I am trying to calculate the numerical gradient of a smooth function in c++. And the parameter value could vary from zero to a very large number(maybe 1e10 to 1e20?)
I used the function f(x,y) = 10*x^3 + y^3 as a testbench, but I found that if x or y is too large, I can't get correct gradient.
Here is my code to calculate the graidient:
#include <iostream>
#include <cmath>
#include <cassert>
using namespace std;
double f(double x, double y)
{
// black box expensive function
return 10 * pow(x, 3) + pow(y, 3);
}
int main()
{
// double x = -5897182590.8347721;
// double y = 269857217.0017581;
double x = 1.13041e+19;
double y = -5.49756e+14;
const double epsi = 1e-4;
double f1 = f(x, y);
double f2 = f(x, y+epsi);
double f3 = f(x, y-epsi);
cout << f1 << endl;
cout << f2 << endl;
cout << f3 << endl;
cout << f1 - f2 << endl; // 0
cout << f2 - f3 << endl; // 0
return 0;
}
If I use the above code to calculate the gradient, the gradient would be zero!
The testbench function, 10*x^3 + y^3, is just a demo, the real problem I need to solve is actually a black box function.
So, is there any "standard" way to calculate the numerical gradient?
In the first place, you should use the central difference scheme, which is more accurate (by cancellation of one more term of the Taylor develoment).
(f(x + h) - f(x - h)) / 2h
rather than
(f(x + h) - f(x)) / h
Then the choice of h is critical and using a fixed constant is the worst thing you can do. Because for small x, h will be too large so that the approximation formula no more works, and for large x, h will be too small, resulting in severe truncation error.
A much better choice is to take a relative value, h = x√ε, where ε is the machine epsilon (1 ulp), which gives a good tradeoff.
(f(x(1 + √ε)) - f(x(1 - √ε))) / 2x√ε
Beware that when x = 0, a relative value cannot work and you need to fall back to a constant. But then, nothing tells you which to use !
You need to consider the precision needed.
At first glance, since |y| = 5.49756e14 and epsi = 1e-4, you need at least ⌈log2(5.49756e14)-log2(1e-4)⌉ = 63 bits of significand precision (that is the number of bits used to encode the digits of your number, also known as mantissa) for y and y+epsi to be considered different.
The double-precision floating-point format only has 53 bits of significand precision (assuming it is 8 bytes). So, currently, f1, f2 and f3 are exactly the same because y, y+epsi and y-epsi are equal.
Now, let's consider the limit : y = 1e20, and the result of your function, 10x^3 + y^3. Let's ignore x for now, so let's take f = y^3. Now we can calculate the precision needed for f(y) and f(y+epsi) to be different : f(y) = 1e60 and f(epsi) = 1e-12. This gives a minimum significand precision of ⌈log2(1e60)-log2(1e-12)⌉ = 240 bits.
Even if you were to use the long double type, assuming it is 16 bytes, your results would not differ : f1, f2 and f3 would still be equal, even though y and y+epsi would not.
If we take x into account, the maximum value of f would be 11e60 (with x = y = 1e20). So the upper limit on precision is ⌈log2(11e60)-log2(1e-12)⌉ = 243 bits, or at least 31 bytes.
One way to solve your problem is to use another type, maybe a bignum used as fixed-point.
Another way is to rethink your problem and deal with it differently. Ultimately, what you want is f1 - f2. You can try to decompose f(y+epsi). Again, if you ignore x, f(y+epsi) = (y+epsi)^3 = y^3 + 3*y^2*epsi + 3*y*epsi^2 + epsi^3. So f(y+epsi) - f(y) = 3*y^2*epsi + 3*y*epsi^2 + epsi^3.
The only way to calculate gradient is calculus.
Gradient is a vector:
g(x, y) = Df/Dx i + Df/Dy j
where (i, j) are unit vectors in x and y directions, respectively.
One way to approximate derivatives is first order differences:
Df/Dx ~ (f(x2, y)-f(x1, y))/(x2-x1)
and
Df/Dy ~ (f(x, y2)-f(x, y1))/(y2-y1)
That doesn't look like what you're doing.
You have a closed form expression:
g(x, y) = 30*x^2 i + 3*y^2 j
You can plug in values for (x, y) and calculate the gradient exactly at any point. Compare that to your differences and see how well your approximation is doing.
How you implement it numerically is your responsibility. (10^19)^3 = 10^57, right?
What is the size of double on your machine? Is it a 64 bit IEEE double precision floating point number?
Use
dx = (1+abs(x))*eps, dfdx = (f(x+dx,y) - f(x,y)) / dx
dy = (1+abs(y))*eps, dfdy = (f(x,y+dy) - f(x,y)) / dy
to get meaningful step sizes for large arguments.
