On Interdisciplinary Research

I have always found the boundaries between academic disciplines slightly artificial. We divide mathematics, physics, computer science, biology and the rest into different departments, courses and communities because specialisation is necessary. There is simply too much to know. But the objects themselves do not respect these boundaries. A function does not know whether the person studying it calls themselves a mathematician or a physicist. An eigenvalue does not become a physics object merely because it appears in a Hamiltonian. A probability distribution does not become statistics simply because someone decided to call the department Statistics.

This becomes particularly interesting when the same mathematical structure appears independently in completely different problems.

One of the examples that I find almost absurdly beautiful is the connection between the distribution of prime numbers, the zeros of the Riemann zeta function, random matrix theory and quantum mechanics. It is the sort of connection that makes the division between "pure" and "applied" mathematics feel rather less fundamental.

The original problem seems to belong entirely to number theory.

And then it doesn't.

The primes hiding inside a complex function

The Riemann zeta function begins innocently enough:

$$\zeta(s)=\sum_{n=1}^{\infty}\frac{1}{n^s}, \qquad \Re(s)>1.$$

But its Euler product is considerably more revealing:

$$\zeta(s) = \prod_{p} \left(1-p^{-s}\right)^{-1},$$

where the product runs over all primes $p$.

This equation is already an interdisciplinary lesson in miniature. The left-hand side is an analytic object; the right-hand side encodes the multiplicative structure of the integers through the primes.

Taking logarithms gives

$$\log \zeta(s) = -\sum_p\log(1-p^{-s})$$

and expanding the logarithm,

$$\log \zeta(s) = \sum_p\sum_{k=1}^{\infty} \frac{1}{k p^{ks}}.$$

Differentiating,

$$-\frac{\zeta'(s)}{\zeta(s)} = \sum_{p}\sum_{k=1}^{\infty} \frac{\log p}{p^{ks}}.$$

If we introduce the von Mangoldt function

$$\Lambda(n)= \begin{cases} \log p,&n=p^k,\quad k\geq 1,\\ 0,&\text{otherwise}, \end{cases}$$

then this becomes

$$-\frac{\zeta'(s)}{\zeta(s)} = \sum_{n=1}^{\infty} \frac{\Lambda(n)}{n^s}.$$

This is where the problem becomes genuinely interesting.

The primes are encoded in the singularities of a complex analytic function.

So a question that initially looks discrete,

$$2,3,5,7,11,13,\ldots$$

has been transformed into a question about the analytic structure of

$$\zeta(s).$$

And the zeros of that function somehow contain information about the primes.

Riemann's great insight was essentially that the apparently chaotic distribution of primes can be studied through the analytic behaviour of $\zeta(s)$. The nontrivial zeros lie in the critical strip

$$0<\Re(s)<1,$$

and the Riemann Hypothesis states that every one of them satisfies

$$\boxed{\Re(s)=\frac12}.$$

Thus every nontrivial zero should have the form

$$\rho_n=\frac12+i\gamma_n.$$

The problem is therefore no longer merely "where are the primes?"

It becomes:

$$\boxed{ \text{What is the structure of } \{\gamma_1,\gamma_2,\gamma_3,\ldots\}? }$$

And this is where things get strange.

From primes to spectra

The zeros are not distributed arbitrarily.

To see their structure, one of the natural things to examine is the spacing between them:

$$\gamma_{n+1}-\gamma_n.$$

But the average spacing itself depends on height. The number of zeros with $0<\gamma<T$ satisfies the asymptotic formula

$$N(T) = \frac{T}{2\pi} \log\frac{T}{2\pi} -\frac{T}{2\pi} +O(\log T).$$

Consequently, the local density of zeros is approximately

$$\frac{dN}{dT} \sim \frac{1}{2\pi}\log\frac{T}{2\pi}.$$

So the average spacing near height $T$ is approximately

$$\Delta(T) \sim \frac{2\pi}{\log(T/2\pi)}.$$

This means that comparing raw quantities

$$\gamma_{n+1}-\gamma_n$$

is not particularly meaningful across different heights. We need to rescale the zeros according to their local density.

Define a normalized coordinate, schematically,

$$\tilde{\gamma} = \frac{\gamma}{2\pi} \log\frac{\gamma}{2\pi}.$$

Then the normalized spacing is approximately

$$s_n = (\gamma_{n+1}-\gamma_n) \frac{\log(\gamma_n/2\pi)}{2\pi}.$$

Now something remarkable happens.

The statistics of these normalized spacings resemble the statistics of eigenvalues of large random Hermitian matrices from the Gaussian Unitary Ensemble, or GUE.

