Drudge tutorial for beginners

Get started

Drudge is a library built on top of the SymPy computer algebra library for noncommutative and tensor alegbras. Usually for these style of problems, the symbolic manipulation and simplification of mathematical expressions requires a lot of context-dependent information, like the specific commutation rules and things like the dummy symbols to be used for different ranges. So the primary entry point for using the library is the Drudge class, which serves as a central repository of all kinds of domain-specific informations. To create a drudge instance, we need to give it a Spark context so that it is capable of parallelize things. For instance, to run things locally with all available cores, we can do

>>> from pyspark import SparkContext  
>>> spark_ctx = SparkContext('local[*]', 'drudge-tutorial')

For using Spark in cluster computing environment, please refer to the Spark documentation and setting of your cluster. With the spark context created, we can make the main entry point for drudge,

>>> import drudge
>>> dr = drudge.Drudge(spark_ctx)

Then from it, we can create the symbolic expressions as Tensor objects, which are basically mathematical expressions containing noncommutative objects and symbolic summations. For the noncommutativity, in spite of the availability of some basic support of it in SymPy, here we have the Vec class to specifically designate the noncommutativity of its multiplication. It can be created with a label and indexed with SymPy expressions.

>>> v = drudge.Vec('v')
>>> import sympy
>>> a = sympy.Symbol('a')
>>> str(v[a])
'v[a]'

For the symbolic summations, we have the Range class, which denotes a symbolic set that a variable could be summed over. It can be created by just a label.

>>> l = drudge.Range('L')

With these, we can create tensor objects by using the Drudge.sum() method,

>>> x = sympy.IndexedBase('x')
>>> tensor = dr.sum((a, l), x[a] * v[a])
>>> str(tensor)
'sum_{a} x[a] * v[a]'

Now we got a symbolic tensor of a sum of vectors modulated by a SymPy IndexedBase. Actually any type of SymPy expression can be used to modulate the noncommutative vectors.

>>> tensor = dr.sum((a, l), sympy.sin(a) * v[a])
>>> str(tensor)
'sum_{a} sin(a) * v[a]'

And we can also have multiple summations and product of the vectors.

>>> b = sympy.Symbol('b')
>>> tensor = dr.sum((a, l), (b, l), x[a, b] * v[a] * v[b])
>>> str(tensor)
'sum_{a, b} x[a, b] * v[a] * v[b]'

Of cause the multiplication of the vectors will not be commutative,

>>> tensor = dr.sum((a, l), (b, l), x[a, b] * v[b] * v[a])
>>> str(tensor)
'sum_{a, b} x[a, b] * v[b] * v[a]'

Normally, for each symbolic range, we have some traditional symbols used as dummies for summations over them, giving these information to drudge objects can be very helpful. Here in this demonstration, we can use the Drudge.set_dumms() method.

>>> dr.set_dumms(l, sympy.symbols('a b c d'))
[a, b, c, d]
>>> dr.add_resolver_for_dumms()

where the call to the Drudge.add_resolver_for_dumms() method could tell the drudge to interpret all the dummy symbols to be over the range that they are set to. By giving drudge object such domain-specific information, we can have a lot convenience. For instance, now we can use Einstein summation convention to create tensor object, without the need to spell all the summations out.

>>> tensor = dr.einst(x[a, b] * v[a] * v[b])
>>> str(tensor)
'sum_{a, b} x[a, b] * v[a] * v[b]'

Also the drudge knows what to do when more dummies are needed in mathematical operations. For instance, when we multiply things,

>>> tensor = dr.einst(x[a] * v[a])
>>> prod = tensor * tensor
>>> str(prod)
'sum_{a, b} x[a]*x[b] * v[a] * v[b]'

Here the dummy b is automatically used since the drudge object knows available dummies for its range. Also the range and the dummies are automatically added to the name archive of the drudge, which can be access by Drudge.names.

