diff --git a/slip39/recovery/entropy.py b/slip39/recovery/entropy.py
index 1d376ef7..6f6dbffb 100644
--- a/slip39/recovery/entropy.py
+++ b/slip39/recovery/entropy.py
@@ -99,28 +99,23 @@ def idft( y: List[complex] ) -> List[complex]:
     """
     N				= len( y )
     #assert N <= 170, f"idft: {N=} too large."  # 512-bit entropy as 3-bit symbols
-    e_sk_N			= idft.e_sk.setdefault( N, [] )
-    if e_sk_N:
-        e_sk_N_i		= iter( e_sk_N )
-        result			= [
-            sum(
-                y_k * e_sk
-                for y_k,e_sk in zip( y, e_sk_N_i )
-            ) / N+0j
+    e_sk_N			= idft.e_sk.get( N )
+    if e_sk_N is None:
+        #print( f"idft: Memoizing {N=} multiplication factors" )
+        e_sk_N			= idft.e_sk[N] = [
+            cmath.exp( 2j * cmath.pi * n / N * k )
             for n in range( N )
+            for k in range( N )
         ]
-    else:
-        #print( f"idft: Memoizing {N=} multiplication factors" )
-        result			= []
-        for n in range( N ):
-            r			= 0
-            s			= 2j * cmath.pi * n / N
-            for k in range( N ):
-                e_sk		= cmath.exp( s * k )
-                e_sk_N.append( e_sk )
-                r	       += y[k] * e_sk
-            r		       /= N+0j
-            result.append( r )
+    # Always accumulate in the same order (not w/ sum(), which since Python 3.12 uses compensated
+    # summation, and would give slightly different results than w/ the same code before and after
+    # the factors are memoized).
+    result			= []
+    for n in range( N ):
+        r			= 0
+        for k in range( N ):
+            r		       += y[k] * e_sk_N[n * N + k]
+        result.append( r / ( N+0j ))
     return result
 idft.e_sk			= {}  # noqa: E305
 
diff --git a/slip39/recovery_test.py b/slip39/recovery_test.py
index a60e34ac..52aa3243 100644
--- a/slip39/recovery_test.py
+++ b/slip39/recovery_test.py
@@ -20,7 +20,7 @@
 
 from .api		import create, account, path_hardened
 from .recovery		import recover, recover_bip39, shannon_entropy, signal_entropy, analyze_entropy
-from .recovery.entropy	import fft, ifft, pfft, dft, dft_on_real, dft_to_rms_mags, entropy_bin_dfts, denoise_mags, signal_draw, signal_recover_real, scan_entropy
+from .recovery.entropy	import fft, ifft, pfft, dft, idft, dft_on_real, dft_to_rms_mags, entropy_bin_dfts, denoise_mags, signal_draw, signal_recover_real, scan_entropy
 from .dependency_test	import substitute, nonrandom_bytes, SEED_XMAS, SEED_ONES, SEED_ZERO
 from .util		import avg, rms, ordinal, commas, round_onto
 
@@ -1127,6 +1127,21 @@ def test_util():
     assert round_onto( None, [-5, -1, None, 5], keep_sign=True  ) is None
 
 
+def test_idft_memoized():
+    """The pure-Python idft (used by ifft when numpy isn't available) must give exactly the same
+    results before and after its multiplication factors are memoized.
+
+    """
+    x				= [ 2, 3, 5, 7, 11 ]
+    y				= dft( x )
+    idft.e_sk.pop( len( y ), None )  # Ensure the factors aren't memoized yet
+    first			= idft( y )
+    assert len( idft.e_sk[len( y )] ) == len( y ) ** 2  # and now they are
+    second			= idft( y )
+    assert first == second
+    assert [ f.real for f in first ] == pytest.approx( x )
+
+
 def test_dft_smoke():
     """Test some basic assumptions on DFTs"""
     print()