Pseudodata in diagonal basis and saving rotation

doc/sphinx/source/n3fit/methodology.rst

@@ -348,53 +348,90 @@ The figure above provides a schematic representation of this feature scaling met
 Diagonal basis
 --------------
 
-Performing the training and validation split without diagonalising the :math:`t_0` covmat :math:`C_{0}` neglects
-any correlations that may be present between training and validation data. To remedy this,
-we rotate to a basis in which the correlation matrix is diagonal before performing any training/validation split.
-Starting from the definition of the :math:`\chi^2` function in the NNPDF methodology, we have
+Training and validation data are obtained by performing a random split of the
+original data set. However, data points in the two sets are not necessarily
+statistically independent, as they may be correlated through the fitting
+covariance matrix :math:`C_{\rm fit}`. Here the fitting covariance matrix is the
+sum of the :math:`t_{0}` experimental covariance matrix :math:`C_{0}` and any
+theory covariance matrix :math:`C_{\rm th}` used in the fit, i.e., :math:`C_{\rm
+fit} = C_{0} + C_{\rm th}`. In order to disentangle the training and
+validation data, we perform the training-validation split in a basis in which
+the correlation matrix is diagonal.
+
+We first compute the correlation matrix :math:`\rho`, which is defined as
 
 .. math::
 
-    \chi^2 &= (D-T)^T C_0^{-1} (D-T) \\
-           &= (D-T)^T R^{-1} R C_0^{-1} R R^{-1} (D-T) \\
-           &= (D-T)^T R^{-1} \left( R^{-1} C_0 R^{-1} \right)^{-1} R^{-1} (D-T) \\
-           &\equiv \tilde{\epsilon}^T \rho^{-1} \tilde{\epsilon} \, ,
+    \rho = \Sigma^{-1} C_{\rm fit} \Sigma^{-1} \, ,
 
-where we have defined :math:`\tilde{\epsilon} \equiv R^{-1}(D-T)` and :math:`\rho = R^{-1} C_0 R^{-1}`.
+where we have defined :math:`\Sigma_{ij} = \sqrt{C_{\rm fit, ii}} \delta_{ij}` and
+:math:`(\Sigma^{-1})_{ij} = \frac{1}{\sqrt{C_{\rm fit, ii}}} \delta_{ij}`. The
+correlation matrix is a symmetric positive-definite matrix, and therefore it can
+be diagonalized by an orthogonal transformation. Therefore we can write
 
-Choosing :math:`R_{ii} = \sqrt{C_{0, ii}}`, we have that :math:`R^{-1} C_0 R^{-1}` coincides with the usual definition of the correlation matrix.
+.. math::
+
+    \rho = V \Lambda V^T \, ,
+
+where :math:`V` is an orthogonal matrix and :math:`\Lambda` is a diagonal matrix
+containing the eigenvalues of :math:`\rho`. The original fitting covariance
+matrix can then be written as
+
+.. math::
+
+    C_{\rm fit} &= \Sigma \rho \Sigma \\
+                &= (\Sigma V) \Lambda (V^T \Sigma) \\
+                &\equiv U \Lambda U^T \, ,
 
-Next, we move to the basis in which :math:`\rho` is diagonal. Writing :math:`\rho = \tilde{U}^T \tilde{\Lambda} \tilde{U}`, we find
+where we have defined the non-orthogonal matrix :math:`U = \Sigma V`. Its
+inverse defines the rotation matrix that diagonalizes the
+:math:`\chi^2`, and is given by :math:`R^T \equiv U^{-1} = V^T \Sigma^{-1}`.
+Therefore, the inverse of the fitting covariance matrix can be written as
 
 .. math::
 
-    \chi^2 &= \tilde{\epsilon}^T \rho^{-1} \tilde{\epsilon} \\
-           &= \tilde{\epsilon}^T (\tilde{U}^T \tilde{\Lambda} \tilde{U})^{-1} \tilde{\epsilon} \\
-           &= \tilde{\epsilon}^T \tilde{U}^T \tilde{\Lambda}^{-1} \tilde{U} \tilde{\epsilon} \\
-           &\equiv \tilde{\tilde{\epsilon}}^T \tilde{\Lambda}^{-1} \tilde{\tilde{\epsilon}} \, ,
+    C_{\rm fit}^{-1} &= (U \Lambda U^T)^{-1} \\
+                     &= (U^T)^{-1} \Lambda^{-1} U^{-1} \\
+                     &= R \Lambda^{-1} R^T \, .
 