>>> p = dr.names
>>> p.L
Range('L')
>>> p.L_dumms
[a, b, c, d]
>>> p.d
d

Here in this example, we set the dummies ourselves by Drudge.set_dumms(). Normally, in subclasses of Drudge for different specific problems, such setting up is already finished within the class. We can just directly get what we need from the names archive. There is also a method Drudge.inject_names() for the convenience of interactive work.

Tensor manipulations

Now with tensors created by Drudge.sum() or Drudge.einst(), a lot of mathematical operations are available to them. In addition to the above example of (noncommutative) multiplication, we can also have the linear algebraic operations of addition and scalar multiplication.

>>> tensor = dr.einst(x[a] * v[a])
>>> y = sympy.IndexedBase('y')
>>> res = tensor + dr.einst(y[a] * v[a])
>>> str(res)
'sum_{a} x[a] * v[a]\n + sum_{a} y[a] * v[a]'

>>> res = 2 * tensor
>>> str(res)
'sum_{a} 2*x[a] * v[a]'

We can also perform some complex substitutions on either the vector or the amplitude part, by using the Drudge.subst() method.

>>> t = sympy.IndexedBase('t')
>>> w = drudge.Vec('w')
>>> substed = tensor.subst(v[a], dr.einst(t[a, b] * w[b]))
>>> str(substed)
'sum_{a, b} x[a]*t[a, b] * w[b]'

>>> substed = tensor.subst(x[a], sympy.sin(a))
>>> str(substed)
'sum_{a} sin(a) * v[a]'

Note that here the substituted vector does not have to match the left-hand side of the substitution exactly, pattern matching is done here. Other mathematical operations are also available, like symbolic differentiation by Tensor.diff() and commutation by | operator Tensor.__or__().

Tensors are purely mathematical expressions, while the utility class TensorDef can be construed as tensor expressions with a left-hand side. They can be easily created by Drudge.define() and Drudge.define_einst().

>>> v_def = dr.define_einst(v[a], t[a, b] * w[b])
>>> str(v_def)
'v[a] = sum_{b} t[a, b] * w[b]'

Their method TensorDef.act() is like a active voice version of Tensor.subst() and could come handy when we need to substitute the same definition in multiple inputs.

>>> res = v_def.act(tensor)
>>> str(res)
'sum_{a, b} x[a]*t[a, b] * w[b]'

More importantly, the definitions can be indexed directly, and the result is designed to work well inside Drudge.sum() or Drudge.einst(). For instance, for the same result, we could have,

>>> res = dr.einst(x[a] * v_def[a])
>>> str(res)
'sum_{b, a} x[a]*t[a, b] * w[b]'

When the only purpose of a vector or indexed base is to be substituted and we never intend to write tensor expressions directly in terms of them, we can just name the definition with a short name directly and put the actual base inside only. For instance,

>>> c = sympy.Symbol('c')
>>> f = dr.define_einst(sympy.IndexedBase('f')[a, b], x[a, c] * y[c, b])
>>> str(f)
'f[a, b] = sum_{c} x[a, c]*y[c, b]'
>>> str(dr.einst(f[a, a]))
'sum_{b, a} x[a, b]*y[b, a]'

which also demonstrates that the tensor definition facility can also be used for scalar quantities. TensorDef is also at the core of the code optimization and generation facility in the gristmill package.

Usually for tensorial problems, full simplification requires the utilization of some symmetries present on the indexed quantities by permutations among their indices. For instance, an anti-symmetric matrix entry changes sign when we transpose the two indices. Such information can be told to drudge by using the Drudge.set_symm() method, by giving generators of the symmetry group by Perm instances. For instance, we can do,

dr.set_symm(x, drudge.Perm([1, 0], drudge.NEG))

Then the master simplification algorithm in Tensor.simplify() is able to take full advantage of such information.

