Derived module from dmdbase.py for hankel dmd.

Reference: - H. Arbabi, I. Mezic, Ergodic theory, dynamic mode decomposition, and computation of spectral properties of the Koopman operator. SIAM Journal on Applied Dynamical Systems, 2017, 16.4: 2096-2126.


Get the reduced Koopman operator A, called A tilde.


Get the timesteps of the reconstructed states.


Get the time evolution of each mode.


Get the eigenvalues of A tilde.


Compute the Dynamic Modes Decomposition to the input data.


Get the matrix containing the DMD modes, stored by column.


Get the timesteps of the original snapshot.


Plot the eigenvalues.


Plot the DMD Modes.


Plot the snapshots.


Build a collection of all the available versions of the given timeindex.


Get the reconstructed data.


Get the original input data.

class HankelDMD(svd_rank=0, tlsq_rank=0, exact=False, opt=False, rescale_mode=None, forward_backward=False, d=1, sorted_eigs=False, reconstruction_method='first')[source]

Bases: pydmd.dmdbase.DMDBase

Hankel Dynamic Mode Decomposition

  • svd_rank (int or float) – the rank for the truncation; If 0, the method computes the optimal rank and uses it for truncation; if positive interger, the method uses the argument for the truncation; if float between 0 and 1, the rank is the number of the biggest singular values that are needed to reach the ‘energy’ specified by svd_rank; if -1, the method does not compute truncation.

  • tlsq_rank (int) – rank truncation computing Total Least Square. Default is 0, that means no truncation.

  • exact (bool) – flag to compute either exact DMD or projected DMD. Default is False.

  • opt (bool or int) – argument to control the computation of DMD modes amplitudes. See DMDBase. Default is False.

  • rescale_mode ({'auto'} or None or numpy.ndarray) – Scale Atilde as shown in 10.1016/j.jneumeth.2015.10.010 (section 2.4) before computing its eigendecomposition. None means no rescaling, ‘auto’ means automatic rescaling using singular values, otherwise the scaling factors.

  • forward_backward (bool) – If True, the low-rank operator is computed like in fbDMD (reference: https://arxiv.org/abs/1507.02264). Default is False.

  • d (int) – the new order for spatial dimension of the input snapshots. Default is 1.

  • sorted_eigs ({'real', 'abs'} or False) – Sort eigenvalues (and modes/dynamics accordingly) by magnitude if sorted_eigs=’abs’, by real part (and then by imaginary part to break ties) if sorted_eigs=’real’. Default: False.

  • reconstruction_method ({'first', 'mean'} or array-like) – Method used to reconstruct the snapshots of the dynamical system from the multiple versions available due to how HankelDMD is conceived. If ‘first’ (default) the first version available is selected (i.e. the nearest to the 0-th row in the augmented matrix). If ‘mean’ we compute the element-wise mean. If reconstruction_method is an array of float values we compute the weighted average (for each snapshots) using the given values as weights (the number of weights must be equal to d).


Return the first occurrence of each snapshot available in the given matrix (which must be the result of self._sub_dmd.reconstructed_data, or have the same shape).


reconstructions (np.ndarray) – A matrix of (higher-order) snapshots having shape (space*self.d, time_instants)


The first snapshot that occurs in reconstructions for each available time instant.

Return type



For a given t such that there is k \in \mathbb{N} such that t = t_0 + k dt, return the index of the first column in Hankel pseudo matrix (see also _pseudo_hankel_matrix()) which contains the snapshot corresponding to t.


time – The time corresponding to the requested snapshot.


The index of the first appeareance of time in the columns of Hankel pseudo matrix.

Return type



Method for arranging the input snapshots X into the (pseudo) Hankel matrix. The attribute d controls the shape of the output matrix. :Example:

>>> from pydmd import HankelDMD
>>> dmd = HankelDMD(d=2)
>>> a = np.array([[1, 2, 3, 4, 5]])
>>> dmd._pseudo_hankel_matrix(a)
array([[1, 2, 3, 4],
       [2, 3, 4, 5]])
>>> dmd = pydmd.hankeldmd.HankelDMD(d=4)
>>> dmd._pseudo_hankel_matrix(a)
array([[1, 2],
       [2, 3],
       [3, 4],
       [4, 5]])

Update the time dictionaries (dmd_time and original_time) of the auxiliary DMD instance HankelDMD._sub_dmd after an update of the time dictionaries of the time dictionaries of this instance of the higher level instance of HankelDMD.

property amplitudes

Get the coefficients that minimize the error between the original system and the reconstructed one. For futher information, see dmdbase._compute_amplitudes.


the array that contains the amplitudes coefficient.

Return type


property d

The new order for spatial dimension of the input snapshots.

property eigs

Get the eigenvalues of A tilde.


the eigenvalues from the eigendecomposition of atilde.

Return type



Compute the Dynamic Modes Decomposition to the input data.


X (numpy.ndarray or iterable) – the input snapshots.

property modes

Get the matrix containing the DMD modes, stored by column.


the matrix containing the DMD modes.

Return type


property modes_activation_bitmask

Get the bitmask which controls which DMD modes are enabled at the moment in this DMD instance.

The DMD instance must be fitted before this property becomes valid. After fit() is called, the defalt value of modes_activation_bitmask is an array of True values of the same shape of amplitudes().

The array returned is read-only (this allow us to react appropriately to changes in the bitmask). In order to modify the bitmask you need to set the field to a brand-new value (see example below).


>>> # this is an error
>>> dmd.modes_activation_bitmask[[1,2]] = False
ValueError: assignment destination is read-only
>>> tmp = np.array(dmd.modes_activation_bitmask)
>>> tmp[[1,2]] = False
>>> dmd.modes_activation_bitmask = tmp

The DMD modes activation bitmask.

Return type


property operator

Get the instance of DMDOperator.


the instance of DMDOperator

Return type


property reconstructed_data

Get the reconstructed data.


the matrix that contains the reconstructed snapshots.

Return type



Build a collection of all the available versions of the given timeindex. The indexing of time instants is the same used for reconstructed_data(). For each time instant there are at least one and at most d versions. If timeindex is None the function returns the whole collection, for all the time instants.


timeindex (int) – The index of the time snapshot.


a collection of all the available versions for the given time snapshot, or for all the time snapshots if timeindex is None (in the second case, time varies along the first dimension of the array returned).

Return type

numpy.ndarray or list

property svd_rank