-where on the last line we have defined
+Considering the definition of the :math:`\chi^2` function in the NNPDF
+methodology, we finally have
 
 .. math::
 
-    \tilde{\tilde{\epsilon}} \equiv \tilde{U}\tilde{\epsilon} = \tilde{U}R^{-1}(D-T).
+    \chi^2 &= (D-T)^T C_{\rm fit}^{-1} (D-T) \\
+           &= (D-T)^T R \Lambda^{-1} R^T (D-T) \\
+           &= \epsilon^T \Lambda^{-1} \epsilon \\
+           &= \lVert \epsilon \rVert^2_{\Lambda^{-1}} \, ,
+
+where we have defined the residuals in the diagonal basis as :math:`\epsilon \equiv R^T(D-T)` or, writing it in index notation,
+
+.. math::
+
+    \epsilon_i = (V^T)_{ij} \frac{(D-T)_j}{\sqrt{C_{\rm fit, jj}}} \,.
+
+In this basis, the :math:`\chi^2` becomes a weighted norm of the residuals,
+where the weights are given by the inverse of the eigenvalues of the correlation
+matrix.
 
-In index notation, this reads
+The transformed data :math:`\epsilon` are statistically independent in the
+diagonal basis of the correlation matrix :math:`\rho`. As a crosscheck, we
+can compute the covariance of :math:`\epsilon`,
 
 .. math::
 
-    \tilde{\tilde{\epsilon_i}} = \tilde{U}_{ij} \frac{(D-T)_j}{\sqrt{C_{0, jj}}}
+    \mathbb{E}[\epsilon \epsilon^T] &= \mathbb{E}[R^T(D-T)(D-T)^T R] \\
+                                    &= R^T \mathbb{E}[(D-T)(D-T)^T] R \\
+                                    &= R^T C_{\rm fit} R \\
+                                    &= R^T U \Lambda U^T R \\
+                                    &= \Lambda \,,
 
-The transformed data :math:`\tilde{\tilde{\epsilon}}` is statistically independent in the diagonal basis of the correlation matrix :math:`\rho`.
-Computing the covariance of :math:`\tilde{\tilde{\epsilon}}`,
+where we have used the fact that :math:`R^T U = I` and the assumption that the
+data are distributed according to the fitting covariance matrix :math:`C_{\rm fit}`
 
 .. math::
 
-    \mathbb{E}[\tilde{\tilde{\epsilon}}\tilde{\tilde{\epsilon}}^T]
-      &= \mathbb{E} \big[ (\tilde{U} R^{-1}(D-T)) (\tilde{U} R^{-1}(D-T))^T \big] \\
-      &= \tilde{U} R^{-1} \mathbb{E}[(D-T)(D-T)^T] R^{-1} \tilde{U}^T \\
-      &= \tilde{U} \rho \tilde{U}^T \\
-      &= \tilde{U}\tilde{U}^T \tilde{\Lambda} \tilde{U}\tilde{U}^T \\
-      &= \tilde{\Lambda} \, ,
+    \mathbb{E}[(D-T)(D-T)^T] = C_{\rm fit} \, .
 
-we find that it is diagonal, which demonstrates that the training/validation data are statistically independent indeed.
+This shows that the correlation is indeed diagonal, and demonstrates that the
+training/validation data are uncorrelated.

doc/sphinx/source/n3fit/runcard_detailed.rst

@@ -429,6 +429,13 @@ according to their experiment. Additionally, the union of these two is saved in
 ``<fit_directory>/replica_<number>/datacuts_theory_fitting_pseudodata_table.csv``
 if one is not interested in the exact nature of the splitting.
 
+When ``diagonal_basis: true`` is used (by default), the saved pseudodata indices are labeled as
+``eigenmode <i>`` corresponding to the diagonal basis used in the fit. In the presence of a theory covariance matrix,
+``vp-setupfit`` writes one file with the eigenvalues of the total correlation matrix and the rotation matrix that diagonalises
+the :math:`\chi2` under
+``<fit_directory>/tables/datacuts_theory_theorycovmatconfig_fitting_covmat_table.csv``.
+Without a theory covariance matrix, ``vp-setupfit`` writes this file instead under
+``<fit_directory>/tables/datacuts_theory_fitting_covmat_table.csv``.
 
