Lau2DCubicSpline.cc

Go to the documentation of this file.

// Copyright University of Warwick 2004 - 2013.
// Distributed under the Boost Software License, Version 1.0.
// (See accompanying file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)

// Authors:
// Thomas Latham
// John Back
// Paul Harrison

#include <cmath>
#include <cstdlib>
#include <iostream>

#include <TH2.h>
#include <TSystem.h>

#include "Lau2DCubicSpline.hh"

Lau2DCubicSpline::~Lau2DCubicSpline()
{ }

inline Double_t Lau2DCubicSpline::histcont(
        const TH2& h, Int_t xbin, Int_t ybin) const
{
        //reflect until we're in range
        while(xbin < 0 || xbin >= nBinsX - 1) {
                if(xbin < 0) xbin = -xbin-1;
                if(xbin >= nBinsX -1) xbin = 2*(nBinsX-1) - xbin - 1;
        }
        while(ybin < 0 || ybin >= nBinsY - 1) {
                if(ybin < 0) ybin = -ybin-1;
                if(ybin >= nBinsY -1) ybin = 2*(nBinsY-1) - ybin - 1;
        }

        return h.GetBinContent(1 + xbin, 1 + ybin);
}

inline Double_t Lau2DCubicSpline::dhistdx(
        const TH2& h, Int_t xbin, Int_t ybin) const
{
    return 0.5 * (histcont(h, xbin + 1, ybin) -
            histcont(h, xbin - 1, ybin));
}

inline Double_t Lau2DCubicSpline::dhistdy(
        const TH2& h, Int_t xbin, Int_t ybin) const
{
    return 0.5 * (histcont(h, xbin, ybin + 1) -
            histcont(h, xbin, ybin - 1));
}

inline Double_t Lau2DCubicSpline::d2histdxdy(
        const TH2& h, Int_t xbin, Int_t ybin) const
{
    return 0.5 * (histcont(h, xbin - 1, ybin - 1) -
            histcont(h, xbin + 1, ybin - 1) +
            histcont(h, xbin + 1, ybin + 1) -
            histcont(h, xbin - 1, ybin + 1));
}

