OCaml has a Random module, I am wondering how it tests itself for randomness. However, i don't have a clue what exactly they are doing. I understand it tries to test for chi-square with two more dependencies tests. Here are the code for the testing part:
chi-square test
(* Return the sum of the squares of v[i0,i1[ *)
let rec sumsq v i0 i1 =
if i0 >= i1 then 0.0
else if i1 = i0 + 1 then Pervasives.float v.(i0) *. Pervasives.float v.(i0)
else sumsq v i0 ((i0+i1)/2) +. sumsq v ((i0+i1)/2) i1
;;
let chisquare g n r =
if n <= 10 * r then invalid_arg "chisquare";
let f = Array.make r 0 in
for i = 1 to n do
let t = g r in
f.(t) <- f.(t) + 1
done;
let t = sumsq f 0 r
and r = Pervasives.float r
and n = Pervasives.float n in
let sr = 2.0 *. sqrt r in
(r -. sr, (r *. t /. n) -. n, r +. sr)
;;
Q1:, why they write sum of squares like that?
It seems it is just summing up all squares. Why not write like:
let rec sumsq v i0 i1 =
if i0 >= i1 then 0.0
else Pervasives.float v.(i0) *. Pervasives.float v.(i0) + (sumsq v (i0+1) i1)
Q2:, why they seem to use different way for chisquare?
From the chi squared test wiki, they formula is
But it seems they are using different formula, what's behind the scene?
Other two dependencies tests
(* This is to test for linear dependencies between successive random numbers.
*)
let st = ref 0;;
let init_diff r = st := int r;;
let diff r =
let x1 = !st
and x2 = int r
in
st := x2;
if x1 >= x2 then
x1 - x2
else
r + x1 - x2
;;
let st1 = ref 0
and st2 = ref 0
;;
(* This is to test for quadratic dependencies between successive random
numbers.
*)
let init_diff2 r = st1 := int r; st2 := int r;;
let diff2 r =
let x1 = !st1
and x2 = !st2
and x3 = int r
in
st1 := x2;
st2 := x3;
(x3 - x2 - x2 + x1 + 2*r) mod r
;;
Q3: I don't really know these two tests, can someone en-light me?
Q1:
It's a question of memory usage. You will notice that for large arrays, your implementation of sumsq will fail with "Stack overflow during evaluation" (on my laptop, it fails for r = 200000). This is because before adding Pervasives.float v.(i0) *. Pervasives.float v.(i0) to (sumsq v (i0+1) i1), you have to compute the latter. So it's not until you have computed the result of the last call of sumsq that you can start "going up the stack" and adding everything. Clearly, sumsq is going to be called r times in your case, so you will have to keep track of r calls.
By contrast, with their approach they only have to keep track of log(r) calls because once sumsq has been computed for half the array, you only need to the result of the corresponding call (you can forget about all the other calls that you had to do to compute that).
However, there are other ways of achieving this result and I'm not sure why they chose this one (maybe somebody will be able to tell ?). If you want to know more on the problems linked to recursion and memory, you should probably check the wikipedia article on tail-recursion. If you want to know more on the technique that they used here, you should check the wikipedia article on divide and conquer algorithms -- be careful though, because here we are talking about memory and the Wikipedia article will probably talk a lot about temporal complexity (speed).
Q2:
You should look more closely at both expressions. Here, all the E_i's are equal to n/r. If you replace this in the expression you gave, you will find the same expression that they use: (r *. t /. n) -. n. I didn't check about the values of the bounds though, but since you have a Chi-squared distribution with parameter r-minus-one-or-two degrees of freedom, and r quite large, it's not surprising to see them use this kind of confidence interval. The Wikipedia article you mentionned should help you figure out what confidence interval they use exactly fairly easily.
Good luck!
Edit: Oops, I forgot about Q3. I don't know these tests either, but I'm sure you should be able to find more about them by googling something like "linear dependency between consecutive numbers" or something. =)
Edit 2: In reply to Jackson Tale's June 29 question about the confidence interval:
They should indeed test it against the Chi-squared distribution -- or, rather, use the Chi-squared distribution to find a confidence interval. However, because of the central limit theorem, the Chi-squared distribution with k degrees of freedom converges to a normal law with mean k and variance 2k. A classical result is that the 95% confidence interval for the normal law is approximately [μ - 1.96 σ, μ + 1.96 σ], where μ is the mean and σ the standard deviation -- so that's roughly the mean ± twice the standard deviation. Here, the number of degrees of freedom is (I think) r - 1 ~ r (because r is large) so that's why I said I wasn't surprised by a confidence interval of the form [r - 2 sqrt(r), r + 2 sqrt(r)]. Nevertheless, now that I think about it I can't see why they don't use ± 2 sqrt(2 r)... But I might have missed something. And anyway, even if I was correct, since sqrt(2) > 1, they get a more stringent confidence interval, so I guess that's not really a problem. But they should document what they're doing a bit more... I mean, the tests that they're using are probably pretty standard so most likely most people reading their code will know what they're doing, but still...