 Imposing sum rules
 ^^^^^^^^^^^^^^^^^^

n3fit/src/n3fit/scripts/n3fit_exec.py

@@ -198,6 +198,9 @@ def from_yaml(cls, o, *args, **kwargs):
             )
             N3FIT_FIXED_CONFIG['point_prescriptions'] = thconfig.get('point_prescriptions')
             N3FIT_FIXED_CONFIG['user_covmat_path'] = thconfig.get('user_covmat_path')
+            # vp-setupfit has already written the theory-covmat CSV. n3fit
+            # should load it instead of rebuilding from scratch.
+            N3FIT_FIXED_CONFIG['load_thcovmat_from_file'] = True
 
         file_content.update(N3FIT_FIXED_CONFIG)
         return cls(file_content, *args, **kwargs)

n3fit/src/n3fit/scripts/vp_setupfit.py

@@ -25,6 +25,7 @@
 # top.
 
 
+import copy
 import hashlib
 import logging
 import pathlib
@@ -54,9 +55,10 @@
     'validphys.results',
     'validphys.theorycovariance.construction',
     'validphys.photon.compute',
+    'validphys.n3fit_data',
 ]
 
-SETUPFIT_DEFAULTS = dict(use_cuts='internal')
+SETUPFIT_DEFAULTS = dict(use_cuts='internal', use_t0=True)
 
 
 log = logging.getLogger(__name__)
@@ -141,6 +143,9 @@ class SetupFitConfig(Config):
 
     @classmethod
     def from_yaml(cls, o, *args, **kwargs):
+        # Create a fresh copy of the fixed config to avoid in-place modifications
+        fixed_config = copy.deepcopy(SETUPFIT_FIXED_CONFIG)
+
         try:
             file_content = yaml_safe.load(o)
         except error.YAMLError as e:
@@ -156,10 +161,10 @@ def from_yaml(cls, o, *args, **kwargs):
             # Use faketheoryid to create the L0 data to be stored into the filter folder
             # (L1 data is stored if fakedata is True)
             if 'faketheoryid' in closuredict:
-                # make sure theory key exists in SETUPFIT_FIXED_CONFIG
-                SETUPFIT_FIXED_CONFIG.setdefault('theory', {})
+                # make sure theory key exists in fixed_config
+                fixed_config.setdefault('theory', {})
                 # overwrite theoryid with the faketheoryid
-                SETUPFIT_FIXED_CONFIG['theory']['theoryid'] = closuredict['faketheoryid']
+                fixed_config['theory']['theoryid'] = closuredict['faketheoryid']
                 # download theoryid since it will be used in the fit
                 try:
                     loader.check_theoryID(file_content['theory']['theoryid'])
@@ -171,8 +176,18 @@ def from_yaml(cls, o, *args, **kwargs):
             filter_action = 'datacuts::theory::fitting filter'
             check_n3fit_action = 'datacuts::theory::fitting n3fit_checks_action'
 
+        # Add rotation action for the total covariance matrix
+
+        # if file_content.get('theorycovmatconfig') is not None:
+        if (thconfig := file_content.get('theorycovmatconfig')) is not None and thconfig.get(
+            'use_thcovmat_in_fitting', True
+        ):
+            rotation_action = 'datacuts::theory::theorycovmatconfig fitting_covmat_table'
+        else:
+            rotation_action = 'datacuts::theory fitting_covmat_table'
+
         # The settings for these actions depend on the presence of closuretest
-        SETUPFIT_FIXED_CONFIG['actions_'] += [check_n3fit_action, filter_action]
+        fixed_config['actions_'] += [check_n3fit_action, filter_action, rotation_action]
 
         # Check theory covariance matrix configuration
         thconfig = file_content.get('theorycovmatconfig', {})
@@ -184,7 +199,7 @@ def from_yaml(cls, o, *args, **kwargs):
                 "`point_prescriptions: ['9 point', '3 point']`"
             )
         if thconfig:
-            SETUPFIT_FIXED_CONFIG['actions_'].append(
+            fixed_config['actions_'].append(
                 'datacuts::theory::theorycovmatconfig nnfit_theory_covmat'
             )
 
