MillerUpdatingRegression.java

/*
 * Licensed to the Apache Software Foundation (ASF) under one or more
 * contributor license agreements.  See the NOTICE file distributed with
 * this work for additional information regarding copyright ownership.
 * The ASF licenses this file to You under the Apache License, Version 2.0
 * (the "License"); you may not use this file except in compliance with
 * the License.  You may obtain a copy of the License at
 *
 *      https://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/*
 * This is not the original file distributed by the Apache Software Foundation
 * It has been modified by the Hipparchus project
 */
package org.hipparchus.stat.regression;

import java.util.Arrays;

import org.hipparchus.exception.LocalizedCoreFormats;
import org.hipparchus.exception.MathIllegalArgumentException;
import org.hipparchus.stat.LocalizedStatFormats;
import org.hipparchus.util.FastMath;
import org.hipparchus.util.MathUtils;
import org.hipparchus.util.Precision;

/**
 * This class is a concrete implementation of the {@link UpdatingMultipleLinearRegression} interface.
 *
 * <p>The algorithm is described in:</p>
 * <pre>
 * Algorithm AS 274: Least Squares Routines to Supplement Those of Gentleman
 * Author(s): Alan J. Miller
 * Source: Journal of the Royal Statistical Society.
 * Series C (Applied Statistics), Vol. 41, No. 2
 * (1992), pp. 458-478
 * Published by: Blackwell Publishing for the Royal Statistical Society
 * Stable URL: <a href="https://www.jstor.org/stable/2347583">Algorithm AS 274:
 * Least Squares Routines to Supplement Those of Gentleman</a> </pre>
 *
 * <p>This method for multiple regression forms the solution to the OLS problem
 * by updating the QR decomposition as described by Gentleman.</p>
 *
 */
public class MillerUpdatingRegression implements UpdatingMultipleLinearRegression {

    /** number of variables in regression */
    private final int nvars;
    /** diagonals of cross products matrix */
    private final double[] d;
    /** the elements of the R`Y */
    private final double[] rhs;
    /** the off diagonal portion of the R matrix */
    private final double[] r;
    /** the tolerance for each of the variables */
    private final double[] tol;
    /** residual sum of squares for all nested regressions */
    private final double[] rss;
    /** order of the regressors */
    private final int[] vorder;
    /** scratch space for tolerance calc */
    private final double[] work_tolset;
    /** number of observations entered */
    private long nobs;
    /** sum of squared errors of largest regression */
    private double sserr;
    /** has rss been called? */
    private boolean rss_set;
    /** has the tolerance setting method been called */
    private boolean tol_set;
    /** flags for variables with linear dependency problems */
    private final boolean[] lindep;
    /** singular x values */
    private final double[] x_sing;
    /** workspace for singularity method */
    private final double[] work_sing;
    /** summation of Y variable */
    private double sumy;
    /** summation of squared Y values */
    private double sumsqy;
    /** boolean flag whether a regression constant is added */
    private final boolean hasIntercept;
    /** zero tolerance */
    private final double epsilon;
    /**
     *  Set the default constructor to private access
     *  to prevent inadvertent instantiation
     */
    @SuppressWarnings("unused")
    private MillerUpdatingRegression() {
        this(-1, false, Double.NaN);
    }

    /**
     * This is the augmented constructor for the MillerUpdatingRegression class.
     *
     * @param numberOfVariables number of regressors to expect, not including constant
     * @param includeConstant include a constant automatically
     * @param errorTolerance  zero tolerance, how machine zero is determined
     * @throws MathIllegalArgumentException if {@code numberOfVariables is less than 1}
     */
    public MillerUpdatingRegression(int numberOfVariables, boolean includeConstant, double errorTolerance)
            throws MathIllegalArgumentException {
        if (numberOfVariables < 1) {
            throw new MathIllegalArgumentException(LocalizedStatFormats.NO_REGRESSORS);
        }
        if (includeConstant) {
            this.nvars = numberOfVariables + 1;
        } else {
            this.nvars = numberOfVariables;
        }
        this.hasIntercept = includeConstant;
        this.nobs = 0;
        this.d = new double[this.nvars];
        this.rhs = new double[this.nvars];
        this.r = new double[this.nvars * (this.nvars - 1) / 2];
        this.tol = new double[this.nvars];
        this.rss = new double[this.nvars];
        this.vorder = new int[this.nvars];
        this.x_sing = new double[this.nvars];
        this.work_sing = new double[this.nvars];
        this.work_tolset = new double[this.nvars];
        this.lindep = new boolean[this.nvars];
        for (int i = 0; i < this.nvars; i++) {
            vorder[i] = i;
        }
        if (errorTolerance > 0) {
            this.epsilon = errorTolerance;
        } else {
            this.epsilon = -errorTolerance;
        }
    }

