public void RootOfEi()
{
double x = FunctionMath.FindZero(AdvancedMath.IntegralEi, Interval.FromEndpoints(0.1, 1.0));
Assert.IsTrue(TestUtilities.IsNearlyEqual(x, 0.37250741078136663446));
Assert.IsTrue(Math.Abs(AdvancedMath.IntegralEi(x)) < TestUtilities.TargetPrecision);
}
public void RootOfPsi()
{
double x = FunctionMath.FindZero(AdvancedMath.Psi, 1.5);
Assert.IsTrue(TestUtilities.IsNearlyEqual(x, 1.46163214496836234126));
Assert.IsTrue(AdvancedMath.Psi(x) < TestUtilities.TargetPrecision);
}
public void IntegralEiZero()
{
// Value of zero documented at https://dlmf.nist.gov/6.13
double x0 = FunctionMath.FindZero(AdvancedMath.IntegralEi, 1.0);
Assert.IsTrue(TestUtilities.IsNearlyEqual(x0, 0.37250741078136663446));
}
public void Bug5505()
{
// finding the root of x^3 would fail with a NonconvergenceException because our termination criterion tested only
// for small relative changes in x; adding a test for small absolute changes too solved the problem
double x0 = FunctionMath.FindZero(delegate(double x) { return(x * x * x); }, 1.0);
Assert.IsTrue(Math.Abs(x0) < TestUtilities.TargetPrecision);
}
public void RootOfJ0()
{
Func <double, double> f = delegate(double x) {
return(AdvancedMath.BesselJ(0, x));
};
double y = FunctionMath.FindZero(f, Interval.FromEndpoints(2.0, 4.0));
Assert.IsTrue(TestUtilities.IsNearlyEqual(y, 2.40482555769577276862));
}
/// <summary>
/// Returns the point at which the right probability function attains the given value.
/// </summary>
/// <param name="Q">The right cumulative probability, which must lie between 0 and 1.</param>
/// <returns>The value x for which <see cref="RightProbability"/> equals Q.</returns>
public virtual double InverseRightProbability(double Q)
{
if ((Q < 0.0) || (Q > 1.0))
{
throw new ArgumentOutOfRangeException(nameof(Q));
}
Func <double, double> f = delegate(double x) {
return(RightProbability(x) - Q);
};
double y = FunctionMath.FindZero(f, Mean);
return(y);
}
/// <summary>
/// Returns the point at which the cumulative distribution function attains a given value.
/// </summary>
/// <param name="P">The left cumulative probability P, which must lie between 0 and 1.</param>
/// <returns>The value x at which <see cref="LeftProbability"/> equals P.</returns>
/// <remarks>
/// <para>The inverse left probability is commonly called the quantile function. Given a quantile,
/// it tells which variable value is the lower border of that quantile.</para>
/// </remarks>
/// <exception cref="ArgumentOutOfRangeException"><paramref name="P"/> lies outside [0,1].</exception>
/// <seealso href="https://en.wikipedia.org/wiki/Quantile_function"/>
public virtual double InverseLeftProbability(double P)
{
// find x where LeftProbability(x) = P
if ((P < 0.0) || (P > 1.0))
{
throw new ArgumentOutOfRangeException(nameof(P));
}
Func <double, double> f = delegate(double x) {
return(LeftProbability(x) - P);
};
double y = FunctionMath.FindZero(f, Mean);
return(y);
// since we have the PDF = CDF', change to a method using Newton's method
}
public static double AiryBiZero(int k)
{
if (k < 1)
{
throw new ArgumentOutOfRangeException(nameof(k));
}
double tm = 3.0 / 8.0 * Math.PI * (4 * k - 4);
double tp = 3.0 / 8.0 * Math.PI * (4 * k - 2);
double bm = (tm == 0.0) ? 0.0 : -T(tm);
double bp = -T(tp);
double b1 = FunctionMath.FindZero(AdvancedMath.AiryBi, Interval.FromEndpoints(bp, bm));
return(b1);
}
// **** Spherical Bessel functions ****
// Functions needed for uniform asymptotic expansions
/// <summary>
/// Computes the requested zero of the regular Bessel function.
/// </summary>
/// <param name="nu">The order, which must be non-negative.</param>
/// <param name="k">The index of the zero, which must be positive.</param>
/// <returns>The <paramref name="k"/>th value of x for which J<sub>ν</sub>(x) = 0.</returns>
/// <exception cref="ArgumentOutOfRangeException"><paramref name="nu"/> is negative or <paramref name="k"/> is non-positive.</exception>
public static double BesselJZero(double nu, int k)
{
if (nu < 0.0)
{
throw new ArgumentOutOfRangeException(nameof(nu));
}
if (k < 1)
{
throw new ArgumentOutOfRangeException(nameof(k));
}
double jMin, jMax;
if (k == 1)
{
// Rayleigh inequalities
// Cited in Ismail and Muldoon, Bounds for the small real and purely imaginary zeros of
// Bessel and related functions, Methods and Applications of Analysis 2 (1995) p. 1-21
// http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.140.288&rep=rep1&type=pdf
// They cite Watson p. 502, but I don't find them there.