Also, you should note that, as is often the case, this kind of test is not conclusive: generally, you want to show that something has some kind of effect. So you formulate two hypothesis : the null hypothesis, "there is no effect", and the alternative hypothesis, "there is an effect". Then, you show that, given your data, the probability that the null hypothesis holds is very low. So you conclude that the alternative hypothesis is (most likely) true -- i.e. that there is some kind of effect. This is conclusive. Here, what we would like to show is that the random number generator is good. So we don't want to show that the numbers it produces differ from some law, but that they conform to it. The only way to do that is to perform as many tests as possible showing that the number produced have the same property as randomly generated ones. But the only conclusion we can draw is "we were not able to find a difference between the actual data and what we would have observed, had they really been randomly generated". But this is not a lack of rigor from the OCaml developers: people always do that (eg, a lot of tests require, say, the normality. So before performing these tests, you try to find a test which would show that your variable is not normally distributed. And when you can't find any, you say "Oh well, the normality of this variable is probably sufficient for my subsequent tests to hold") -- simply because there is no other way to do it...
Anyway, I'm no statistician and the considerations above are simply my two cents, so you should be careful. For instance, I'm sure there is a better reason why they're using this particular confidence interval. I also think you should be able to figure it out if you write everything down carefully to make sure about what they're doing exactly.
Related
Determining the square root through successive approximation is implemented using the following algorithm:
Begin by guessing that the square root is x / 2. Call that guess g.
The actual square root must lie between g and x/g. At each step in the successive approximation, generate a new guess by averaging g and x/g.
Repeat step 2 until the values of g and x/g are as close together as the precision of the hardware allows. In Java, the best way to check for this condition is to test whether the average is equal to either of the values used to generate it.
What really confuses me is the last statement of step 3. I interpreted it as follows:
private double sqrt(double x) {
double g = x / 2;
while(true) {
double average = (g + x/g) / 2;
if(average == g || average == x/g) break;
g = average;
}
return g;
}
This seems to just cause an infinite loop. I am following the algorithm exactly, if the average equals either g or x/g (the two values used to generate it) then we have our answer ?
Why would anyone ever use that approach, when they could simply use the formulas for (2n^2) = 4n^2 and (n + 1)^2 = n^2 + 2n + 1, to populate each bit in the mantissa, and divide the exponent by two, multiplying the mantissa by two iff the the mod of the exponent with two equals 1?
To check if g and x/g are as close as the HW allow, look at the relative difference and compare
it with the epsilon for your floating point format. If it is within a small integer multiple of epsilon, you are OK.
Relative difference of x and y, see https://en.wikipedia.org/wiki/Relative_change_and_difference
The epsilon for 32-bit IEEE floats is about 1.0e-7, as in one of the other answers here, but that answer used the absolute rather than the relative difference.
In practice, that means something like:
Math.abs(g-x/g)/Math.max(Math.abs(g),Math.abs(x/g)) < 3.0e-7
Never compare floating point values for equality. The result is not reliable.
Use a epsilon like so:
if(Math.abs(average-g) < 1e-7 || Math.abs(average-x/g) < 1e-7)
You can change the epsilon value to be whatever you need. Probably best is something related to the original x.
Suppose I have a real number. I want to approximate it with something of the form a+sqrt(b) for integers a and b. But I don't know the values of a and b. Of course I would prefer to get a good approximation with small values of a and b. Let's leave it undefined for now what is meant by "good" and "small". Any sensible definitions of those terms will do.
Is there a sane way to find them? Something like the continued fraction algorithm for finding fractional approximations of decimals. For more on the fractions problem, see here.
EDIT: To clarify, it is an arbitrary real number. All I have are a bunch of its digits. So depending on how good of an approximation we want, a and b might or might not exist. Brute force is naturally not a particularly good algorithm. The best I can think of would be to start adding integers to my real, squaring the result, and seeing if I come close to an integer. Pretty much brute force, and not a particularly good algorithm. But if nothing better exists, that would itself be interesting to know.
EDIT: Obviously b has to be zero or positive. But a could be any integer.
No need for continued fractions; just calculate the square-root of all "small" values of b (up to whatever value you feel is still "small" enough), remove everything before the decimal point, and sort/store them all (along with the b that generated it).
Then when you need to approximate a real number, find the radical whose decimal-portion is closet to the real number's decimal-portion. This gives you b - choosing the correct a is then a simple matter of subtraction.