    /**
     * Primary constructor for the MillerUpdatingRegression.
     *
     * @param numberOfVariables maximum number of potential regressors
     * @param includeConstant include a constant automatically
     * @throws MathIllegalArgumentException if {@code numberOfVariables is less than 1}
     */
    public MillerUpdatingRegression(int numberOfVariables, boolean includeConstant)
            throws MathIllegalArgumentException {
        this(numberOfVariables, includeConstant, Precision.EPSILON);
    }

    /**
     * A getter method which determines whether a constant is included.
     * @return true regression has an intercept, false no intercept
     */
    @Override
    public boolean hasIntercept() {
        return this.hasIntercept;
    }

    /**
     * Gets the number of observations added to the regression model.
     * @return number of observations
     */
    @Override
    public long getN() {
        return this.nobs;
    }

    /**
     * Adds an observation to the regression model.
     * @param x the array with regressor values
     * @param y  the value of dependent variable given these regressors
     * @exception MathIllegalArgumentException if the length of {@code x} does not equal
     * the number of independent variables in the model
     */
    @Override
    public void addObservation(final double[] x, final double y)
            throws MathIllegalArgumentException {

        if ((!this.hasIntercept && x.length != nvars) ||
               (this.hasIntercept && x.length + 1 != nvars)) {
            throw new MathIllegalArgumentException(LocalizedStatFormats.INVALID_REGRESSION_OBSERVATION,
                    x.length, nvars);
        }
        if (!this.hasIntercept) {
            include(x.clone(), 1.0, y);
        } else {
            final double[] tmp = new double[x.length + 1];
            System.arraycopy(x, 0, tmp, 1, x.length);
            tmp[0] = 1.0;
            include(tmp, 1.0, y);
        }
        ++nobs;

    }

    /**
     * Adds multiple observations to the model.
     * @param x observations on the regressors
     * @param y observations on the regressand
     * @throws MathIllegalArgumentException if {@code x} is not rectangular, does not match
     * the length of {@code y} or does not contain sufficient data to estimate the model
     */
    @Override
    public void addObservations(double[][] x, double[] y) throws MathIllegalArgumentException {
        MathUtils.checkNotNull(x, LocalizedCoreFormats.INPUT_ARRAY);
        MathUtils.checkNotNull(y, LocalizedCoreFormats.INPUT_ARRAY);
        MathUtils.checkDimension(x.length, y.length);
        if (x.length == 0) {  // Must be no y data either
            throw new MathIllegalArgumentException(LocalizedCoreFormats.NO_DATA);
        }
        if (x[0].length + 1 > x.length) {
            throw new MathIllegalArgumentException(
                  LocalizedStatFormats.NOT_ENOUGH_DATA_FOR_NUMBER_OF_PREDICTORS,
                  x.length, x[0].length);
        }
        for (int i = 0; i < x.length; i++) {
            addObservation(x[i], y[i]);
        }
    }

    /**
     * The include method is where the QR decomposition occurs. This statement forms all
     * intermediate data which will be used for all derivative measures.
     * According to the miller paper, note that in the original implementation the x vector
     * is overwritten. In this implementation, the include method is passed a copy of the
     * original data vector so that there is no contamination of the data. Additionally,
     * this method differs slightly from Gentleman's method, in that the assumption is
     * of dense design matrices, there is some advantage in using the original gentleman algorithm
     * on sparse matrices.
     *
     * @param x observations on the regressors
     * @param wi weight of the this observation (-1,1)
     * @param yi observation on the regressand
     */
    private void include(final double[] x, final double wi, final double yi) {
        int nextr = 0;
        double w = wi;
        double y = yi;
        double xi;
        double di;
        double wxi;
        double dpi;
        double xk;
        double _w;
        this.rss_set = false;
        sumy = smartAdd(yi, sumy);
        sumsqy = smartAdd(sumsqy, yi * yi);
        for (int i = 0; i < x.length; i++) {
            if (w == 0.0) {
                return;
            }
            xi = x[i];