@@ -214,14 +229,14 @@ def from_yaml(cls, o, *args, **kwargs):
             if compute_in_setupfit:
                 log.info("Forcing photon computation with FiatLux during setupfit.")
                 # Since the photon will be computed, check that the luxset and additional_errors exist
-                SETUPFIT_FIXED_CONFIG['actions_'].append('fiatlux check_luxset_exists')
+                fixed_config['actions_'].append('fiatlux check_luxset_exists')
                 if fiatlux.get("additional_errors"):
-                    SETUPFIT_FIXED_CONFIG['actions_'].append('fiatlux check_additional_errors')
-                SETUPFIT_FIXED_CONFIG['actions_'].append('fiatlux::theory compute_photon_to_disk')
+                    fixed_config['actions_'].append('fiatlux check_additional_errors')
+                fixed_config['actions_'].append('fiatlux::theory compute_photon_to_disk')
 
         # Check positivity bound
         if file_content.get('positivity_bound') is not None:
-            SETUPFIT_FIXED_CONFIG['actions_'].append('positivity_bound check_unpolarized_bc')
+            fixed_config['actions_'].append('positivity_bound check_unpolarized_bc')
 
         # Check hyperscan
         trials_config = file_content.get('trial_specs', {})
@@ -233,7 +248,7 @@ def from_yaml(cls, o, *args, **kwargs):
             file_content.setdefault(k, v)
 
         # Update file content with fixed configuration
-        file_content.update(SETUPFIT_FIXED_CONFIG)
+        file_content.update(fixed_config)
 
         return cls(file_content, *args, **kwargs)
 

n3fit/src/n3fit/tests/test_fit.py

@@ -283,8 +283,11 @@ def test_parallel_against_sequential(tmp_path, rep_from=6, rep_to=8):
         "ATLAS_TTBAR_8TEV_TOT_X-SEC",
         "CMS_SINGLETOP_13TEV_TCHANNEL-XSEC",
     ]
-    dataset_inputs = [{"dataset": d, "frac": 0.6, "variant": "legacy"} for d in datasets]
+    dataset_inputs = [{"dataset": d, "variant": "legacy"} for d in datasets]
     n3fit_input["dataset_inputs"] = dataset_inputs
+    # Using diaogonal basis
+    n3fit_input["diagonal_basis"] = True
+    n3fit_input["diagonal_frac"] = 0.5
     # Exit inmediately
     n3fit_input["parameters"]["epochs"] = 1
     # Save pseudodata
@@ -311,8 +314,9 @@ def test_parallel_against_sequential(tmp_path, rep_from=6, rep_to=8):
     for csvfile_seq in folder_seq.glob("*/*.csv"):
         csvfile_par = folder_par / csvfile_seq.relative_to(folder_seq)
 
-        result_seq = pd.read_csv(csvfile_seq, sep="\t", index_col=[0, 1, 2], header=0)
-        result_par = pd.read_csv(csvfile_par, sep="\t", index_col=[0, 1, 2], header=0)
+        # Diagonal basis writes single-index csv files
+        result_seq = pd.read_csv(csvfile_seq, sep="\t", index_col=[0], header=0)
+        result_par = pd.read_csv(csvfile_par, sep="\t", index_col=[0], header=0)
         pd.testing.assert_frame_equal(result_seq, result_par)
 
     # Check the rest of the fit, while numerical differences are expected between sequential

validphys2/src/validphys/config.py

@@ -805,16 +805,36 @@ def produce_experiment_from_input(self, experiment_input, theoryid, use_cuts, fi
         }
 
     @configparser.explicit_node
-    def produce_dataset_inputs_fitting_covmat(self, use_thcovmat_in_fitting=False):
+    def produce_dataset_inputs_fitting_covmat(
+        self, use_thcovmat_in_fitting=False, load_thcovmat_from_file=False
+    ):
         """
-        Produces the correct covmat to be used in fitting_data_dict according
-        to some options: whether to include the theory covmat, whether to
-        separate the multiplcative errors and whether to compute the
-        experimental covmat using the t0 prescription.
+        Dispatcher node for the covmat used in ``fitting_data_dict``.
+
+        Returns the experimental t0 covmat (``dataset_inputs_t0_exp_covmat``) when
+        no theory covmat is requested. When ``use_thcovmat_in_fitting=True``,
+        returns the total (experimental + theory) covmat — either rebuilt from
+        scratch via ``nnfit_theory_covmat`` (``load_thcovmat_from_file=False``,
+        the default) or with the theory part loaded from a CSV previously
+        written by ``vp-setupfit`` (``load_thcovmat_from_file=True``).
+
+        The two contexts:
+
+        * **vp-setupfit** — leaves ``load_thcovmat_from_file`` at the default. It
+          *writes* the theory covmat CSV, so the load variant would raise
+          FileNotFoundError on a file that does not yet exist. The result feeds
+          ``setupfit_fitting_covmat``, which serialises either the full fitting
+          covmat or its diagonal-basis rotation table.
+        * **n3fit** — sets ``load_thcovmat_from_file=True`` (see
+          ``n3fit_exec.py``). The result feeds ``_inv_covmat_prepared``, which
+          prepares the inverse for the fit.
         """
+
         from validphys import covmats
 
