MATLAB Indexing type dependence (bug?)

Please see this strange example, where the datatype (single vs double) of the starting index affects how many data points are returned:

y = rand(67347456+50,1);
y = single(y);
N = single(50);
size(y(N+1:end))

MATLAB returns: >> 67347454 1 (incorrect, missing 2 elements)

size([N+1:length(y)])

MATLAB returns: >> 1 67347456 (which is correct)

and

size(y(double(N)+1:end))

MATLAB returns >> 67347456 1 (also correct)

Other datatypes like uint32 also work fine.

Please help me understand: If N is only 50, what does it matter to the indexing whether it is a single or a double? Also, aren't these ending index values well within the presicion of single?

Answer 1

[Stephen23]

"what does it matter to the indexing whether it is a single or a double?"

Because single precision cannot represent all of those indices. The largest index is:

y = rand(67347456+50,1);
x = numel(y)

x = 67347506

which cannot be represented using SINGLE. The closest SINGLE value is:

single(x)

ans = single
67347504

"...aren't these ending index values well within the presicion of single?"

Nope, those indices are well above FLINTMAX:

flintmax('single')

ans = single
16777216

"MATLAB Indexing type dependence (bug?)"

I do not see anything that indicates a bug.

Answer 2

[Walter Roberson]

eps(single(67347456+50))

ans = single
8

single precision numbers are 8 apart by the time of 67347456