//jMin = 2.0 * Math.Sqrt((nu + 1.0) * Math.Sqrt(nu + 2.0));
// Lee Lorch, "Some Inequalities for the First Positive Zeros of the Bessel Functions",
// SIAM J. Math. Anal. 24 (1993) 814
jMin = Math.Sqrt((nu + 1.0) * (nu + 5.0));
jMax = ApproximateBesselJZero(nu + 1, k);
double tt = Math.Sqrt(2.0 * (nu + 1.0) * (nu + 3.0));
if (tt < jMax)
{
jMax = tt;
}
}
else
{
// Use the interlacing property j_{\nu + 1, k - 1} < j_{\nu, k} < j_{\nu + 1, k}
// and the approximation function to get a bound. The interlacing property is exact,
// but our approximation functions are not, so it is concievable this could go wrong.
jMin = ApproximateBesselJZero(nu + 1, k - 1);
jMax = ApproximateBesselJZero(nu + 1, k);
}
Debug.Assert(jMin < jMax);
// We should use internal method that takes best guess. Also, add derivative and use Newton's method.
double j = FunctionMath.FindZero(x => BesselJ(nu, x), Interval.FromEndpoints(jMin, jMax));
return(j);
}
public void ComplexRiemannZetaZeros()
{
// Zeros from http://www.dtc.umn.edu/~odlyzko/zeta_tables/zeros2
double[] zeros = new double[] {
14.134725141734693790,
21.022039638771554993,
25.010857580145688763,
30.424876125859513210
};
foreach (double zero in zeros)
{
double rho = FunctionMath.FindZero((double x) => {
Complex f = AdvancedComplexMath.RiemannZeta(new Complex(0.5, x));
return(f.Re + f.Im);
}, Math.Round(zero));
Assert.IsTrue(TestUtilities.IsNearlyEqual(rho, zero));
}
}
public static double AiryAiZero(int k)
{
if (k < 1)
{
throw new ArgumentOutOfRangeException(nameof(k));
}
double tm = 3.0 / 8.0 * Math.PI * (4 * k - 2);
//double t0 = 3.0 / 8.0 * Math.PI * (4 * k - 1);
double tp = 3.0 / 8.0 * Math.PI * (4 * k - 0);
double am = -T(tm);
//double a0 = -T(t0);
double ap = -T(tp);
double a1 = FunctionMath.FindZero(AdvancedMath.AiryAi, Interval.FromEndpoints(ap, am));
return(a1);
}
// routines for maximum likelyhood fitting
/// <summary>
/// Computes the Gamma distribution that best fits the given sample.
/// </summary>
/// <param name="sample">The sample to fit.</param>
/// <returns>The best fit parameters.</returns>
/// <remarks>
/// <para>The returned fit parameters are the <see cref="ShapeParameter"/> and <see cref="ScaleParameter"/>, in that order.
/// These are the same parameters, in the same order, that are required by the <see cref="GammaDistribution(double,double)"/> constructor to
/// specify a new Gamma distribution.</para>
/// </remarks>
/// <exception cref="ArgumentNullException"><paramref name="sample"/> is null.</exception>
/// <exception cref="InvalidOperationException"><paramref name="sample"/> contains non-positive values.</exception>
/// <exception cref="InsufficientDataException"><paramref name="sample"/> contains fewer than three values.</exception>
public static FitResult FitToSample(Sample sample)
{
if (sample == null)
{
throw new ArgumentNullException("sample");
}
if (sample.Count < 3)
{
throw new InsufficientDataException();
}
// The log likelyhood of a sample given k and s is
// \log L = (k-1) \sum_i \log x_i - \frac{1}{s} \sum_i x_i - N \log \Gamma(k) - N k \log s
// Differentiating,
// \frac{\partial \log L}{\partial s} = \frac{1}{s^2} \sum_i x_i - \frac{Nk}{s}
// \frac{\partial \log L}{\partial k} = \sum_i \log x_i - N \psi(k) - N \log s
// Setting the first equal to zero gives
// k s = N^{-1} \sum_i x_i = <x>
// \psi(k) + \log s = N^{-1} \sum_i \log x_i = <log x>
// Inserting the first into the second gives a single equation for k
// \log k - \psi(k) = \log <x> - <\log x>
// Note the RHS need only be computed once.
// \log k > \psi(k) for all k, so the RHS had better be positive. They get
// closer for large k, so smaller RHS will produce a larger k.
double s = 0.0;
foreach (double x in sample)
{
if (x <= 0.0)
{
throw new InvalidOperationException();
}
s += Math.Log(x);
}
s = Math.Log(sample.Mean) - s / sample.Count;
// We can get an initial guess for k from the method of moments
// \frac{\mu^2}{\sigma^2} = k
double k0 = MoreMath.Sqr(sample.Mean) / sample.Variance;
// Since 1/(2k) < \log(k) - \psi(k) < 1/k, we could get a bound; that
// might be better to avoid the solver running into k < 0 territory
double k1 = FunctionMath.FindZero(k => (Math.Log(k) - AdvancedMath.Psi(k) - s), k0);
double s1 = sample.Mean / k1;
// Curvature of the log likelyhood is straightforward
// \frac{\partial^2 \log L}{\partial s^2} = -\frac{2}{s^3} \sum_i x_i + \frac{Nk}{s^2} = - \frac{Nk}{s^2}
// \frac{\partial^2 \log L}{\partial k \partial s} = - \frac{N}{s}
// \frac{\partial^2 \log L}{\partial k^2} = - N \psi'(k)
// This gives the curvature matrix and thus via inversion the covariance matrix.
SymmetricMatrix B = new SymmetricMatrix(2);
B[0, 0] = sample.Count * AdvancedMath.Psi(1, k1);
B[0, 1] = sample.Count / s1;
B[1, 1] = sample.Count * k1 / MoreMath.Sqr(s1);
SymmetricMatrix C = B.CholeskyDecomposition().Inverse();
// Do a KS test for goodness-of-fit
TestResult test = sample.KolmogorovSmirnovTest(new GammaDistribution(k1, s1));
return(new FitResult(new double[] { k1, s1 }, C, test));
}
/// <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));
}