This is actually more of a math problem than a computer problem, but to answer the question I think you are right that you can use continued fractions. What you do is first represent the target number as a continued fraction. For example, if you want to approximate pi (3.14159265) then the CF is:
3: 7, 15, 1, 288, 1, 2, 1, 3, 1, 7, 4 ...
The next step is create a table of CFs for square roots, then you compare the values in the table to the fractional part of the target value (here: 7, 15, 1, 288, 1, 2, 1, 3, 1, 7, 4...). For example, let's say your table had square roots for 1-99 only. Then you would find the closest match would be sqrt(51) which has a CF of 7: 7,14 repeating. The 7,14 is the closest to pi's 7,15. Thus your answer would be:
sqrt(51)-4
As the closest approximation given a b < 100 which is off by 0.00016. If you allow larger b's then you could get a better approximation.
The advantage of using CFs is that it is faster than working in, say, doubles or using floating point. For example, in the above case you only have to compare two integers (7 and 15), and you can also use indexing to make finding the closest entry in the table very fast.
This can be done using mixed integer quadratic programming very efficiently (though there are no run-time guarantees as MIQP is NP-complete.)
Define:
d := the real number you wish to approximate
b, a := two integers such that a + sqrt(b) is as "close" to d as possible
r := (d - a)^2 - b, is the residual of the approximation
The goal is to minimize r. Setup your quadratic program as:
x := [ s b t ]
D := | 1 0 0 |
| 0 0 0 |
| 0 0 0 |
c := [0 -1 0]^T
with the constraint that s - t = f (where f is the fractional part of d)
and b,t are integers (s is not)
This is a convex (therefore optimally solvable) mixed integer quadratic program since D is positive semi-definite.
Once s,b,t are computed, simply derive the answer using b=b, s=d-a and t can be ignored.
Your problem may be NP-complete, it would be interesting to prove if so.
Some of the previous answers use methods that are of time or space complexity O(n), where n is the largest “small number” that will be accepted. By contrast, the following method is O(sqrt(n)) in time, and O(1) in space.
Suppose that positive real number r = x + y, where x=floor(r) and 0 ≤ y < 1. We want to approximate r by a number of the form a + √b. If x+y ≈ a+√b then x+y-a ≈ √b, so √b ≈ h+y for some integer offset h, and b ≈ (h+y)^2. To make b an integer, we want to minimize the fractional part of (h+y)^2 over all eligible h. There are at most √n eligible values of h. See following python code and sample output.
import math, random
def findb(y, rhi):
bestb = loerror = 1;
for r in range(2,rhi):
v = (r+y)**2
u = round(v)
err = abs(v-u)
if round(math.sqrt(u))**2 == u: continue
if err < loerror:
bestb, loerror = u, err
return bestb
#random.seed(123456) # set a seed if testing repetitively
f = [math.pi-3] + sorted([random.random() for i in range(24)])
print (' frac sqrt(b) error b')
for frac in f:
b = findb(frac, 12)
r = math.sqrt(b)
t = math.modf(r)[0] # Get fractional part of sqrt(b)
print ('{:9.5f} {:9.5f} {:11.7f} {:5.0f}'.format(frac, r, t-frac, b))
(Note 1: This code is in demo form; the parameters to findb() are y, the fractional part of r, and rhi, the square root of the largest small number. You may wish to change usage of parameters. Note 2: The
if round(math.sqrt(u))**2 == u: continue
line of code prevents findb() from returning perfect-square values of b, except for the value b=1, because no perfect square can improve upon the accuracy offered by b=1.)
Sample output follows. About a dozen lines have been elided in the middle. The first output line shows that this procedure yields b=51 to represent the fractional part of pi, which is the same value reported in some other answers.
frac sqrt(b) error b
0.14159 7.14143 -0.0001642 51
0.11975 4.12311 0.0033593 17
0.12230 4.12311 0.0008085 17
0.22150 9.21954 -0.0019586 85
0.22681 11.22497 -0.0018377 126
0.25946 2.23607 -0.0233893 5
0.30024 5.29150 -0.0087362 28
0.36772 8.36660 -0.0011170 70
0.42452 8.42615 0.0016309 71
...
0.93086 6.92820 -0.0026609 48
0.94677 8.94427 -0.0024960 80
0.96549 11.95826 -0.0072333 143
0.97693 11.95826 -0.0186723 143
With the following code added at the end of the program, the output shown below also appears. This shows closer approximations for the fractional part of pi.
