private static void FitToNormalInternal(IEnumerable <double> sample,
out double m, out double dm, out double s, out double ds)
{
// We factor out this method because it is used by both the normal and the log-normal fit methods.
// Maximum likelihood estimate is straightforward.
// p_i = \frac{1}{\sqrt{2\pi}\sigma} \exp \left[ -\frac{1}{2} \left( \frac{x_i - \mu}{\sigma} \right)^2 \right]
// \ln p_i = -\ln (\sqrt{2\pi} \sigma) - \frac{1}{2} \left( \frac{x_i - \mu}{\sigma} \right)^2
// \ln L = \sum_i
// so
// \frac{\partial \ln L}{\partial \mu} = \sum_i \frac{x_i - \mu}{\sigma^2}
// \frac{\partial \ln L}{\partial \sigma} = -\frac{n}{\sigma} - \frac{1}{\sigma^2} \sum_i (x_i - \mu)^2
// Setting equal to zero and solving gives the unsurprising result
// \mu = n^{-1} \sum_i x_i
// \sigma^2 = n^{-1} \sum_i (x_i - \mu)^2
// that MLE says to estimate the model mean and variance by the sample mean and variance.
// MLE estimators are guaranteed to be asymptotically unbiased, but they can be biased for finite n.
// You can see that must be the case for \sigma because the denominator has n instead of n-1.
// To un-bias our estimators, we will derive exact distributions for these quantities.
// First the mean estimator. Start from x_i \sim N(\mu, \sigma). By the addition of normal deviates,
// \sum_i x_i \sim N(n \mu, \sqrt{n} \sigma). So
// m = \frac{1}{n} \sum_i x_i \sim N(\mu, \sigma / \sqrt{n}).
// which means the estimator m is normally distributed with mean \mu and standard deviation
// \sigma / \sqrt{n}. Now we know that m is unbiased and we know its variance.
int n;
double ss;
Univariate.ComputeMomentsUpToSecond(sample, out n, out m, out ss);
dm = Math.Sqrt(ss) / n;
// Next the variance estimator. By the definition of the chi squared distribution and a bit of algebra that
// reduces the degrees of freedom by one, u^2 = \sum_i ( \frac{x_i - m}{\sigma} )^2 \sim \chi^2(n - 1), which has
// mean n - 1 and variance 2(n-1). Therefore the estimator
// v = \sigma^2 u^2 / (n-1) = \frac{1}{n-1} \sum_i ( x_i - m )^2
// has mean \sigma^2 and variance 2 \sigma^4 / (n-1).
// If we consider \sigma^2 the parameter, we are done -- we have derived an estimator that is unbiased and
// know its variance. But we don't consider the parameter \sigma^2, we consider it \sigma.
// The mean of the square root is not the square root of the mean, so the square root of an unbiased
// estimator of \sigma^2 will not be an unbiased estimator of \sigma. If we want an unbiased estimator of \sigma
// itself, we need to go a bit further. Since u^2 ~ \chi^2(n-1), u ~ \chi(n-1). It's mean is a complicated ratio
// of Gamma functions and it's variance is an even more complicated difference whose evaluation can be delicate,
// but our machinery in the ChiDistribution class handles that. To get an unbiased estimator of \sigma, we just
// need to apply the same principal of dividing by the mean of this distribution.
// s = \sigma u / <u> = \sqrt{\sum_i (x_i - m)^2} / <u>
// to get an estimator with mean \sigma and known variance.
ChiDistribution d = new ChiDistribution(n - 1);
s = Math.Sqrt(ss) / d.Mean;
ds = d.StandardDeviation / d.Mean * s;
}
internal MultiLinearRegressionResult(IReadOnlyList <double> yColumn, IReadOnlyList <IReadOnlyList <double> > xColumns, IReadOnlyList <string> xNames) : base()
{
Debug.Assert(yColumn != null);
Debug.Assert(xColumns != null);
Debug.Assert(xColumns.Count > 0);
Debug.Assert(xNames.Count == xColumns.Count);
n = yColumn.Count;
m = xColumns.Count;
if (n <= m)
{
throw new InsufficientDataException();
}
// Compute the design matrix X.
