# Copyright 2021 The NetKet Authors - All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from typing import Callable, Optional
from functools import partial
import jax
import jax.numpy as jnp
from netket import jax as nkjax
from netket.utils.types import Array, PyTree
[docs]
@partial(jax.jit, static_argnames=("apply_fun"))
def is_probably_holomorphic(
apply_fun: Callable[[PyTree, Array], Array],
parameters: PyTree,
samples: Array,
model_state: Optional[PyTree] = None,
) -> bool:
r"""
Check if a function :math:`\psi` is likely to be holomorphic almost
everywhere.
The check is done by verifying if the function satisfies on the provided
samples the
`Cauchy-Riemann equations <https://en.wikipedia.org/wiki/Cauchy–Riemann_equations>`_.
In particular, assuming that the function is complex-valued
:math:`\psi(\theta, x) = \psi_r(\theta, x) + i \psi_i(\theta, x)`
where :math:`\psi_r` and :math:`\psi_i` are it's real and imaginary part,
and the parameters can also be split in real and imaginary part according to
:math:`\theta = \theta_r + i\theta_i`, the conditions real
.. math::
\frac{\partial \psi_r(\theta, x)}{\partial \theta_r} =
\frac{\partial \psi_i(\theta, x)}{\partial \theta_i}
\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
\frac{\partial \psi_r(\theta, x)}{\partial \theta_i} =
-\frac{\partial \psi_i(\theta, x)}{\partial \theta_r}
Those conditions are always not verified if the parameters are
real (:math:`\theta_i=0`), in which case we automatically return False.
We verify the Cauchy-Riemann equations on the provided samples, so you should
provide a sufficiently large number of points to be relatively sure if your
function is holomorphic, because there are some functions that are locally non
holomorphic, however that is rare and in general you can trust the result of this
function.
.. Note::
If you are working with NetKet, you are most likely pretty good at math. In that
case, instead of relying on the computer to tell you if the function is holomorphic
or not you can do it by yourself! It's quite easy: the Cauchy-Riemann conditions
can be rewritten using `Wirtinger derivatives <https://en.wikipedia.org/wiki/Wirtinger_derivatives>`_
(where we treat the complex conjugate variables :math:`\theta` and :math:`\theta^\star`
as linearly independent) as
.. math::
\frac{\partial \psi(\theta, x)]}{\partial \theta^\star} = 0 .
Practically, this can be read as *if the function :math:`\psi` depends on
:math:`\theta^\star` it will not be holomorphic.* Therefore, a very
easy way to spot non-holomorphic functions is to verify if any operation
among the following appear in the function:
* complex conjugation
* absolute value
* real or imaginary part
.. Note::
In most Machine-Learning applications we are only interested in holomorphicity
almost-everywhere, therefore if the function is not holomorphic on one particular
point, such as :math:`\theta=0`, :math:`\theta=1` or similar this will in general
not matter because that is a set of dimension zero that we will never cross.
This is similar to the way the derivative of locally non-differentiable functions like
the modulus is obtained: we assume that we never enter those regions.
Args:
apply_fun: The forward pass of the variational function. This should be a Callable
accepting two inputs, the first being the parameters as a dictionary with at least
a key :code:`params`, and possibly other keys given by the :code:`model_state`. The
second argument to :code:`apply_fun` will be the samples.
parameters : a pytree of parameters to be passed as the first argument to the :code:`apply_fun`.
We will check if the ansatz is holomoprhic with respect to the derivatives of those
parameters.
samples : a set of samples.
model_state: optional dictionary of state parameters of the model.
"""
if nkjax.tree_leaf_isreal(parameters):
return False
samples = samples.reshape(-1, samples.shape[-1])
jacs = nkjax.jacobian(
apply_fun, parameters, samples, model_state=model_state, mode="complex"
)
# ∂ᵣψᵣ
dr_dpr = jax.tree_util.tree_map(lambda x: x[:, 0, ...], jacs.real)
# ∂ᵣψᵢ
dr_dpi = jax.tree_util.tree_map(lambda x: x[:, 1, ...], jacs.real)
# ∂ᵢψᵣ
di_dpr = jax.tree_util.tree_map(lambda x: x[:, 0, ...], jacs.imag)
# ∂ᵢψᵢ
di_dpi = jax.tree_util.tree_map(lambda x: x[:, 1, ...], jacs.imag)
# verify that ∂ᵣψᵣ == ∂ᵢψᵢ
cond1 = jax.tree_util.tree_map(lambda x, y: jnp.allclose(x, y), dr_dpr, di_dpi)
# verify that ∂ᵣψᵢ == -∂ᵢψᵣ
cond2 = jax.tree_util.tree_map(lambda x, y: jnp.allclose(x, -y), dr_dpi, di_dpr)
return jax.tree_util.tree_reduce(jnp.bitwise_and, (cond1, cond2))