>>> tensor = dr.einst(x[a, b] * v[a] * v[b] + x[b, a] * v[a] * v[b])
>>> str(tensor)
'sum_{a, b} x[a, b] * v[a] * v[b]\n + sum_{a, b} x[b, a] * v[a] * v[b]'
>>> str(tensor.simplify())
'0'

Normally, drudge subclasses for specific problems add symmetries for some important indexed bases in the problem. And some drudge subclasses have helper methods for the setting of such symmetries, like FockDrudge.set_n_body_base() and FockDrudge.set_dbbar_base().

For the simplification of the noncommutative vector parts, the base Drudge class does not consider any commutation rules among the vectors. It works on the free algebra, while the subclasses could have the specific commutation rules added for the algebraic system. For instance, WickDrudge add abstract commutation rules where all the commutators have scalar values. Based on it, its special subclass FockDrudge implements the canonical commutation relations for bosons and the canonical anti-commutation relations for fermions. Also based on it, the subclass CliffordDrudge is capable of treating all kinds of Clifford algebras, like geometric algebra, Pauli matrices, Dirac matrices, and Majorana fermion operators. For algebraic systems where the commutator is not always a scalar, the abstract base class GenQuadDrudge can be used for basically all kinds of commutation rules. For instance, its subclass SU2LatticeDrudge can be used for su(2) algebra in Cartan-Weyl form.

These drudge subclasses only has the mathematical commutation rules implemented, for convenience in solving problems, many drudge subclasses are built-in with a lot of domain-specific information like the ranges and dummies, which are listed in Direct support of different problems. For instance, we can easily see the commutativity of two particle-hole excitation operators by using the PartHoleDrudge.

>>> phdr = drudge.PartHoleDrudge(spark_ctx)
>>> t = sympy.IndexedBase('t')
>>> u = sympy.IndexedBase('u')
>>> p = phdr.names
>>> a, i = p.a, p.i
>>> excit1 = phdr.einst(t[a, i] * p.c_dag[a] * p.c_[i])
>>> excit2 = phdr.einst(u[a, i] * p.c_dag[a] * p.c_[i])
>>> comm = excit1 | excit2
>>> str(comm)
'sum_{i, a, j, b} t[a, i]*u[b, j] * c[CR, a] * c[AN, i] * c[CR, b] * c[AN, j]\n + sum_{i, a, j, b} -t[a, i]*u[b, j] * c[CR, b] * c[AN, j] * c[CR, a] * c[AN, i]'
>>> str(comm.simplify())
'0'

Note that here basically all things related to the problem, like the vector for creation and annihilation operator, the conventional dummies a and i for particle and hole labels, are directly read from the name archive of the drudge. Problem-specific drudges are supposed to give such convenience.

In addition to providing context-dependent information for general tensor operations, drudge subclasses could also provide additional operations on tensors created from them. For instance, for the above commutator, we can directly compute the expectation value with respect to the Fermi vacuum by

>>> str(comm.eval_fermi_vev())
'0'

These additional operations are called tensor methods and are documented in the drudge subclasses.

Drudge scripts

For maximum flexibility, drudge has been designed to be a Python library from the beginning. However, in a lot of cases, like for small tasks or for users unfamiliar with the Python language or the Spark environment, a domain-specific language capable of making simple tasks simple can be desired. Drudge script is such a language for this purpose.

A drudge script is essentially a Python script heavily tweaked to be executed inside a special environment. So all Python lexicographical and syntactical rules apply. For a technical description of the pre-processing and execution drudge scripts, please see Drudge.exec_drs(). To execute a drudge script, we first need a Drudge object, such that the domain specific information about the current problem can be available. For this, we can either have a normal Python script, where a Drudge object is created with its Drudge.exec_drs() called with the source code for the drudge script, and execute it normally as Python scripts. Or drudge can also be used as the main program, either by python3 -m drudge or drudge. Then two files needs to be given as arguments. The first one is a configuration script, which is a normal Python script with a Drudge object assigned to a special variable DRUDGE. Then this Drudge object will be used for the execution of the actual drudge script given in the second argument.

