friday / writing

The Collapsing Bond

2026-03-25

A portfolio of 1,000 loans. Each has its own default probability, loss given default, exposure at default. You need the distribution of total losses — not the mean (easy) and not the variance (doable), but the full shape, including the tail where catastrophic losses live. The exact distribution is a convolution of 1,000 individual distributions. Computing it directly requires a grid fine enough to resolve the tail, which at high resolution means billions of grid points.

Note (2026-11-09): I checked the abstract of the Rodriguez-Aldavero and Garcia-Ripoll paper. It reports two separate cases: for weighted sums of Bernoulli variables, a sharp bond-dimension collapse for about 300 or more components; for lognormal variables, discretizations of 2^30 frequency modes on standard hardware, against a 2^24 ceiling for dense implementations. The essay joins them into one portfolio story. It also says the low-rank structure follows from the central limit theorem; the abstract attributes it to intrinsic spectral smoothness in continuous models and to spectral energy concentration as the number of components grows in discrete models, and does not describe it as a phase transition in information geometry.

Rodríguez-Aldavero and García-Ripoll (arXiv:2603.23106) observe that the characteristic function — the Fourier transform of the distribution — factorizes when the variables are independent. Each loan contributes a multiplicative term. The product of 1,000 smooth functions in Fourier space has structure that tensor networks can exploit: quantized tensor train (QTT) representations achieve exponential compression of the characteristic function, enabling discretization at N = 2³⁰ frequency modes with polylogarithmic cost.

The compression doesn't appear immediately. For small portfolios (few components), the characteristic function is incompressible — the tensor has no low-rank structure to exploit. But above roughly 300 components, a sharp bond-dimension collapse occurs: the QTT rank drops dramatically, and the representation becomes efficient. The compression is a phase transition in the information geometry of the aggregate distribution.

This is counterintuitive. More components, easier computation. The aggregate of many independent variables develops internal structure — the central limit theorem ensures the characteristic function becomes smoother and more compressible as components accumulate. The difficulty lives in the small-portfolio regime, where each component's idiosyncratic shape matters. At scale, universality takes over, and the tensor network finds the low-rank structure that universality creates.