Lau2DCubicSpline::Lau2DCubicSpline(const TH2& h) :
    nBinsX(1 + h.GetNbinsX()), nBinsY(1 + h.GetNbinsY()),
    binSizeX(h.GetXaxis()->GetBinWidth(1)),
    binSizeY(h.GetYaxis()->GetBinWidth(1)),
    xmin(h.GetXaxis()->GetBinCenter(1) - binSizeX),
    xmax(h.GetXaxis()->GetBinCenter(nBinsX - 1) + binSizeX),
    ymin(h.GetYaxis()->GetBinCenter(1) - binSizeY),
    ymax(h.GetYaxis()->GetBinCenter(nBinsY - 1) + binSizeY),
    coeffs(CoeffRecLen * nBinsX * nBinsY)
{
    TAxis *xaxis = h.GetXaxis(), *yaxis = h.GetYaxis();
    // verify that all bins have same size
    for (Int_t i = 1; i < nBinsX; ++i) {
        if (std::abs(xaxis->GetBinWidth(i) / binSizeX - 1.) > 1e-9) {
            std::cerr << "ERROR in Lau2DCubicSpline constructor : the histogram has variable bin sizes." << std::endl;
            gSystem->Exit(EXIT_FAILURE);
        }
    }
    for (Int_t i = 1; i < nBinsY; ++i) {
        if (std::abs(yaxis->GetBinWidth(i) / binSizeY - 1.) > 1e-9) {
            std::cerr << "ERROR in Lau2DCubicSpline constructor : the histogram has variable bin sizes." << std::endl;
            gSystem->Exit(EXIT_FAILURE);
        }
    }
    // ok, go through histogram to precalculate the interpolation coefficients
    // in rectangles between bin centres
    //
    // for that purpose, we map each of those rectangles to the unit square
    for (Int_t j = -1; j < nBinsY - 1; ++j) {
        for (Int_t i = -1; i < nBinsX - 1; ++i) {
            const Double_t rhs[NCoeff] = {
                // function values in bin centres
                histcont(h, i, j),
                histcont(h, i + 1, j),
                histcont(h, i, j + 1),
                histcont(h, i + 1, j + 1),
                // df/dx in bin centres (finite difference approximation)
                dhistdx(h, i, j),
                dhistdx(h, i + 1, j),
                dhistdx(h, i, j + 1),
                dhistdx(h, i + 1, j + 1),
                // df/dy in bin centres (finite difference approximation)
                dhistdy(h, i, j),
                dhistdy(h, i + 1, j),
                dhistdy(h, i, j + 1),
                dhistdy(h, i + 1, j + 1),
                // d^2f/dxdy in bin centres (finite difference approximation)
                d2histdxdy(h, i, j),
                d2histdxdy(h, i + 1, j),
                d2histdxdy(h, i, j + 1),
                d2histdxdy(h, i + 1, j + 1)
            };
            // work out solution - strange array placement is due to the fact
            // that terms with x/y to high powers can be small, so they should
            // be added up first during evaluation to avoid cancellation
            // issues; at the same time you want to access them in order to
            // not confuse the CPU cache, so they're stored back to front
            //
            // a_00 ... a_30
            coeff(1 + i, 1 + j, 15) = rhs[0];
            coeff(1 + i, 1 + j, 14) = rhs[4];
            coeff(1 + i, 1 + j, 13) =
                3. * (-rhs[0] + rhs[1]) - 2. * rhs[4] - rhs[5];
            coeff(1 + i, 1 + j, 12) =
                2. * (rhs[0] - rhs[1]) + rhs[4] + rhs[5];
            // a_31 ... a_31
            coeff(1 + i, 1 + j, 11) = rhs[8];
            coeff(1 + i, 1 + j, 10) = rhs[12];
            coeff(1 + i, 1 + j, 9) =
                3. * (-rhs[8] + rhs[9]) - 2. * rhs[12] - rhs[13];
            coeff(1 + i, 1 + j, 8) =
                2. * (rhs[8] - rhs[9]) + rhs[12] + rhs[13];
            // a_02 ... a_32
            coeff(1 + i, 1 + j, 7) =
                3. * (-rhs[0] + rhs[2]) - 2. * rhs[8] - rhs[10];
            coeff(1 + i, 1 + j, 6) =
                3. * (-rhs[4] + rhs[6]) - 2. * rhs[12] - rhs[14];
            coeff(1 + i, 1 + j, 5) =
                9. * (rhs[0] - rhs[1] - rhs[2] + rhs[3]) +
                6. * (rhs[4] - rhs[6] + rhs[8] - rhs[9]) + 4. * rhs[12] +
                3. * (rhs[5] - rhs[7] + rhs[10] - rhs[11]) +
                2. * (rhs[13] + rhs[14]) + rhs[15];
            coeff(1 + i, 1 + j, 4) =
                6. * (-rhs[0] + rhs[1] + rhs[2] - rhs[3]) +
                4. * (-rhs[8] + rhs[9]) +
                3. * (-rhs[4] - rhs[5] + rhs[6] + rhs[7]) +
                2. * (-rhs[10] + rhs[11] - rhs[12] - rhs[13]) -
                rhs[14] - rhs[15];
            // a_03 ... a_33
            coeff(1 + i, 1 + j, 3) =
                2. * (rhs[0] - rhs[2]) + rhs[8] + rhs[10];
            coeff(1 + i, 1 + j, 2) =
                2. * (rhs[4] - rhs[6]) + rhs[12] + rhs[14];
            coeff(1 + i, 1 + j, 1) =
                6. * (-rhs[0] + rhs[1] + rhs[2] - rhs[3]) +
                4. * (-rhs[4] + rhs[6]) +
                3. * (-rhs[8] + rhs[9] - rhs[10] + rhs[11]) +
                2. * (- rhs[5] + rhs[7] - rhs[12] - rhs[14]) -
                rhs[13] - rhs[15];
            coeff(1 + i, 1 + j, 0) =
                4. * (rhs[0] - rhs[1] - rhs[2] + rhs[3]) +
                2. * (rhs[4] + rhs[5] - rhs[6] - rhs[7] +
                        rhs[8] - rhs[9] + rhs[10] - rhs[11]) +
                rhs[12] + rhs[13] + rhs[14] + rhs[15];
            // coeff(1 + i, 1 + j, 17) contains integral of function over the
            // square in "unit square coordinates", i.e. neglecting bin widths
            // this is done to help speed up calculations of 2D integrals
            Double_t sum = 0.;
            for (Int_t k = 0; k < NCoeff; ++k)
                sum += coeff(1 + i, 1 + j, k) /
                    Double_t((4 - (k % 4)) * (4 - (k / 4)));
            coeff(1 + i, 1 + j, NCoeff) = sum;
        }
    }
}