interceptIndex = -1;
RectangularMatrix X = new RectangularMatrix(n, m);
for (int c = 0; c < m; c++)
{
IReadOnlyList <double> xColumn = xColumns[c];
if (xColumn == null)
{
Debug.Assert(xNames[c] == "Intercept");
Debug.Assert(interceptIndex < 0);
for (int r = 0; r < n; r++)
{
X[r, c] = 1.0;
}
interceptIndex = c;
}
else
{
Debug.Assert(xNames[c] != null);
if (xColumn.Count != n)
{
throw new DimensionMismatchException();
}
for (int r = 0; r < n; r++)
{
X[r, c] = xColumn[r];
}
}
}
Debug.Assert(interceptIndex >= 0);
ColumnVector v = new ColumnVector(yColumn);
// Use X = QR to solve X b = y and compute C.
QRDecomposition.SolveLinearSystem(X, v, out b, out C);
// For ANOVA, we will need mean and variance of y
int yn;
double ym;
Univariate.ComputeMomentsUpToSecond(yColumn, out yn, out ym, out SST);
// Compute residuals
SSR = 0.0;
SSF = 0.0;
ColumnVector yHat = X * b;
residuals = new List <double>(n);
for (int i = 0; i < n; i++)
{
double z = yColumn[i] - yHat[i];
residuals.Add(z);
SSR += z * z;
SSF += MoreMath.Sqr(yHat[i] - ym);
}
sigma2 = SSR / (n - m);
// Scale up C by \sigma^2
// (It sure would be great to be able to overload *=.)
for (int i = 0; i < m; i++)
{
for (int j = i; j < m; j++)
{
C[i, j] = C[i, j] * sigma2;
}
}
names = xNames;
}
/// <summary>
/// Performs a Spearman rank-order test of association between the two variables.
/// </summary>
/// <param name="x">The values of the first variable.</param>
/// <param name="y">The values of the second variable.</param>
/// <returns>The result of the test.</returns>
/// <remarks>
/// <para>The Spearman rank-order test of association is a non-parametric test for association between
/// two variables. The test statistic rho is the correlation coefficient of the <em>rank</em> of
/// each entry in the sample. It is thus invariant over monotonic re-parameterizations of the data,
/// and will, for example, detect a quadratic or exponential association just as well as a linear
/// association.</para>
/// <para>The Spearman rank-order test requires O(N log N) operations.</para>
/// </remarks>
/// <exception cref="ArgumentNullException"><paramref name="x"/> or <paramref name="y"/> is <see langword="null"/>.</exception>
/// <exception cref="DimensionMismatchException"><paramref name="x"/> and <paramref name="y"/> do not contain the same number of entries.</exception>
/// <exception cref="InsufficientDataException">There are fewer than three data points.</exception>
/// <seealso cref="PearsonRTest(IReadOnlyList{double},IReadOnlyList{double})"/>
/// <seealso cref="KendallTauTest"/>
/// <seealso href="http://en.wikipedia.org/wiki/Spearman%27s_rank_correlation_coefficient"/>
public static TestResult SpearmanRhoTest(IReadOnlyList <double> x, IReadOnlyList <double> y)
{
if (x == null)
{
throw new ArgumentNullException(nameof(x));
}
if (y == null)
{
throw new ArgumentNullException(nameof(y));
}
if (x.Count != y.Count)
{
throw new DimensionMismatchException();
}
int n = x.Count;
if (n < 3)
{
throw new InsufficientDataException();
}
// Find the ranks.
int[] rx = Univariate.GetRanks(x);
int[] ry = Univariate.GetRanks(y);
// Compute the statistic and its null distribution.
// Use analytic expressions for the mean M and variance V of the ranks.
// C is the covariance of the ranks, rho is just the corresponding correlation coefficient.
// S encodes the same information, but as an integer that varies in steps of one, so
// its null distribution can be described by a DiscreteDistribution.
double M = (n - 1) / 2.0;
double V = (n + 1) * (n - 1) / 12.0;
int S = 0;
double C = 0.0;
for (int i = 0; i < n; i++)
{
// Statisticians define S using 1-based ranks, so add 1 to each
// rank when computing S. This isn't important for C, because
// we are subtracting off mean.