Use eps = 1e-8 for one-sided difference formulas, eps = 1e-5 for central difference quotients.
Explore automatic differentiation (see autodiff.org) for derivatives without difference quotients and thus much smaller numerical errors.
We can examine the behaviour of the error in the derivative using the following program - it calculates the 1-sided derivative and the central difference based derivative using a varying step size. Here I'm using x and y ~ 10^10, which is smaller than what you were using, but should illustrate the same point.
#include <iostream>
#include <cmath>
#include <cassert>
using namespace std;
double f(double x, double y) {
return 10 * pow(x, 3) + pow(y, 3);
}
double f_x(double x, double y) {
return 3 * 10 * pow(x,2);
}
double f_y(double x, double y) {
return 3 * pow(y,2);
}
int main()
{
// double x = -5897182590.8347721;
// double y = 269857217.0017581;
double x = 1.13041e+10;
double y = -5.49756e+10;
//double x = 10.1;
//double y = -5.2;
double epsi = 1e8;
for(int i=0; i<60; ++i) {
double dfx_n = (f(x+epsi,y) - f(x,y))/epsi;
double dfx_cd = (f(x+epsi,y) - f(x-epsi,y))/(2*epsi);
double dfx = f_x(x,y);
cout<<epsi<<" "<<fabs(dfx-dfx_n)<<" "<<fabs(dfx - dfx_cd)<<std::endl;
epsi/=1.5;
}
return 0;
}
The output shows that a 1-sided difference gets us an optimal error of about 1.37034e+13 at a step length of about 100.0. Note that while this error looks large, as a relative error it is 3.5746632302764072e-09 (since the exact value is 3.833e+21)
In comparison the 2-sided difference gets an optimal error of about 1.89493e+10 with a step size of about 45109.3. This is three-orders of magnitude better, (with a much larger step-size).
How can we work out the step size? The link in the comments of Yves Daosts answer gives us a ballpark value:
h=x_c sqrt(eps) for 1-Sided, and h=x_c cbrt(eps) for 2-Sided.
But either way, if the required step size for decent accuracy at x ~ 10^10 is 100.0, the required step size with x ~ 10^20 is going to be 10^10 larger too. So the problem is simply that your step size is way too small.
This can be verified by increasing the starting step-size in the above code and resetting the x/y values to the original values.
Then expected derivative is O(1e39), best 1-sided error of about O(1e31) occurs near a step length of 5.9e10, best 2-sided error of about O(1e29) occurs near a step length of 6.1e13.
As numerical differentiation is ill conditioned (which means a small error could alter your result significantly) you should consider to use Cauchy's integral formula. This way you can calculate the n-th derivative with an integral. This will lead to less problems with considering accuracy and stability.
a * x = b
I have a seemingly rather complicated multiplication / imul problem: if I have a and I have b, how can I calculate x if they're all 32-bit dwords (e.g. 0-1 = FFFFFFFF, FFFFFFFF+1 = 0)?
For example:
0xcb9102df * x = 0x4d243a5d
In that case, x is 0x1908c643. I found a similar question but the premises were different and I'm hoping there's a simpler solution than those given.
Numbers have a modular multiplicative inverse modulo a power of two precisely iff they are odd. Everything else is a bit-shifted odd number (even zero, which might be anything, with all bits shifted out). So there are a couple of cases:
Given a * x = b
tzcnt(a) > tzcnt(b) no solution
tzcnt(a) <= tzcnt(b) solvable, with 2tzcnt(a) solutions
The second case has a special case with 1 solution, for odd a, namely x = inverse(a) * b
More generally, x = inverse(a >> tzcnt(a)) * (b >> tzcnt(a)) is a solution, because you write a as (a >> tzcnt(a)) * (1 << tzcnt(a)), so we cancel the left factor with its inverse, we leave the right factor as part of the result (cannot be cancelled anyway) and then multiply by what remains to get it up to b. Still only works in the second case, obviously. If you wanted, you could enumerate all solutions by filling in all possibilities for the top tzcnt(a) bits.
The only thing that remains is getting the inverse, you've probably seen it in the other answer, whatever it was, but for completeness you can compute it as follows: (not tested)
; input x
dword y = (x * x) + x - 1;
dword t = y * x;
y *= 2 - t;
t = y * x;
y *= 2 - t;
t = y * x;
y *= 2 - t;
; result y
I stumbled across code like
double x,y = ...;
double n = sqrt(x*x+y*y);
if (n > 0)
{
double d1 = (x*x)/n;
double d2 = (x*y)/n;
}
and I am wondering about the numerical stability of such an expression for small values of x and y.