Double_t Lau2DCubicSpline::evaluate(Double_t x, Double_t y) const
{
    // protect against NaN and out of range
    if (x <= xmin || x >= xmax || y <= ymin || y >= ymax || x != x || y != y)
        return 0.;
    // find the bin in question
    const Int_t binx = Int_t(Double_t(nBinsX) * (x - xmin) / (xmax - xmin));
    const Int_t biny = Int_t(Double_t(nBinsY) * (y - ymin) / (ymax - ymin));
    // get low edge of bin
    const Double_t xlo = Double_t(nBinsX - binx) / Double_t(nBinsX) * xmin +
        Double_t(binx) / Double_t(nBinsX) * xmax;
    const Double_t ylo = Double_t(nBinsY - biny) / Double_t(nBinsY) * ymin +
        Double_t(biny) / Double_t(nBinsY) * ymax;
    // normalise to coordinates in unit sqare
    const Double_t hx = (x - xlo) / binSizeX;
    const Double_t hy = (y - ylo) / binSizeY;
    // monomials
    const Double_t hxton[4] = { hx * hx * hx, hx * hx, hx, 1. };
    const Double_t hyton[4] = { hy * hy * hy, hy * hy, hy, 1. };
    // sum up
    Double_t retVal = 0.;
    for (Int_t k = 0; k < NCoeff; ++k)
        retVal += coeff(binx, biny, k) * hxton[k % 4] * hyton[k / 4];

    return retVal;
}

Double_t Lau2DCubicSpline::analyticalIntegral() const
{
    return evalXY(xmin,xmax,ymin,ymax);
}

Double_t Lau2DCubicSpline::analyticalIntegral(Double_t x1, Double_t x2, Double_t y1, Double_t y2) const
{
    if(y1==y2) return evalX(x1, x2, y1);
    if(x1==x2) return evalY(x1, y1, y2);
    return evalXY(x1, x2, y1, y2);
}

Double_t Lau2DCubicSpline::evalX(Double_t x1, Double_t x2, Double_t y) const
{
    // protect against NaN
    if (x1 != x1 || x2 != x2 || y != y) return 0.;
    // find the bin in question
    const Int_t biny = Int_t(Double_t(nBinsY) * (y - ymin) / (ymax - ymin));
    // get low edge of bin
    const Double_t ylo = Double_t(nBinsY - biny) / Double_t(nBinsY) * ymin +
        Double_t(biny) / Double_t(nBinsY) * ymax;
    // normalise to coordinates in unit sqare
    const Double_t hy = (y - ylo) / binSizeY;
    // monomials
    const Double_t hyton[4] = { hy * hy * hy, hy * hy, hy, 1. };
    // integral
    Double_t sum = 0.;
    for (Int_t binx = 0; binx < nBinsX; ++binx) {
        // get low/high edge of bin
        const Double_t xhi = Double_t(nBinsX - binx - 1) / Double_t(nBinsX) * xmin +
            Double_t(binx + 1) / Double_t(nBinsX) * xmax;
        if (xhi < x1) continue;
        const Double_t xlo = Double_t(nBinsX - binx) / Double_t(nBinsX) * xmin +
            Double_t(binx) / Double_t(nBinsX) * xmax;
        if (xlo > x2) break;
        // work out integration range
        const Double_t a = ((xlo > x1) ? 0. : (x1 - xlo)) / binSizeX;
        const Double_t b = ((xhi < x2) ? binSizeX : (x2 - xlo)) / binSizeX;
        // integrated monomials
        const Double_t hxton[4] = { 0.25 * (b * b * b * b - a * a * a * a),
            (b * b * b - a * a * a) / 3., 0.5 * (b * b - a * a), b - a };
        Double_t lsum = 0.;
        for (Int_t k = 0; k < NCoeff; ++k)
            lsum += coeff(binx, biny, k) * hxton[k % 4] * hyton[k / 4];
        sum += lsum;
    }
    // move from unit square coordinates to user coordinates
    return sum * binSizeX;
}

