Polarised proton fits

doc/sphinx/source/tutorials/index.rst

@@ -14,6 +14,7 @@ Running fits
    ./run-fit.md
    ./run-iterated-fit.rst
    ./run-qed-fit.rst
+   ./polarized_fits.rst
    ./general_th_covmat.rst
    ./thcov_tutorial.rst
 

doc/sphinx/source/tutorials/polarized_fits.rst

@@ -0,0 +1,122 @@
+.. _polarized:
+
+How to run a Polarized fit
+==========================
+
+The user should first refer to the :ref:`detailed guide <n3fit-usage>` on how to run
+a standard unpolarized PDF fit. Most of the steps in that guide are still apply
+and in the following section we only highlight the differences.
+
+
+Preparing a runcard and running a fit
+-------------------------------------
+
+The writing of a polarized fit runcard follows exactly the guideline described
+in :ref:`Preparing a fit runcard <prepare-fits>`. The user only need to modify 
+and adjust various entries.
+
+In polarized fits, the positivity of the polarized PDFs are enforced using as
+a boundary condition an unpolarized PDF set. The information on how such a
+constraint is enforced is defined under the key ``positivity_bound`` as follows:
+
+.. code-block:: yaml
+
+  # Define the unpolarized PDF set to be used as BC for positivity
+  ...
+  positivity_bound:
+    unpolarized_bc: NNPDF40_nnlo_pch_as_01180
+    n_std: 1.00 # Standard Deviation to be added as Shift
+  ...
+
+where ``pdfbc`` specifies the name of the unpolarized PDF set to be used as a
+boundary condition while ``stderr`` specifies the shift in terms of the standard
+deviation to be applied to the PDF central values. If ``stderr=0.0`` then the
+PDF central values will be used to constrain their polarized counterparts.
+
+Given that polarized fits require different fitting bases and different theory
+constraints, the fields under the ``fitting`` key require adjustments.
+Specifically:
+
+.. code-block:: yaml
+
+  ...
+  fitting:
+    fitbasis: EVOL_POL
+    sum_rules: TSR
+    basis:
+    - {fl: sng, trainable: false, smallx: [1.094, 1.118], largex: [1.46, 3.003]}
+    - {fl: g, trainable: false, smallx: [0.8189, 1.844], largex: [2.591, 5.697]}
+    - {fl: t3, trainable: false, smallx: [-0.4401, 0.9163], largex: [1.773, 3.333]}
+    - {fl: t8, trainable: false, smallx: [0.5852, 0.8537], largex: [1.533, 3.436]}
+  ...
+
+where ``TSR`` specifies that only sum rules on the :math:`T_3` and :math:`T_8`
+distributions. If any of these values are specified different the program will
+raise an error.
+
+.. note::
+
+   While the treatment of integrability follows exactly the same concept as in the 
+   default unpolarized fits, the treatment of positivity in the polarized case 
+   requires specific treatment. Similar to the unpolarised fits, the cost function 
+   used to enforce the positivity is defined by the following observable:
+
+   .. math::
+     \chi_{\mathrm{tot}}^2 \rightarrow \chi_{\mathrm{tot}}^2+ \Lambda_{\rm POS} \sum_{k=1}^8 \sum_{i=1}^{n_i} \operatorname{ReLU}\left(-\mathcal{C}_k\left(x_i, Q^2\right)\right)
+
+   where:
+
+   .. math::
+     \mathrm{ReLU}(t)= \begin{cases}t & \text { if } t>0 \\ 0 & \text { if } t \leq 0\end{cases}
+
+   with :math:`n_i=20` and :math:`Q^2=5~\mathrm{GeV}^2` chosen to be the same as in the unpolarised 
+   case. In the polarized case, omitting the :math:`Q^2`-dependence, the expression of :math:`\mathcal{C}_k` 
+   is rather given by:
+
+     .. math::
+       \mathcal{C}_k(x_i) = \mathcal{F}_k(x_i) + \Sigma_k(x_i) - | \Delta \mathcal{F}_k(x_i)  |
+
+   where :math:`\Sigma_k(x_i)` represents the one standard deviation error computed at 
+   :math:`x_i` for the flavour :math:`k`. :math:`\mathcal{F}_k` can be a (p)PDF of 
+   flavour :math:`k` or the longitudinally (polarized) structure functions :math:`(k=1)`.
+   The way in which the unpolarized prediction uncertainties are accounted for during 
+   the fit is by sampling according to a normal distribution and ought to enforce the 
+   following inequality:
+
+   .. math::
+     \mathcal{N}_r \left( \mathcal{F}_k, \Sigma_k^2 \right) - | \Delta \mathcal{F}_k | \geq 0
+
+   where the subscript :math:`r` indicates that the random seed per replica is always 
+   different. In practice, when imposing the positivity at the level of PDFs, we enforce 
+   the constraints on the flavor combination :math:`\left( \Delta f_i + \Delta \bar{f}_i \right)`, 
+   that is :math:`(\Delta) \mathcal{F}_k \equiv (\Delta) f_k + (\Delta) \bar{f}_k`.
+
+Once the runcard is ready, the user can follow the guidelines :ref:`here <run-n3fit-fits>` 
+to set up and run fits.
+
+
+Evolving the fit
+----------------
+
+In order to evolve the fitted replicas, we have to use the polarized DGLAP evolution. This
+can simply be done by supplementing a flag to the ``evolven3fit```:
+
+.. code-block:: bash
+
+  evolven3fit evolve $runcard_folder --use_polarized
+
+Alternatively, the user can explicitly specify the path to the EKO using the flag ``--load``.
+
+
+Comparing polarized fits
+------------------------
+
+Additionally, a specific report template should be used when comparing two polarized
+fits. This can be done by simply using the ``--use_polarized`` when using ``vp-comparefits``: 
+
+.. code-block:: bash
+
+  vp-comparefits -i --use_polarized
+
+To read in details how to compare two fits, head to the :ref:`following <compare-fits>` 
+documentation.

