Technical Overview

Three Central Classes

From a bird’s eye perspective, pyMOR is a collection of generic algorithms operating on objects of the following types:


Vector arrays are ordered collections of vectors. Each vector of the array must be of the same dimension. Vectors can be copied to a new array, appended to an existing array or deleted from the array. Basic linear algebra operations can be performed on the vectors of the array: vectors can be scaled in-place, the BLAS axpy operation is supported and inner products between vectors can be formed. Linear combinations of vectors can be formed using the lincomb method. Moreover, various norms can be computed and selected dofs of the vectors can be extracted for empirical interpolation. To act on subsets of vectors of an array, arrays can be indexed with an integer, a list of integers or a slice, in each case returning a new VectorArray which acts as a modifiable view onto the respective vectors in the original array. As a convenience, many of Python’s math operators are implemented in terms of the interface methods.

Note that there is not the notion of a single vector in pyMOR. The main reason for this design choice is to take advantage of vectorized implementations like NumpyVectorArray which internally store the vectors as two-dimensional NumPy arrays. As an example, the application of a linear matrix based operator to an array via the apply method boils down to a call to NumPy’s optimized dot method. If there were only lists of vectors in pyMOR, the above matrix-matrix multiplication would have to be expressed by a loop of matrix-vector multiplications. However, when working with external solvers, vector arrays will often be given as lists of individual vector objects. For this use-case we provide ListVectorArray, a VectorArray based on a Python list of vectors.

Associated to each vector array is a VectorSpace which acts as a factory for new arrays of a given type. New vector arrays can be created using the zeros and empty methods. To wrap the raw objects of the underlying linear algebra backend into a new VectorArray, make_array is used.

The data needed to define a new VectorSpace largely depends on the implementation of the underlying backend. For NumpyVectorSpace, the only required datum is the dimension of the contained vectors. VectorSpaces for other backends could, e.g., hold a socket for communication with a specific PDE solver instance. Additionally, each VectorSpace has a string id, defaulting to None, which is used to signify the mathematical identity of the given space.

Two arrays in pyMOR are compatible (e.g. can be added) if they are from the same VectorSpace. If a VectorArray is contained in a given VectorSpace can be tested with the in operator.


The main property of operators in pyMOR is that they can be applied to VectorArrays resulting in a new VectorArray. For this operation to be allowed, the operator’s source VectorSpace must be identical with the VectorSpace of the given array. The result will be a vector array from the range space. An operator can be linear or not. The apply_inverse method provides an interface for (linear) solvers.

Operators in pyMOR are also used to represent bilinear forms via the apply2 method. A functional in pyMOR is simply an operator with NumpyVectorSpace(1) as range. Dually, a vector-like operator is an operator with NumpyVectorSpace(1) as source. Such vector-like operators are used in pyMOR to represent Parameter-dependent vectors such as the initial data of an InstationaryModel. For linear functionals and vector-like operators, the as_vector method can be called to obtain a vector representation of the operator as a VectorArray of length 1.

Linear combinations of operators can be formed using a LincombOperator. When such a linear combination is assembled, _assemble_lincomb is called to ensure that, for instance, linear combinations of operators represented by a matrix lead to a new operator holding the linear combination of the matrices.

For many interface methods default implementations are provided which may be overridden with operator-specific code. Base classes for NumPy-based operators can be found in pymor.operators.numpy. Several methods for constructing new operators from existing ones are contained in pymor.operators.constructions.


Models in pyMOR encode the mathematical structure of a given discrete problem by acting as container classes for Operators. Each model object has Operators and the products dictionary of Operators which appear in the formulation of the discrete problem. The keys in the products dictionary describe the role of the respective product in the discrete problem.

Apart from describing the discrete problem, models also implement algorithms for solving the given problem, returning VectorArrays from the solution_space. The solution can be cached, s.t. subsequent solving of the problem for the same parameter values reduces to looking up the solution in pyMOR’s cache.

While special model classes may be implemented which make use of the specific types of operators they contain (e.g. using some external high-dimensional solver for the problem), it is generally favorable to implement the solution algorithms only through the interfaces provided by the operators contained in the model, as this allows to use the same model class to solve high-dimensional and reduced problems. This has been done for the simple stationary and instationary models found in pymor.models.basic.