As an example illustrating the basic principles and ease of drudge scripts, we assume that we are working on a drudge with a single range registered in the name archive as R. To create a symbolic definition of a matrix as a product of two matrices, suppose the drudge object can be accessed by a variable dr, we need to write something like:

p = dr.names
r = sympy.IndexedBase('r')
x = sympy.IndexedBase('x')
y = sympy.IndexedBase('y')
i, j, k = sympy.symbols('i j k')
def_ = dr.define(r, (i, p.R), (j, p.R), dr.sum((k, p.R), x[i, k] * y[k, j]))

which can be quite cumbersome for such a simple task. Suppose the drudge has a resolver capable of resolving any index to the range, we can write:

r = sympy.IndexedBase('r')
x = sympy.IndexedBase('x')
y = sympy.IndexedBase('y')
i, j, k = sympy.symbols('i j k')
def_ = dr.define_einst(r[i, j], x[i, k] * y[k, j])

which although is simplified a lot, still contains quite a lot of noise. Because of the Python execution model and scoping rules, the indexed bases and symbols must be explicitly created before they can be used.

Inside a drudge script, names in the name archive, all methods of the current drudge object, as well as names from the drudge, gristmill (if installed), and the SymPy package can directly be used without any qualification. More importantly, Symbol objects and IndexedBase objects are no longer needed to be explicitly created. All undefined names will be resolved as an atomic symbol, which can be construed as both a SymPy symbol and a SymPy IndexedBase. With these, the above definition can be simplified into:

def_ = define_einst(r[i, j], x[i, k] * y[k, j])

Due to the ubiquity of tensor definitions in common drudge tasks, a special operator <<= (Python left-shift augmented assignment operator) is introduced for the making definitions. With this, the above definition can be written as:

r[i, j] <<= sum((k, R), x[i, k] * y[k, j])

which makes the definition and put the definition in the name archive by Drudge.set_name(). So by default, the definition is put into the name archive under name r as a TensorDef object, and the base of the definition is put under name _r. Since names in the name archive do not need to be qualified in drudge scripts:

sum((k, R), r[i, k] * r[k, j])

directly gives us the chain product XYXY. And symbolic references to the r tensor without the concrete definition substituted in can still be made by using _r, like:

s = sum((k, R), _r[i, k] * _r[k, j])

which gives us the product RR. For this, the actual definition can be substituted explicitly when desired, for example, by:

s.subst(r)

which gives us XYXY.

Note that the definition by <<= is made by using the Drudge.def_() method. As a result, when the drudge property Drudge.default_einst() is set, Einstein summation convention is going to be automatically applied to the right-hand side. So we can simply write:

r[i, j] <<= x[i, k] * y[k, j]

when the ranges of i,j,k can be resolved by the drudge.

In cases where tainting of the global name archive is undesired for a tensor definition, we can use the <= operator, which simply returns the definition object without adding it to the name archive. For instance, to store the tensor definition in a variable def_, we can use:

def_ = r[i, j] <= x[i, k] * y[k, j]

This can be useful in functions inside drudge scripts.

Additionally, drudges could have more functions specifically to be used inside drudge scripts. For instance, in the base Drudge class, we have a simple constructor S, for converting strings to the special kind of symbols that can be indexed and used in <<= in drudge scripts. Also have sum_ for the actual Python built-in sum function, which is shadowed by the Drudge.sum() method.

For the taste of users without much object-oriented programming, inside drudge scripts, method calling like obj.meth(args) can also be written as meth(obj, args). For instance, for a tensor tensor:

simplify(tensor)

is equivalent to:

tensor.simplify()

Attribute access can be done in the same way, for instance,:

n_terms(tensor)

is equivalent to:

tensor.n_terms

Note that a caveat of this syntactic sugar is that the method name cannot be defined to be anything else before the calling. For instance,:

n_terms = 10
n_terms(tensor)

does not work, since n_terms is already defined to the integer 10, thus cannot be called any more. Another caveat is that static methods cannot be called in this way, which fortunately does not appear a lot in common usages of drudge.