That statement is not simply a visual similarity. Montgomery's pair-correlation work produced a formula involving

$$1- \left( \frac{\sin \pi u}{\pi u} \right)^2,$$

which is precisely the two-point correlation function appearing in the GUE setting, within the relevant asymptotic framework. Dyson recognised this connection when Montgomery showed him the result in 1972.

And this is the point where I think interdisciplinary research becomes genuinely interesting.

Because now we have

$$\boxed{ \text{Number Theory} \longleftrightarrow \text{Random Matrix Theory}. }$$

There is no obvious reason this should happen.

What is a random matrix actually doing here?

Take an $N\times N$ Hermitian matrix

$$H=H^\dagger.$$

Its eigenvalues are real:

$$\lambda_1,\lambda_2,\ldots,\lambda_N.$$

For the Gaussian Unitary Ensemble, one can assign a probability density to the matrix entries of the form

$$P(H)\propto e^{-\frac{N}{2}\operatorname{Tr}(H^2)}.$$

The remarkable part comes when we transform from matrix entries to eigenvalues. The joint probability density of the eigenvalues takes the form

$$P(\lambda_1,\ldots,\lambda_N) \propto e^{-\frac{N}{2}\sum_i\lambda_i^2} \prod_{i<j} |\lambda_i-\lambda_j|^2.$$

Look carefully at the second factor:

$$\prod_{i<j}|\lambda_i-\lambda_j|^2.$$

It means that eigenvalues strongly disfavor occupying the same location.

There is an effective repulsion.

For two eigenvalues,

$$|\lambda_1-\lambda_2|^2$$

goes to zero as

$$\lambda_1\rightarrow\lambda_2.$$

So the eigenvalues of a random matrix are random, but they are not independent random numbers.

Their randomness has structure.

This is exactly the sort of thing that makes the Riemann connection so interesting. The zeros of $\zeta(s)$ also exhibit a kind of statistical repulsion. Their local statistics look like the eigenvalue statistics of GUE matrices.

We have therefore gone from

$$\zeta(s)=0$$

to

$$\det(H-\lambda I)=0.$$

On the left, we have zeros of a complex analytic function.

On the right, we have eigenvalues of a matrix.

And statistically, they appear to be speaking the same language.

Why should a number theorist care about a matrix?

This is where the usual idea of interdisciplinarity becomes too weak.

It isn't merely that number theory can borrow a technique from physics.

The random matrix model changes what we expect the zeta function to do.

For example, consider the characteristic polynomial

$$Z_N(U)=\det(I-U),$$

where $U$ is a random unitary matrix. More generally,

$$Z_N(\theta) = \det(I-e^{-i\theta}U).$$

Random matrix theory allows us to calculate quantities such as

$$\mathbb{E}\left[ |Z_N(\theta)|^{2k} \right].$$

This becomes relevant because the zeta function itself has moments along the critical line:

$$M_k(T) = \frac{1}{T} \int_0^T \left| \zeta\left(\frac12+it\right) \right|^{2k} \,dt.$$

The random matrix model suggests asymptotic forms for these moments. Keating and Snaith famously developed this connection using characteristic polynomials of random unitary matrices, producing predictions for moments of $\zeta(1/2+it)$.

The remarkable thing is that the random matrix is no longer merely a metaphor.

It becomes a computational model.

A difficult arithmetic quantity is replaced by an object from probability and linear algebra whose statistics can actually be calculated.

That is an enormous conceptual shift.

The Hilbert–Pólya idea

And this leads to perhaps the most beautiful part of the entire story.

Suppose there existed some self-adjoint operator $H$ such that its eigenvalues were precisely the imaginary parts of the nontrivial zeros:

$$\operatorname{Spec}(H) = \{\gamma_n\}.$$

Then

$$H\psi_n=\gamma_n\psi_n.$$

Because $H$ is self-adjoint,

$$H=H^\dagger,$$

its eigenvalues are real.

But the $\gamma_n$ are real precisely because the zeros are being written as

$$\rho_n=\frac12+i\gamma_n.$$

The missing part is the $\frac12$.

The dream is that the zeros can be interpreted spectrally:

$$\rho_n = \frac12+iE_n$$

for eigenvalues $E_n$ of some appropriate self-adjoint operator.

If such an operator could be constructed with the correct spectrum, the Riemann Hypothesis would follow.

This is the essence of the Hilbert–Pólya idea.

And now the random matrix connection suddenly makes more sense.

Random matrix theory is fundamentally a theory of spectra.

Quantum mechanics is fundamentally full of spectra.

The Riemann zeros appear to possess spectral statistics.

