issue_comments

I think this is fixed now thanks to @johnomotani

I think it would be fine to add a special case to preserve coordinates corresponding to min/max values with xarray's min() and max() methods, but I don't feel strongly about this. The exact coordinates could be surprising if there are multiple min/max values.

I agree that it does not make sense to preserve coordinates along aggregated dimensions for argmin/argmax, but we can preserve other coordinates. So I like @fujiisoup's example behavior above.

I suppose we now have two candidate APIs for returning multiple indices from a method like argmin/argmax: 1. Add an additional dimension, e.g., argmaxdim for keeping track of multiple indices. 2. Return a dict or Dataset with multiple indices.

I think my favorite option is (2) with da.argmin_indices() returning a Dataset, which will allow da[da.argmin_indices()] to work after we finish switching __iter__ to only iterate over data variables (https://github.com/pydata/xarray/issues/884#issuecomment-338445322). One downside of this approach is that it is not obvious how we woudl define argmin_indices() to work on a Dataset, but that's probably OK given that you cannot use a Dataset (yet) for indexing.

My concern with adding an additional dimension is that it is always a little surprising and error-prone when we invent new dimension names not supplied by the user (for example, this can lead to conflicting names). Also, consolidating indices will not work as well with idxmin(), which might put indices of different dtypes in the same array.

Either way, I would like a separate dedicated method for returning multiple indexing arrays. It's convenient (and what users expect) for argmax to return a single array if taking the max only over one dimension. However, if we switch to add an argmaxdim or return a dict/Dataset for multiple dimensions, then we will end up with an annoying inconsistency between the 1D and N-D versions. It would be better to say argmax(dim) is only for one dimension (and raise an error if this is not true) and have the separate argmax_indices(dims) that is consistently defined for any number of dimensions.

@gajomi

Sorry for my late response and thank you for the proposal.

But aside from my previous proposal, I was thinking whether such aggregation methods (including argmin) should propagate the coordinate. For example, as you pointed out, in theory, we may be able to track x-coordinate at the argmin index after da.argmin(dim='x'). But it is not reasonable for da.mean(dim='x'). It may be reasonable for da.max(dim='x') but not for da.median(dim='x').

Such specific rules may be confusing and bring additional complexity. I think the rule we do not track coordinates after aggregations would be much simpler and easier to understand.

If we adopt the above rule, I think the argmin would give just an array of indices, ```python In [1]: import xarray as xr ...: da = xr.DataArray([[0, 3, 2], [2, 1, 4]], dims=['x', 'y'], ...: coords={'x': [1, 2], 'y': ['a', 'b', 'c']}) ...:

In [4]: da.argmin(dim='x') Out[4]: <xarray.DataArray (y: 3)> array([0, 1, 0]) Coordinates: * y (y) <U1 'a' 'b' 'c'

In [3]: da.isel(x=da.argmin(dim='x')) Out[3]: <xarray.DataArray (y: 3)> array([0, 1, 2]) Coordinates: x (y) int64 1 2 1 * y (y) <U1 'a' 'b' 'c'

```

I think your logic would be useful even we do not track the coordinate.

I would appreciate any feedback.

I am thinking again how argmin should work with our new vectorizing indexing #1639 . It would be great if arr.isel(**arr.argmin(dim)) == arr.min(dim) could be satisfied even with a multi-dimensional array, although the behavior is different from numpy.argmin. (Maybe our current min should be replaced by arr.isel(**arr.argmin(dim)) so that it preserves the coordinates.)

(We discussed the name for this new method in #1469 but here I just use argmin for the simplicity.)

For example with a three dimensional array with dims=['x', 'y', 'z'], such as arr = xr.DataArray(np.random.randn(4, 3, 2), dims=['x', 'y', 'z']) I am thinking that... + arr.argmin() would return a xr.Dataset which contains 'x', 'y', 'z' as its data_vars. 1. ds = arr.argmin(dims=None) case: - ds['x'], ds['y'], ds['z'] would be 0d-integers. 2. ds = arr.argmin(dims=['x', 'y']) case: - ds['x'], ds['y'], ds['z'] would be 1d-integer-arrays. - The dimension of these three arrays would be 'z_argmin', where ds['z_argmin'] == arr['z']. 3. ds = arr.argmin(dims='x') case: - ds['x'], ds['y'], ds['z'] would be 2d-integer-arrays. - The dimensions of these three arrays are 'y_argmin' and 'z_argmin', where ds['y_argmin'] == arr['y'] and ds['z_argmin'] == arr['z'].

The above proposal for ii (and iii) is not quite clean, as if it is used as an argument of isel, it appends a new coordinate 'z_argmin', which is just a duplicate of 'arr['z']', i.e. arr.isel(**arr.argmin(dims=['x', 'y']))['z_argmin'] == arr['z'].

Any thoughts are welcome.

I'm working to fix this and I would like to make some design decisions;

1. What should max() look like? I guess this method should work also for multi-dimensional data. To satisfy the arr.isel_points(**arr.argmin_indices(dim)) == arg.min(dim) relation, the result array should have proper coordinates?

2. Multiple dim arguments Currently, doc says argmin accepts multiple axes, but np.argmin does not. Can we limit argmin's arguments only str not sequence of strs?

Edit:

1. Multi-dimensional array to isel_points Currently, isel_points only accepts 1-dimensional array, while the result of argmin_indexes can be multi-dimensional, e.g. python xr.DataArray(np.random.randn(4, 3, 2), dims=['x', 'y', 'z']).argmin_indexes(dims=['x']) Do we need special treatment for this (maybe in isel_points) or just raise an Error (current behavior)?

I agree that arr[arr.argmin(dim)] == arr.min(dim) is a useful invariant, but currently xarray's indexing works a little differently from NumPy. Probably arr.isel_points(**arr.argmin(dim)) == arg.min(dim) is the better invariant for now, and in the future arr[arr.argmin(dim)] will change to work consistently (#974).

The main downside of returning a tuple or dict from argmin() is that it makes the common case of taking the max/min over one dimension a little harder. So possibly it would be better to write two methods:

• argmin would work like it does currently (returning an xarray object), but error if reducing over multiple dimensions.
• argmin_indices would return a dict suitable for use in indexing.

Agreed, the current implementation of argmin() only gives the correct result when given one dimension. It's not entirely obvious what argmin() should yield when done to multiple dimensions, certainly this is not very useful. I would prefer raising an error to this current behavior.