real sqrtx2y2(const real& x, const real& y) throw()
// calculating sqrt(x^2 + y^2) in high accuracy. Blomquist 01.12.02
{
real a,b,r;
dotprecision dot;
int exa,exb,ex;
a = x; b = y;
exa = expo(a); exb = expo(b);
if (exb > exa)
{ // Permutation of a,b:
r = a; a = b; b = r;
ex = exa; exa = exb; exb = ex;
} // |a| >= |b|
// a = |a| >= |b| > 0:
ex = 511 - exa; // scaling a with 2^ex --> expo(a) == 511, --> a*a and
times2pown(a,ex); // b*b without overflow. An underflow of b*b will not
times2pown(b,ex); // affect the accuracy of the result!
dot = 0;
accumulate(dot,a,a);
accumulate(dot,b,b);
r = rnd(dot);
r = sqrt(r); // sqrt(...) declared in rmath.hpp
times2pown(r,-ex); // back-scaling
return r;
} // sqrtx2y2
void sqr2uv(const real& x, real& u, real& v)
// Liefert u,v für: x2 = u + v; EXAKTE Darstellung, falls kein overflow
// auftritt und v im normalisierten Bereich liegt. u > |v|
// Vorsicht: Funktioniert zunächst nur auf INTEL-Systemen!!!
{
real a,b,t,y1,y2;
a = Cut26(x);
b = x-a; // x = a+b;
u = x*x;
t = u - a*a; // exakte Auswertung!
y2 = a*b;
times2pown(y2,1); // y2 = 2*a*b, exakt!
t -= y2;
// Jetzt fehlt noch: t2 - b*b, aber b*b wird nicht immer korrekt berechnet,
// daher nochmalige Aufspaltung von b in y1+y2!!
y1 = Cut25(b);
y2 = b - y1; // b = y1+y2, exakt;
t -= y1*y1;
if (sign(y2)!=0)
{
a = y1*y2;
times2pown(a,1); // a = 2*y1*y2, exakt!
t -= a;
t -= y2*y2;
}
v = -t;
} // sqr2uv
real sqrt1mx2(const real& x) throw(STD_FKT_OUT_OF_DEF)
// sqrt(1-x2); rel. Fehlerschranke: eps = 3.700747E-16 = e(f)
// Blomquist, 19.06.04;
{
real t=x,res;
int ex;
if (sign(t)<0) t = -t; // Argument t >=0;
if (t>1) cxscthrow(STD_FKT_OUT_OF_DEF("real sqrt1mx2(const real&)"));
// For argument t now it holds: 0 <= t <=1;
ex = expo(t);
if (ex<=-26) res = 1; // t < 2^(-26) --> res = 1
else if (ex<=-15) { // t < 2^(-15) --> res = 1-x2/2
res = t*t;
times2pown(res,-1);
res = 1 - res;
} else {
if (ex>=0)
{ // ex>=0 --> t>=0.5;
res = 1-t; // res: delta = 1-t;
t = res * res;
times2pown(res,1); // res: 2*delta
res = res - t; // res: 1-x2 = 2*delta - delta2
} else res = 1-t*t; // res: Maschinenwert von 1-x2
res = sqrt(res); // res: Nullte Naeherung
}
return res;
} // sqrt1mx2
real expmx2(const real& x) throw()
// e^(-x^2); rel. Fehlerschranke: eps = 4.618919E-16 = e(f) gilt
// fuer alle |x| <= expmx2_x0 = 26.61571750925....
// Fuer |x| > expmx2_x0 --> expmx2(x) = 0;
// Blomquist, 05.07.04;
{
real t=x,u,v,res=0;
int ex;
if (t<0) t = -t; // t >= 0;
ex = expo(t);
if (ex<=-26) res = 1; // t < 2^(-26)
else if (ex<=-6) // t < 2^(-6)
{
u = t*t; v = u;
times2pown(v,-1); // v: 0.5*x2
res = 1-u*( 1-v*(1-u/3) );
} else if (t<=expmx2_x0) {
sqr2uv(x,u,v); // u:= x*x,v aus S(2,53); x2 = u+v (exakt!)
res = exp(-u);
if (v!=0) {
times2pown(res,500); // Die Skalierung verhindert, dass
v *= res; // v*exp(-u) in den denormalisierten Bereich faellt
res -= v;
times2pown(res,-500); // Rueckskalierung
}
}
return res;
} // expmx2
static void xscalbln(xcomplex *z, long int a)
{
cxsc::real re = Re(*z), im = Im(*z);
times2pown(re,a);
times2pown(im,a);
*z = cxsc::complex(re,im);
}
real expx2(const real& x)
// e^(+x^2); rel. Fehlerschranke: eps = 4.618958E-16 = e(f) gilt
// fuer alle |x| <= x0 = 26.64174755704632....