So perhaps we should stop thinking of

$$\gamma_1,\gamma_2,\gamma_3,\ldots$$

as merely a sequence of mysterious numbers and start thinking of them as the spectrum of something.

We don't currently know the required operator.

That is the enormous gap.

But the statistical evidence is sufficiently compelling that the spectral viewpoint has become an important part of the subject. The AMS explicitly describes the RMT connection as evidence that the Hilbert–Pólya spectral interpretation has merit, while also noting that it has not supplied a proof of the Riemann Hypothesis.

Quantum mechanics enters

This is where physics becomes unavoidable.

In quantum mechanics, an observable is represented by an operator $A$. The possible measured values correspond to eigenvalues satisfying

$$A|\psi\rangle=a|\psi\rangle.$$

For the Hamiltonian,

$$H|\psi_n\rangle=E_n|\psi_n\rangle,$$

and therefore

$$\{E_n\}$$

is the energy spectrum.

Now compare that with

$$\zeta\left(\frac12+iE_n\right)=0.$$

The notation itself begins to look suggestive.

A physical system has

$$H|\psi_n\rangle=E_n|\psi_n\rangle.$$

The Riemann zeros give us

$$\rho_n=\frac12+i\gamma_n.$$

The speculative bridge is therefore

$$E_n\longleftrightarrow\gamma_n.$$

Again, this is not a proof, and there is no accepted physical Hamiltonian whose spectrum has been shown to be exactly the Riemann zeros.

But it gives us a radically different question:

$$\boxed{ \text{What quantum system, if any, has the Riemann zeros as its spectrum?} }$$

That is a much more interesting question than simply asking someone to prove a statement about $\zeta(s)$.

And this is where physics has contributed something more than computational assistance. It has suggested an ontology for the numbers.

Maybe they are not just numbers.

Maybe they are spectral data.

The explicit formula makes this even stranger

The connection becomes even more interesting when one looks at the explicit formulas connecting zeros and primes.

Very roughly, the von Mangoldt explicit formula relates weighted prime-counting information to sums over the zeros:

$$\psi(x) = x - \sum_{\rho} \frac{x^\rho}{\rho} +\cdots,$$

where

$$\psi(x) = \sum_{n\leq x}\Lambda(n).$$

The omitted terms involve contributions from the pole of $\zeta$, the trivial zeros and other correction terms.

The important point is the structure:

$$\boxed{ \text{Prime distribution} \longleftrightarrow \text{Zero distribution}. }$$

Now imagine replacing the zeros by a spectral sequence

$$\{\gamma_n\}.$$

Then we have something that looks suspiciously like a trace formula in spectral theory: information about the spectrum can be related to information about periodic or classical structures.

This is one reason the analogy with quantum chaos became so attractive.

In semiclassical quantum mechanics, trace formulas can connect quantum energy levels with classical periodic orbits. Schematically,

$$d(E) \sim \bar d(E) + \sum_p A_p(E) \cos\!\left( \frac{S_p(E)}{\hbar} -\frac{\pi\mu_p}{2} \right).$$

Here $p$ indexes periodic orbits, $S_p$ is an action, $A_p$ is an amplitude and $\mu_p$ is a phase index.

Compare that with the zeta situation.

The primes behave almost like primitive periodic orbits.

The zeros behave almost like quantum energy levels.

And the explicit formula behaves, in a broad structural sense, like a trace formula.

That analogy is one of the reasons quantum-chaotic interpretations of the zeta zeros have been so influential.

It is an extraordinary example of mathematics providing a bridge between apparently unrelated descriptions of order.

But there is a trap here

This is also where I think interdisciplinary research can go horribly wrong.

Once you find one beautiful correspondence, it is incredibly tempting to start seeing correspondences everywhere.

We need to distinguish

$$\text{analogy} \neq \text{equivalence} \neq \text{proof}.$$

The fact that

$$R_2(u) = 1- \left( \frac{\sin\pi u}{\pi u} \right)^2$$

appears in both the GUE eigenvalue problem and the pair correlation of zeta zeros is extraordinary.

But it does not imply

$$\text{Riemann Hypothesis} \iff \text{GUE statistics}.$$

Likewise, finding a plausible Hamiltonian is not enough.

One needs to establish the exact spectral correspondence and all the necessary analytic properties.

This is an important part of interdisciplinary research that I think gets overlooked. The danger is not merely that disciplines fail to communicate. The opposite danger is that they communicate so enthusiastically that an analogy gets mistaken for a theorem.

A physicist may say, "This behaves like a quantum spectrum."

A mathematician should immediately ask,

$$\text{In what precise sense?}$$

What is the operator?

