Can argmin act as a "projection"?

270 Views Asked by At

My professor is teaching us about a clustering algorithm called K-means clustering. He used the term projection in a way I'm unfamiliar with.

Assume all vectors in this example are in $\mathbb{R}^d$. Suppose we are given a collection of vectors $C = \lbrace c_1,...,c_K\rbrace$. Now consider a vector $x$ and let $d$ be the Euclidean metric. Then, suppose I want to find the $c_k \in C$ such that

$$d(x,c_k)\leq d(x,c_j), \forall j\ne k$$

My professor chose to denote this

$$\Pi_C(x)=\underset{c_k\in C}{\text{argmin}}\lbrace \lVert x-c_k\rVert_2 \rbrace$$

He called this the "projection" of $x$ onto $C$.

In what sense is this a projection? I thought a projection was a linear transformation $T:V\rightarrow W$ where $V,W$ are both vector spaces. What sense of the term projection is he using?

1

There are 1 best solutions below

0
On

Wikipedia's page on projections implicitly covers the type of map your professor describes, since under "Definition" it says:

A projection may also refer to a mapping which has a right inverse

and having a right inverse is the same thing as being surjective. Your arg min map is obviously surjective as a map from $\mathbb R^d$ to $C \subset \mathbb R^d$.

For a more geometric perspective, note that $\lVert x - c_k \rVert_2$ is the distance between $x$ and $c_k$, as induced by the norm $\lVert \cdot \rVert_2$. So your professor is asking for the point in $C$ that is closest to $x$. This is very much consistent with what it intuitively means geometrically to project a point $y$ onto a set $A$, like the $x$-$y$ plane for example.