frac, rhi = math.pi-3, 16
print (' frac sqrt(b) error b bMax')
while rhi < 1000:
b = findb(frac, rhi)
r = math.sqrt(b)
t = math.modf(r)[0] # Get fractional part of sqrt(b)
print ('{:11.7f} {:11.7f} {:13.9f} {:7.0f} {:7.0f}'.format(frac, r, t-frac, b,rhi**2))
rhi = 3*rhi/2
frac sqrt(b) error b bMax
0.1415927 7.1414284 -0.000164225 51 256
0.1415927 7.1414284 -0.000164225 51 576
0.1415927 7.1414284 -0.000164225 51 1296
0.1415927 7.1414284 -0.000164225 51 2916
0.1415927 7.1414284 -0.000164225 51 6561
0.1415927 120.1415831 -0.000009511 14434 14641
0.1415927 120.1415831 -0.000009511 14434 32761
0.1415927 233.1415879 -0.000004772 54355 73441
0.1415927 346.1415895 -0.000003127 119814 164836
0.1415927 572.1415909 -0.000001786 327346 370881
0.1415927 911.1415916 -0.000001023 830179 833569
I do not know if there is any kind of standard algorithm for this kind of problem, but it does intrigue me, so here is my attempt at developing an algorithm that finds the needed approximation.
Call the real number in question r. Then, first I assume that a can be negative, in that case we can reduce the problem and now only have to find a b such that the decimal part of sqrt(b) is a good approximation of the decimal part of r. Let us now write r as r = x.y with x being the integer and y the decimal part.
Now:
b = r^2
= (x.y)^2
= (x + .y)^2
= x^2 + 2 * x * .y + .y^2
= 2 * x * .y + .y^2 (mod 1)
We now only have to find an x such that 0 = .y^2 + 2 * x * .y (mod 1) (approximately).
Filling that x into the formulas above we get b and can then calculate a as a = r - b. (All of these calculations have to be carefully rounded of course.)
Now, for the time being I am not sure if there is a way to find this x without brute forcing it. But even then, one can simple use a simple loop to find an x good enough.
I am thinking of something like this(semi pseudo code):
max_diff_low = 0.01 // arbitrary accuracy
max_diff_high = 1 - max_diff_low
y = r % 1
v = y^2
addend = 2 * y
x = 0
while (v < max_diff_high && v > max_diff_low)
x++;
v = (v + addend) % 1
c = (x + y) ^ 2
b = round(c)
a = round(r - c)
Now, I think this algorithm is fairly efficient, while even allowing you to specify the wished accuracy of the approximation. One thing that could be done that would turn it into an O(1) algorithm is calculating all the x and putting them into a lookup table. If one only cares about the first three decimal digits of r(for example), the lookup table would only have 1000 values, which is only 4kb of memory(assuming that 32bit integers are used).
Hope this is helpful at all. If anyone finds anything wrong with the algorithm, please let me know in a comment and I will fix it.
EDIT:
Upon reflection I retract my claim of efficiency. There is in fact as far as I can tell no guarantee that the algorithm as outlined above will ever terminate, and even if it does, it might take a long time to find a very large x that solves the equation adequately.
One could maybe keep track of the best x found so far and relax the accuracy bounds over time to make sure the algorithm terminates quickly, at the possible cost of accuracy.
These problems are of course non-existent, if one simply pre-calculates a lookup table.
I have two vectors that represents a function f(x), and another vector f(ax+b) i.e. a scaled and shifted version of f(x). I would like to find the best scale and shift factors.
*best - by means of least squares error , maximum likelihood, etc.
any ideas?
for example:
f1 = [0;0.450541598502498;0.0838213779969326;0.228976968716819;0.91333736150167;0.152378018969223;0.825816977489547;0.538342435260057;0.996134716626885;0.0781755287531837;0.442678269775446;0];
f2 = [-0.029171964726699;-0.0278570165494982;0.0331454732535324;0.187656956432487;0.358856370923984;0.449974662483267;0.391341738643094;0.244800719791534;0.111797007617227;0.0721767235173722;0.0854437239807415;0.143888234591602;0.251750993723227;0.478953530572365;0.748209818420035;0.908044924557262;0.811960826711455;0.512568916956487;0.22669198638799;0.168136111568694;0.365578085161896;0.644996661336714;0.823562159983554;0.792812945867018;0.656803251999341;0.545799498053254;0.587013303815021;0.777464637372241;0.962722388208354;0.980537136457874;0.734416947254272;0.375435649393553;0.106489547770962;0.0892376361668696;0.242467741982851;0.40610516900965;0.427497319032133;0.301874099075184;0.128396341665384;0.00246347624097456;-0.0322120242872125]
*note that f(x) may be irreversible...
