Memorisation vs Proof

Let random variable \(X\) be distributed like a univariate Gaussian \(N(\mu, \sigma^2)\). It is very well known that the random variable \(aX\) is distributed like \(N(a \mu, a^{2} \sigma^2)\). Let us call this statement Lemma 1. Lemma 1 is so useful that it’s given without proof as part of the ‘hypothesis testing’ topic in A-level maths. Students who continue engaging with maths after A-level might find themselves making use Lemma 1, without ever having seen a proof. This certainly applied to me.

Memorising a piece of mathematics without knowing (at least) a sketch of a proof can be dangerous, as I found out recently when requiring a very slight generalisation of Lemma 1. I needed to understand the following:

Question: Let random vector \(X\) be distributed like an \(n\) dimensional multivariate Gaussian \(N(\mu, \Sigma)\). Let \(A\) be a real \(m\) by \(n\) matrix, how is the \(m\) dimensional random vector \(AX\) distributed?

Answering this question took me a little while, very embarrassing. Perhaps it would have been immediately obvious if I’d seen a proof of Lemma 1. Moreover, perhaps I would have saved some time and confusion if I’d known that Lemma 1 was only memorised, and I’d never proven it for myself. After some thought I arrived at Lemma 2:

Lemma: Let random vector \(X\) be distributed like an \(n\) dimensional multivariate Gaussian \(N(\mu, \Sigma)\). Let \(A\) be a real \(m\) by \(n\) matrix. The \(m\) dimensional random vector \(AX\) distributed like \(N(A \mu, A \Sigma A^T)\).

Proof (Sketch): We can see this with multivariate moment generating functions. Recall that for a multivariate Gaussian distribution \(N(\mu, \Sigma )\) the moment generating function is given by:

\[ MGF(X) = \mathbb{E}(exp({t^T X})) = exp(\mu^Tt + \frac{1}{2}t^T \Sigma t) \]

Now, we can write \(MGF(AX) = \mathbb{E}(exp({t^T AX})) = \mathbb{E}(exp({(A^Tt)^T X}))\). This tells us that the moment generating function of \(AX\) can be written as:

\[ MGF(AX) = exp(\mu^TA^Tt + \frac{1}{2}(A^Tt)^T \Sigma A^Tt) = exp((A \mu)^Tt + \frac{1}{2} t^T A \Sigma A^Tt). \]

We conclude that \(AX\) is distributed like \(N(A\mu, A \Sigma A^T)\), as moment generating functions uniquely determine distributions.

Closing Thoughts: Every modern mathematician uses results they don’t know how to prove, and it would be extremist to recommend only using a result if one can sketch a proof. I do think we should all try to be aware of when a result we’ve used is only memorised however, and in the case of simple results like Lemma 1, remedy that state of affiars before embarrassment might ensue.

The same sort of discussion doesn’t just apply to mathematical results, but also to other things, like opinions.