Error analysis of BDF Compound-Fast multirate method for
differential-algebraic equations

Arie A. Verhoeven
HG 8.47 (Technische Universiteit Eindhoven)
Den Dolech 2, 5600 MB Eindhoven, The Netherlands
averhoev@win.tue.nl
Jan ter Maten, Bob Mattheij, Theo Beelen,
Ahmed El Guennouni, Bratislav Tasic

Analogue electrical circuits are usually modeled by differential-algebraic equations (DAE) of type:

$\displaystyle \qquad \qquad \qquad
\frac{d}{dt}\left[
\mathbf{q}(t,\mathbf{x})\right]+
\mathbf{j}(t,\mathbf{x})=\mathbf{0},
\qquad \qquad \qquad (1)
$

where $ \mathbf{x} \in \mathbb{R}^d$ represents the state of the circuit. A common analysis is the transient analysis, which computes the solution $ \mathbf{x}(t)$ of this non-linear DAE along the time interval $ [0,T]$ for a given initial state. Often, parts of electrical circuits have latency or multirate behaviour.

For a multirate method it is necessary to partition the variables and equations into an active (A) and a latent (L) part. The active and latent parts can be expressed by $ \mathbf{x}_A = \mathbf{B}_A
\mathbf{x},\mathbf{x}_L = \mathbf{B}_L \mathbf{x}$ where $ \mathbf{B}_A,\mathbf{B}_L$ are permutation matrices. Then equation (1) is written as the following partitioned system:

\begin{displaymath}
\begin{array}{c} \frac{d}{dt}\left
[\mathbf{q}_{A}(t,\mathbf...
...hbf{j}_{L}(t,\mathbf{x}_A,\mathbf{x}_L)=\mathbf{0}.
\end{array}\end{displaymath}

In contradiction to classical integration methods, multirate methods integrate both parts with different stepsizes. Besides the coarse time-grid $ \{T_n,0\leq n \leq N\}$ with stepsizes $ H_n = T_n -
T_{n-1}$, also a refined time-grid $ \{t_{n-1,m},1\leq n \leq
N, 0 \leq m \leq q_n\}$ is used with stepsizes $ h_{n,m} =
t_{n,m} - t_{n,m-1}$ and multirate factors $ q_n$. If the two time-grids are synchronized, $ T_n= t_{n,0} = t_{n-1,q_n}$ holds for all $ n$. There are a lot of multirate approaches for partitioned systems but we will consider the Compound-Fast version of the BDF methods. This method performs the following four steps:
  1. The complete system is integrated at the coarse time-grid.
  2. The latent interface variables are interpolated at the refined time-grid.
  3. The active part is integrated at the refined time-grid, using the interpolated values at the latent interface.
  4. The active solution at the coarse time-grid is updated.

The local discretization error $ \delta^n$ of the compound phase still has the same behaviour $ \delta^n = O(H_n^{K+1})$. Let $ \bar{\mathbf{P}}^n,\bar{\mathbf{Q}}^n$ be the Nordsieck vectors which correspond to the predictor and corrector polynomials of $ \mathbf{q}$. Then the error $ \delta^n$ can be estimated by $ \hat{\delta}^n$:

$\displaystyle \hat{\delta}^n = \frac{-H_n}{T_n - T_{n-K-1}} \left
[\bar{\mathbf{Q}}_{1}^{n}-\bar{\mathbf{P}}_{1}^{n} \right ].
$

Now $ \hat r_C^n = \Vert\mathbf{B}_L
\hat{\delta}^n\Vert + \tau\Vert\mathbf{B}_A \hat{\delta}^n\Vert $ is the used weighted error norm, which must satisfy $ \hat r_C^n
<$ TOL$ _C$.
The local discretization error $ \delta^{n,m}$ is defined as the residue after inserting the exact solution in the refinement BDF scheme. During the refinement instead of $ \delta^{n,m}$ the perturbed local error $ \tilde{\delta}^{n,m}$ is estimated. A tedious analysis yields the following asymptotic behaviour:

$\displaystyle \mathbf{B}_A\delta^{n-1,m} \doteq
\mathbf{B}_A\tilde{\delta}^{n-1,m} + \frac{1}{4}h
\mathbf{K}_{n-1,m}\mathbf{B}_L\rho^{n-1,m}.
$

Here $ \rho^{n-1,m}$ is the interpolation error at the refined grid and $ \mathbf{K}_{n-1,m}$ is the coupling matrix. The perturbed local discretization error $ \mathbf{B}_A\tilde{\delta}^{n,m}$ behaves as $ O(h_{n-1,m}^{k+1})$ and can be estimated in a similar way as $ \delta^n$. Thus the active error estimate $ \mathbf{B}_A\hat\delta^{n-1,m}$ satisfies $ \mathbf{B}_A\hat\delta^{n-1,m} \doteq
\mathbf{B}_A\hat{\tilde{\delta}}^{n-1,m} + \frac{1}{4}h
\hat{\mathbf{K}}_{n-1,m}\mathbf{B}_L\hat\rho^{n-1,m}$. Let $ L$ be the interpolation order, then it can be shown that $ \frac{1}{4}\Vert\hat{\mathbf{K}}_{n}\mathbf{B}_L\rho^{n-1,m}\Vert$ is less than

$\displaystyle \hat{r}_I^{n} =
\frac{1}{4}\frac{H_n}{T_n - T_{n-L-1}}
\Vert\hat{...
...mathbf{B}_L
\left[\bar{\mathbf{X}}_{1}^n
-\bar{\mathbf{Y}}_{1}^n\right] \Vert.
$

Here $ \bar{\mathbf{Y}}^n,\bar{\mathbf{X}}^n$ are the Nordsieck vectors which correspond to the predictor and corrector polynomials of $ \mathbf{x}$. This error estimate $ \hat{r}_I^n$ has the asymptotic behaviour $ \hat{r}_I^{n} =
O(H_n^{L+1})$. It follows that $ \Vert\mathbf{B}_A\hat{\delta}^{n,m}\Vert$ satisfies:

\begin{displaymath}
\begin{array}{rcl}
\Vert\mathbf{B}_A\hat{\delta}^{n-1,m}\Ver...
...}}_A^{n-1,m} + h\hat{r}_I^{n} =:
\hat{r}_A^{n-1,m}. \end{array}\end{displaymath}

If $ \hat{r}_I^{n} \leq $ TOL $ _I = \sigma$TOL$ _A$ and $ \hat{\tilde{r}}_A^{n-1,m} \leq \tilde{\mbox{TOL}}_A =
(1-\sigma h)$TOL$ _A$ then $ \hat{r}_A^{n-1,m} \leq
\tilde{\mbox{TOL}}_A + h$TOL$ _I =$ TOL$ _A$.

We tested a circuit with $ 5 \times 10$ inverters. The location of the active part is controlled by the connecting elements and the voltage sources. The connecting elements were chosen such that the active part consists of 3 inverters. We did an Euler Backward Compound-Fast multirate simulation on $ [0,10^{-8}]$ with $ \sigma = 0.5,\tau = 0$. We get accurate results combined with a speedup factor 13.