FIX: Add on_few_samples parameter to core rank estimation (#13350)

Signed-off-by: Emmanuel Ferdman <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Eric Larson <[email protected]>

Author: 3 people

doc/changes/dev/13350.apichange.rst

@@ -0,0 +1 @@
+Add ``on_few_samples`` parameter to :func:`mne.compute_covariance` and :func:`mne.compute_raw_covariance` for controlling behavior when there are fewer samples than channels, which can lead to inaccurate covariance estimates, by :newcontrib:`Emmanuel Ferdman`.

doc/changes/names.inc

@@ -76,6 +76,7 @@
 .. _Eduard Ort: https://github.com/eort
 .. _Emily Stephen: https://github.com/emilyps14
 .. _Emma Bailey: https://www.cbs.mpg.de/employees/bailey
+.. _Emmanuel Ferdman: https://github.com/emmanuel-ferdman
 .. _Emrecan Çelik: https://github.com/emrecncelik
 .. _Enrico Varano: https://github.com/enricovara/
 .. _Enzo Altamiranda: https://www.linkedin.com/in/enzoalt

mne/cov.py

@@ -553,16 +553,17 @@ def make_ad_hoc_cov(info, std=None, *, verbose=None):
     return Covariance(data, ch_names, info["bads"], info["projs"], nfree=0)
 
 
-def _check_n_samples(n_samples, n_chan):
+def _check_n_samples(n_samples, n_chan, on_few_samples="warn"):
     """Check to see if there are enough samples for reliable cov calc."""
     n_samples_min = 10 * (n_chan + 1) // 2
     if n_samples <= 0:
         raise ValueError("No samples found to compute the covariance matrix")
     if n_samples < n_samples_min:
-        warn(
-            f"Too few samples (required : {n_samples_min} got : {n_samples}), "
-            "covariance estimate may be unreliable"
+        msg = (
+            f"Too few samples (required {n_samples_min} but got {n_samples} for "
+            f"{n_chan} channels), covariance estimate may be unreliable"
         )
+        _on_missing(on_few_samples, msg, "on_few_samples")
 
 
 @verbose
@@ -574,6 +575,8 @@ def compute_raw_covariance(
     reject=None,
     flat=None,
     picks=None,
+    *,
+    on_few_samples="warn",
     method="empirical",
     method_params=None,
     cv=3,
@@ -623,6 +626,12 @@ def compute_raw_covariance(
         are floats that set the minimum acceptable peak-to-peak amplitude.
         If flat is None then no rejection is done.
     %(picks_good_data_noref)s
+    on_few_samples : str
+        Can be 'warn' (default), 'ignore', or 'raise' to control behavior when
+        there are fewer samples than channels, which can lead to inaccurate
+        covariance or rank estimates.
+
+        .. versionadded:: 1.11
     method : str | list | None (default 'empirical')
         The method used for covariance estimation.
         See :func:`mne.compute_covariance`.
@@ -736,7 +745,7 @@ def compute_raw_covariance(
             mu += raw_segment.sum(axis=1)
             data += np.dot(raw_segment, raw_segment.T)
             n_samples += raw_segment.shape[1]
-        _check_n_samples(n_samples, len(picks))
+        _check_n_samples(n_samples, len(picks), on_few_samples)
         data -= mu[:, None] * (mu[None, :] / n_samples)
         data /= n_samples - 1.0
         logger.info("Number of samples used : %d", n_samples)
@@ -864,6 +873,8 @@ def compute_covariance(
     tmin=None,
     tmax=None,
     projs=None,
+    *,
+    on_few_samples="warn",
     method="empirical",
     method_params=None,
     cv=3,
@@ -909,6 +920,12 @@ def compute_covariance(
         List of projectors to use in covariance calculation, or None
         to indicate that the projectors from the epochs should be
         inherited. If None, then projectors from all epochs must match.
+    on_few_samples : str
+        Can be 'warn' (default), 'ignore', or 'raise' to control behavior when
+        there are fewer samples than channels, which can lead to inaccurate
+        covariance or rank estimates.
+
+        .. versionadded:: 1.11
     method : str | list | None (default 'empirical')
         The method used for covariance estimation. If 'empirical' (default),
         the sample covariance will be computed. A list can be passed to
@@ -1144,7 +1161,7 @@ def _unpack_epochs(epochs):
 
     epochs = np.hstack(epochs)
     n_samples_tot = epochs.shape[-1]
-    _check_n_samples(n_samples_tot, len(picks_meeg))
+    _check_n_samples(n_samples_tot, len(picks_meeg), on_few_samples)
 
     epochs = epochs.T  # sklearn | C-order
     cov_data = _compute_covariance_auto(
@@ -1158,6 +1175,7 @@ def _unpack_epochs(epochs):
         picks_list=picks_list,
         scalings=scalings,
         rank=rank,
+        on_few_samples=on_few_samples,
     )
 
     if keep_sample_mean is False:
@@ -1221,7 +1239,7 @@ def _eigvec_subspace(eig, eigvec, mask):
 