S += (rx[i] + 1) * (ry[i] + 1);
C += (rx[i] - M) * (ry[i] - M);
}
C = C / n;
double rho = C / V;
// Compute the null distribution.
if (n < 12)
{
// For small enough samples, use the exact distribution.
// It would be nice to do this for at least slightly higher n, but the time to compute the exact
// distribution grows dramatically with n. I would like to return in less than about 100ms.
// Current timings are n = 10 35ms, n = 11, 72ms, n = 12 190ms.
DiscreteDistribution sDistribution = new SpearmanExactDistribution(n);
ContinuousDistribution rhoDistribution = new DiscreteAsContinuousDistribution(new SpearmanExactDistribution(n), Interval.FromEndpoints(-1.0, 1.0));
return(new TestResult("s", S, sDistribution, "ρ", rho, rhoDistribution, TestType.TwoTailed));
}
else
{
// For larger samples, use the normal approximation.
// It would be nice to fit support and/or fourth cumulant.
// I was not happy with an Edgeworth expansion, which can fit the fourth cumulant, but screws up the tails
// badly, even giving negative probabilities for extreme values, which are quite likely for null-violating samples.
// Look into bounded quasi-normal distributions such as the logit-normal and truncated normal.
ContinuousDistribution rhoDistribution = new NormalDistribution(0.0, 1.0 / Math.Sqrt(n - 1));
return(new TestResult("ρ", rho, rhoDistribution, TestType.TwoTailed));
}
}
internal PolynomialRegressionResult(IReadOnlyList <double> x, IReadOnlyList <double> y, int degree) : base()
{
Debug.Assert(x != null);
Debug.Assert(y != null);
Debug.Assert(x.Count == y.Count);
Debug.Assert(degree >= 0);
m = degree;
n = x.Count;
if (n < (m + 1))
{
throw new InsufficientDataException();
}
// Construct the n X m design matrix X_{ij} = x_{i}^{j}
RectangularMatrix X = new RectangularMatrix(n, m + 1);
ColumnVector Y = new ColumnVector(n);
for (int i = 0; i < n; i++)
{
double x_i = x[i];
X[i, 0] = 1.0;
for (int j = 1; j <= m; j++)
{
X[i, j] = X[i, j - 1] * x_i;
}
double y_i = y[i];
Y[i] = y_i;
}
// Use X = QR to solve X b = y and compute C
QRDecomposition.SolveLinearSystem(X, Y, out b, out C);
// Compute mean and total sum of squares.
// This could be done inside loop above, but this way we get to re-use code from Univariate.
double yMean;
Univariate.ComputeMomentsUpToSecond(y, out n, out yMean, out SST);
// Compute residuals
SSR = 0.0;
SSF = 0.0;
ColumnVector yHat = X * b;
residuals = new List <double>(n);
for (int i = 0; i < n; i++)
{
double z = y[i] - yHat[i];
residuals.Add(z);
SSR += z * z;
SSF += MoreMath.Sqr(yHat[i] - yMean);
}
sigma2 = SSR / (n - (m + 1));
// Scale up C by \sigma^2
// (It sure would be great to be able to overload *=.)
for (int i = 0; i <= m; i++)
{
for (int j = i; j <= m; j++)
{
C[i, j] = C[i, j] * sigma2;
}
}
}
/// <summary>
/// Find the Gumbel distribution that best fit the given sample.
/// </summary>
/// <param name="sample">The sample to fit.</param>
/// <returns>The fit result.</returns>
/// <exception cref="ArgumentNullException"><paramref name="sample"/> is <see langword="null"/>.</exception>
/// <exception cref="InsufficientDataException"><paramref name="sample"/> contains fewer than three values.</exception>
public static GumbelFitResult FitToGumbel(this IReadOnlyList <double> sample)
{
if (sample == null)
{
throw new ArgumentNullException(nameof(sample));
}
if (sample.Count < 3)
{
throw new InsufficientDataException();
}
// To do a maximum likelihood fit, start from the log probability of each data point and aggregate to
// obtain the log likelihood of the sample
// z_i = \frac{x_i - m}{s}
// -\ln p_i = \ln s + ( z_i + e^{-z_i})
// \ln L = \sum_i \ln p_i
// Take derivatives wrt m and s.