Double_t Lau2DCubicSpline::evalY(Double_t x, Double_t y1, Double_t y2) const
{
    // protect against NaN
    if (x != x || y1 != y1 || y2 != y2) return 0.;
    // find the bin in question
    const Int_t binx = Int_t(Double_t(nBinsX) * (x - xmin) / (xmax - xmin));
    // get low edge of bin
    const Double_t xlo = Double_t(nBinsX - binx) / Double_t(nBinsX) * xmin +
        Double_t(binx) / Double_t(nBinsX) * xmax;
    // normalise to coordinates in unit sqare
    const Double_t hx = (x - xlo) / binSizeX;
    // monomials
    const Double_t hxton[4] = { hx * hx * hx, hx * hx, hx, 1. };
    // integral
    Double_t sum = 0.;
    for (Int_t biny = 0; biny < nBinsY; ++biny) {
        // get low/high edge of bin
        const Double_t yhi = Double_t(nBinsY - biny - 1) / Double_t(nBinsY) * ymin +
            Double_t(biny + 1) / Double_t(nBinsY) * ymax;
        if (yhi < y1) continue;
        const Double_t ylo = Double_t(nBinsY - biny) / Double_t(nBinsY) * ymin +
            Double_t(biny) / Double_t(nBinsY) * ymax;
        if (ylo > y2) break;
        // work out integration range
        const Double_t a = ((ylo > y1) ? 0. : (y1 - ylo)) / binSizeY;
        const Double_t b = ((yhi < y2) ? binSizeY : (y2 - ylo)) / binSizeY;
        // integrated monomials
        const Double_t hyton[4] = { 0.25 * (b * b * b * b - a * a * a * a),
            (b * b * b - a * a * a) / 3., 0.5 * (b * b - a * a), b - a };
        Double_t lsum = 0.;
        for (Int_t k = 0; k < NCoeff; ++k)
            lsum += coeff(binx, biny, k) * hxton[k % 4] * hyton[k / 4];
        sum += lsum;
    }
    // move from unit square coordinates to user coordinates
    return sum * binSizeY;
}

Double_t Lau2DCubicSpline::evalXY(
        Double_t x1, Double_t x2, Double_t y1, Double_t y2) const
{
    // protect against NaN
    if (x1 != x1 || x2 != x2 || y1 != y1 || y2 != y2) return 0.;
    // integral
    Double_t sum = 0.;
    for (Int_t biny = 0; biny < nBinsY; ++biny) {
        // get low/high edge of bin
        const Double_t yhi = Double_t(nBinsY - biny - 1) / Double_t(nBinsY) * ymin +
            Double_t(biny + 1) / Double_t(nBinsY) * ymax;
        if (yhi < y1) continue;
        const Double_t ylo = Double_t(nBinsY - biny) / Double_t(nBinsY) * ymin +
            Double_t(biny) / Double_t(nBinsY) * ymax;
        if (ylo > y2) break;
        // work out integration range
        const Double_t ay = ((ylo > y1) ? 0. : (y1 - ylo)) / binSizeY;
        const Double_t by = ((yhi < y2) ? binSizeY : (y2 - ylo)) / binSizeY;
        const Bool_t yFullyContained = std::abs(by - ay - 1.0) < 1e-15;
        // integrated monomials
        const Double_t hyton[4] = {
            0.25 * (by * by * by * by - ay * ay * ay * ay),
            (by * by * by - ay * ay * ay) / 3., 0.5 * (by * by - ay * ay),
            by - ay };
        for (Int_t binx = 0; binx < nBinsX; ++binx) {
            // get low/high edge of bin
            const Double_t xhi = Double_t(nBinsX - binx - 1) / Double_t(nBinsX) * xmin +
                Double_t(binx + 1) / Double_t(nBinsX) * xmax;
            if (xhi < x1) continue;
            const Double_t xlo = Double_t(nBinsX - binx) / Double_t(nBinsX) * xmin +
                Double_t(binx) / Double_t(nBinsX) * xmax;
            if (xlo > x2) break;
            // work out integration range
            const Double_t ax = ((xlo > x1) ? 0. : (x1 - xlo)) / binSizeX;
            const Double_t bx = ((xhi < x2) ? binSizeX : (x2 - xlo)) / binSizeX;
            const Bool_t xFullyContained = std::abs(bx - ax - 1.0) < 1e-15;
            if (xFullyContained && yFullyContained) {
                // for fully contained bins, we have cached the integral
                sum += coeff(binx, biny, NCoeff);
                continue;
            }
            // integrated monomials
            const Double_t hxton[4] = {
                0.25 * (bx * bx * bx * bx - ax * ax * ax * ax),
                (bx * bx * bx - ax * ax * ax) / 3., 0.5 * (bx * bx - ax * ax),
                bx - ax };
            // integrate over bin

            Double_t lsum = 0.;
            for (Int_t k = 0; k < NCoeff; ++k)
                lsum += coeff(binx, biny, k) * hxton[k % 4] * hyton[k / 4];
            sum += lsum;
        }
    }
    // move from unit square coordinates to user coordinates
    return sum * binSizeX * binSizeY;
}