What is its domain?

Is it self-adjoint?

What is its spectrum?

What boundary condition is being imposed?

What is the exact correspondence?

What is proved and what is conjectured?

That friction is useful.

Why this matters beyond the Riemann Hypothesis

The reason I keep coming back to this example is not because I expect to solve the Riemann Hypothesis.

I very obviously don't.

What interests me is the method of thought.

A difficult problem can sometimes become more tractable when you stop asking,

$$\text{"How do I solve this problem?"}$$

and instead ask,

$$\boxed{ \text{"What other problem does this secretly resemble?"} }$$

That is a completely different question.

Suppose I encounter a complicated matrix. I might ask about its eigenvalues.

Then I notice that the eigenvalue statistics are universal.

Then I learn that similar statistics occur in quantum chaotic systems.

Then I find the same kernel appearing in a number-theoretic correlation problem.

At every stage, I am importing a language.

The progression might look like

$$\text{object} \rightarrow \text{invariant} \rightarrow \text{statistic} \rightarrow \text{analogy} \rightarrow \text{model} \rightarrow \text{new conjecture}.$$

The final step is particularly important.

Interdisciplinary research does not necessarily give you the answer.

It can give you a better question.

And sometimes that is more valuable.

Depth versus breadth

There is an obvious objection to all of this.

If I keep learning mathematics, physics, statistics, computer science, biology and everything else, when do I actually become good at anything?

This is a legitimate concern.

There is a real trade-off between depth and breadth.

If I know

$$D=\text{depth}$$

and

$$B=\text{breadth},$$

then it is tempting to imagine that research ability is something like

$$R=f(D,B).$$

But I don't think the useful relationship is additive:

$$R\neq D+B.$$

A person with enormous breadth but no depth may know many facts but lack the technical machinery to do anything with them.

Likewise, someone with enormous depth but no exposure outside their field may possess extraordinary technical ability but have fewer opportunities to recognise structures that appear elsewhere.

I would instead think of the useful part as multiplicative:

$$R\sim D\times C,$$

where $C$ is the ability to form meaningful connections.

If

$$D=0,$$

there is nothing substantial to connect.

If

$$C=0,$$

everything remains isolated.

The interesting region is where both exist.

That is why I don't think interdisciplinary research means becoming mediocre at five subjects.

It means becoming deep enough in at least one subject that you can recognise when something from another subject matters.

The boundaries are ours

Perhaps this is the thing I find most beautiful about mathematics.

We divide it into number theory, analysis, algebra, geometry, topology and so on. Then someone discovers that a problem in one area is secretly governed by an object from another.

The boundary moves.

Physics does the same thing.

A matrix is algebra until it becomes a Hamiltonian.

A differential equation is mathematics until it becomes the equation of motion.

A probability distribution is statistics until it describes a physical ensemble.

The equations do not care.

We care.

We create the categories because categories help us think.

But occasionally the categories begin preventing us from thinking.

The Riemann Hypothesis is a particularly extreme example.

It begins with

$$\zeta(s) = \prod_p(1-p^{-s})^{-1},$$

which connects analysis to prime numbers.

Its zeros satisfy a conjectured symmetry

$$\rho\mapsto 1-\rho,$$

and the critical line

$$\Re(\rho)=\frac12$$

becomes the central object.

Then the zeros acquire statistical structure.

Their correlations resemble

$$R_2(u) = 1- \left( \frac{\sin\pi u}{\pi u} \right)^2.$$

That is the same local correlation structure appearing in GUE random matrices.

Random matrices describe spectra.

Spectra describe quantum systems.

And suddenly the chain becomes

$$\boxed{ \text{Primes} \rightarrow \zeta \rightarrow \text{Zeros} \rightarrow \text{Correlations} \rightarrow \text{Random Matrices} \rightarrow \text{Spectra} \rightarrow \text{Quantum Chaos}. }$$

None of these arrows, individually, solves the Riemann Hypothesis.

But together they tell us something profound.

They tell us that the problem has more than one natural language.

And perhaps that is what interdisciplinary research really is.

Not putting two disciplines next to each other.

Not learning physics because you are a mathematician and therefore want to call yourself interdisciplinary.

Not collecting terminology from five fields.

It is the ability to recognise when an object has a structure that another discipline already knows how to study.

The most interesting research may happen precisely at those points where our categories stop working.

Because sometimes the problem is not difficult because we have not worked hard enough inside the existing framework.

Sometimes the problem is difficult because we are looking at it in the wrong language.

And perhaps the real advantage of learning across disciplines is not that you know more.

It is that you have more languages in which to be confused.