Is it true that the probability of n hyperplanes with a maximum of K-2 dimensions intersecting in K-dimensional ambient space is 0?

15 Views Asked by At

For back-ground, I'm not well-schooled in higher-dimensional geometry, but I'm currently learning statistical data-science in which many methods rely on the properties of higher-dimensional space.

In Support-Vector Machine (SVM) clustering, the goal is to find a hyper-plane which will cleanly divide the data-points (vectors of n-dimensions) into labelled groups, and to use that dividing line to predict future data. SVM is especially good for binary classification.

However, it might be the case that no such hyper-plane exists in n-dimensions (especially common if n is low, like 2 or 3). One interesting technique used to address this problem is to randomly assign each point a value in an additional dimension (e.g. in 2-d space, assign each data-point a random z-value, transforming the 2-d space into a new 3-d space in which it is more likely that a hyper-plane exists which will cleanly divide the data in two). I was wondering how effective this process is, and after thinking for awhile, I concluded that projecting the data into K+2 dimensions, where K is the size of the largest cluster of data-points, should guarantee that a hyperplane can cleanly divide the data in 2.

My reasoning is this:

Given n points in n+1 dimensional space, and that no set of 3 or more of the points are colinear (or I guess that no set of n+1 points are co-planar in n-dimensions), those points define a unique n-dimensional hyperplane.

Given a set of n hyperplanes with <K dimensions embedded in K+2 space, the probability of those hyperplanes intersecting is 0.

Both of these claims make sense to me intuitively, but I have no idea how to even begin to check the reasoning. The only support I can give for these claims is intuition and trial and error (e.g. it seems like the probability of two points intersecting in 2-d space is 0, and the probability of two lines intersecting in 3d space is 0). I'm annoyed because this reasoning (and my intuition) is not generalizable to higher-dimensions.

So yeah, can any of you prove (or disprove) this or perhaps tell me about what branch of geometry it would fall into? Or just explain how to formally state this.

1

There are 1 best solutions below

0
On

It's nice to think of codimension for this sort of thing. The codimension of a subspace of $\mathbb R^n$ is defined to be $n$ minus the dimension of the subspace. So a line in 3-space has co-dimension 2, and a plane in 4-space has dimension 2 and codimension 2.

The great theorem is that if you have subspaces $U$ and $V$, the codimension of $U \cap V$ is the sum of the codimensions of $U$ and $V$ (in the "generic" case, i.e., avoiding things like the case in 3-space where $U$ is a plane and $V$ is a line that just happens to be completely contained in $U$ -- that's the non-generic case). [Special case: if the codimension is larger than the dimension of the ambient space, the intersection can still be the 0-dimensional subspace consisting of just the origin.] The big idea of non-generic cases is that they have probability zero under a uniform distribution on the (compact) set of $k$-planes in $n$ space. (The more general theorem is that in all cases, the codimension of $U \cap V$ is at most the sum of their codimensions.)

In your problem, you have $n$ codimension-two subspaces in $\mathbb R^K$, and you're assuming the "generic" case. So their intersection has codimension at least $2n$. If $n$ is at least $\lceil K/2 \rceil$, then the codimension of the intersection is at least $K$, so the intersection has dimension at most $0$, i.e., it consists of the single point $0$.

If $n$ is less than that number, then the codimension is no more than $2n$, so there's some nontrivial intersection in general.