|
| 1 | +# Rosenbrock methods |
| 2 | + |
| 3 | +In this page, we introduce Rosenbrock-type methods to solve ordinary |
| 4 | +differential equations. In doing do, we roughly follow Chapter IV.7 of "Solving |
| 5 | +Ordinary Differential Equations II" by Hairer and Wanner. (Beware, notation is |
| 6 | +not the same.). In this spirit, let us introduce the need for Rosenbrock-type |
| 7 | +methods in the same way as Hairer and Wanner, by quoting Rosenbrock himself: |
| 8 | + |
| 9 | +> When the functions are non-linear, implicit equations can in general be solved |
| 10 | +> only by iteration. This is a severe drawback, as it adds to the problem of |
| 11 | +> stability, that of convergence of the iterative process. An alternative, which |
| 12 | +> avoids this difficulty, is ... |
| 13 | +
|
| 14 | +Rosenbrock method! |
| 15 | + |
| 16 | +Before reading this page, we recommend reading the page on ODE solvers first |
| 17 | +[TODO: Add link] |
| 18 | + |
| 19 | +## Introduction to the formalism |
| 20 | + |
| 21 | +Let us consider an ordinary differential equation of the form |
| 22 | + |
| 23 | +```math |
| 24 | +\frac{d}{dt}u(t) = T(u(t), t)\,, |
| 25 | +``` |
| 26 | + |
| 27 | +where $u$ is the state, $T$ is the tendency, |
| 28 | +and $t$ the time. For the sake of simplicity, let us ignore the explicit time |
| 29 | +dependency in the tendency (we will get back to that at the end). |
| 30 | + |
| 31 | +The simplest way to introduce the Rosenbrock method is to start from a |
| 32 | +diagonally implicit Runge-Kutta scheme (see page on DIRK). In an implicit |
| 33 | +Runge-Kutta method with $s$ stages and tableau $a, b, c$, the updated value |
| 34 | +$u_1$ for a step of size $\Delta t$ is obtained starting from the known value $u_0$ |
| 35 | +with |
| 36 | + |
| 37 | +```math |
| 38 | +u_1 = u_0 + \sum_{i=1}^s b_i k_i\,, |
| 39 | +``` |
| 40 | +with |
| 41 | +```math |
| 42 | +k_i = \Delta t T ( u_0 + \sum_{j=1}^{i-1}\alpha_{ij}k_j + \alpha_{ii} k_i)\,. |
| 43 | +``` |
| 44 | +$\alpha_{ij}$, $b_i$ are carefully chosen coefficients. |
| 45 | + |
| 46 | +Rosenbrock's idea consists in linearizing $T$. This operation can be interpreted |
| 47 | +as taking one iteration for Netwon solver and yields |
| 48 | + |
| 49 | +```math |
| 50 | +k_i = \Delta t T(g_i) + \Delta t T'(g_i) \alpha_{ii} k_i |
| 51 | +``` |
| 52 | + |
| 53 | +with $J(g_i)$ Jacobian of the tendency and with |
| 54 | + |
| 55 | +```math |
| 56 | +g_i = u_0 + \sum_{j=1}^{i-1}\alpha_{ij} k_j |
| 57 | +``` |
| 58 | + |
| 59 | +In Rosenbrock-type methods, the Jacobian $T'$ is replaced with linear combinations |
| 60 | +of the Jacobian evaluated at $u_0$, $J$: |
| 61 | +```math |
| 62 | +T'(g_i) \alpha_{ii} k_i \approx J \sum_{j=1}^{i-1}\gamma_{ij} k_j + J \gamma_{ii} k_i\,, |
| 63 | +``` |
| 64 | +with $\gamma_{ij}$ other coefficients that can be chosen. |
| 65 | + |
| 66 | +Now, each stage consists of solving a system of linear equations in $k_i$ with |
| 67 | +matrix $I - \Delta t \gamma_{ii}$: |
| 68 | +```math |
| 69 | +(I - J \Delta t \gamma_{ii}) k_i = \Delta |
| 70 | +t T(g_i) + J \sum_{j=1}^{i-1}\gamma_{ij} k_j |
| 71 | +``` |
| 72 | +for each $i$ in $1, \dots, s$. Once $k_i$ are known, $u_1$ can is easily computed and the process repeated for |
| 73 | +the next step. |
| 74 | + |
| 75 | +In practice, there are computational advantages at implementing a slight |