            if (xi == 0.0) {
                nextr += nvars - i - 1;
                continue;
            }
            di = d[i];
            wxi = w * xi;
            _w = w;
            if (di != 0.0) {
                dpi = smartAdd(di, wxi * xi);
                final double tmp = wxi * xi / di;
                if (FastMath.abs(tmp) > Precision.EPSILON) {
                    w = (di * w) / dpi;
                }
            } else {
                dpi = wxi * xi;
                w = 0.0;
            }
            d[i] = dpi;
            for (int k = i + 1; k < nvars; k++) {
                xk = x[k];
                x[k] = smartAdd(xk, -xi * r[nextr]);
                if (di != 0.0) {
                    r[nextr] = smartAdd(di * r[nextr], (_w * xi) * xk) / dpi;
                } else {
                    r[nextr] = xk / xi;
                }
                ++nextr;
            }
            xk = y;
            y = smartAdd(xk, -xi * rhs[i]);
            if (di != 0.0) {
                rhs[i] = smartAdd(di * rhs[i], wxi * xk) / dpi;
            } else {
                rhs[i] = xk / xi;
            }
        }
        sserr = smartAdd(sserr, w * y * y);
    }

    /**
     * Adds to number a and b such that the contamination due to
     * numerical smallness of one addend does not corrupt the sum.
     * @param a - an addend
     * @param b - an addend
     * @return the sum of the a and b
     */
    private double smartAdd(double a, double b) {
        final double _a = FastMath.abs(a);
        final double _b = FastMath.abs(b);
        if (_a > _b) {
            final double eps = _a * Precision.EPSILON;
            if (_b > eps) {
                return a + b;
            }
            return a;
        } else {
            final double eps = _b * Precision.EPSILON;
            if (_a > eps) {
                return a + b;
            }
            return b;
        }
    }

    /**
     * As the name suggests,  clear wipes the internals and reorders everything in the
     * canonical order.
     */
    @Override
    public void clear() {
        Arrays.fill(this.d, 0.0);
        Arrays.fill(this.rhs, 0.0);
        Arrays.fill(this.r, 0.0);
        Arrays.fill(this.tol, 0.0);
        Arrays.fill(this.rss, 0.0);
        Arrays.fill(this.work_tolset, 0.0);
        Arrays.fill(this.work_sing, 0.0);
        Arrays.fill(this.x_sing, 0.0);
        Arrays.fill(this.lindep, false);
        for (int i = 0; i < nvars; i++) {
            this.vorder[i] = i;
        }
        this.nobs = 0;
        this.sserr = 0.0;
        this.sumy = 0.0;
        this.sumsqy = 0.0;
        this.rss_set = false;
        this.tol_set = false;
    }

    /**
     * This sets up tolerances for singularity testing.
     */
    private void tolset() {
        int pos;
        double total;
        final double eps = this.epsilon;
        for (int i = 0; i < nvars; i++) {
            this.work_tolset[i] = FastMath.sqrt(d[i]);
        }
        tol[0] = eps * this.work_tolset[0];
        for (int col = 1; col < nvars; col++) {
            pos = col - 1;
            total = work_tolset[col];
            for (int row = 0; row < col; row++) {
                total += FastMath.abs(r[pos]) * work_tolset[row];
                pos += nvars - row - 2;
            }
            tol[col] = eps * total;
        }
        tol_set = true;
    }

    /**
     * The regcf method conducts the linear regression and extracts the
     * parameter vector. Notice that the algorithm can do subset regression
     * with no alteration.
     *
     * @param nreq how many of the regressors to include (either in canonical
     * order, or in the current reordered state)
     * @return an array with the estimated slope coefficients
     * @throws MathIllegalArgumentException if {@code nreq} is less than 1
     * or greater than the number of independent variables
     */
    private double[] regcf(int nreq) throws MathIllegalArgumentException {
        int nextr;
        if (nreq < 1) {
            throw new MathIllegalArgumentException(LocalizedStatFormats.NO_REGRESSORS);
        }
        if (nreq > this.nvars) {
            throw new MathIllegalArgumentException(
                    LocalizedStatFormats.TOO_MANY_REGRESSORS, nreq, this.nvars);
        }
        if (!this.tol_set) {
            tolset();
        }
        final double[] ret = new double[nreq];
        boolean rankProblem = false;
        for (int i = nreq - 1; i > -1; i--) {
            if (FastMath.sqrt(d[i]) < tol[i]) {
                ret[i] = 0.0;
                d[i] = 0.0;
                rankProblem = true;
            } else {
                ret[i] = rhs[i];
                nextr = i * (nvars + nvars - i - 1) / 2;
                for (int j = i + 1; j < nreq; j++) {
                    ret[i] = smartAdd(ret[i], -r[nextr] * ret[j]);
                    ++nextr;
                }
            }
        }
        if (rankProblem) {
            for (int i = 0; i < nreq; i++) {
                if (this.lindep[i]) {
                    ret[i] = Double.NaN;
                }
            }
        }
        return ret;
    }