For the convenience of symbolic computation, all integer literals inside drudge scripts are automatically resolved to SymPy integer values, rather than the built-in integer values. As a result, we can directly write:

1 / 2

for the rational value of one-half, without having to worry about the truncation or degradation to finite-precision floating-point numbers for Python integers. To access built-in integers, which is normally unnecessary, we can explicitly write something like int(1).

For convenience of users, some drudge functions has got slightly different behaviour inside drudge scripts. For instance, the Tensor.simplify() method will eagerly compute the result and repartition the terms among the workers. And tensors also have more readable string representation inside drudge scripts.

Examples on real-world applications

In this tutorial, some simple examples are run directly inside a Python interpreter. Actually drudge is designed to work inside Jupyter notebooks as well. By calling the Tensor.display() method, tensor objects can be mathematically displayed in Jupyter sessions. An example of interactive usage of drudge, we have a sample notebook in docs/examples/ccsd.ipynb in the project source. Also included is a general script gencc.py for the automatic derivation of coupled-cluster theories, mostly to demonstrate using drudge programmatically. And we also have a script for RCCSD theory to demonstrate its usage in large-scale spin-explicit coupled-cluster theories.

For drudge scripts, we have two example scripts both deriving the classical CCD theory. Both of them is based on the following configuration script conf_ph.py,

"""Configures a simple drudge for particle-hole model."""

from dummy_spark import SparkContext
from drudge import PartHoleDrudge

ctx = SparkContext()
dr = PartHoleDrudge(ctx)
dr.full_simplify = False

DRUDGE = dr

Here we only set a simple PartHoleDrudge without much modification. To illustrate the most basic usage of drudge scripts, we have example ccd.drs,

# A simple example on using drudge script for CCD theory
#
# In this example, the most basic aspects of using drudge scripts is
# illustrated.  It should be understandable for new-comers without much
# previous Python background.

# Define the cluster excitation operator.  Note that we need to inform the
# drudge that the t amplitude tensor has the double bar symmetry of
# t_{abij} = -t_{baij} = -t_{abji} = t_{baji}
set_dbbar_base(t, 2)

# Einstein summation convention can be used for easy tensor creation.
t2 = einst(
    t[a, b, i, j] * c_dag[a] * c_dag[b] * c_[j] * c_[i] / 4
)

# Get the similarity-transformed Hamiltonian.  Note that ``|`` operator
# computes the commutator between operators.
c0 = ham
c1 = simplify(c0 | t2)
c2 = simplify(c1 | t2)
c3 = simplify(c2 | t2)
c4 = simplify(c3 | t2)
h_bar = simplify(
    c0 + c1 + (1/2) * c2 + (1/6) * c3 + (1/24) * c4
)

print('Similarity-transformed Hamiltonian has {} terms'.format(
    n_terms(h_bar)
))

# Derive the working equations by projection.
en_eqn = simplify(eval_fermi_vev(h_bar))
proj = c_dag[i] * c_dag[j] * c_[b] * c_[a]
t2_eqn = simplify(eval_fermi_vev(proj * h_bar))

print('Working equation derived!')

with report('ccd.html', 'CCD theory') as rep:
    rep.add('Energy equation', en_eqn)
    rep.add('Doubles amplitude equation', t2_eqn)

With the comment described in the above script, we can see that drudge script can bare a lot of resemblance to the mathematical notation. To make a derivation of the many-body theory, we basically just use the operators like +, *, and | to do arithmetic operations on the tensors and use simplify to get the result simplified.