Thanks,
Ohad
For each f(x), take the absolute value of f(x) and normalize it such that it can be considered a probability mass function over its support. Calculate the expected value E[x] and variance of Var[x]. Then, we have that
E[a x + b] = a E[x] + b
Var[a x + b] = a^2 Var[x]
Use the above equations and the known values of E[x] and Var[x] to calculate a and b. Taking your values of f1 and f2 from your example, the following Octave script performs this procedure:
% Octave script
% f1, f2 are defined as given in your example
f1 = [zeros(length(f2) - length(f1), 1); f1];
save_f1 = f1; save_f2 = f2;
f1 = abs( f1 ); f2 = abs( f2 );
f1 = f1 ./ sum( f1 ); f2 = f2 ./ sum( f2 );
mean = #(x)sum(((1:length(x))' .* x));
var = #(x)sum((((1:length(x))'-mean(x)).^2) .* x);
m1 = mean(f1); m2 = mean(f2);
v1 = var(f1); v2 = var(f2)
a = sqrt( v2 / v1 ); b = m2 - a * m1;
plot( a .* (1:length( save_f1 )) + b, save_f1, ...
1:length( save_f2 ), save_f2 );
axis([0 length( save_f1 )];
And the output is
Here's a simple, effective, but perhaps somewhat naive approach.
First make sure you make a generic interpolator through both functions. That way you can evaluate both functions in between the given data points. I used a cubic-splines interpolator, since that seems general enough for the type of smooth functions you provided (and does not require additional toolboxes).
Then you evaluate the source function ("original") at a large number of points. Use this number also as a parameter in an inline function, that takes as input X, where
X = [a b]
(as in ax+b). For any input X, this inline function will compute
the function values of the target function at the same x-locations, but then scaled and offset by a and b, respectively.
The sum of the squared-differences between the resulting function values, and the ones of the source function you computed earlier.
Use this inline function in fminsearch with some initial estimate (one that you have obtained visually or by via automatic means). For the example you provided, I used a few random ones, which all converged to near-optimal fits.
All of the above in code:
function s = findScaleOffset
%% initialize
f2 = [0;0.450541598502498;0.0838213779969326;0.228976968716819;0.91333736150167;0.152378018969223;0.825816977489547;0.538342435260057;0.996134716626885;0.0781755287531837;0.442678269775446;0];
f1 = [-0.029171964726699;-0.0278570165494982;0.0331454732535324;0.187656956432487;0.358856370923984;0.449974662483267;0.391341738643094;0.244800719791534;0.111797007617227;0.0721767235173722;0.0854437239807415;0.143888234591602;0.251750993723227;0.478953530572365;0.748209818420035;0.908044924557262;0.811960826711455;0.512568916956487;0.22669198638799;0.168136111568694;0.365578085161896;0.644996661336714;0.823562159983554;0.792812945867018;0.656803251999341;0.545799498053254;0.587013303815021;0.777464637372241;0.962722388208354;0.980537136457874;0.734416947254272;0.375435649393553;0.106489547770962;0.0892376361668696;0.242467741982851;0.40610516900965;0.427497319032133;0.301874099075184;0.128396341665384;0.00246347624097456;-0.0322120242872125];
figure(1), clf, hold on
h(1) = subplot(2,1,1); hold on
plot(f1);
legend('Original')
h(2) = subplot(2,1,2); hold on
plot(f2);
linkaxes(h)
axis([0 max(length(f1),length(f2)), min(min(f1),min(f2)),max(max(f1),max(f2))])
%% make cubic interpolators and test points
pp1 = spline(1:numel(f1), f1);
pp2 = spline(1:numel(f2), f2);
maxX = max(numel(f1), numel(f2));
N = 100 * maxX;
x2 = linspace(1, maxX, N);
y1 = ppval(pp1, x2);
%% search for parameters
s = fminsearch(#(X) sum( (y1 - ppval(pp2,X(1)*x2+X(2))).^2 ), [0 0])
%% plot results
y2 = ppval( pp2, s(1)*x2+s(2));
figure(1), hold on
subplot(2,1,2), hold on
plot(x2,y2, 'r')
legend('before', 'after')
end
Results:
s =
2.886234493867320e-001 3.734482822175923e-001
Note that this computes the opposite transformation from the one you generated the data with. Reversing the numbers:
>> 1/s(1)
ans =
3.464721948700991e+000 % seems pretty decent
>> -s(2)
ans =
-3.734482822175923e-001 % hmmm...rather different from 7/11!
(I'm not sure about the 7/11 value you provided; using the exact values you gave to make a plot results in a less accurate approximation to the source function...Are you sure about the 7/11?)
Accuracy can be improved by either
using a different optimizer (fmincon, fminunc, etc.)
demanding a higher accuracy from fminsearch through optimset
having more sample points in both f1 and f2 to improve the quality of the interpolations
Using a better initial estimate
Anyway, this approach is pretty general and gives nice results. It also requires no toolboxes.
It has one major drawback though -- the solution found may not be the global optimizer, e.g., the quality of the outcomes of this method could be quite sensitive to the initial estimate you provide. So, always make a (difference) plot to make sure the final solution is accurate, or if you have a large number of such things to do, compute some sort of quality factor upon which you decide to re-start the optimization with a different initial estimate.