    /**
     * The method which checks for singularities and then eliminates the offending
     * columns.
     */
    private void singcheck() {
        int pos;
        for (int i = 0; i < nvars; i++) {
            work_sing[i] = FastMath.sqrt(d[i]);
        }
        for (int col = 0; col < nvars; col++) {
            // Set elements within R to zero if they are less than tol(col) in
            // absolute value after being scaled by the square root of their row
            // multiplier
            final double temp = tol[col];
            pos = col - 1;
            for (int row = 0; row < col - 1; row++) {
                if (FastMath.abs(r[pos]) * work_sing[row] < temp) {
                    r[pos] = 0.0;
                }
                pos += nvars - row - 2;
            }
            // If diagonal element is near zero, set it to zero, set appropriate
            // element of LINDEP, and use INCLUD to augment the projections in
            // the lower rows of the orthogonalization.
            lindep[col] = false;
            if (work_sing[col] < temp) {
                lindep[col] = true;
                if (col < nvars - 1) {
                    Arrays.fill(x_sing, 0.0);
                    int _pi = col * (nvars + nvars - col - 1) / 2;
                    for (int _xi = col + 1; _xi < nvars; _xi++, _pi++) {
                        x_sing[_xi] = r[_pi];
                        r[_pi] = 0.0;
                    }
                    final double y = rhs[col];
                    final double weight = d[col];
                    d[col] = 0.0;
                    rhs[col] = 0.0;
                    this.include(x_sing, weight, y);
                } else {
                    sserr += d[col] * rhs[col] * rhs[col];
                }
            }
        }
    }

    /**
     * Calculates the sum of squared errors for the full regression
     * and all subsets in the following manner: <pre>
     * rss[] ={
     * ResidualSumOfSquares_allNvars,
     * ResidualSumOfSquares_FirstNvars-1,
     * ResidualSumOfSquares_FirstNvars-2,
     * ..., ResidualSumOfSquares_FirstVariable} </pre>
     */
    private void ss() {
        double total = sserr;
        rss[nvars - 1] = sserr;
        for (int i = nvars - 1; i > 0; i--) {
            total += d[i] * rhs[i] * rhs[i];
            rss[i - 1] = total;
        }
        rss_set = true;
    }

    /**
     * Calculates the cov matrix assuming only the first nreq variables are
     * included in the calculation. The returned array contains a symmetric
     * matrix stored in lower triangular form. The matrix will have
     * ( nreq + 1 ) * nreq / 2 elements. For illustration <pre>
     * cov =
     * {
     *  cov_00,
     *  cov_10, cov_11,
     *  cov_20, cov_21, cov22,
     *  ...
     * } </pre>
     *
     * @param nreq how many of the regressors to include (either in canonical
     * order, or in the current reordered state)
     * @return an array with the variance covariance of the included
     * regressors in lower triangular form
     */
    private double[] cov(int nreq) {
        if (this.nobs <= nreq) {
            return null; // NOPMD
        }
        double rnk = 0.0;
        for (int i = 0; i < nreq; i++) {
            if (!this.lindep[i]) {
                rnk += 1.0;
            }
        }
        final double var = rss[nreq - 1] / (nobs - rnk);
        final double[] rinv = new double[nreq * (nreq - 1) / 2];
        inverse(rinv, nreq);
        final double[] covmat = new double[nreq * (nreq + 1) / 2];
        Arrays.fill(covmat, Double.NaN);
        int pos2;
        int pos1;
        int start = 0;
        for (int row = 0; row < nreq; row++) {
            pos2 = start;
            if (!this.lindep[row]) {
                for (int col = row; col < nreq; col++) {
                    if (!this.lindep[col]) {
                        pos1 = start + col - row;
                        double total;
                        if (row == col) {
                            total = 1.0 / d[col];
                        } else {
                            total = rinv[pos1 - 1] / d[col];
                        }
                        for (int k = col + 1; k < nreq; k++) {
                            if (!this.lindep[k]) {
                                total += rinv[pos1] * rinv[pos2] / d[k];
                            }
                            ++pos1;
                            ++pos2;
                        }
                        covmat[ (col + 1) * col / 2 + row] = total * var;
                    } else {
                        pos2 += nreq - col - 1;
                    }
                }
            }
            start += nreq - row - 1;
        }
        return covmat;
    }

