Minor clarifications

docs/jwst/emicorr/description.rst

@@ -63,7 +63,7 @@ the EMI noise are fit simultaneously. It works by choosing pixels with modest
 scatter among the reads, and then finding the amplitude and phase of a supplied
 reference waveform that, when subtracted, makes these pixels' ramps as straight
 as possible. The straightness of the ramps is measured by a chi-squared metric, after
-fitting lines to each one.  As for the sequential algorithm, the noise at each
+fitting lines to each one.  As for the sequential algorithm, the EMI signal at each
 frequency is fit and removed iteratively, for each desired frequency.
 
 The optimal choice for the fitting algorithm depends on the input.

jwst/emicorr/emicorr.py

@@ -64,8 +64,8 @@ def apply_emicorr(
     scatter among the reads, and then finding the amplitude and phase of a supplied
     reference waveform that, when subtracted, makes these pixels' ramps as straight
     as possible. The straightness of the ramps is measured by chi squared after
-    fitting lines to each one.  As for the sequential algorithm, the noise at each
-    frequency is fit and removed iteratively, for each desired frequency.
+    fitting lines to each one.  As for the sequential algorithm, the EMI signal at
+    each frequency is fit and removed iteratively, for each desired frequency.
 
     For either algorithm, removing the highest amplitude EMI first will probably
     give best results.
@@ -246,8 +246,10 @@ def _run_joint_algorithm(
     Returns
     -------
     DataModel
-        The output datamodel, with noise fit and subtracted.
+        The input datamodel, with noise fit and subtracted.
     """
+    # Additional frame time to account for the extra frame reset between
+    # MIRI integrations
     readpatt = str(input_model.meta.exposure.readpatt).upper()
     if readpatt == "FASTR1" or readpatt == "SLOWR1":
         _frameclocks = frameclocks
@@ -941,17 +943,15 @@ def emicorr_refwave(
     all_n = np.zeros((nints, nphases))
 
     # "Good" pixel here has no more than twice the median standard
-    # deviation among group differences and is not flagged in the pdq
+    # deviation among group values and is not flagged in the pdq
     # array.  This should discard most bad and high-flux pixels.
 
     pixel_std = np.std(data, axis=1)
     pixel_ok = (pixel_std < 2 * np.median(pixel_std)) & (pdq == 0)
 
     for j in range(ny):
         for k in range(nx4):
-            # Bad reference column?  I am not sure whether this is needed,
-            # but I think so. I carried it from the current routine.
-
+            # Bad reference column
             if k == 1 or k == 2:
                 continue
 
@@ -1051,6 +1051,10 @@ def get_best_phase(phases, chisq):
 
     # Calculate the vertex of the parabola.
     chisq_opt = [chisq[ibest_m1], chisq[ibest], chisq[ibest_p1]]
+    if np.any(~np.isfinite(chisq_opt)):
+        log.warning("Chi squared values in final optimization are unexpectedly NaN")
+        return phases[ibest]
+
     x = [phases[ibest_m1], phases[ibest], phases[ibest_p1]]
 
     # Just in case all of these chi squared values are equal
@@ -1177,7 +1181,14 @@ def __init__(self, all_y, all_n, phasefunc, phases_template, phaselist, dphase_f
         """
         Compute and store quantities needed for chi2, amplitude calculation.
 
-        Some of the math from the writeup goes in here.
+        This class precomputes a series of arrays so that the sums in
+        `calc_chisq_amplitudes` do not need to be fully recomputed at every
+        possible phase of the EMI signal.
+
+        Sums of pixel read values are precomputed over pixels that share the
+        same EMI phase to avoid double sums over pixels and reads later. Sums over
+        the EMI waveform itself are also precomputed at each trial phase offset for
+        the same reason.
 
         Parameters
         ----------