doc/sphinx/source/tutorials/run-fit.md

@@ -15,6 +15,10 @@ The fitting methodology is detailed in the [Methodology](methodology) page.
 - [Upload and analyse the fit](#upload-and-analyse-the-fit)
 
 
+```eval_rst
+.. _prepare-fits:
+```
+
 Preparing a fit runcard
 -----------------------
 

extra_tests/regression_checks.py

@@ -26,6 +26,7 @@
     "trainable_prepro": 61,
     "no_lagrange": 27,
     "no_csr": 613,
+    "polarized_evol": 34,
 }
 
 

extra_tests/regression_fits/polarized_evol.yml

@@ -0,0 +1,80 @@
+############################################################
+# Configuration file for N3FIT regression Test
+############################################################
+description: N3FIT regression test for polarised fits
+
+############################################################
+dataset_inputs:
+- {dataset: E143_NC_4GEV_EP_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E143_NC_8GEV_ED_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E154_NC_9GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+
+############################################################
+datacuts:
+  t0pdfset: NNPDFpol10_100   # PDF set to generate t0 covmat
+  q2min: 1.00                # Q2 minimum
+  w2min: 4.00                # W2 minimum
+
+############################################################
+# Define the unpolarized PDF set to be used as BC for POS
+positivity_bound:
+  unpolarized_bc: NNPDF40_nnlo_pch_as_01180
+  n_stdv: 1.00 # Standard Deviation to be added as Error
+
+############################################################
+theory:
+  theoryid: 822
+
+############################################################
+trvlseed: 14613416
+nnseed: 1865683875
+mcseed: 746898422
+genrep: true
+
+save: "weights_polevol.h5"
+
+parameters:
+  nodes_per_layer: [25, 20, 4]
+  activation_per_layer: [tanh, tanh, linear]
+  initializer: glorot_normal
+  optimizer:
+    clipnorm: 6.073e-6
+    learning_rate: 2.621e-3
+    optimizer_name: Nadam
+  epochs: 20000
+  positivity:
+    initial: 1
+    multiplier:
+    threshold: 1e-7
+  integrability:
+    initial: 1
+    multiplier:
+    threshold: 1e-2
+  stopping_patience: 0.1
+  layer_type: dense
+  dropout: 0.0
+  threshold_chi2: 3.5
+
+fitting:
+  fitbasis: EVOL_POL
+  sum_rules: TSR
+  basis:
+  - {fl: sng, trainable: false, smallx: [1.094, 1.118], largex: [1.46, 3.003]}
+  - {fl: g, trainable: false, smallx: [0.8189, 1.844], largex: [2.591, 5.697]}
+  - {fl: t3, trainable: false, smallx: [-0.4401, 0.9163], largex: [1.773, 3.333]}
+  - {fl: t8, trainable: false, smallx: [0.5852, 0.8537], largex: [1.533, 3.436]}
+
+###########################################################
+positivity:
+  posdatasets:
+  - {dataset: NNPDF_POS_5GEV_XDELTAG, maxlambda: 1e8}
+  - {dataset: NNPDF_POS_5GEV_XDELTADQ, maxlambda: 1e8}
+
+############################################################
+integrability:
+  integdatasets:
+  - {dataset: NNPDF_INTEG_1GEV_XDELTAG, maxlambda: 1e2}
+  - {dataset: NNPDF_INTEG_1GEV_XDELTASIGMA, maxlambda: 1e2}
+
+############################################################
+debug: true