    /**
     * This internal method calculates the inverse of the upper-triangular portion
     * of the R matrix.
     * @param rinv  the storage for the inverse of r
     * @param nreq how many of the regressors to include (either in canonical
     * order, or in the current reordered state)
     */
    private void inverse(double[] rinv, int nreq) {
        int pos = nreq * (nreq - 1) / 2 - 1;
        Arrays.fill(rinv, Double.NaN);
        for (int row = nreq - 1; row > 0; --row) {
            if (!this.lindep[row]) {
                final int start = (row - 1) * (nvars + nvars - row) / 2;
                for (int col = nreq; col > row; --col) {
                    int pos1 = start;
                    int pos2 = pos;
                    double total = 0.0;
                    for (int k = row; k < col - 1; k++) {
                        pos2 += nreq - k - 1;
                        if (!this.lindep[k]) {
                            total += -r[pos1] * rinv[pos2];
                        }
                        ++pos1;
                    }
                    rinv[pos] = total - r[pos1];
                    --pos;
                }
            } else {
                pos -= nreq - row;
            }
        }
    }

    /**
     * In the original algorithm only the partial correlations of the regressors
     * is returned to the user. In this implementation, we have <pre>
     * corr =
     * {
     *   corrxx - lower triangular
     *   corrxy - bottom row of the matrix
     * }
     * Replaces subroutines PCORR and COR of:
     * ALGORITHM AS274  APPL. STATIST. (1992) VOL.41, NO. 2 </pre>
     *
     * <p>Calculate partial correlations after the variables in rows
     * 1, 2, ..., IN have been forced into the regression.
     * If IN = 1, and the first row of R represents a constant in the
     * model, then the usual simple correlations are returned.</p>
     *
     * <p>If IN = 0, the value returned in array CORMAT for the correlation
     * of variables Xi &amp; Xj is:</p><pre>
     * sum ( Xi.Xj ) / Sqrt ( sum (Xi^2) . sum (Xj^2) )</pre>
     *
     * <p>On return, array CORMAT contains the upper triangle of the matrix of
     * partial correlations stored by rows, excluding the 1's on the diagonal.
     * e.g. if IN = 2, the consecutive elements returned are:
     * (3,4) (3,5) ... (3,ncol), (4,5) (4,6) ... (4,ncol), etc.
     * Array YCORR stores the partial correlations with the Y-variable
     * starting with YCORR(IN+1) = partial correlation with the variable in
     * position (IN+1). </p>
     *
     * @param in how many of the regressors to include (either in canonical
     * order, or in the current reordered state)
     * @return an array with the partial correlations of the remainder of
     * regressors with each other and the regressand, in lower triangular form
     */
    public double[] getPartialCorrelations(int in) {
        final double[] output = new double[(nvars - in + 1) * (nvars - in) / 2];
        int pos;
        int pos1;
        int pos2;
        final int rms_off = -in;
        final int wrk_off = -(in + 1);
        final double[] rms = new double[nvars - in];
        final double[] work = new double[nvars - in - 1];
        double sumxx;
        double sumxy;
        double sumyy;
        final int offXX = (nvars - in) * (nvars - in - 1) / 2;
        if (in < -1 || in >= nvars) {
            return null; // NOPMD
        }
        final int nvm = nvars - 1;
        final int base_pos = r.length - (nvm - in) * (nvm - in + 1) / 2;
        if (d[in] > 0.0) {
            rms[in + rms_off] = 1.0 / FastMath.sqrt(d[in]);
        }
        for (int col = in + 1; col < nvars; col++) {
            pos = base_pos + col - 1 - in;
            sumxx = d[col];
            for (int row = in; row < col; row++) {
                sumxx += d[row] * r[pos] * r[pos];
                pos += nvars - row - 2;
            }
            if (sumxx > 0.0) {
                rms[col + rms_off] = 1.0 / FastMath.sqrt(sumxx);
            } else {
                rms[col + rms_off] = 0.0;
            }
        }
        sumyy = sserr;
        for (int row = in; row < nvars; row++) {
            sumyy += d[row] * rhs[row] * rhs[row];
        }
        if (sumyy > 0.0) {
            sumyy = 1.0 / FastMath.sqrt(sumyy);
        }
        pos = 0;
        for (int col1 = in; col1 < nvars; col1++) {
            sumxy = 0.0;
            Arrays.fill(work, 0.0);
            pos1 = base_pos + col1 - in - 1;
            for (int row = in; row < col1; row++) {
                pos2 = pos1 + 1;
                for (int col2 = col1 + 1; col2 < nvars; col2++) {
                    work[col2 + wrk_off] += d[row] * r[pos1] * r[pos2];
                    pos2++;
                }
                sumxy += d[row] * r[pos1] * rhs[row];
                pos1 += nvars - row - 2;
            }
            pos2 = pos1 + 1;
            for (int col2 = col1 + 1; col2 < nvars; col2++) {
                work[col2 + wrk_off] += d[col1] * r[pos2];
                ++pos2;
                output[ (col2 - 1 - in) * (col2 - in) / 2 + col1 - in] =
                        work[col2 + wrk_off] * rms[col1 + rms_off] * rms[col2 + rms_off];
                ++pos;
            }
            sumxy += d[col1] * rhs[col1];
            output[col1 + rms_off + offXX] = sumxy * rms[col1 + rms_off] * sumyy;
        }

