Problem 7 - A First Course in the Numerical Analysis of Differential Equations

Question

The solution of the linear matrix ODE $$ Y^{\prime}=A(t) Y, \quad t \geq 0, \quad Y(0)=Y_{0} \in \mathcal{G}, $$ evolves in the Lie group \(\mathcal{G}\), subject to the assumption that \(A(t) \in \mathfrak{g}, t \geq 0\), where \(\mathfrak{g}\) is the corresponding Lie algebra. a Consider the method $$ Y_{n+1}=\exp \left(\int_{t_{n}}^{t_{n+1}} A(\tau) \mathrm{d} \tau\right) Y_{n}, \quad n=0,1, \ldots, $$ where \(\exp (\cdots)\) is the standard matrix exponential. Prove that the method is of order 2 and that \(Y_{n} \in \mathcal{G}, n=0,1, \ldots\) b Suppose that the integral above is discretized by Gaussian quadrature with a single node, $$ \int_{t_{n}}^{t_{n+1}} A(\tau) \mathrm{d} \tau \approx h A\left(t_{n}+\frac{1}{2} h\right) . $$ Prove that the new method is also of order 2 and that it evolves in \(\mathcal{G}\). c * Prove that $$ Y_{n+1}=\exp \left(\int_{t_{n}}^{t_{n+1}} A(\tau) \mathrm{d} \tau-\frac{1}{2} \int_{t_{n}}^{t_{n+1}}\left(\int_{t_{n}}^{\tau} A(\zeta) \mathrm{d} \zeta, A(\tau)\right) \mathrm{d} \tau\right) Y_{n} $$ for \(n=0,1, \ldots\) is a fourth-order method and, again, \(Y_{n} \in \mathcal{G}, n=0,1, \ldots\)

Step-by-Step Solution

Answer

Each method ensures \(Y_n \in \mathcal{G}\), with orders of 2 and 4 demonstrated by Taylor expansions and structure-preserving properties.

Step 1: Analyze the given method

The method given for the evolution of the linear matrix ODE is: \[ Y_{n+1} = \exp\left(\int_{t_n}^{t_{n+1}} A(\tau) \mathrm{d} \tau\right) Y_n. \] This expression shows that we evolve from \(Y_n\) to \(Y_{n+1}\) using a matrix exponential of the integral of \(A(t)\) over the interval \([t_n, t_{n+1}]\). Our job is to prove that this method has an order 2 accuracy and that every \(Y_n\) evolves within the Lie group \(\mathcal{G}\).

Step 2: Order of accuracy for the first method

To prove the order of accuracy, we need to consider the Taylor expansion of the matrix exponential. The first-order Taylor expansion of \(Y(t+h)\) is given by \(Y(t+h) = (I + hA(t)) Y(t) + O(h^2)\), and expanding further gives the next order term as \(\frac{1}{2}h^2A'(t) Y(t)\). Thus, the order of the method \[ Y_{n+1} = \exp(\int_{t_n}^{t_{n+1}} A(\tau) \mathrm{d} \tau) Y_n \] is 2 because the error term starts from \(O(h^2)\).

Step 3: Lie group evolution in the first method

Since \(A(t) \in \mathfrak{g}\) (the Lie algebra), \(\exp(A(t))\) results in an element within \(\mathcal{G}\) by properties of the exponential map of Lie algebras to Lie groups. This is because of the closure of these operations within the group structure, preserving the evolution \(Y_n \in \mathcal{G}\).

Step 4: Discretization with Gaussian quadrature

The integral \(\int_{t_{n}}^{t_{n+1}} A(\tau) \mathrm{d} \tau\) is approximated using a single-node Gaussian quadrature as \(h A\left(t_n + \frac{1}{2}h\right)\). Since Gaussian quadrature with just one node effectively approximates the integral to second order, the order of the method is preserved as 2 even with this discretization method.

Step 5: Lie group evolution in Gaussian quadrature method

The structure of the transformation \(Y_{n+1} = \exp(h A(t_n + \frac{1}{2}h)) Y_n\) ensures that even the approximate integral via Gaussian quadrature retains the element within \(\mathcal{G}\). This is because \(A(t_n + \frac{1}{2}h)\) is in the Lie algebra \(\mathfrak{g}\), ensuring \(\exp(h A(t_n + \frac{1}{2}h))\) is an element in \(\mathcal{G}\).

