I study isomorphism in algebra, which asks how do we tell two algebras apart. Over more than a century, researchers have puzzled over this unsolved question with various motives. It is an exciting time in the subject with advances that have improved exponentially over what we could do even 5 years ago.
What follows is one present day application with a substantial pay off. I hope it will be a gentel introduction, appropriate for students and anyone curious, into some of the work I do.
= What is really changing. | |
Our isomorphism tests |
Measurement makes observations into numbers. To do that, we seem to need coordiantes and scales. E.g. fix xyz directions, choosing meters or feet, assigning each color a number, etc.. In real data, repeated measurements cannot guarantee the same coordinates. This creates a problem.
How can we tell if data has changed, or if we just recorded the same event but in different coordinates?
I think of data as uniform tables of numbers (you may need to fill in missing numbers with 0). E.g. a digital picture measures the color of point $(x,y)$ in an image, recording it as pixel $T_{xy}$. We might also add coordinates $RGB$ to interpet color as "Red", "Blue", "Green", which we do by adding a 3rd dimension to our frame of reference, e.g. $T_{xy}^{RGB}$.
Predictable coordinate changes include changing lengths of $x$ and $y$, e.g. centimeters to inches. Or swapping Red-Blue-Green (RGB) with Cyan-Yellow-Grayscale (CYK). Here official standards take care of of organizing the converstion of coordinates. So data recorded in on coordinate system is easy to compare with data recorded in another, so long as we are told about the system of coordinates.
Difficult coordinate changes result when we cannot even know what coordinates are involved. Consider the following.
I convert data into algebra where the effects of coordinates have predictable effects, known as isomorphism.
Algebra is the study of equation solving. You have two problems:
So Different number systems give different solutions to the same equations.
However, if we change our coordinates or write our algebras with completely different symbols, for example as follows, the solutions to our equations remain the same!
$1$ | $i$ | $j$ | $k$ |
$(1,0,0,0)$ | $(0,1,0,0)$ | $(0,0,1,0)$ | $(0,0,0,1)$ |
$\begin{bmatrix} 1 & 0 \\ 0 & 1\end{bmatrix}$ | $\begin{bmatrix} \sqrt{-1} & 0 \\ 0 & \sqrt{-1}\end{bmatrix}$ | $\begin{bmatrix} 0 & 1 \\ -1 & 0 \end{bmatrix}$ | $\begin{bmatrix} 0 & \sqrt{-1} \\ -\sqrt{-1} & 0\end{bmatrix}$ |
The moral:
For my purposes, all data is assumed to come as a uniform grid of numbers, such as $[2,3,0,5]$, $\begin{bmatrix} 1 & -1 & 0\\ -1 & 1 & 2\end{bmatrix}$, or
Examples.
Given measurements $\mathcal{M}(\mathcal{O}_1,\mathcal{F}_1)$ and $\mathcal{M}(\mathcal{O}_2,\mathcal{F}_2)$, decide if the observations agree, i.e. $\mathcal{O}_1=\mathcal{O}_2$.
Changing coordinates breaks down into scaling, re-ordering, or adding together entries in the data. These steps are collectively known as tensor contractions but you may know them by the name of matrix multiplication or row and column operations which is the name we use when our data is 2-dimensional grids.
The simplest contraction is between tensors with a single coordinate chart, e.g. $u=[2,5,0,7]$ and $v=[1,0.5,9,-7]$. A contraction here is a function that replaces a pair $(u,v)$ of coordinates with a single number, i.e. it contracts the two tensors. There are some requirements to maintain the geometry. For instance if we scale $u$ or $v$ the result should scale, and if we translate the the result should distribution. So int total we need the following. $$(u+u',v)=(u,v)+(u',v)$$ $$(u,v+v')=(u,v)+(u',v')$$ $$(\lambda u,v)=\lambda(u,v)=(u,\lambda v)$$ Such functions are called bilinear forms.
For contractions of general tensors $T_1$ and $T_2$ having many coordinate charts, we first a chart $U$ in the frame $\mathcal{F}_1$ for $T_1$ and $V$ in the frame $\mathcal{F}_2$ of $T_2$, and then contract $(U,V)$ producing a new tensor $T_1*T_2$ framed by $$(\mathcal{F}_1-\{U\})\cup(\mathcal{F}_2-\{V\}).$$ For example a $(2\times 3)$-matrices times a $(3\times 4)$ matrix is the contraction along $3$ and returns a tensor (a matrix) on the remaining terms in the frame, namely a $(2\times 4)$-matrix.
Given two tensors (illustrated below as the two volumes of numbers), guess a change in coordinates $X_0,X_1,\dots,X_{\ell}$ (below $\ell=2$) to change the frame of the first to the second and see if the resulting data agree.
Getting algebras out of tensors is on the surface quite easy. We can just insert the data to describe a multiplication. See below for one way to do this.
Suppose we have a tensor $T$ framed by $(\mathbb{R}^a, \mathbb{R}^b,\mathbb{R}^c)$, that is a grid of numbers arranged in a $3$-dimensional array that is $(a\times b\times c)$. Then just just by using contraction we get a multiplication $\mathbb{R}^a\times \mathbb{R}^b\to \mathbb{R}^c$ as follows. $$ (u_1,\dots,u_a)\bullet (v_1,\dots,v_b) = \left(\sum_{i=1}^a\sum_{j=1}^b T_{ij1} u_i v_j, \cdots, \sum_{i=1}^a\sum_{j=1}^b T_{ij1} u_i v_j\right). $$ We can do the same with any numbers we want, including binary numbers $\{0,1\}$.
This product is highly unusual. First, unless $a=b=c$, we cannot even work out what a cube would mean, for example the following expression would not make sense. $$(u\bullet v)\bullet w.$$ Even if such an express did make sense (e.g. if $a=b=c$) we still could not count on assumption we usually depend on such as associativity, commutativity, and other properities.
The problem is that whatever algebras we might cook up from data directly it will be a rare situation that that algebra is something we understand. We need a leap into algebras we do understand. For that we define a surprising correspondence summarized as follows.
Through careful study we have found precisely which polynomials pair with tensors to make good algebras of operators. For example,