| 76 | +variation of what presented here. |
| 77 | + |
| 78 | +Let us define $\tilde{k}_{i} = \sum_{j=1}^{i} \gamma_{ij} k_j$. If the matrix |
| 79 | +$\gamma_{ij}$ is invertible, we can move freely from $k_i$ to $\tilde{k}_i$. A |
| 80 | +convenient step is to also define $C$, as |
| 81 | + |
| 82 | +```math |
| 83 | +C = diag(\gamma_{11}^{-1}, \gamma_{22}^{-1}, \dots, \gamma_{ss}^{-1}) - \Gamma^{-1} |
| 84 | +``` |
| 85 | + |
| 86 | +Which establishes that |
| 87 | + |
| 88 | +```math |
| 89 | +k_i = \frac{\tilde{k}_i}{\gamma_{ii}} - \sum_{j=1}^{i -1} c_{ij} \tilde{k}_j |
| 90 | +``` |
| 91 | +Substituting this, we obtain the following equations |
| 92 | + |
| 93 | +```math |
| 94 | +(J \Delta t \gamma_{ii} - 1) \tilde{k}_i = - \Delta |
| 95 | +t \gamma_{ii} T(g_i) - \gamma_{ii} J \sum_{j=1}^{i-1}c_{ij} \tilde{k}_j \,, |
| 96 | +``` |
| 97 | + |
| 98 | +```math |
| 99 | +g_i = u_0 + \sum_{j=1}^{i-1}a_{ij} \tilde{k}_j \,, |
| 100 | +``` |
| 101 | + |
| 102 | +```math |
| 103 | +u_1 = u_0 + \sum_{j=1}^{s} m_j \tilde{k}_j\,, |
| 104 | +``` |
| 105 | +with |
| 106 | +```math |
| 107 | +(a_{ij}) = (\alpha_{ij}) \Gamma^{-1}\,, |
| 108 | +``` |
| 109 | +and |
| 110 | +```math |
| 111 | +(m_i) = (b_i) \Gamma^{-1} |
| 112 | +``` |
| 113 | + |
| 114 | +Finally, small changes are required to add support for explicit time derivative: |
| 115 | + |
| 116 | +```math |
| 117 | +(J \Delta t \gamma_{ii} - 1) \tilde{k}_i = - \Delta |
| 118 | +t \gamma_{ii} T( t_0 + \alpha_i \Delta t, g_i) - \gamma_{ii} J \sum_{j=1}^{i-1}c_{ij} \tilde{k}_j - \Delta |
| 119 | +t \gamma_{ii} \gamma_i \Delta |
| 120 | +t \frac{\partial T}{\partial t}(t_0, u_0) |
| 121 | +``` |
| 122 | + |
| 123 | +where we defined |
| 124 | + |
| 125 | +$\alpha_{i} = \sum_{j=1}^{i-1}\alpha_{ij} $ |
| 126 | + |
| 127 | +$ \gamma _{i} = \sum_{j=1}^{i}\gamma_{ij} $ |
| 128 | + |
| 129 | +> :note: You may wonder, what is up with all these ugly minus signs? Why don't |
| 130 | +> we cancel them out? The reason for this is because we want to have the |
| 131 | +> prefactor $(J \Delta t \gamma_{ii} - 1)$. This is called `Wfact` in the |
| 132 | +> language of `DifferentialEquations.jl`. Most other algorithms specify this |
| 133 | +> quantity, so it is convenient to be consistent. |
| 134 | +
|
| 135 | +## Implementation |
| 136 | + |
| 137 | +In `ClimaTimeSteppers.jl`, we implement the last equation presented in the |
| 138 | +previous section. Currently, the only Rosenbrock-type algorithm implemented is |
| 139 | +`SSPKnoth`. `SSPKnoth` is 3-stage, second-order accurate Rosenbrock with |
| 140 | + |
| 141 | +```math |
| 142 | +\alpha = \begin{bmatrix} |
| 143 | + 0 & 0 & 0 \\ |
| 144 | + 1 & 0 & 0 \\ |
| 145 | + \frac{1}{4} & \frac{1}{4} & 0 \\ |
| 146 | + \end{bmatrix} |
| 147 | +``` |
| 148 | + |
| 149 | +```math |
| 150 | +\Gamma = \begin{bmatrix} |
| 151 | + 1 & 0 & 0 \\ |
| 152 | + 1 & 1 & 0 \\ |
| 153 | + -\frac{3}{4} & \frac{3}{4} & 1 \\ |
| 154 | + \end{bmatrix} |
| 155 | +``` |
| 156 | + |
| 157 | +and |
| 158 | +```math |
| 159 | +b = (\frac{1}{6}, \frac{1}{6}, \frac{2}{3}) |
| 160 | +``` |
| 161 | + |
| 162 | +At each stage, the state is updated to $g_i$ and precomputed quantities are updated. Tendencies are computed and, unless the stage is the last, DSS is applied to the sum of all the tendencies. After the last stage, DSS is applied to the incremented state. |
0 commit comments