n3fit/runcards/examples/nnpdfpol20_dis_nnlo.yaml

@@ -0,0 +1,109 @@
+############################################################
+# Configuration file for N3FIT
+############################################################
+description: NNPDFpol2.0 NNLO DIS fit with ID822 (W2min=4)
+
+############################################################
+dataset_inputs:
+# -------------- Superseeded by COMPASS15 ---------------------
+# - {dataset: COMPASS_NC_17GEV_MUP_G1, frac: 0.60, cfac: [NRM]}
+# - {dataset: COMPASS_NC_25GEV_MUD_G1, frac: 0.60, cfac: [NRM]}
+# -------------------------------------------------------------
+- {dataset: COMPASS15_NC_17GEV_MUP_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: COMPASS15_NC_25GEV_MUD_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E142_NC_58GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E143_NC_4GEV_EP_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E143_NC_8GEV_ED_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E154_NC_9GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: E155_NC_9GEV_EN_G1F1RATIO, frac: 0.60}
+- {dataset: E155_NC_9GEV_EP_G1F1RATIO, frac: 0.60}
+- {dataset: EMC_NC_16GEV_MUP_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: HERMES_NC_7GEV_ED_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: HERMES_NC_7GEV_EP_G1, frac: 0.60, cfac: [NRM]}
+# ------------- Same as P+D due to Isospin --------------------
+# - {dataset: HERMES_NC_7GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+# -------------------------------------------------------------
+- {dataset: HERMES97_NC_7GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: JLABE06_NC_3GEV_EN_G1F1RATIO, frac: 0.60}
+- {dataset: JLABEG1B_NC_3GEV_EP_G1F1RATIO, frac: 0.60}
+- {dataset: JLABEG1B_NC_4GEV_ED_G1F1RATIO, frac: 0.60}
+- {dataset: JLABE97_NC_3GEV_EN_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: JLABE99_NC_3GEV_EN_G1F1RATIO, frac: 0.60}
+- {dataset: JLABEG1DVCS_NC_5GEV_ED_G1F1RATIO, frac: 0.60}
+- {dataset: JLABEG1DVCS_NC_3GEV_EP_G1F1RATIO, frac: 0.60}
+- {dataset: SMC_NC_24GEV_MUD_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: SMC_NC_17GEV_MUP_G1, frac: 0.60, cfac: [NRM]}
+- {dataset: SMCSX_NC_24GEV_MUD_G1F1RATIO, frac: 0.60}
+- {dataset: SMCSX_NC_17GEV_MUP_G1F1RATIO, frac: 0.60}
+
+############################################################
+datacuts:
+  t0pdfset: NNPDFpol10_100   # PDF set to generate t0 covmat
+  q2min: 1.00                # Q2 minimum
+  w2min: 4.00                # W2 minimum
+
+############################################################
+# Define the unpolarized PDF set to be used as BC for POS
+positivity_bound:
+  unpolarized_bc: NNPDF40_nnlo_pch_as_01180
+  n_stdv: 1.00 # Standard Deviation to be added as Error
+
+############################################################
+theory:
+  theoryid: 822
+
+############################################################
+trvlseed: 14613416
+nnseed: 1865683875
+mcseed: 746898422
+save: false
+genrep: true
+
+parameters:
+  nodes_per_layer: [25, 20, 4]
+  activation_per_layer: [tanh, tanh, linear]
+  initializer: glorot_normal
+  optimizer:
+    clipnorm: 6.073e-6
+    learning_rate: 2.621e-3
+    optimizer_name: Nadam
+  epochs: 20000
+  positivity:
+    initial: 1
+    multiplier:
+    threshold: 1e-7
+  integrability:
+    initial: 1
+    multiplier:
+    threshold: 1e-2
+  stopping_patience: 0.1
+  layer_type: dense
+  dropout: 0.0
+  threshold_chi2: 3.5
+
+fitting:
+  fitbasis: EVOL_POL
+  sum_rules: TSR
+  basis:
+  - {fl: sng, trainable: false, smallx: [1.094, 1.118], largex: [1.46, 3.003]}
+  - {fl: g, trainable: false, smallx: [0.8189, 1.844], largex: [2.591, 5.697]}
+  - {fl: t3, trainable: false, smallx: [-0.4401, 0.9163], largex: [1.773, 3.333]}
+  - {fl: t8, trainable: false, smallx: [0.5852, 0.8537], largex: [1.533, 3.436]}
+
+###########################################################
+positivity:
+  posdatasets:
+  - {dataset: NNPDF_POS_5GEV_XDELTAG, maxlambda: 1e8}
+  - {dataset: NNPDF_POS_5GEV_XDELTADQ, maxlambda: 1e8}
+  - {dataset: NNPDF_POS_5GEV_XDELTASQ, maxlambda: 1e8}
+  - {dataset: NNPDF_POS_5GEV_XDELTAUQ, maxlambda: 1e8}
+
+############################################################
+integrability:
+  integdatasets:
+  - {dataset: NNPDF_INTEG_1GEV_XDELTAG, maxlambda: 1e2}
+  - {dataset: NNPDF_INTEG_1GEV_XDELTASIGMA, maxlambda: 1e2}
+
+############################################################
+debug: false
+maxcores: 4

n3fit/src/n3fit/backends/keras_backend/operations.py

@@ -290,6 +290,21 @@ def pow(tensor, power):
     return tf.pow(tensor, power)
 
 
+@tf.function
+def absolute(tensor):
+    """
+    Compute the absolute value of a tensor
+    """
+    return K.abs(tensor)
+
+
+def multiply_minusone(tensor):
+    """
+    Multiply the elements of a given tensor by (-1)
+    """
+    return keras_Lambda(lambda x: -1 * x)(tensor)
+
+
 @tf.function(reduce_retracing=True)
 def op_log(o_tensor, **kwargs):
     """