        return output;
    }

    /**
     * ALGORITHM AS274 APPL. STATIST. (1992) VOL.41, NO. 2.
     * Move variable from position FROM to position TO in an
     * orthogonal reduction produced by AS75.1.
     *
     * @param from initial position
     * @param to destination
     */
    private void vmove(int from, int to) {
        double d1;
        double d2;
        double X;
        double d1new;
        double d2new;
        double cbar;
        double sbar;
        double Y;
        int first;
        int inc;
        int m1;
        int m2;
        int mp1;
        int pos;
        boolean bSkipTo40 = false;
        if (from == to) {
            return;
        }
        if (!this.rss_set) {
            ss();
        }
        final int count;
        if (from < to) {
            first = from;
            inc = 1;
            count = to - from;
        } else {
            first = from - 1;
            inc = -1;
            count = from - to;
        }

        int m = first;
        int idx = 0;
        while (idx < count) {
            m1 = m * (nvars + nvars - m - 1) / 2;
            m2 = m1 + nvars - m - 1;
            mp1 = m + 1;

            d1 = d[m];
            d2 = d[mp1];
            // Special cases.
            if (d1 > this.epsilon || d2 > this.epsilon) {
                X = r[m1];
                if (FastMath.abs(X) * FastMath.sqrt(d1) < tol[mp1]) {
                    X = 0.0;
                }
                if (d1 < this.epsilon || FastMath.abs(X) < this.epsilon) {
                    d[m] = d2;
                    d[mp1] = d1;
                    r[m1] = 0.0;
                    for (int col = m + 2; col < nvars; col++) {
                        ++m1;
                        X = r[m1];
                        r[m1] = r[m2];
                        r[m2] = X;
                        ++m2;
                    }
                    X = rhs[m];
                    rhs[m] = rhs[mp1];
                    rhs[mp1] = X;
                    bSkipTo40 = true;
                    //break;
                } else if (d2 < this.epsilon) {
                    d[m] = d1 * X * X;
                    r[m1] = 1.0 / X;
                    for (int _i = m1 + 1; _i < m1 + nvars - m - 1; _i++) {
                        r[_i] /= X;
                    }
                    rhs[m] /= X;
                    bSkipTo40 = true;
                    //break;
                }
                if (!bSkipTo40) {
                    d1new = d2 + d1 * X * X;
                    cbar = d2 / d1new;
                    sbar = X * d1 / d1new;
                    d2new = d1 * cbar;
                    d[m] = d1new;
                    d[mp1] = d2new;
                    r[m1] = sbar;
                    for (int col = m + 2; col < nvars; col++) {
                        ++m1;
                        Y = r[m1];
                        r[m1] = cbar * r[m2] + sbar * Y;
                        r[m2] = Y - X * r[m2];
                        ++m2;
                    }
                    Y = rhs[m];
                    rhs[m] = cbar * rhs[mp1] + sbar * Y;
                    rhs[mp1] = Y - X * rhs[mp1];
                }
            }
            if (m > 0) {
                pos = m;
                for (int row = 0; row < m; row++) {
                    X = r[pos];
                    r[pos] = r[pos - 1];
                    r[pos - 1] = X;
                    pos += nvars - row - 2;
                }
            }
            // Adjust variable order (VORDER), the tolerances (TOL) and
            // the vector of residual sums of squares (RSS).
            m1 = vorder[m];
            vorder[m] = vorder[mp1];
            vorder[mp1] = m1;
            X = tol[m];
            tol[m] = tol[mp1];
            tol[mp1] = X;
            rss[m] = rss[mp1] + d[mp1] * rhs[mp1] * rhs[mp1];