It is of course very possible to use the results of the Fourier+Mellin transforms (as suggested by chaohuang below) as an initial estimate to this method. That might be overkill for the simple example you provide, but I can easily imagine situations where this could indeed be very useful.
For the scale factor a, you can estimate it by computing the ratio of the amplitude spectra of the two signals since the Fourier transform is invariant to shift.
Similarly, you can estimate the shift factor b by using the Mellin transform, which is scale invariant.
Here's a super simple approach to estimate the scale a that works on your example data:
a = length(f2) / length(f1)
This gives 3.4167 which is close to your stated value of 3.4. If that estimate is good enough, you can use correlation to estimate the shift.
I realize that this is not exactly what you asked, but it may be an acceptable alternative depending on the data.
Both Rody Oldenhuis and jstarr's answers are correct. I'm adding my own answer just to sum things up, and connect between them.
I've messed up Rody's code a little bit and ended up with the following:
function findScaleShift
load f1f2
x0 = [length(f1)/length(f2) 0]; %initial guess, can do better
n=length(f1);
costFunc = #(z) sum((eval_f1(z,f2,n)-f1).^2);
opt.TolFun = eps;
xopt=fminsearch(costFunc,x0,opt);
f1r=eval_f1(xopt,f2,n);
subplot(211);
plot(1:n,f1,1:n,f1r,'--','linewidth',5)
title(xopt);
subplot(212);
plot(1:n,(f1-f1r).^2);
title('squared error')
end
function y = eval_f1(x,f2,n)
t = maketform('affine',[x(1) 0 x(2); 0 1 0 ; 0 0 1]');
y=imtransform(f2',t,'cubic','xdata',[1 n ],'ydata',[1 1])';
end
This gives zero results:
This method is accurate but exhaustive and may take some time. Another disadvantage is that it finds only a local minima, and may give false results if initial guess (x0) is far.
On the other hand, jstarr method gave the following results:
xopt = [ 3.49655562549115 -0.676062367063033]
which is 10% deviation from the correct answer. Pretty fast solution, but not as accurate as I requested, but still should be noted.
I think in order to get the best results jstarr method should be used as an initial guess for the method purposed by Rody, giving an accurate solution.
Ohad
John Carmack has a special function in the Quake III source code which calculates the inverse square root of a float, 4x faster than regular (float)(1.0/sqrt(x)), including a strange 0x5f3759df constant. See the code below. Can someone explain line by line what exactly is going on here and why this works so much faster than the regular implementation?
float Q_rsqrt( float number )
{
long i;
float x2, y;
const float threehalfs = 1.5F;
x2 = number * 0.5F;
y = number;
i = * ( long * ) &y;
i = 0x5f3759df - ( i >> 1 );
y = * ( float * ) &i;
y = y * ( threehalfs - ( x2 * y * y ) );
#ifndef Q3_VM
#ifdef __linux__
assert( !isnan(y) );
#endif
#endif
return y;
}
FYI. Carmack didn't write it. Terje Mathisen and Gary Tarolli both take partial (and very modest) credit for it, as well as crediting some other sources.
How the mythical constant was derived is something of a mystery.
To quote Gary Tarolli:
Which actually is doing a floating
point computation in integer - it took
a long time to figure out how and why
this works, and I can't remember the
details anymore.
A slightly better constant, developed by an expert mathematician (Chris Lomont) trying to work out how the original algorithm worked is:
float InvSqrt(float x)
{
float xhalf = 0.5f * x;
int i = *(int*)&x; // get bits for floating value
i = 0x5f375a86 - (i >> 1); // gives initial guess y0
x = *(float*)&i; // convert bits back to float
x = x * (1.5f - xhalf * x * x); // Newton step, repeating increases accuracy
return x;
}
In spite of this, his initial attempt a mathematically 'superior' version of id's sqrt (which came to almost the same constant) proved inferior to the one initially developed by Gary despite being mathematically much 'purer'. He couldn't explain why id's was so excellent iirc.
Of course these days, it turns out to be much slower than just using an FPU's sqrt (especially on 360/PS3), because swapping between float and int registers induces a load-hit-store, while the floating point unit can do reciprocal square root in hardware.
It just shows how optimizations have to evolve as the nature of underlying hardware changes.
Greg Hewgill and IllidanS4 gave a link with excellent mathematical explanation.
I'll try to sum it up here for ones who don't want to go too much into details.
Any mathematical function, with some exceptions, can be represented by a polynomial sum:
y = f(x)
can be exactly transformed into:
y = a0 + a1*x + a2*(x^2) + a3*(x^3) + a4*(x^4) + ...
Where a0, a1, a2,... are constants. The problem is that for many functions, like square root, for exact value this sum has infinite number of members, it does not end at some x^n. But, if we stop at some x^n we would still have a result up to some precision.
