Why approximate pairwise distances only over lower triangle of distance matrix?

112 Views Asked by At

In the context of dimensionality reduction.

Why approximate pairwise distances only over lower triangle of distance matrix?

$$\min_{\{\hat{x_i}\}} I = \sum_{i <j} ({ \hat{ d_{ij} } - d_{ij} })^2$$

(also known as raw stress)

Where $d_{ij}$ are distances of original data points and $\hat{d_{ij}}$ contain new coordinates that would approximate pairwise distances.

(as my notes give)

Is it perhaps that the upper triangle is a duplicate of the lower, since surely $d_{ij}=d_{ji}$ ($i \not= j$)?

1

There are 1 best solutions below

0
On

$\require{cancel}$

You are right

\begin{eqnarray} \sum_{i,j} (\hat{d}_{ij} - d_{ij}) &=& \color{blue}{\sum_{i < j}(\hat{d}_{ij} - d_{ij})} + \color{red}{\sum_{i > j}(\hat{d}_{ij} - d_{ij})} + \color{orange}{\cancelto{0}{\sum_{i = j}(\hat{d}_{ij} - d_{ij})}} \\ &=& \color{blue}{\sum_{i < j}(\hat{d}_{ij} - d_{ji})} + \color{red}{\sum_{i > j}(\hat{d}_{ji} - d_{ji})} \\ &=& \color{blue}{\sum_{i < j}(\hat{d}_{ij} - d_{ji})} + \color{red}{\sum_{j < i}(\hat{d}_{ji} - d_{ji})} \\ &=& 2\color{blue}{\sum_{i < j}(\hat{d}_{ij} - d_{ji})} \end{eqnarray}