// \frac{\partial \ln L}{\partial m} = \frac{1}{s} \sum_i ( 1 - e^{-z_i} )
// \frac{\partial \ln L}{\partial s} = \frac{1}{s} \sum_i ( -1 + z_i - z_i e^{-z_i} )
// Set derivatives to zero to get a system of equations for the maximum.
// n = \sum_i e^{-z_i}
// n = \sum_i ( z_i - z_i e^{-z_i} )
// that is, <e^z> = 1 and <z> - <z e^z> = 1.
// To solve this system, pull e^{m/s} out of the sum in the first equation and solve for m
// n = e^{m / s} \sum_i e^{-x_i / s}
// m = -s \ln \left( \frac{1}{n} \sum_i e^{-x_i / s} \right) = -s \ln <e^{-x/s}>
// Substituting this result into the second equation gets us to
// s = \bar{x} - \frac{ <x e^{-x/s}> }{ <e^{x/s}> }
// which involves only s. We can use a one-dimensional root-finder to determine s, then determine m
// from the first equation.
// To avoid exponentiating potentially large x_i, it's better to write the problem in terms
// of d_i, where x_i = \bar{x} + d_i.
// m = \bar{x} - s \ln <e^{-d/s}>
// s = -\frac{ <d e^{-d/s}> }{ <e^{-d/s}> }
// To get the covariance matrix, we need the curvature matrix at the minimum, so take more derivatives
// \frac{\partial^2 \ln L}{\partial m^2} = - \frac{1}{s} \sum_i e^{-z_i} = - \frac{n}{s^2}
// \frac{\partial^2 \ln L}{\partial m \partial s} = - \frac{n}{s^2} <z e^{-z}>
// \frac{\partial^2 \ln L}{\partial s^2} = - \frac{n}{s^2} ( <z^2 e^{-z}> + 1 )
// Several crucial pieces of this analysis are taken from Mahdi and Cenac, "Estimating Parameters of Gumbel Distribution
// "using the method of moments, probability weighted moments, and maximum likelihood", Revista de Mathematica:
// Teoria y Aplicaciones 12 (2005) 151-156 (http://revistas.ucr.ac.cr/index.php/matematica/article/viewFile/259/239)
// We will be needed the sample mean and standard deviation
int n;
double mean, stdDev;
Univariate.ComputeMomentsUpToSecond(sample, out n, out mean, out stdDev);
stdDev = Math.Sqrt(stdDev / n);
// Use the method of moments to get an initial estimate of s.
double s0 = Math.Sqrt(6.0) / Math.PI * stdDev;
// Define the function to zero
Func <double, double> fnc = (double s) => {
double u, v;
MaximumLikelihoodHelper(sample, n, mean, s, out u, out v);
return(s + v / u);
};
// Zero it to compute the best-fit s
double s1 = FunctionMath.FindZero(fnc, s0);
// Compute the corresponding best-fit m
double u1, v1;
MaximumLikelihoodHelper(sample, n, mean, s1, out u1, out v1);
double m1 = mean - s1 * Math.Log(u1);
// Compute the curvature matrix
double w1 = 0.0;
double w2 = 0.0;
foreach (double x in sample)
{
double z = (x - m1) / s1;
double e = Math.Exp(-z);
w1 += z * e;
w2 += z * z * e;
}
w1 /= sample.Count;
w2 /= sample.Count;
SymmetricMatrix C = new SymmetricMatrix(2);
C[0, 0] = (n - 2) / (s1 * s1);
C[0, 1] = (n - 2) / (s1 * s1) * w1;
C[1, 1] = (n - 2) / (s1 * s1) * (w2 + 1.0);
SymmetricMatrix CI = C.CholeskyDecomposition().Inverse();
// The use of (n-2) here in place of n is a very ad hoc attempt to increase accuracy.
// Compute goodness-of-fit
GumbelDistribution dist = new GumbelDistribution(m1, s1);
TestResult test = sample.KolmogorovSmirnovTest(dist);
return(new GumbelFitResult(m1, s1, CI[0, 0], CI[1, 1], CI[0, 1], test));
}