 @verbose
 def _compute_rank_raw_array(
-    data, info, rank, scalings, *, log_ch_type=None, verbose=None
+    data, info, rank, scalings, *, log_ch_type=None, on_few_samples="warn", verbose=None
 ):
     from .io import RawArray
 
@@ -1231,6 +1249,7 @@ def _compute_rank_raw_array(
         scalings,
         info,
         log_ch_type=log_ch_type,
+        on_few_samples=on_few_samples,
     )
 
 
@@ -1249,6 +1268,7 @@ def _compute_covariance_auto(
     cov_kind="",
     log_ch_type=None,
     log_rank=True,
+    on_few_samples="warn",
 ):
     """Compute covariance auto mode."""
     # rescale to improve numerical stability
@@ -1258,6 +1278,7 @@ def _compute_covariance_auto(
         info,
         rank=rank,
         scalings=scalings,
+        on_few_samples=on_few_samples,
         verbose=_verbose_safe_false(),
     )
     with _scaled_array(data.T, picks_list, scalings):
@@ -1729,6 +1750,7 @@ def prepare_noise_cov(
     rank=None,
     scalings=None,
     on_rank_mismatch="ignore",
+    *,
     verbose=None,
 ):
     """Prepare noise covariance matrix.
@@ -2119,6 +2141,9 @@ def _regularized_covariance(
     log_ch_type=None,
     log_rank=None,
     cov_kind="",
+    # backward-compat default for decoding (maybe someday we want to expose this but
+    # it's likely too invasive and since it's usually regularized, unnecessary):
+    on_few_samples="ignore",
     verbose=None,
 ):
     """Compute a regularized covariance from data using sklearn.
@@ -2166,6 +2191,7 @@ def _regularized_covariance(
         cov_kind=cov_kind,
         log_ch_type=log_ch_type,
         log_rank=log_rank,
+        on_few_samples=on_few_samples,
     )[reg]["data"]
     return cov
 

mne/decoding/_covs_ged.py

@@ -88,6 +88,7 @@ def _csp_estimate(X, y, reg, cov_method_params, cov_est, info, rank, norm_trace)
             rank=rank,
             scalings=None,
             log_ch_type="data",
+            on_few_samples="ignore",
         )
 
     covs = []
@@ -158,6 +159,7 @@ def _xdawn_estimate(
             rank=rank,
             scalings=None,
             log_ch_type="data",
+            on_few_samples="ignore",
         )
     return covs, C_ref, info, rank, dict()
 
@@ -280,5 +282,6 @@ def _spoc_estimate(X, y, reg, cov_method_params, info, rank):
             rank=rank,
             scalings=None,
             log_ch_type="data",
+            on_few_samples="ignore",
         )
     return covs, C_ref, info, rank, dict()

mne/decoding/tests/test_csp.py

@@ -431,7 +431,7 @@ def test_spoc():
     # check y
     pytest.raises(ValueError, spoc.fit, X, y * 0)
 
-    # Check that doesn't take CSP-spcific input
+    # Check that doesn't take CSP-specific input
     pytest.raises(TypeError, SPoC, cov_est="epoch")
 
     # Check mixing matrix on simulated data

mne/rank.py

@@ -130,14 +130,27 @@ def _estimate_rank_from_s(s, tol="auto", tol_kind="absolute"):
 
 
 def _estimate_rank_raw(
-    raw, picks=None, tol=1e-4, scalings="norm", with_ref_meg=False, tol_kind="absolute"
+    raw,
+    picks=None,
+    tol=1e-4,
+    scalings="norm",
+    with_ref_meg=False,
+    tol_kind="absolute",
+    on_few_samples="warn",
 ):
     """Aid the transition away from raw.estimate_rank."""
     if picks is None:
         picks = _picks_to_idx(raw.info, picks, with_ref_meg=with_ref_meg)
     # conveniency wrapper to expose the expert "tol" option + scalings options
     return _estimate_rank_meeg_signals(
-        raw[picks][0], pick_info(raw.info, picks), scalings, tol, False, tol_kind
+        raw[picks][0],
+        pick_info(raw.info, picks),
+        scalings,
+        tol,
+        False,
+        tol_kind,
+        log_ch_type=None,
+        on_few_samples=on_few_samples,
     )
 
 
@@ -150,6 +163,7 @@ def _estimate_rank_meeg_signals(
     return_singular=False,
     tol_kind="absolute",
     log_ch_type=None,
+    on_few_samples="warn",
 ):
     """Estimate rank for M/EEG data.
 
@@ -173,6 +187,10 @@ def _estimate_rank_meeg_signals(
         to determine the rank.
     tol_kind : str
         Tolerance kind. See ``estimate_rank``.
+    on_few_samples : str
+        Can be 'warn' (default), 'ignore', or 'raise' to control behavior when
+        there are fewer samples than channels, which can lead to inaccurate rank
+        estimates.
 