// x0 = 7498985273150791.0 / 281474976710656.0;
// Fuer |x| > x0 --> Programmabbruch
// Ausfuehrlich getestet; Blomquist, 26.07.06;
{
real t=x,u,v,res;
int ex;
if (t<0) t = -t; // t >= 0;
ex = expo(t);
if (ex<=-26) res = 1; // t < 2^(-26)
else
if (ex<=-6) // t < 2^(-6)
{
u = t*t; v = u;
times2pown(v,-1); // v: 0.5*x^2
res = 1+u*( 1+v*(1+u/3) );
}
else
{
sqr2uv(x,u,v); // u := x*x und v aus S(2,53);
// x^2 = u+v ist exakt!
res = exp(u);
v *= res; // v*exp(+u)
res += v;
}
return res;
} // expx2
real sqrtx2m1(const real& x)
// sqrt(x^2-1); rel. Fehlerschranke: eps = 2.221305E-16 = e(f)
// Blomquist, 13.04.04;
{ const real c1 = 1.000732421875,
c2 = 44000.0,
c3 = 1024.0; // c1,c2,c3 werden exakt gespeichert!
real res,ep,ep2,s1,s2,x1,x2,arg=x;
if (sign(arg)<0) arg = -arg; // arg = |x| >= 0
if (arg <= c1) { // x = 1+ep; x^2-1 = 2*ep + ep^2;
ep = x - 1; // Differenz rundungsfehlerfrei!
ep2 = ep*ep; // ep2 i.a. fehlerbehaftet!
times2pown(ep,1); // ep = 2*ep;
res = sqrt(ep+ep2); // res=y0: Startwert;
// x - y0^2 = (2*eps - s1^2) + [eps^2 - s2*(y0 + s1)]
s1 = Cut26(res); s2 = res - s1; // Startwert y0 = s1 + s2;
arg = ep - s1*s1; // arg = 2*eps - s1^2;
arg += (ep2 - s2*(res+s1)); // arg = x - y0^2
if (sign(arg)>0) {
arg = arg / res;
times2pown(arg,-1);
res += arg; // 1. Newton-Schritt beendet; eps = 2.221261E-16
}
} else if (arg<c2) {
// x-y0^2 = [(x1^2-1)-s1^2] + [x2*(x+x1)-s2*(y0+s1)]
x1 = Cut26(arg); x2 = arg - x1; // arg = x = x1 + x2;
ep2 = x2*(arg+x1); // ep2 ist fehlerbehaftet
x2 = x1*x1; ep = x2-1;
res = sqrt(ep+ep2); // res ist Startwert für Newton-Verfahren
s1 = Cut26(res); s2 = res - s1; // Startwert = s1 + s2;
ep2 = ep2 - s2 * (res+s1); // ep2 = [x2*(x+x1)-s2*(y0+s1)]
if (arg<c3) ep -= s1*s1; // ep = (x1^2-1) - s1^2;
else {
x2 -= s1*s1; // x2 = x1^2-s1^2
ep = x2 - 1; } // ep = (x1^2-s1^2) - 1;
ep += ep2; // ep = x - y0^2;
ep /= res;
times2pown(ep,-1);
res = res + ep; // 1. Newton-Schritt in hoher Genauigkeit
// beendet; eps = 2.221305E-16
} else { // arg = |x| >= 44000;
res = -1/arg;
times2pown(res,-1); // Multiplikation mit 0.5;
res += arg; // res = x - 1/(2*x); eps = 2.221114E-16
}
return res;
} // sqrtx2m1 (Punktargumente)
real sqrtp1m1(const real& x) throw()
// sqrtp1m1(x) = sqrt(x+1)-1;
// Blomquist, 05.08.03;
{
real y = x;
int ex = expo(x);
if (ex<=-50) times2pown(y,-1); // |x|<2^(-50); fast division by 2
else if (ex>=105) y = sqrt(x); // x >= 2^(+104) = 2.02824...e+31
else if (ex>=53) y = sqrt(x)-1; // x >= 2^(+52) = 4.50359...e+15
else if (x>-0.5234375 && x<=sqrtp1m1_s ) y = x / (sqrt(x+1) + 1);
else y = sqrt(x+1)-1;
return y;
}
real expx2m1(const real& x)
// e^(+x^2)-1; rel. Fehlerschranke: eps = 4.813220E-16 = e(f) gilt
// fuer alle x, mit: 2^(-511) <= x <= x0 = 26.64174755704632....
// x0 = 7498985273150791.0 / 281474976710656.0;
// Fuer x > x0 --> Programmabbruch wegen Overflow;
// Fuer 0 < x < 2^(-511) --> Programmabbruch, denorm. Bereich!!
