Comments on: Matlab numerical gotchas

By: Sami Varjo

Sami Varjo — Thu, 09 Jan 2014 13:35:06 +0000

It was "cool" to note that the weak typing favors integers in arrays. One would expect that the more weak variable would be promoted as the floating point presentation is usually default, however:

 
>> myUint8Var = uint8(0);
>> newVector = [myUint8Var 510.14 45.24 1.2]
newVector =
    0  255   45    1

This actually happens also if you preallocate an array for a number of vectors where you assign the values:

>> array = zeros(3,4,'double');
>> array(1,:) = [myUint8Var 510.14 45.24 1.2]
array =
0   255    45     1
0     0     0     0
0     0     0     0

So also here explicit type cast for myUint8Var is required for ~~expected~~ assumed behavior (R2012a). This one haunted me quite a while...

By: Malcolm Lidierth

Malcolm Lidierth — Mon, 18 Nov 2013 16:04:31 +0000

Take a Java method with signatures accepting either: a scalar int a double[] as input. MATLAB's Java reflection mechanism will pass a scalar flint as a double[] to the second method by default. To invoke the method with an int input, it needs to be cast explicitly to an int.

By: Daniel E. Shub

Daniel E. Shub — Mon, 11 Nov 2013 12:06:31 +0000

Despite the documentation for the colon function I cannot help but expect x = 0:0.1:1; to be equal to [0*0.1, 1*0.1, 2*0.1, …, 10*0.1] as that is how I describe the function working. This can cause a gotcha when checking for subsets

>> x = 0:0.1:1;
>> y = 0:0.1:0.5;
>> isequal(x(1:6), y(1:6))

By: Yair Altman

Yair Altman — Wed, 06 Nov 2013 19:39:06 +0000

In reply to Guillaume.

@Guillaume – thanks for the correction. It serves me right for copying wiki math without double checking. I have now corrected both the main article and the referenced FAQ one.

BTW – it appears that like me, you have also gotten an error in your comment: I believe that the correct roots are 2e^iπ, 2e^iπ/3 and 2e^-iπ/3, and not as you have stated.
🙂

By: Guillaume

Guillaume — Wed, 06 Nov 2013 09:24:10 +0000

Hmm, 2e^2iπ/3 is not a cube root of -8. The correct formula for nth roots of Ae^iθ is A^1/n * e^i(θ/n + 2kπ/n). Hence, cube roots of -8 are 2e^iπ (k=1), 2e^-iπ (k=2) and 2e^iπ/3 (k=3).

By: the cyclist

the cyclist — Fri, 01 Nov 2013 20:56:31 +0000

I find it surprising that min(0,NaN) returns a result of 0 rather than NaN.

There is some discussion of this here:

http://www.mathworks.com/matlabcentral/answers/38999-what-is-the-reasoning-behind-the-fact-that-min-0-nan-is-0

By: Wesley Brodsky

Wesley Brodsky — Fri, 01 Nov 2013 17:28:15 +0000

One problem I have noticed in the past, which might still exist, is one of the “creeping imaginary part”. After a long series of calculations with all real numbers, a teeny-weenie imaginary part (<10^-10 times real part) appears. This causes an error if input to a function that only accepts real numbers. This is not the same as the cube-root issue, since the real part is what I expect. I place x = real(x) at random places in the calculations to avoid this. The problem might no longer exist, but I have not experimented to find out.

Has anyone else seen this?

By: Chandrakanth

Chandrakanth — Thu, 31 Oct 2013 15:58:09 +0000

I think I have come across quite such 'gotachas', but this is the only one i can recollect now.

 
m>=3  % is fine.
m>= 3 % notice the space before 3. This is fine as well.
m >= 3 % notice space before 3 and after m. This too is fine.
m >=3 % space only after m. This complains with an error.

By: Walter Roberson on Facebook

Walter Roberson on Facebook — Wed, 30 Oct 2013 21:08:28 +0000

Let A and B be positive. Then mod(-A,B) will return B itself, if A < eps(B), instead of returning something in [0,B) as would normally be expected for mathematical modulus operator.