Sometimes I am asked questions like "why did you introduce function xyz into MATLAB? And sometimes I have a compelling answer, even for something that looks simple on the surface. Consider the function hypot.
hypot essentially computes the square root of the sum of the squares of 2 inputs. So what's the big deal, right? For "reasonable" input values, there is no big issue.
fhypot = @(a,b) sqrt(abs(a).^2+abs(b).^2);
Let's create a set of values and compute hypot 3 ways for these: with hypot, fhypot, and sqrt(2)*x. These should all give the same answers, provide a is real and positive.
a = logspace(0,5,6) format long aHypot = hypot(a,a) aFhypot = fhypot(a,a) aSqrt2 = sqrt(2)*a
a = 1 10 100 1000 10000 100000 aHypot = 1.0e+005 * Columns 1 through 3 0.000014142135624 0.000141421356237 0.001414213562373 Columns 4 through 6 0.014142135623731 0.141421356237310 1.414213562373095 aFhypot = 1.0e+005 * Columns 1 through 3 0.000014142135624 0.000141421356237 0.001414213562373 Columns 4 through 6 0.014142135623731 0.141421356237310 1.414213562373095 aSqrt2 = 1.0e+005 * Columns 1 through 3 0.000014142135624 0.000141421356237 0.001414213562373 Columns 4 through 6 0.014142135623731 0.141421356237310 1.414213562373095
The results are the same, to within round-off, for the 3 methods here. But what happens if the magnitude of a is larger, near realmax perhaps?
ans = 1.797693134862316e+308
For my computer, a 32-bit Windows machine, realmax for doubles is on the order of 10^308. Let's see what happens if a value nearly that large is used with the different versions of hypot.
a = 1e308 aHypot = hypot(a,a) aFhypot = fhypot(a,a) aSqrt2 = sqrt(2)*a
a = 1.000000000000000e+308 aHypot = 1.414213562373095e+308 aFhypot = Inf aSqrt2 = 1.414213562373095e+308
What you can see is that the straight-forward method returns Inf instead of a finite answer. And that's why hypot was added to MATLAB, to compute the hypotenuse robustly, avoiding both underflow and overflow.
Get the MATLAB code
Published with MATLAB® 7.5
7 CommentsOldest to Newest
That was a pretty interesting post Loren. I’m not sure if it will affect the way I do things around here, but it’s something that could come in useful in the future. Thanks for the tip!
Before hypot existed, I used a=abs(complex(a,b)); In doing so, I was able to tap the robustness of the abs() function.
Thanks for pointing out that abs is also implemented robustly.
Does hypot(a,a) any different from norm([a a])?
norm is also careful to scale results, like abs and hypot.
I have a binary image and I want to find the distance between every object. I have already computed the centroid for every object. Can Hypot be used for many objects ???
You can use hypot if that is what makes sense. Here’s the help page. And here’s a few details from it:
c = hypot(a,b) returns the element-wise result of the following equation, computed to avoid underflow and overflow:
c = sqrt(abs(a).^2 + abs(b).^2)
Inputs a and b must follow these rules:
* Both a and b must be single- or double-precision, floating-point arrays.
* The sizes of the a and b arrays must either be equal, or one a scalar and the other nonscalar. In the latter case, hypot expands the scalar input to match the size of the nonscalar input.