
    return 0;
  },
  gaussian: function(u) {
    return 1 / Math.sqrt(2 * Math.PI) * Math.exp(-.5 * u * u);
  },
  cosine: function(u) {
    if (u <= 1 && u >= -1) return Math.PI / 4 * Math.cos(Math.PI / 2 * u);
    return 0;
  }
};
// http://exploringdata.net/den_trac.htm
science.stats.kde = function() {
  var kernel = science.stats.kernel.gaussian,
      sample = [],
      bandwidth = science.stats.bandwidth.nrd;

  function kde(points, i) {
    var bw = bandwidth.call(this, sample);
    return points.map(function(x) {
      var i = -1,
          y = 0,
          n = sample.length;
      while (++i < n) {
        y += kernel((x - sample[i]) / bw);
      }
      return [x, y / bw / n];
    });
  }

  kde.kernel = function(x) {
    if (!arguments.length) return kernel;
    kernel = x;
    return kde;
  };

  kde.sample = function(x) {
    if (!arguments.length) return sample;
    sample = x;
    return kde;
  };

  kde.bandwidth = function(x) {
    if (!arguments.length) return bandwidth;
    bandwidth = science.functor(x);
    return kde;
  };

  return kde;
};
// Based on figue implementation by Jean-Yves Delort.
// http://code.google.com/p/figue/
science.stats.kmeans = function() {
  var distance = science.stats.distance.euclidean,
      maxIterations = 1000,
      k = 1;

  function kmeans(vectors) {
    var n = vectors.length,
        assignments = [],
        clusterSizes = [],
        repeat = 1,
        iterations = 0,
        centroids = science_stats_kmeansRandom(k, vectors),
        newCentroids,
        i,
        j,
        x,
        d,
        min,
        best;

    while (repeat && iterations < maxIterations) {
      // Assignment step.
      j = -1; while (++j < k) {
        clusterSizes[j] = 0;
      }

      i = -1; while (++i < n) {
        x = vectors[i];
        min = Infinity;
        j = -1; while (++j < k) {
          d = distance.call(this, centroids[j], x);
          if (d < min) {
            min = d;
            best = j;
          }
        }
        clusterSizes[assignments[i] = best]++;
      }

      // Update centroids step.
      newCentroids = [];
      i = -1; while (++i < n) {
        x = assignments[i];
        d = newCentroids[x];
        if (d == null) newCentroids[x] = vectors[i].slice();
        else {
          j = -1; while (++j < d.length) {
            d[j] += vectors[i][j];
          }
        }
      }
      j = -1; while (++j < k) {
        x = newCentroids[j];
        d = 1 / clusterSizes[j];
        i = -1; while (++i < x.length) x[i] *= d;
      }

      // Check convergence.
      repeat = 0;
      j = -1; while (++j < k) {
        if (!science_stats_kmeansCompare(newCentroids[j], centroids[j])) {
          repeat = 1;
          break;
        }
      }
      centroids = newCentroids;
      iterations++;
    }
    return {assignments: assignments, centroids: centroids};
  }

  kmeans.k = function(x) {
    if (!arguments.length) return k;
    k = x;
    return kmeans;
  };

  kmeans.distance = function(x) {
    if (!arguments.length) return distance;
    distance = x;
    return kmeans;
  };

  return kmeans;
};

function science_stats_kmeansCompare(a, b) {
  if (!a || !b || a.length !== b.length) return false;
  var n = a.length,
      i = -1;
  while (++i < n) if (a[i] !== b[i]) return false;
  return true;
}

