Understanding the difference between the principal components construction and the Karhunen-Loeve expansion.

137 Views Asked by At

I'm currently studying the book Monte Carlo Methods in Financial Engineering by Glasserman. When covering the simulation of Brownian Motions, Glasserman explains the Principal Component Construction (PCC) and the Karhunen-Loeve (KL) expansion.

In the PCC, the Brownian path is approximated by $$\begin{pmatrix}W(t_1)\\W(t_2)\\\vdots\\W(t_n)\end{pmatrix} = \begin{pmatrix}a_{11}\\a_{21}\\\vdots\\ a_{n1}Z_1\end{pmatrix}Z_1 + \ldots + \begin{pmatrix}a_{1n} \\a_{2n}\\\vdots\\a_{nn}\end{pmatrix}Z_n$$ with $Z_i$ independent random $N(0,1)$ variables. This is a valid construction of the discrete Brownian path if $AA^T$ is the covariance matrix $C$ of $W = (W(t_1),\ldots,W(t_n))^T$. Specifically, $a_i = \sqrt{\lambda_i}v_i, i = 1,\ldots, n$, where $\lambda_1 >\lambda_2>\ldots >\lambda_n$ are the eigenvalues of $C$ and the $v_i$ are the corresponding eigenvectors.

Alternatively, in the KL expansion the Brownian path is equal to $$W(t) = \sum_{i = 0}^\infty \sqrt{\lambda_i}\psi(t)Z_i, 0\leq t \leq 1, $$ with $Z_i$ again independent random $N(0,1)$ variables. This infinite series an exact representation of the continuous Brownian path. By taking just the first $k$ terms in this series, we arrive at an approximation to the continuous path $\{W(t), 0\leq t\leq1\}$ that apparently is optimal in the sense of explained variability.

I'm trying to compare these methods, and basically the only difference I can think of is the coefficients. For PCC we have that $$\lambda_i = \dfrac{\Delta t}{4}\sin^{-2}\bigg(\dfrac{2i - 1}{2n + 1}\dfrac{\pi}{2}\bigg), v_i(j) = \dfrac{2}{\sqrt{2n + 1}}\sin\bigg(\dfrac{2i - 1}{2n + 1} j\pi\bigg), j = 1,\ldots, n$$ and for KL we have: $$\lambda_i = \bigg(\dfrac{2}{(2i +1)\pi}\bigg)^2, \psi_i = \sqrt{2}\sin\bigg(\dfrac{(2i + 1)\pi t}{2}\bigg)$$

Question: What are the advantages/disadvantages of these methods? When comparing the differences between these approximations at time $t$, what do we get? That is, what is $PCC_k(t) - KL_k(t)$ in general, where $k$ is the number of random normals used?

Edit: Perhaps, instead of answering my initial question, could somehow give me some hints on how I should analyse the coefficients of the different methods? I've tried looking at what they might be for different values for $i$ and $t$, but I didn't get very far.