Models can also implement estimate_error and visualize methods to estimate the discretization or model reduction error of a computed solution and create graphic representations of VectorArrays from the solution_space.

Base Classes

While VectorArrays are mutable objects, both Operators and Models are immutable in pyMOR: the application of an Operator to the same VectorArray will always lead to the same result, solving a Model for the same parameter will always produce the same solution array. This has two main benefits:

  1. If multiple objects/algorithms hold references to the same Operator or Model, none of the objects has to worry that the referenced object changes without their knowledge.

  2. The return value of a method of an immutable object only depends on its arguments, allowing reliable caching of these return values.

A class can be made immutable in pyMOR by deriving from ImmutableObject, which ensures that write access to the object’s attributes is prohibited after __init__ has been executed. However, note that changes to private attributes (attributes whose name starts with _) are still allowed. It lies in the implementors responsibility to ensure that changes to these attributes do not affect the outcome of calls to relevant interface methods. As an example, a call to enable_caching will set the objects private __cache_region attribute, which might affect the speed of a subsequent solve call, but not its result.

Of course, in many situations one may wish to change properties of an immutable object, e.g. the number of timesteps for a given model. This can be easily achieved using the with_ method every immutable object has: a call of the form o.with_(a=x, b=y) will return a copy of o in which the attribute a now has the value x and the attribute b the value y. It can be generally assumed that calls to with_ are inexpensive. The set of allowed arguments can be found in the _init_arguments attribute.

All immutable classes in pyMOR and most other classes derive from BasicObject which, through its meta class, provides several convenience features for pyMOR. Most notably, every subclass of BasicObject obtains its own logger instance with a class specific prefix.

Creating Models

pyMOR ships a small (and still quite incomplete) framework for creating finite element or finite volume discretizations based on the NumPy/Scipy software stack. To end up with an appropriate Model, one starts by instantiating an analytical problem which describes the problem we want to discretize. analytical problems contain Functions which define the analytical data functions associated with the problem and a DomainDescription that provides a geometrical definition of the domain the problem is posed on and associates a boundary type to each part of its boundary.

To obtain a Model from an analytical problem we use a discretizer. A discretizer will first mesh the computational domain by feeding the DomainDescription into a domaindiscretizer which will return the Grid along with a BoundaryInfo associating boundary entities with boundary types. Next, the Grid, BoundaryInfo and the various data functions of the analytical problem are used to instatiate finite element or finite volume operators. Finally these operators are used to instatiate one of the provided Model classes.

In pyMOR, analytical problems, Functions, DomainDescriptions, BoundaryInfos and Grids are all immutable, enabling efficient disk caching for the resulting Models, persistent over various runs of the applications written with pyMOR.

While pyMOR’s internal discretizations are useful for getting started quickly with model reduction experiments, pyMOR’s main goal is to allow the reduction of models provided by external solvers. In order to do so, all that needs to be done is to provide VectorArrays, Operators and Models which interact appropriately with the solver. pyMOR makes no assumption on how the communication with the solver is managed. For instance, communication could take place via a network protocol or job files. In particular it should be stressed that in general no communication of high-dimensional data between the solver and pyMOR is necessary: VectorArrays can merely hold handles to data in the solver’s memory or some on-disk database. Where possible, we favor, however, a deep integration of the solver with pyMOR by linking the solver code as a Python extension module. This allows Python to directly access the solver’s data structures which can be used to quickly add features to the high-dimensional code without any recompilation. A minimal example for such an integration using pybind11 can be found in the src/pymordemos/minimal_cpp_demo directory of the pyMOR repository. Bindings for FEnicS and NGSolve packages are available in the bindings.fenics and bindings.ngsolve modules. The pymor-deal.II repository contains bindings for deal.II.


pyMOR classes implement dependence on a parameter by deriving from the ParametricObject base class. This class gives each instance a parameters attribute describing the Parameters the object and its relevant methods (apply, solve, evaluate, etc.) depend on. Each Parameter in pyMOR has a name and a fixed dimension, i.e. the number of scalar components of the Parameter. Scalar parameters are simply represented by one-dimensional Parameters. To assign concrete values to Parameters the specialized dict-like class Mu is used. In particular, it ensures, that all of its values are one-dimensional NumPy arrays.