// Returns an array of k distinct vectors randomly selected from the input
// array of vectors. Returns null if k > n or if there are less than k distinct
// objects in vectors.
function science_stats_kmeansRandom(k, vectors) {
  var n = vectors.length;
  if (k > n) return null;
  
  var selected_vectors = [];
  var selected_indices = [];
  var tested_indices = {};
  var tested = 0;
  var selected = 0;
  var i,
      vector,
      select;

  while (selected < k) {
    if (tested === n) return null;
    
    var random_index = Math.floor(Math.random() * n);
    if (random_index in tested_indices) continue;
    
    tested_indices[random_index] = 1;
    tested++;
    vector = vectors[random_index];
    select = true;
    for (i = 0; i < selected; i++) {
      if (science_stats_kmeansCompare(vector, selected_vectors[i])) {
        select = false;
        break;
      }
    }
    if (select) {
      selected_vectors[selected] = vector;
      selected_indices[selected] = random_index;
      selected++;
    }
  }
  return selected_vectors;
}
science.stats.hcluster = function() {
  var distance = science.stats.distance.euclidean,
      linkage = "simple"; // simple, complete or average

  function hcluster(vectors) {
    var n = vectors.length,
        dMin = [],
        cSize = [],
        distMatrix = [],
        clusters = [],
        c1,
        c2,
        c1Cluster,
        c2Cluster,
        p,
        root,
        i,
        j;

    // Initialise distance matrix and vector of closest clusters.
    i = -1; while (++i < n) {
      dMin[i] = 0;
      distMatrix[i] = [];
      j = -1; while (++j < n) {
        distMatrix[i][j] = i === j ? Infinity : distance(vectors[i] , vectors[j]);
        if (distMatrix[i][dMin[i]] > distMatrix[i][j]) dMin[i] = j;
      }
    }

    // create leaves of the tree
    i = -1; while (++i < n) {
      clusters[i] = [];
      clusters[i][0] = {
        left: null,
        right: null,
        dist: 0,
        centroid: vectors[i],
        size: 1,
        depth: 0
      };
      cSize[i] = 1;
    }

    // Main loop
    for (p = 0; p < n-1; p++) {
      // find the closest pair of clusters
      c1 = 0;
      for (i = 0; i < n; i++) {
        if (distMatrix[i][dMin[i]] < distMatrix[c1][dMin[c1]]) c1 = i;
      }
      c2 = dMin[c1];

      // create node to store cluster info 
      c1Cluster = clusters[c1][0];
      c2Cluster = clusters[c2][0];

      newCluster = {
        left: c1Cluster,
        right: c2Cluster,
        dist: distMatrix[c1][c2],
        centroid: calculateCentroid(c1Cluster.size, c1Cluster.centroid,
          c2Cluster.size, c2Cluster.centroid),
        size: c1Cluster.size + c2Cluster.size,
        depth: 1 + Math.max(c1Cluster.depth, c2Cluster.depth)
      };
      clusters[c1].splice(0, 0, newCluster);
      cSize[c1] += cSize[c2];

      // overwrite row c1 with respect to the linkage type
      for (j = 0; j < n; j++) {
        switch (linkage) {
          case "single":
            if (distMatrix[c1][j] > distMatrix[c2][j])
              distMatrix[j][c1] = distMatrix[c1][j] = distMatrix[c2][j];
            break;
          case "complete":
            if (distMatrix[c1][j] < distMatrix[c2][j])
              distMatrix[j][c1] = distMatrix[c1][j] = distMatrix[c2][j];
            break;
          case "average":
            distMatrix[j][c1] = distMatrix[c1][j] = (cSize[c1] * distMatrix[c1][j] + cSize[c2] * distMatrix[c2][j]) / (cSize[c1] + cSize[j]);
            break;
        }
      }
      distMatrix[c1][c1] = Infinity;

      // infinity ­out old row c2 and column c2
      for (i = 0; i < n; i++)
        distMatrix[i][c2] = distMatrix[c2][i] = Infinity;

      // update dmin and replace ones that previous pointed to c2 to point to c1
      for (j = 0; j < n; j++) {
        if (dMin[j] == c2) dMin[j] = c1;
        if (distMatrix[c1][j] < distMatrix[c1][dMin[c1]]) dMin[c1] = j;
      }