Step 6: Fourth-order method expression

The method described by \[ Y_{n+1}=exp\left(\int_{t_{n}}^{t_{n+1}} A(\tau) \mathrm{d} \tau-\frac{1}{2} \int_{t_{n}}^{t_{n+1}}\left(\int_{t_{n}}^{\tau} A(\zeta) \mathrm{d} \zeta, A(\tau)\right) \mathrm{d} \tau\right) Y_n \] is constructed to negate specific higher-order error terms, achieving fourth-order accuracy. The commutator \(\left(\int_{t_n}^{\tau} A(\zeta) \mathrm{d} \zeta, A(\tau)\right)\) corrects the earlier terms, yielding significantly better accuracy than the second-order method.

Step 7: Lie group evolution in fourth-order method

As before, the integral terms and the commutator involved all start with \(A(t) \in \mathfrak{g}\). As a result, \(\exp(\cdots)\) in the complex method still results in elements within \(\mathcal{G}\), preserving the group's evolution structure. This is due to the properties of Lie algebra exponentiation into Lie groups.

Key Concepts

Matrix Exponential

The matrix exponential is a powerful concept in numerical analysis and differential equations. It generalizes the exponential function for matrices, allowing us to solve systems of linear differential equations. If you have a square matrix, say \( A \), the exponential is given by

• \( \exp(A) = I + A + \frac{1}{2!}A^2 + \frac{1}{3!}A^3 + \cdots \)

This series is analogous to the Taylor series of the scalar exponential function. One of the key applications of matrix exponential is in solving matrix differential equations of the form:

• \( Y'(t) = A(t) Y(t) \)

In this equation, \( A(t) \) acts as a matrix that evolves over time. To find \( Y(t) \), we use the matrix exponential:

• \( Y(t) = \exp\left(\int_0^t A(\tau) \, d\tau\right) Y_0 \)

The ability of the matrix exponential to encapsulate the evolution of such systems makes it invaluable, especially in fields like quantum mechanics and control theory.

Gaussian Quadrature

Gaussian quadrature is an essential technique in numerical integration, offering a way to approximate the definite integrals of functions. It is known for its accuracy, especially when dealing with polynomials. Essentially, it transforms the task of integration into a weighted sum of function values evaluated at specific points, known as nodes.

• The integral \( \int_{a}^{b} f(x) \, dx \) is approximated as \( \sum_{i=1}^{n} w_i f(x_i) \),
• where \( x_i \) are the nodes and \( w_i \) are the weights.

In the context of numerical solutions to differential equations, Gaussian quadrature can be used to approximate the integral of the matrix function \( A(t) \) over an interval:

• \( \int_{t_n}^{t_{n+1}} A(\tau) \, d\tau \approx h A\left(t_n + \frac{1}{2}h\right) \)

This single-node Gaussian quadrature used in the exercise ensures second-order accuracy, making it suitable for applications needing precision in fewer steps. Its efficiency and precision make it a popular choice in numerical analysis.

Lie Groups and Lie Algebras

Lie groups and Lie algebras are mathematical structures that play a critical role in understanding continuous symmetry in various systems. A Lie group is a set equipped with operations of multiplication and inversion, satisfying group properties, and is a differentiable manifold, meaning you can perform calculus on it. Common examples include:

• The rotations in Euclidean space \( \text{SO}(3) \).
• The group of non-singular matrices \( \text{GL}(n) \).

Lie algebras, on the other hand, are related to Lie groups but instead explore their tangent spaces at the identity element. They define the structure and local "directions" in which the Lie group can evolve. Functionally, any element \( A \) from a Lie algebra \( \mathfrak{g} \) can generate a one-parameter subgroup in the Lie group \( \mathcal{G} \) via the exponential map:

• \( \exp: \mathfrak{g} \to \mathcal{G} \)

This means that exponentiating elements of a Lie algebra maps them into the Lie group, preserving its structure. These properties ensure the method stays within \( \mathcal{G} \) for each \( Y_n \) when evolving the differential equation. This is crucial for many applications in physics and engineering, where ensuring such group properties are preserved is vital.