The Parameters of a ParametricObject are usually automatically derived as the union of all Parameters of the objects that are passed to it’s __init__ method. For instance, an Operator that implements the L2-product with some user-provided Function will automatically inherit all Parameters of that Function. Additional Parameters can be easily added by setting the parameters_own attribute.


pyMOR offers a convenient mechanism for handling default values such as solver tolerances, cache sizes, log levels, etc. Each default in pyMOR is the default value of an optional argument of some function. Such an argument is made a default by decorating the function with the defaults decorator:

def some_algorithm(x, y, tolerance=1e-5)

Default values can be changed by calling set_defaults. By calling print_defaults a summary of all defaults in pyMOR and their values can be printed. A configuration file with all defaults can be obtained with write_defaults_to_file. This file can then be loaded, either programmatically or automatically by setting the PYMOR_DEFAULTS environment variable.

As an additional feature, if None is passed as value for a function argument which is a default, its default value is used instead of None. This allows writing code of the following form:

def method_called_by_user(U, V, tolerance_for_algorithm=None):
    algorithm(U, V, tolerance=tolerance_for_algorithm)

See the defaults module for more information.


Many algorithms in pyMOR can be seen as transformations acting on trees of Operators. One example is the structure-preserving (Petrov-)Galerkin projection of Operators performed by the project method. For instance, a LincombOperator is projected by replacing all its children (the Operators forming the affine decomposition) with projected Operators.

During development of pyMOR, it turned out that using inheritance for selecting the action to be taken to project a specific operator (i.e. single dispatch based on the class of the to-be-projected Operator) is not sufficiently flexible. With pyMOR 0.5 we have introduced algorithms which are based on RuleTables instead of inheritance. A RuleTable is simply an ordered list of rules, i.e. pairs of conditions to match with corresponding actions. When a RuleTable is applied to an object (e.g. an Operator), the action associated with the first matching rule in the table is executed. As part of the action, the RuleTable can be easily applied recursively to the children of the given object.

This approach has several advantages over an inheritance-based model:

  • Rules can match based on the class of the object, but also on more general conditions, i.e. the name of the Operator or being linear and non-parametric.

  • The entire mathematical algorithm can be specified in a single file even when the definition of the possible classes the algorithm can be applied to is scattered over various files.

  • The precedence of rules is directly apparent from the definition of the RuleTable.

  • Generic rules (e.g. the projection of a linear non-parametric Operator by simply applying the basis) can be easily scheduled to take precedence over more specific rules.

  • Users can implement or modify RuleTables without modification of the classes shipped with pyMOR.

The Reduction Process

The reduction process in pyMOR is handled by so called reductors, where each reductor takes arbitrary Models of associated type, and additional data (e.g. the reduced basis) to create reduced Models. For instance, the StationaryRBReductor and InstationaryRBReductor reductors can reduce any StationaryModel and InstationaryModel, while the BTReductor reduces LTIModel (just to name a few).

If proper offline/online decomposition is achieved by the reductor, the reduced Model will not store any high-dimensional data. Note that there is no inherent distinction between low- and high-dimensional Models in pyMOR. The only difference lies in the different types of operators, the Model contains.

In particular, in most projection-based MOR methods, only the Operators of a given Model are projected onto the reduced-basis space, whereas the structure of the problem (i.e. the type of the Model) stays the same. The actual projection of an individual Operator is performed by the generic project algorithm, which is able to perform a Petrov-Galerkin projection of any Operator available to pyMOR. The algorithm automatically exploits the structure of the given Operator (linearity, parameter separability, etc.) where possible to decouple the evaluation of the projected Operator from any high-dimensional spaces.

In addition to the projection of the Model, reductors may also assemble efficient offline-online decomposed a posterior error estimates (available via the estimate_error method of the resulting reduced order Model), if more information about the underlying problem yielding the full order Model is available (a popular example is the CoerciveRBReductor for Models representing stationary coercive problems).

If you want to further dive into the inner workings of pyMOR, we recommend to study the Projecting a Model tutorial.