     Returns
     -------
@@ -183,11 +201,14 @@ def _estimate_rank_meeg_signals(
         thresholded to determine the rank are also returned.
     """
     picks_list = _picks_by_type(info)
-    if data.shape[1] < data.shape[0]:
-        ValueError(
-            "You've got fewer samples than channels, your "
-            "rank estimate might be inaccurate."
+    assert data.ndim == 2, data.shape
+    n_channels, n_samples = data.shape
+    if n_samples < n_channels:
+        msg = (
+            f"Too few samples ({n_samples=} is less than {n_channels=}), "
+            "rank estimate may be unreliable"
         )
+        _on_missing(on_few_samples, msg, "on_few_samples")
     with _scaled_array(data, picks_list, scalings):
         out = estimate_rank(
             data,
@@ -214,6 +235,7 @@ def _estimate_rank_meeg_cov(
     return_singular=False,
     *,
     log_ch_type=None,
+    on_few_samples="warn",
     verbose=None,
 ):
     """Estimate rank of M/EEG covariance data, given the covariance.
@@ -236,6 +258,10 @@ def _estimate_rank_meeg_cov(
     return_singular : bool
         If True, also return the singular values that were used
         to determine the rank.
+    on_few_samples : str
+        Can be 'warn' (default), 'ignore', or 'raise' to control behavior when
+        there are fewer samples than channels, which can lead to inaccurate rank
+        estimates.
 
     Returns
     -------
@@ -249,10 +275,11 @@ def _estimate_rank_meeg_cov(
     scalings = _handle_default("scalings_cov_rank", scalings)
     _apply_scaling_cov(data, picks_list, scalings)
     if data.shape[1] < data.shape[0]:
-        ValueError(
+        msg = (
             "You've got fewer samples than channels, your "
             "rank estimate might be inaccurate."
         )
+        _on_missing(on_few_samples, msg, "on_few_samples")
     out = estimate_rank(data, tol=tol, norm=False, return_singular=return_singular)
     rank = out[0] if isinstance(out, tuple) else out
     if log_ch_type is None:
@@ -325,7 +352,7 @@ def _compute_rank_int(inst, *args, **kwargs):
     # XXX eventually we should unify how channel types are handled
     # so that we don't need to do this, or we do it everywhere.
     # Using pca=True in compute_whitener might help.
-    return sum(compute_rank(inst, *args, **kwargs).values())
+    return sum(compute_rank(inst, *args, on_few_samples="ignore", **kwargs).values())
 
 
 @verbose
@@ -335,9 +362,11 @@ def compute_rank(
     scalings=None,
     info=None,
     tol="auto",
+    *,
     proj=True,
     tol_kind="absolute",
     on_rank_mismatch="ignore",
+    on_few_samples=None,
     verbose=None,
 ):
     """Compute the rank of data or noise covariance.
@@ -363,6 +392,13 @@ def compute_rank(
         considered when ``rank=None`` or ``rank='info'``.
     %(tol_kind_rank)s
     %(on_rank_mismatch)s
+    on_few_samples : str | None
+        Can be 'warn', 'ignore', or 'raise' to control behavior when
+        there are fewer samples than channels, which can lead to inaccurate rank
+        estimates. None (default) means "ignore" if ``inst`` is a
+        :class:`mne.Covariance` or ``rank in ("info", "full")``, and "warn" otherwise.
+
+        .. versionadded:: 1.11
     %(verbose)s
 
     Returns
@@ -384,6 +420,7 @@ def compute_rank(
         proj=proj,
         tol_kind=tol_kind,
         on_rank_mismatch=on_rank_mismatch,
+        on_few_samples=on_few_samples,
     )
 
 
@@ -398,6 +435,7 @@ def _compute_rank(
     proj=True,
     tol_kind="absolute",
     on_rank_mismatch="ignore",
+    on_few_samples=None,
     log_ch_type=None,
     verbose=None,
 ):
@@ -441,6 +479,12 @@ def _compute_rank(
         if rank is None:
             rank = dict()
 
+    if on_few_samples is None:
+        if inst_type != "covariance" and rank_type == "estimated":
+            on_few_samples = "warn"
+        else:
+            on_few_samples = "ignore"
+
     simple_info = _simplify_info(info)
     picks_list = _picks_by_type(info, meg_combined=True, ref_meg=False, exclude="bads")
     for ch_type, picks in picks_list:
@@ -503,6 +547,7 @@ def _compute_rank(
                     False,
                     tol_kind,
                     log_ch_type=log_ch_type,
+                    on_few_samples=on_few_samples,
                 )
             else:
                 assert isinstance(inst, Covariance)
@@ -520,6 +565,7 @@ def _compute_rank(
                         tol,
                         return_singular=True,
                         log_ch_type=log_ch_type,
+                        on_few_samples=on_few_samples,
                         verbose=est_verbose,
                     )
                     if ch_type in rank:

mne/simulation/tests/test_evoked.py

@@ -137,7 +137,8 @@ def test_add_noise():
         if inst is raw:
             cov_new = compute_raw_covariance(inst, picks=picks)
         else:
-            cov_new = compute_covariance(inst)
+            with pytest.warns(RuntimeWarning, match=".*Too few samples.*"):
+                cov_new = compute_covariance(inst)
         assert cov["names"] == cov_new["names"]
         r = np.corrcoef(cov["data"].ravel(), cov_new["data"].ravel())[0, 1]
         assert r > 0.99