She had been staring at eigenvalues for three years before she thought to look at the holes.
Not holes in the data—holes in the topology. The kind you find when you wrap a quadratic form around a sphere and ask what shape the level sets make. Most people in her department studied the eigenvalues directly: their spacings, their correlations, the way they repelled each other like charged particles on a line. She'd done that too. It was fine work. Clean. Known.
The holes were different.
She started with GOE matrices—the symmetric ones, the ones that model time-reversal-invariant quantum systems, the ones whose eigenvalue spacing distribution Wigner had guessed correctly on a napkin in 1955. She computed the persistence diagram: born here, dies there, each topological feature a point in the plane. A scatter of lifetimes for the hills and valleys of a random landscape on a sphere.
The diagram had structure. Not the kind you impose—the kind that was already there, waiting. The birth-death pairs clustered along curves that depended only on the dimension and the symmetry class. Change the matrix, resample, repeat: the same curves. The same fingerprint.
Then she tried GUE—the Hermitian ones, no time-reversal symmetry—and the fingerprint shifted. Same sphere, same quadratic form, different symmetry class, different topology. The eigenvalue spacings were different too, of course, everyone knew that. But the persistence diagram caught the difference better. She computed the persistence entropy—one number summarizing the topological complexity—and it discriminated GOE from GUE more cleanly than the level spacing ratio that had been the standard tool for decades.
She sat with that for a long time.
The standard tool, the one everyone used, the one in every textbook—it was looking at the gaps between adjacent eigenvalues. One-dimensional information. The persistence diagram was looking at the global shape of a function on a high-dimensional sphere. It was seeing structure that the gaps couldn't see.
Her advisor said, "This is nice, but what does it mean physically?"
She didn't know yet. She knew it meant that the topology of randomness was not random. That the holes in a random landscape carried information about the symmetry of the system that created it. That you could read the symmetry class from the shape of the emptiness.
She wrote the paper. She called it a "new spectral diagnostic," because that's what journals accept. What she meant was: we've been reading the text when we should have been reading the margins. The spaces between the eigenvalues are a story, yes. But the spaces between the spaces—the voids, the tunnels, the cavities in the landscape of a random quadratic form—those are the deeper story. The one that was always there, in the shape of what's missing.