         if use_thcovmat_in_fitting:
+            if load_thcovmat_from_file:
+                return covmats.dataset_load_inputs_t0_total_covmat
             return covmats.dataset_inputs_t0_total_covmat
         return covmats.dataset_inputs_t0_exp_covmat
 
@@ -830,8 +850,13 @@ def produce_dataset_inputs_sampling_covmat(
         """
         Produces the correct MC replica method sampling covmat to be used in
         make_replica according to some options: whether to sample using a t0
-        covariance matrix, include the theory covmat and whether to
-        separate the multiplcative errors.
+        covariance matrix, include the theory covmat and whether to separate the
+        multiplcative errors.
+
+        This node is never invoked by setupfit, but is used in n3fit when
+        sampling the MC replicas for the fit (``make_replica``). It routes to
+        the load variants under ``use_thcovmat_in_sampling=True``, which load
+        the theory covmat from the CSV file generated by setupfit.
 
         Parameters
         ----------
@@ -851,9 +876,9 @@ def produce_dataset_inputs_sampling_covmat(
         if use_t0_sampling:
             if use_thcovmat_in_sampling:
                 if sep_mult:
-                    return covmats.dataset_inputs_t0_total_covmat_separate
+                    return covmats.dataset_load_inputs_t0_total_covmat_separate
                 else:
-                    return covmats.dataset_inputs_t0_total_covmat
+                    return covmats.dataset_load_inputs_t0_total_covmat
             else:
                 if sep_mult:
                     return covmats.dataset_inputs_t0_exp_covmat_separate
@@ -863,15 +888,28 @@ def produce_dataset_inputs_sampling_covmat(
         else:
             if use_thcovmat_in_sampling:
                 if sep_mult:
-                    return covmats.dataset_inputs_total_covmat_separate
+                    return covmats.dataset_load_inputs_total_covmat_separate
                 else:
-                    return covmats.dataset_inputs_total_covmat
+                    return covmats.dataset_load_inputs_total_covmat
             else:
                 if sep_mult:
                     return covmats.dataset_inputs_exp_covmat_separate
                 else:
                     return covmats.dataset_inputs_exp_covmat
 
+    def produce_fitting_covmat_name(self, fit):
+        """Produce the name of the covmat to be used in fitting,
+        according to how it was generated by vp-setupfit.
+        """
+        runcard = fit.as_input()
+        use_thcovmat = runcard.get("theorycovmatconfig", {}).get("use_thcovmat_in_fitting", False)
+        if use_thcovmat:
+            covmat_name = "datacuts_theory_theorycovmatconfig_fitting_covmat_table.csv"
+        else:
+            covmat_name = "datacuts_theory_fitting_covmat_table.csv"
+        path = fit.path / "tables" / covmat_name
+        return path
+
     def produce_loaded_theory_covmat(
         self,
         output_path,
@@ -885,6 +923,7 @@ def produce_loaded_theory_covmat(
         Loads the theory covmat from the correct file according to how it
         was generated by vp-setupfit.
         """
+
         if not use_thcovmat_in_sampling and not use_thcovmat_in_fitting:
             return 0.0
         # Load correct file according to how the thcovmat was generated by vp-setupfit
@@ -1328,6 +1367,7 @@ def produce_nnfit_theory_covmat(
         This function is only used in vp-setupfit to store the necessary covmats as .csv files in
         the tables directory.
         """
+
         if point_prescriptions is not None:
             if user_covmat_path is not None:
                 # Both scalevar and user uncertainties
@@ -1336,9 +1376,9 @@ def produce_nnfit_theory_covmat(
                 f = total_theory_covmat_fitting
             else:
                 # Only scalevar uncertainties
-                from validphys.theorycovariance.construction import theory_covmat_custom
+                from validphys.theorycovariance.construction import theory_covmat_custom_fitting
 
-                f = theory_covmat_custom
+                f = theory_covmat_custom_fitting
         elif user_covmat_path is not None:
             # Only user uncertainties
             from validphys.theorycovariance.construction import user_covmat_fitting