      // keep track of the last added cluster
      root = newCluster;
    }

    return root;
  }

  hcluster.distance = function(x) {
    if (!arguments.length) return distance;
    distance = x;
    return hcluster;
  };

  return hcluster;
};

function calculateCentroid(c1Size, c1Centroid, c2Size, c2Centroid) {
  var newCentroid = [],
      newSize = c1Size + c2Size,
      n = c1Centroid.length,
      i = -1;
  while (++i < n) {
    newCentroid[i] = (c1Size * c1Centroid[i] + c2Size * c2Centroid[i]) / newSize;
  }
  return newCentroid;
}
science.stats.iqr = function(x) {
  var quartiles = science.stats.quantiles(x, [.25, .75]);
  return quartiles[1] - quartiles[0];
};
// Based on org.apache.commons.math.analysis.interpolation.LoessInterpolator
// from http://commons.apache.org/math/
science.stats.loess = function() {    
  var bandwidth = .3,
      robustnessIters = 2,
      accuracy = 1e-12;

  function smooth(xval, yval, weights) {
    var n = xval.length,
        i;

    if (n !== yval.length) throw {error: "Mismatched array lengths"};
    if (n == 0) throw {error: "At least one point required."};

    if (arguments.length < 3) {
      weights = [];
      i = -1; while (++i < n) weights[i] = 1;
    }

    science_stats_loessFiniteReal(xval);
    science_stats_loessFiniteReal(yval);
    science_stats_loessFiniteReal(weights);
    science_stats_loessStrictlyIncreasing(xval);

    if (n == 1) return [yval[0]];
    if (n == 2) return [yval[0], yval[1]];

    var bandwidthInPoints = Math.floor(bandwidth * n);

    if (bandwidthInPoints < 2) throw {error: "Bandwidth too small."};

    var res = [],
        residuals = [],
        robustnessWeights = [];

    // Do an initial fit and 'robustnessIters' robustness iterations.
    // This is equivalent to doing 'robustnessIters+1' robustness iterations
    // starting with all robustness weights set to 1.
    i = -1; while (++i < n) {
      res[i] = 0;
      residuals[i] = 0;
      robustnessWeights[i] = 1;
    }

    var iter = -1;
    while (++iter <= robustnessIters) {
      var bandwidthInterval = [0, bandwidthInPoints - 1];
      // At each x, compute a local weighted linear regression
      var x;
      i = -1; while (++i < n) {
        x = xval[i];

        // Find out the interval of source points on which
        // a regression is to be made.
        if (i > 0) {
          science_stats_loessUpdateBandwidthInterval(xval, weights, i, bandwidthInterval);
        }

        var ileft = bandwidthInterval[0],
            iright = bandwidthInterval[1];

        // Compute the point of the bandwidth interval that is
        // farthest from x
        var edge = (xval[i] - xval[ileft]) > (xval[iright] - xval[i]) ? ileft : iright;

        // Compute a least-squares linear fit weighted by
        // the product of robustness weights and the tricube
        // weight function.
        // See http://en.wikipedia.org/wiki/Linear_regression
        // (section "Univariate linear case")
        // and http://en.wikipedia.org/wiki/Weighted_least_squares
        // (section "Weighted least squares")
        var sumWeights = 0,
            sumX = 0,
            sumXSquared = 0,
            sumY = 0,
            sumXY = 0,
            denom = Math.abs(1 / (xval[edge] - x));

        for (var k = ileft; k <= iright; ++k) {
          var xk   = xval[k],
              yk   = yval[k],
              dist = k < i ? x - xk : xk - x,
              w    = science_stats_loessTricube(dist * denom) * robustnessWeights[k] * weights[k],
              xkw  = xk * w;
          sumWeights += w;
          sumX += xkw;
          sumXSquared += xk * xkw;
          sumY += yk * w;
          sumXY += yk * xkw;
        }

        var meanX = sumX / sumWeights,
            meanY = sumY / sumWeights,
            meanXY = sumXY / sumWeights,
            meanXSquared = sumXSquared / sumWeights;

        var beta = (Math.sqrt(Math.abs(meanXSquared - meanX * meanX)) < accuracy)
            ? 0 : ((meanXY - meanX * meanY) / (meanXSquared - meanX * meanX));

        var alpha = meanY - beta * meanX;

        res[i] = beta * x + alpha;
        residuals[i] = Math.abs(yval[i] - res[i]);
      }

      // No need to recompute the robustness weights at the last
      // iteration, they won't be needed anymore
      if (iter === robustnessIters) {
        break;
      }

      // Recompute the robustness weights.