// Ausfuehrlich getestet; Blomquist, 10.08.2006;
{
real t(x),u,v,y,res(0);
int ex;
if (t<0) t = -t; // t >= 0;
if (t>=6.5) res = expx2(t);
else
{
ex = expo(t);
sqr2uv(x,u,v); // u := x*x und v aus S(2,53);
if (ex>=2) // g4(x)
{
y = exp(u);
res = 1 - v*y;
res = y - res;
}
else
if (ex>=-8) res = expm1(u) + v*exp(u); // g3(x)
else
if (ex>=-25) { // g2(x)
y = u*u;
times2pown(y,-1);
res = (1+u/3)*y + u;
}
else
if(ex>=-510) res = u; // g1(x)
else if (ex>=-1073)
{
std::cerr << "expx2m1: denormalized range!"
<< std::endl; exit(1);
}
}
return res;
} // expx2m1
real ln_sqrtx2y2(const real& x, const real& y)
throw(STD_FKT_OUT_OF_DEF)
// ln( sqrt(x^2+y^2) ) == 0.5*ln(x^2+y^2); Blomquist, 21.11.03;
// Relative error bound: 5.160563E-016;
// Absolute error bound: 2.225075E-308; if x=1 and 0<=y<=b0;
{
int j,N;
real a,b,r,r1;
dotprecision dot;
a = sign(x)<0 ? -x : x; // a = |x| >= 0;
b = sign(y)<0 ? -y : y; // b = |y| >= 0;
int exa=expo(a), exb=expo(b), ex;
if (b > a)
{
r = a; a = b; b = r;
ex = exa; exa = exb; exb = ex;
}
// It holds now: 0 <= b <= a
if (sign(a)==0)
cxscthrow(STD_FKT_OUT_OF_DEF
("real ln_sqrtx2y2(const real&, const real&)"));
if (exa>20) // to avoid overflow by calculating a^2 + b^2
{ // a>=2^(20):
j = Interval_Nr(B_lnx2y2_1,21,exa); // j: No. of subinterval
N = B_lnx2y2_N1[j]; // N: Optimal int value
if (exb-exa > -25)
{ // For (exb-exa>-25) we use the complete term:
// N*ln(2) + [ln(2^(-N)*a)+0.5*ln(1+(b/a)^2)]
b = b/a; // a > 0
b = lnp1(b*b);
times2pown(b,-1); // exact division by 2
times2pown(a,-N);
r = b + ln(a); // [ ... ] calculated!
r += B_lnx2y2_c1[j];
}
else { // For (exb-exa<=-25) only two summands!:
times2pown(a,-N);
r = ln(a) + B_lnx2y2_c1[j];
}
}
else // exa<=20 or a<2^(20):
{ // Now calculation of a^2+b^2 without overflow:
if (exa<=-20) // to avoid underflow by calculating a^2+b^2
if (exa<=-1022) // a in the denormalized range
{
r = b/a;
r = lnp1(r*r); times2pown(r,-1); // r: 0.5*ln(1+..)
times2pown(a,1067);
r += ln(a); // [ .... ] ready
r -= ln2_1067; // rel. error = 2.459639e-16;
}
else // MinReal=2^(-1022) <= a < 2^(-20)
{ // Calculating the number j of the subinterval:
j = 20 - Interval_Nr(B_lnx2y2_2,21,exa);
r = a; times2pown(r,B_lnx2y2_N1[j]);
r = ln(r); // r: ln(2^N*a);
if (exb-exa > -25) { // calculating the complete term
b = b/a;
a = lnp1(b*b);
times2pown(a,-1);
r += a; // [ ... ] ready now
}
// We now have: exb-exa<=-25, ==> b/a <= 2^(-24);
r -= B_lnx2y2_c1[j]; // 0.5*ln(1+(b/a)^2) neglected!
// relative error = 4.524090e-16 in both cases;
}
else // calculation of a^2+b^2 without overflow or underflow:
{ // exa>-20 respective a>=2^(-20):
dot = 0;
accumulate(dot,a,a);
accumulate(dot,b,b); // dot = a^2+b^2, exact!
real s = rnd(dot); // s = a^2 + b^2, rounded!
if (s>=0.25 && s<=1.75)
if (s>=0.828125 && s<=1.171875)
{ // Series:
if (a==1 && exb<=-28)
{
r = b; times2pown(r,-1);
r *= b;
}
else {
dot -= 1;
r = rnd(dot); // r = a^2+b^2-1 rounded!
r = lnp1(r);
times2pown(r,-1);
}
}
else { // Reading dot = a^2+b^2 twice:
r = rnd(dot);
dot -= r;
r1 = rnd(dot); // a^2+b^2 = r+r1, rounded!
r1 = lnp1(r1/r);
r = ln(r) + r1;
times2pown(r,-1); // exact division by 2
}
else { // calculating straight from: 0.5*ln(x^2+y^2)
r = ln(s);
times2pown(r,-1);
}
}
}
return r;
} // ln_sqrtx2y2