            m += inc;
            ++idx;
        }
    }

    /**
     * ALGORITHM AS274  APPL. STATIST. (1992) VOL.41, NO. 2
     *
     * <p> Re-order the variables in an orthogonal reduction produced by
     * AS75.1 so that the N variables in LIST start at position POS1,
     * though will not necessarily be in the same order as in LIST.
     * Any variables in VORDER before position POS1 are not moved.
     * Auxiliary routine called: VMOVE. </p>
     *
     * <p>This internal method reorders the regressors.</p>
     *
     * @param list the regressors to move
     * @param pos1 where the list will be placed
     * @return -1 error, 0 everything ok
     */
    private int reorderRegressors(int[] list, int pos1) {
        int next;
        int i;
        int l;
        if (list.length < 1 || list.length > nvars + 1 - pos1) {
            return -1;
        }
        next = pos1;
        i = pos1;
        while (i < nvars) {
            l = vorder[i];
            for (int j = 0; j < list.length; j++) {
                if (l == list[j] && i > next) {
                    this.vmove(i, next);
                    ++next;
                    if (next >= list.length + pos1) {
                        return 0;
                    } else {
                        break;
                    }
                }
            }
            ++i;
        }
        return 0;
    }

    /**
     * Gets the diagonal of the Hat matrix also known as the leverage matrix.
     *
     * @param  row_data returns the diagonal of the hat matrix for this observation
     * @return the diagonal element of the hatmatrix
     */
    public double getDiagonalOfHatMatrix(double[] row_data) {
        double[] wk = new double[this.nvars];
        int pos;
        double total;

        if (row_data.length > nvars) {
            return Double.NaN;
        }
        double[] xrow;
        if (this.hasIntercept) {
            xrow = new double[row_data.length + 1];
            xrow[0] = 1.0;
            System.arraycopy(row_data, 0, xrow, 1, row_data.length);
        } else {
            xrow = row_data;
        }
        double hii = 0.0;
        for (int col = 0; col < xrow.length; col++) {
            if (FastMath.sqrt(d[col]) < tol[col]) {
                wk[col] = 0.0;
            } else {
                pos = col - 1;
                total = xrow[col];
                for (int row = 0; row < col; row++) {
                    total = smartAdd(total, -wk[row] * r[pos]);
                    pos += nvars - row - 2;
                }
                wk[col] = total;
                hii = smartAdd(hii, (total * total) / d[col]);
            }
        }
        return hii;
    }

    /**
     * Gets the order of the regressors, useful if some type of reordering
     * has been called. Calling regress with int[]{} args will trigger
     * a reordering.
     *
     * @return int[] with the current order of the regressors
     */
    public int[] getOrderOfRegressors(){
        return vorder.clone();
    }

    /**
     * Conducts a regression on the data in the model, using all regressors.
     *
     * @return RegressionResults the structure holding all regression results
     * @exception  MathIllegalArgumentException - thrown if number of observations is
     * less than the number of variables
     */
    @Override
    public RegressionResults regress() throws MathIllegalArgumentException {
        return regress(this.nvars);
    }

    /**
     * Conducts a regression on the data in the model, using a subset of regressors.
     *
     * @param numberOfRegressors many of the regressors to include (either in canonical
     * order, or in the current reordered state)
     * @return RegressionResults the structure holding all regression results
     * @exception  MathIllegalArgumentException - thrown if number of observations is
     * less than the number of variables or number of regressors requested
     * is greater than the regressors in the model
     */
    public RegressionResults regress(int numberOfRegressors) throws MathIllegalArgumentException {
        if (this.nobs <= numberOfRegressors) {
           throw new MathIllegalArgumentException(
                   LocalizedStatFormats.NOT_ENOUGH_DATA_FOR_NUMBER_OF_PREDICTORS,
                   this.nobs, numberOfRegressors);
        }
        if( numberOfRegressors > this.nvars ){
            throw new MathIllegalArgumentException(
                    LocalizedStatFormats.TOO_MANY_REGRESSORS, numberOfRegressors, this.nvars);
        }

        tolset();
        singcheck();

        double[] beta = this.regcf(numberOfRegressors);

        ss();

        double[] cov = this.cov(numberOfRegressors);

        int rnk = 0;
        for (int i = 0; i < this.lindep.length; i++) {
            if (!this.lindep[i]) {
                ++rnk;
            }
        }

        boolean needsReorder = false;
        for (int i = 0; i < numberOfRegressors; i++) {
            if (this.vorder[i] != i) {
                needsReorder = true;
                break;
            }
        }
        if (!needsReorder) {
            return new RegressionResults(
                    beta, new double[][]{cov}, true, this.nobs, rnk,
                    this.sumy, this.sumsqy, this.sserr, this.hasIntercept, false);
        } else {
            double[] betaNew = new double[beta.length];
            double[] covNew = new double[cov.length];

            int[] newIndices = new int[beta.length];
            for (int i = 0; i < nvars; i++) {
                for (int j = 0; j < numberOfRegressors; j++) {
                    if (this.vorder[j] == i) {
                        betaNew[i] = beta[ j];
                        newIndices[i] = j;
                    }
                }
            }