For both expressions, lim (x->0, y->0) (...) = 0, so from a mathematical point of view, it looks safe (the nominator O(x²) whereas the denominator is O(x)).
Nevertheless my question is: Are there any possible numerical problems with this code?
EDIT: If possible I'd like to avoid re-writing the expressions because n is actually used more than twice and to keep readability (it's relatively clear in the context what happens).
If x and y are very close to DBL_MIN, the calculations are
succeptible to underflow or extreme loss of precision: if x is
very close to DBL_MIN, for example x * x may be 0.0, or
(for somewhat larger values) it may result in what is called
gradual underflow, with extreme loss of precision: e.g. with
IEEE double (most, if not all desktop and laptop PCs), 1E-300
* 1E-300 will be 0.0. Obviously, if this happens for both
* x and y, you'll end up with n == 0.0, even if x and
y are both positive.
In C++11, there is a function hypot, which will solve the
problem for n; if x * x is 0.0, however, d1 will still
be 0.0; you'll probably get better results with (x / n) * x
(but I think that there still may be limit cases where you'll
end up with 0.0 or gradual underflow—I've not analyzed it sufficiently to be sure). A better solution
would be to scale the data differently, to avoid such limit
cases.
double MyClass::dx = ?????;
double MyClass::f(double x)
{
return 3.0*x*x*x - 2.0*x*x + x - 5.0;
}
double MyClass::fp(double x) // derivative of f(x), that is f'(x)
{
return (f(x + dx) - f(x)) / dx;
}
When using finite difference method for derivation, it is critical to choose an optimum dx value. Mathematically, dx must be as small as possible. However, I'm not sure if it is a correct choice to choose it the smallest positive double precision number (i.e.; 2.2250738585072014 x 10−308).
Is there an optimal numeric interval or exact value to choose a dx in to make the calculation error as small as possible?
(I'm using 64-bit compiler. I will run my program on a Intel i5 processor.)
Choosing the smallest possible value is almost certainly wrong: if dx were that smallest number, then f(x + dx) would be exactly equal to f(x) due to rounding.
So you have a tradeoff: Choose dx too small, and you lose precision to rounding errors. Choose it too large, and your result will be imprecise due to changes in the derivative as x changes.
To judge the numeric errors, consider (f(x + dx) - f(x))/f(x)1 mathematically. The numerator denotes the difference you want to compute, but the denominator denotes the magnitude of numbers you're dealing with. If that fraction is about 2‒k, then you can expect approximately k bits of precision in your result.
If you know your function, you can compute what error you'd get from choosing dx too large. You can then balence things, so that the error incurred from this is about the same as the error incurred from rounding. But if you know the function, you might be better off by providing a function that directly computes the derivative, like in your example with the polygonal f.
The Wikipedia section that pogorskiy pointed out suggests a value of sqrt(ε)x, or approximately 1.5e-8 * x. Without any more detailed knowledge about the function, such a rule of thumb will provide a reasonable default. Also note that that same section suggests not dividing by dx, but instead by (x + dx) - x, as this takes rounding errors incurred by computing x + dx into account. But I guess that whole article is full of suggestions you might use.
1 This formula really should divide by f(x), not by dx, even though a past editor thought differently. I'm attempting to compare the amount of significant bits remaining after the division, not the slope of the tangent.
Why not just use the Power Rule to derive the derivative, you'll get an exact answer:
f(x) = 3x^3 - 2x^2 + x - 5
f'(x) = 9x^2 - 4x + 1
Therefore:
f(x) = 3.0 * x * x * x - 2.0 * x * x + x - 5.0
fp(x) = 9.0 * x * x - 4.0 * x + 1.0
Vincent answered Fast Arc Cos algorithm by suggesting this function.
float arccos(float x)
{
x = 1 - (x + 1);
return pi * x / 2;
}
The question is, why x = 1 - (x + 1) and not x = -x?
It returns a different result only when (x + 1) causes a loss of precision, that is, x is many orders of magnitude larger or smaller than one.
But I don't think this is tricky or sleight of hand, I think it's just plain wrong.
cos(0) = 1 but f(1) = -pi/2
cos(pi/2) = 0 but f(0) = 0
cos(pi) = -1 but f(-1) = pi/2
where f(x) is Vincent's arccos implementation. All of them are off by pi/2, a linear approximation that gets at least these three points correct would be
g(x) = (1 - x) * pi / 2
I don't see the details instantly, but think about what happens as x approaches 1 or -1 from either side, and consider roundoff error.
Addition causes that both numbers are normalized (in this case, relevant for x). IIRC, in Knuth's volume 2, in the chapter on floating-point arithmetic, you can even see an expression like x+0.