For another more advanced example, we have the ccd_adv.drs script,

# An advanced example on using drudge script for CCD theory
#
# In this example, it is emphasized that drudge scripts are just Python scripts
# with special execution.  So all Python constructions can be used for our
# convenience.  At the same time, by using drudge scripts, we can have all the
# syntactical sugar for making symbolic computation easy.
#

set_dbbar_base(t, 2)
t2 = einst(
    t[a, b, i, j] * c_dag[a] * c_dag[b] * c_[j] * c_[i] / 4
)

def compute_h_bar():
    """Compute the similarity transformed Hamiltonian."""
    # Here we use a Python loop to get the nested commutators.
    curr = ham
    h_bar = ham
    for order in range(0, 4):
        curr = simplify(curr | t2) / (order + 1)
        h_bar += curr
    return simplify(h_bar)

# By using the `memoise` function, the result can be automatically dumped into
# the given pickle file, and read from it if it is already available.  This can
# be convenient for large multi-step jobs.
h_bar = memoize(compute_h_bar, 'h_bar.pickle')
print('H-bar has {} terms'.format(n_terms(h_bar)))

# Derive the working equations by projection.  Here we make them into tensor
# definition with explicit left-hand side, so that they can be used for
# optimization.
e <<= simplify(eval_fermi_vev(h_bar))
proj = c_dag[i] * c_dag[j] * c_[b] * c_[a]
r2[a, b, i, j] <<= simplify(eval_fermi_vev(proj * h_bar))

print('Working equation derived!')

# When the gristmill package is also installed, the evaluation of the working
# equations can also be optimized with it.
eval_seq = optimize(
    [e, r2], substs={no: 1000, nv: 10000}
)

# In addition to HTML report, we can also have LaTeX report.  Note that the
# report can be structured into sections with descriptions.  For LaTeX output,
# the `dmath` environment from the `breqn` package can be used to break lines
# automatically inside large equations.

# Long descriptions of contents can be put in Python multi-line strings.
opt_description = """
The optimization is based on 1000 occupied orbitals and 10000 virtual orbitals,
which should be representative of common problems for CCD theory.
"""

with report('ccd.tex', 'CCD theory') as rep:
    rep.add(title='Working equations')
    rep.add(content=e, description='The energy equation')
    rep.add(content=r2, description='Doubles amplitude equation', env='dmath')
    rep.add(title='Optimized evaluation', description=opt_description)
    for step in eval_seq:
        rep.add(content=step, env='dmath')

In the example ccd.drs, it is attempted to be emphasized that drudge scripts are very similar to common mathematical notation and should be easy to get started. In this ccd_adv.drs example, the power and flexibility of drudge scripts being actually Python scripts is emphasized. Foremost, rather than spelling each order of commutation out, here the similarity-transformed Hamiltonian ˉH is computed by using a Python loop. This can be helpful for repetitive tasks. Also the computation of ˉH is put inside a function. Being able to define and execute functions makes it easy to reuse code inside drudge scripts. Here, the function is given to the Drudge.memoize() function. So its result is automatically dumped into the given pickle file. When the file is already there, the result will be directly read and used with the execution of the function skipped. This can be helpful for large multi-step jobs.

Note that <<= is used to make the working equations as tensor definitions of class TensorDef. In drudge scripts,:

variable = tensor

assigns the tensor tensor to the variable variable. The variable is a normal Python variable and works in the normal Python way. And the tensor is just a static expression of its mathematical content, with all the free symbols being free. At the same time,:

lhs <<= tensor

defines the lhs as the tensor, with the definition pushed into the name archive of the drudge. By using TensorDef objects, we also have a left-hand side, which enables the accompanying gristmill package to optimize the evaluation of the entire array by its advanced algorithms.

For the result, here they are written into a very structured LaTeX output, which can be easily compiled into PDF files. Note that by using the Report.add() function with different arguments, we can create structured report with sections and descriptions for the equations.

Note about importing drudge

In this tutorial, import drudge and import sympy is used and we need to give fully-qualified name to refer to objects in them. Normally, it can be convenient to use from drudge import * to import everything from drudge. For these cases, it needs to be careful that the importation of all objects from drudge needs to follow the importation of all objects from SymPy, or the SymPy Range class will shallow the actual class for symbolic range in drudge.