diff --git a/RELEASES.md b/RELEASES.md index 03093ac6f..cedd7ab48 100644 --- a/RELEASES.md +++ b/RELEASES.md @@ -19,6 +19,7 @@ #### Closed issues - Fix ordinary JAX gradients for dense `ot.emd2` and preserve the magnitude and higher-order dependence of gradients supplied to `JaxBackend.set_gradients`, including the envelope gradients in `ot.solve` (Issue #625) +- Center floating-point inputs in `ot.utils.proj_simplex` to prevent large common offsets from erasing the simplex mass (Issue #879, PR #880). - Fix `ot.semidiscrete_wasserstein2_unif_circle` pairing sorted sample locations with unsorted weights, which broke permutation and rotation invariance for nonuniform weights (Issue #921) - Fix an out-of-bounds read in `ot.bsp.merge_bijections`: the number of plans was taken from `perms.shape[1]` (the number of samples) instead of `perms.shape[0]`, so merging fewer plans than samples read past the array and merging more silently dropped plans. A `perms` outside the documented `(n_plans, n_samples)` layout, an empty plan stack and rows that are not permutations are now rejected (PR #911, Issue #910) - Fix overflow to NaN in `ot.stochastic.coordinate_grad_semi_dual` by shifting before exponentiating, as `ot.stochastic.c_transform_entropic` already does. The factor cancels in the normalisation, so results are unchanged where the old code did not overflow (PR #866, Issue #264) diff --git a/ot/utils.py b/ot/utils.py index 69a4854b5..159b442d2 100644 --- a/ot/utils.py +++ b/ot/utils.py @@ -111,6 +111,10 @@ def proj_simplex(v, z=1): .. note:: This function is backend-compatible and will work on arrays from all compatible backends. + Floating-point inputs are centered by their column maximum before projection. + A common offset does not change the simplex projection; removing it prevents + a large offset from erasing the simplex size in the cumulative sums. + Parameters ---------- v : {array-like}, shape (n, d) @@ -123,6 +127,9 @@ def proj_simplex(v, z=1): Array of projections on the simplex """ nx = get_backend(v) + # The simplex projection is invariant to a common offset per column. + if nx.is_floating_point(v): + v = v - nx.max(v, axis=0) n = v.shape[0] if v.ndim == 1: d1 = 1 diff --git a/test/test_utils.py b/test/test_utils.py index 9c93d520a..17c8e8318 100644 --- a/test/test_utils.py +++ b/test/test_utils.py @@ -99,6 +99,24 @@ def test_proj_simplex(nx): np.testing.assert_allclose(l1, l2, atol=1e-5) +@pytest.mark.parametrize( + "dtype, offset", + [ + (np.float32, 0), + (np.float32, 1e8), + (np.float64, 0), + (np.float64, 1e16), + (np.int64, 0), + ], +) +@pytest.mark.parametrize("matrix", [False, True]) +def test_proj_simplex_large_translation(nx, dtype, offset, matrix): + values = np.full((3, 2) if matrix else (3,), offset, dtype=dtype) + result = nx.to_numpy(ot.utils.proj_simplex(nx.from_numpy(values))) + np.testing.assert_allclose(result, np.full(values.shape, 1 / 3), rtol=1e-6) + np.testing.assert_allclose(result.sum(axis=0), 1, rtol=1e-6) + + def test_projection_sparse_simplex(): def double_sort_projection_sparse_simplex(X, max_nz, z=1, axis=None): r"""This is an equivalent but less efficient version