FPR Bugfixes and Feature Enhancements

jupyter_notebooks/Tutorials/algorithms/advanced/GST-FiducialPairReduction.ipynb

@@ -11,9 +11,11 @@
     "\n",
     "where $F_i$ is a \"preparation fiducial\" sequence, $F_j$ is a \"measurement fiducial\" sequence, and \"g_k\" is a \"germ\" sequence.  The repeated germ sequence $(g_k)^n$ we refer to as a \"germ-power\".  There are currently three different ways to reduce a standard set of GST operation sequences within pyGSTi, each of which removes certain $(F_i,F_j)$ fiducial pairs for certain germ-powers.\n",
     "\n",
-    "- **Global fiducial pair reduction (GFPR)** removes the same intelligently-selected set of fiducial pairs for all germ-powers.  This is a conceptually simple method of reducing the operation sequences, but it is the most computationally intensive since it repeatedly evaluates the number of amplified parameters for en *entire germ set*.  In practice, while it can give very large sequence reductions, its long run can make it prohibitive, and the \"per-germ\" reduction discussed next is used instead.\n",
+    "- **Global fiducial pair reduction (GFPR)** removes the same intelligently-selected set of fiducial pairs for all germ-powers.  This is a conceptually simple method of reducing the operation sequences, but it is the most computationally intensive since it repeatedly evaluates the number of amplified parameters for en *entire germ set*.  In practice, while it can give very large sequence reductions, its long run can make it prohibitive, and the \"per-germ\" reduction discussed next is used instead. \n",
+    "<span style=\"color:red\">Note: this form of FPR is deprecated on the latest versions of pygsti's develop branch. We now recommend using per-germ FPR instead. Also note that the current implementation of per-germ FPR will in most cases return smaller experiment designs than the legacy global FPR does.</span>\n",
+    "\n",
     "- **Per-germ fiducial pair reduction (PFPR)** removes the same intelligently-selected set of fiducial pairs for all powers of a given germ, but different sets are removed for different germs.  Since different germs amplify different directions in model space, it makes intuitive sense to specify different fiducial pair sets for different germs.  Because this method only considers one germ at a time, it is less computationally intensive than GFPR, and thus more practical.  Note, however, that PFPR usually results in less of a reduction of the operation sequences, since it does not (currently) take advantage overlaps in the amplified directions of different germs (i.e. if $g_1$ and $g_3$ both amplify two of the same directions, then GST doesn't need to know about these from both germs).\n",
-    "- **Random fiducial pair reduction (RFPR)** randomly chooses a different set of fiducial pairs to remove for each germ-power.  It is extremly fast to perform, as pairs are just randomly selected for removal, and in practice works well (i.e. does not impair Heisenberg-scaling) up until some critical fraction of the pairs are removed.  This reflects the fact that the direction detected by a fiducial pairs usually has some non-negligible overlap with each of the directions amplified by a germ, and it is the exceptional case that an amplified direction escapes undetected.  As such, the \"critical fraction\" which can usually be safely removed equals the ratio of amplified-parameters to germ-process-matrix-elements (typically $\\approx 1/d^2$ where $d$ is the Hilbert space dimension, so $1/4 = 25\\%$ for 1 qubit and $1/16 = 6.25\\%$ for 2 qubits).  RFPR can be combined with GFPR or PFPR so that some number of randomly chosen pairs can be added on top of the \"intelligently-chosen\" pairs of GFPR or PFPR.  In this way, one can vary the amount of sequence reduction (in order to trade off speed vs. robustness to non-Markovian noise) without inadvertently selecting too few or an especially bad set of random fiducial pairs.\n",
+    "- **Random per-germ power fiducial pair reduction (RFPR)** randomly chooses a different set of fiducial pairs to remove for each germ-power.  It is extremly fast to perform, as pairs are just randomly selected for removal, and in practice works well (i.e. does not impair Heisenberg-scaling) up until some critical fraction of the pairs are removed.  This reflects the fact that the direction detected by a fiducial pairs usually has some non-negligible overlap with each of the directions amplified by a germ, and it is the exceptional case that an amplified direction escapes undetected.  As such, the \"critical fraction\" which can usually be safely removed equals the ratio of amplified-parameters to germ-process-matrix-elements (typically $\\approx 1/d^2$ where $d$ is the Hilbert space dimension, so $1/4 = 25\\%$ for 1 qubit and $1/16 = 6.25\\%$ for 2 qubits).  RFPR can be combined with GFPR or PFPR so that some number of randomly chosen pairs can be added on top of the \"intelligently-chosen\" pairs of GFPR or PFPR.  In this way, one can vary the amount of sequence reduction (in order to trade off speed vs. robustness to non-Markovian noise) without inadvertently selecting too few or an especially bad set of random fiducial pairs.\n",
     "\n",
     "## Preliminaries\n",
     "\n",
@@ -129,7 +131,11 @@
     "fid_pairsDict = pygsti.alg.find_sufficient_fiducial_pairs_per_germ(\n",
     "                target_model, prep_fiducials, meas_fiducials, germs,\n",
     "                search_mode=\"random\", constrain_to_tp=True,\n",
-    "                n_random=100, seed=1234, verbosity=1,\n",
+    "                n_random=100, min_iterations=50,\n",
+    "                base_loweig_tol= .25, num_soln_returned=1,\n",
+    "                type_soln_returned= 'best',\n",
+    "                retry_for_smaller=True,\n",
+    "                seed=1234, verbosity=1,\n",
     "                mem_limit=int(2*(1024)**3))\n",
     "print(\"\\nPer-germ FPR to keep the pairs:\")\n",
     "for germ,pairsToKeep in fid_pairsDict.items():\n",