      // Find the median residual.
      var sortedResiduals = residuals.slice();
      sortedResiduals.sort();
      var medianResidual = sortedResiduals[Math.floor(n / 2)];

      if (Math.abs(medianResidual) < accuracy)
        break;

      var arg,
          w;
      i = -1; while (++i < n) {
        arg = residuals[i] / (6 * medianResidual);
        robustnessWeights[i] = (arg >= 1) ? 0 : ((w = 1 - arg * arg) * w);
      }
    }

    return res;
  }

  smooth.bandwidth = function(x) {
    if (!arguments.length) return x;
    bandwidth = x;
    return smooth;
  };

  smooth.robustnessIterations = function(x) {
    if (!arguments.length) return x;
    robustnessIters = x;
    return smooth;
  };

  smooth.accuracy = function(x) {
    if (!arguments.length) return x;
    accuracy = x;
    return smooth;
  };

  return smooth;
};

function science_stats_loessFiniteReal(values) {
  var n = values.length,
      i = -1;

  while (++i < n) if (!isFinite(values[i])) return false;

  return true;
}

function science_stats_loessStrictlyIncreasing(xval) {
  var n = xval.length,
      i = 0;

  while (++i < n) if (xval[i - 1] >= xval[i]) return false;

  return true;
}

// Compute the tricube weight function.
// http://en.wikipedia.org/wiki/Local_regression#Weight_function
function science_stats_loessTricube(x) {
  return (x = 1 - x * x * x) * x * x;
}

// Given an index interval into xval that embraces a certain number of
// points closest to xval[i-1], update the interval so that it embraces
// the same number of points closest to xval[i], ignoring zero weights.
function science_stats_loessUpdateBandwidthInterval(
  xval, weights, i, bandwidthInterval) {

  var left = bandwidthInterval[0],
      right = bandwidthInterval[1];

  // The right edge should be adjusted if the next point to the right
  // is closer to xval[i] than the leftmost point of the current interval
  var nextRight = science_stats_loessNextNonzero(weights, right);
  if ((nextRight < xval.length) && (xval[nextRight] - xval[i]) < (xval[i] - xval[left])) {
    var nextLeft = science_stats_loessNextNonzero(weights, left);
    bandwidthInterval[0] = nextLeft;
    bandwidthInterval[1] = nextRight;
  }
}

function science_stats_loessNextNonzero(weights, i) {
  var j = i + 1;
  while (j < weights.length && weights[j] === 0) j++;
  return j;
}
// Welford's algorithm.
science.stats.mean = function(x) {
  var n = x.length;
  if (n === 0) return NaN;
  var m = 0,
      i = -1;
  while (++i < n) m += (x[i] - m) / (i + 1);
  return m;
};
science.stats.median = function(x) {
  return science.stats.quantiles(x, [.5])[0];
};
science.stats.mode = function(x) {
  x = x.slice().sort(science.ascending);
  var mode,
      n = x.length,
      i = -1,
      l = i,
      last = null,
      max = 0,
      tmp,
      v;
  while (++i < n) {
    if ((v = x[i]) !== last) {
      if ((tmp = i - l) > max) {
        max = tmp;
        mode = last;
      }
      last = v;
      l = i;
    }
  }
  return mode;
};
// Uses R's quantile algorithm type=7.
science.stats.quantiles = function(d, quantiles) {
  d = d.slice().sort(science.ascending);
  var n_1 = d.length - 1;
  return quantiles.map(function(q) {
    if (q === 0) return d[0];
    else if (q === 1) return d[n_1];

    var index = 1 + q * n_1,
        lo = Math.floor(index),
        h = index - lo,
        a = d[lo - 1];

    return h === 0 ? a : a + h * (d[lo] - a);
  });
};
// Unbiased estimate of a sample's variance.
// Also known as the sample variance, where the denominator is n - 1.
science.stats.variance = function(x) {
  var n = x.length;
  if (n < 1) return NaN;
  if (n === 1) return 0;
  var mean = science.stats.mean(x),
      i = -1,
      s = 0;
  while (++i < n) {
    var v = x[i] - mean;
    s += v * v;
  }
  return s / (n - 1);
};
})()