So, if we have:
y = 1/sqrt(x)
In this particular case they decided to discard all polynomial members above second, probably because of calculation speed:
y = a0 + a1*x + [...discarded...]
And the task has now came down to calculate a0 and a1 in order for y to have the least difference from the exact value. They have calculated that the most appropriate values are:
a0 = 0x5f375a86
a1 = -0.5
So when you put this into equation you get:
y = 0x5f375a86 - 0.5*x
Which is the same as the line you see in the code:
i = 0x5f375a86 - (i >> 1);
Edit: actually here y = 0x5f375a86 - 0.5*x is not the same as i = 0x5f375a86 - (i >> 1); since shifting float as integer not only divides by two but also divides exponent by two and causes some other artifacts, but it still comes down to calculating some coefficients a0, a1, a2... .
At this point they've found out that this result's precision is not enough for the purpose. So they additionally did only one step of Newton's iteration to improve the result accuracy:
x = x * (1.5f - xhalf * x * x)
They could have done some more iterations in a loop, each one improving result, until required accuracy is met. This is exactly how it works in CPU/FPU! But it seems that only one iteration was enough, which was also a blessing for the speed. CPU/FPU does as many iterations as needed to reach the accuracy for the floating point number in which the result is stored and it has more general algorithm which works for all cases.
So in short, what they did is:
Use (almost) the same algorithm as CPU/FPU, exploit the improvement of initial conditions for the special case of 1/sqrt(x) and don't calculate all the way to precision CPU/FPU will go to but stop earlier, thus gaining in calculation speed.
I was curious to see what the constant was as a float so I simply wrote this bit of code and googled the integer that popped out.
long i = 0x5F3759DF;
float* fp = (float*)&i;
printf("(2^127)^(1/2) = %f\n", *fp);
//Output
//(2^127)^(1/2) = 13211836172961054720.000000
It looks like the constant is "An integer approximation to the square root of 2^127 better known by the hexadecimal form of its floating-point representation, 0x5f3759df" https://mrob.com/pub/math/numbers-18.html
On the same site it explains the whole thing. https://mrob.com/pub/math/numbers-16.html#le009_16
According to this nice article written a while back...
The magic of the code, even if you
can't follow it, stands out as the i =
0x5f3759df - (i>>1); line. Simplified,
Newton-Raphson is an approximation
that starts off with a guess and
refines it with iteration. Taking
advantage of the nature of 32-bit x86
processors, i, an integer, is
initially set to the value of the
floating point number you want to take
the inverse square of, using an
integer cast. i is then set to
0x5f3759df, minus itself shifted one
bit to the right. The right shift
drops the least significant bit of i,
essentially halving it.
It's a really good read. This is only a tiny piece of it.
The code consists of two major parts. Part one calculates an approximation for 1/sqrt(y), and part two takes that number and runs one iteration of Newton's method to get a better approximation.
Calculating an approximation for 1/sqrt(y)
i = * ( long * ) &y;
i = 0x5f3759df - ( i >> 1 );
y = * ( float * ) &i;
Line 1 takes the floating point representation of y and treats it as an integer i. Line 2 shifts i over one bit and subtracts it from a mysterious constant. Line 3 takes the resulting number and converts it back to a standard float32. Now why does this work?
Let g be a function that maps a floating point number to its floating point representation, read as an integer. Line 1 above is setting i = g(y).
The following good approximation of g exists(*):
g(y) ≈ Clog_2 y + D for some constants C and D. An intuition for why such a good approximation exists is that the floating point representation of y is roughly linear in the exponent.
The purpose of line 2 is to map from g(y) to g(1/sqrt(y)), after which line 3 can use g^-1 to map that number to 1/sqrt(y). Using the approximation above, we have g(1/sqrt(y)) ≈ Clog_2 (1/sqrt(y)) + D = -C/2 log_2 y + D. We can use these formulas to calculate the map from g(y) to g(1/sqrt(y)), which is g(1/sqrt(y)) ≈ 3D/2 - 1/2 * g(y). In line 2, we have 0x5f3759df ≈ 3D/2, and i >> 1 ≈ 1/2*g(y).
The constant 0x5f3759df is slightly smaller than the constant that gives the best possible approximation for g(1/sqrt(y)). That is because this step is not done in isolation. Due to the direction that Newton's method tends to miss in, using a slightly smaller constant tends to yield better results. The exact optimal constant to use in this setting depends on your input distribution of y, but 0x5f3759df is one such constant that gives good results over a fairly broad range.
A more detailed description of this process can be found on Wikipedia: https://en.wikipedia.org/wiki/Fast_inverse_square_root#Algorithm
(*) More explicitly, let y = 2^e*(1+f). Taking the log of both sides, we get log_2 y = e + log_2(1+f), which can be approximated as log_2 y ≈ e + f + σ for a small constant sigma. Separately, the float32 encoding of y expressed as an integer is g(y) ≈ 2^23 * (e+127) + f * 2^23. Combining the two equations, we get g(y) ≈ 2^23 * log_2 y + 2^23 * (127 - σ).
