cannot get cut() to display desired bins label

[qAp]

srs1 are some 1000 random numbers to be binned using the boundaries in absths:

srs1 = pd.Series(np.random.uniform(low=0, high=3e-17, size=(1000,)))
absths = np.array([0., 1.e-22, 1.e-18, 1.e-16])

Bin them, print out the boundaries, and the results for the first 5 numbers in srs1:

ncut1 = pd.cut(srs1, 
              bins=absths, 
              include_lowest=True, 
              precision=16,
              retbins=True)

print(ncut1[1])
print(ncut1[0][:5])

gives

[  0.00000000e+00   1.00000000e-22   1.00000000e-18   1.00000000e-16]
0    (1.0000000000000001e-18, 9.9999999999999998e-17]
1    (1.0000000000000001e-18, 9.9999999999999998e-17]
2    (1.0000000000000001e-18, 9.9999999999999998e-17]
3    (1.0000000000000001e-18, 9.9999999999999998e-17]
4    (1.0000000000000001e-18, 9.9999999999999998e-17]
dtype: category
Categories (3, object): [[0, 1] < (1, 1.0000000000000001e-18] < (1.0000000000000001e-18, 9.9999999999999998e-17]]

The boundary that is meant to be 1e-22 is displayed as 1 in Categories. The keyword argument precision is already set to 16 to display many decimals. Is this a bug or am I not using the function correctly?

Thanks

[jorisvandenbossche]

There seems to go something wrong with the conversion of the bin edges to a string I think.
Given that we are reworking this to be based on IntervalIndex, this may be fixed by that and @jreback this is maybe a case to test there in the PR? (#15309)

[jreback]

if you try with a higher precision, e.g. 22 this works, though I suspect you are actually hitting machine precision limits anyhow. Comparing numbers beyond 1e-15 can be somewhat arbitrary. So sometimes it will work, and if there are too many significant digits it might not work. I would either add a doc-note or raise if precision is too large here (IOW > 15). Yes this might work in #15309 because we are not stringifying but using actual values, but I think the same caveats apply.

So i'll mark this a doc-issue.

[anujloomba]

I was wondering if there was any resolution on this? I am facing a similar issue with labels.

[jorisvandenbossche]

@anujloomba can you provide a reproducible example that shows your problem?
The problem illustrated above is in fact mostly fixed on master (although the 0 is now not represented correctly)

[jorisvandenbossche]

@qAp on master I now get:

In [8]: print(ncut1[1])
[  0.00000000e+00   1.00000000e-22   1.00000000e-18   1.00000000e-16]

In [9]: print(ncut1[0][:5])
0    (1e-18, 1.0000000000000001e-16]
1    (1e-18, 1.0000000000000001e-16]
2    (1e-18, 1.0000000000000001e-16]
3    (1e-18, 1.0000000000000001e-16]
4    (1e-18, 1.0000000000000001e-16]
dtype: category
Categories (3, interval[float64]): [(-1e-16, 1.0000000000000002e-22] < (1.0000000000000002e-22, 1e-18] <
                                    (1e-18, 1.0000000000000001e-16]]

Which seems better as it displays the 1e-22 correctly. Although now the left bound of 0 is represented as -1e-16.