            int idx1 = 0;
            int idx2;
            int _i;
            int _j;
            for (int i = 0; i < beta.length; i++) {
                _i = newIndices[i];
                for (int j = 0; j <= i; j++, idx1++) {
                    _j = newIndices[j];
                    if (_i > _j) {
                        idx2 = _i * (_i + 1) / 2 + _j;
                    } else {
                        idx2 = _j * (_j + 1) / 2 + _i;
                    }
                    covNew[idx1] = cov[idx2];
                }
            }
            return new RegressionResults(
                    betaNew, new double[][]{covNew}, true, this.nobs, rnk,
                    this.sumy, this.sumsqy, this.sserr, this.hasIntercept, false);
        }
    }

    /**
     * Conducts a regression on the data in the model, using regressors in array
     * Calling this method will change the internal order of the regressors
     * and care is required in interpreting the hatmatrix.
     *
     * @param  variablesToInclude array of variables to include in regression
     * @return RegressionResults the structure holding all regression results
     * @exception  MathIllegalArgumentException - thrown if number of observations is
     * less than the number of variables, the number of regressors requested
     * is greater than the regressors in the model or a regressor index in
     * regressor array does not exist
     */
    @Override
    public RegressionResults regress(int[] variablesToInclude) throws MathIllegalArgumentException {
        if (variablesToInclude.length > this.nvars) {
            throw new MathIllegalArgumentException(
                    LocalizedStatFormats.TOO_MANY_REGRESSORS, variablesToInclude.length, this.nvars);
        }
        if (this.nobs <= this.nvars) {
            throw new MathIllegalArgumentException(
                    LocalizedStatFormats.NOT_ENOUGH_DATA_FOR_NUMBER_OF_PREDICTORS,
                    this.nobs, this.nvars);
        }
        Arrays.sort(variablesToInclude);
        int iExclude = 0;
        for (int i = 0; i < variablesToInclude.length; i++) {
            if (i >= this.nvars) {
                throw new MathIllegalArgumentException(
                        LocalizedCoreFormats.INDEX_LARGER_THAN_MAX, i, this.nvars);
            }
            if (i > 0 && variablesToInclude[i] == variablesToInclude[i - 1]) {
                variablesToInclude[i] = -1;
                ++iExclude;
            }
        }
        int[] series;
        if (iExclude > 0) {
            int j = 0;
            series = new int[variablesToInclude.length - iExclude];
            for (int i = 0; i < variablesToInclude.length; i++) {
                if (variablesToInclude[i] > -1) {
                    series[j] = variablesToInclude[i];
                    ++j;
                }
            }
        } else {
            series = variablesToInclude;
        }

        reorderRegressors(series, 0);
        tolset();
        singcheck();

        double[] beta = this.regcf(series.length);

        ss();

        double[] cov = this.cov(series.length);

        int rnk = 0;
        for (int i = 0; i < this.lindep.length; i++) {
            if (!this.lindep[i]) {
                ++rnk;
            }
        }

        boolean needsReorder = false;
        for (int i = 0; i < this.nvars; i++) {
            if (this.vorder[i] != series[i]) {
                needsReorder = true;
                break;
            }
        }
        if (!needsReorder) {
            return new RegressionResults(
                    beta, new double[][]{cov}, true, this.nobs, rnk,
                    this.sumy, this.sumsqy, this.sserr, this.hasIntercept, false);
        } else {
            double[] betaNew = new double[beta.length];
            int[] newIndices = new int[beta.length];
            for (int i = 0; i < series.length; i++) {
                for (int j = 0; j < this.vorder.length; j++) {
                    if (this.vorder[j] == series[i]) {
                        betaNew[i] = beta[ j];
                        newIndices[i] = j;
                    }
                }
            }
            double[] covNew = new double[cov.length];
            int idx1 = 0;
            int idx2;
            int _i;
            int _j;
            for (int i = 0; i < beta.length; i++) {
                _i = newIndices[i];
                for (int j = 0; j <= i; j++, idx1++) {
                    _j = newIndices[j];
                    if (_i > _j) {
                        idx2 = _i * (_i + 1) / 2 + _j;
                    } else {
                        idx2 = _j * (_j + 1) / 2 + _i;
                    }
                    covNew[idx1] = cov[idx2];
                }
            }
            return new RegressionResults(
                    betaNew, new double[][]{covNew}, true, this.nobs, rnk,
                    this.sumy, this.sumsqy, this.sserr, this.hasIntercept, false);
        }
    }
}