Using Newton's method
y = y * ( threehalfs - ( x2 * y * y ) );
Consider the function f(y) = 1/y^2 - num. The positive zero of f is y = 1/sqrt(num), which is what we are interested in calculating.
Newton's method is an iterative algorithm for taking an approximation y_n for the zero of a function f, and calculating a better approximation y_n+1, using the following equation: y_n+1 = y_n - f(y_n)/f'(y_n).
Calculating what that looks like for our function f gives the following equation: y_n+1 = y_n - (-y_n+y_n^3*num)/2 = y_n * (3/2 - num/2 * y_n * y_n). This is exactly what the line of code above is doing.
You can learn more about the details of Newton's method here: https://en.wikipedia.org/wiki/Newton%27s_method
I have five values, A, B, C, D and E.
Given the constraint A + B + C + D + E = 1, and five functions F(A), F(B), F(C), F(D), F(E), I need to solve for A through E such that F(A) = F(B) = F(C) = F(D) = F(E).
What's the best algorithm/approach to use for this? I don't care if I have to write it myself, I would just like to know where to look.
EDIT: These are nonlinear functions. Beyond that, they can't be characterized. Some of them may eventually be interpolated from a table of data.
There is no general answer to this question. A solver finding the solution to any equation does not exist. As Lance Roberts already says, you have to know more about the functions. Just a few examples
If the functions are twice differentiable, and you can compute the first derivative, you might try a variant of Newton-Raphson
Have a look at the Lagrange Multiplier Method for implementing the constraint.
If the function F is continuous (which it probably is, if it is an interpolant), you could also try the Bisection Method, which is a lot like binary search.
Before you can solve the problem, you really need to know more about the function you're studying.
As others have already posted, we do need some more information on the functions. However, given that, we can still try to solve the following relaxation with a standard non-linear programming toolbox.
min k
st.
A + B + C + D + E = 1
F1(A) - k = 0
F2(B) - k = 0
F3(C) -k = 0
F4(D) - k = 0
F5(E) -k = 0
Now we can solve this in any manner we wish, such as penalty method
min k + mu*sum(Fi(x_i) - k)^2
st
A+B+C+D+E = 1
or a straightforward SQP or interior-point method.
More details and I can help advise as to a good method.
m
The functions are all monotonically increasing with their argument. Beyond that, they can't be characterized. The approach that worked turned out to be:
1) Start with A = B = C = D = E = 1/5
2) Compute F1(A) through F5(E), and recalculate A through E such that each function equals that sum divided by 5 (the average).
3) Rescale the new A through E so that they all sum to 1, and recompute F1 through F5.
4) Repeat until satisfied.
It converges surprisingly fast - just a few iterations. Of course, each iteration requires 5 root finds for step 2.
One solution of the equations
A + B + C + D + E = 1
F(A) = F(B) = F(C) = F(D) = F(E)
is to take A, B, C, D and E all equal to 1/5. Not sure though whether that is what you want ...
Added after John's comment (thanks!)
Assuming the second equation should read F1(A) = F2(B) = F3(C) = F4(D) = F5(E), I'd use the Newton-Raphson method (see Martijn's answer). You can eliminate one variable by setting E = 1 - A - B - C - D. At every step of the iteration you need to solve a 4x4 system. The biggest problem is probably where to start the iteration. One possibility is to start at a random point, do some iterations, and if you're not getting anywhere, pick another random point and start again.
Keep in mind that if you really don't know anything about the function then there need not be a solution.
ALGENCAN (part of TANGO) is really nice. There are Python bindings, too.
http://www.ime.usp.br/~egbirgin/tango/codes.php - " general nonlinear programming that does not use matrix manipulations at all and, so, is able to solve extremely large problems with moderate computer time. The general algorithm is of Augmented Lagrangian type ... "
http://pypi.python.org/pypi/TANGO%20Project%20-%20ALGENCAN/1.0
Google OPTIF9 or ALLUNC. We use these for general optimization.
You could use standard search technic as the others mentioned. There are a few optimization you could make use of it while doing the search.
First of all, you only need to solve A,B,C,D because 1-E = A+B+C+D.
Second, you have F(A) = F(B) = F(C) = F(D), then you can search for A. Once you get F(A), you could solve B, C, D if that is possible. If it is not possible to solve the functions, you need to continue search each variable, but now you have a limited range to search for because A+B+C+D <= 1.
If your search is discrete and finite, the above optimizations should work reasonable well.
I would try Particle Swarm Optimization first. It is very easy to implement and tweak. See the Wiki page for it.