schur

Schur Complementary Portfolios

Bridges between heuristic and optimization-based portfolio construction.

Portfolio construction has long had two camps. One is global optimization: a single objective over the whole covariance matrix, with estimation error dealt with before the optimizer sees it, by shrinkage, factor structure and the like. The other is heuristic, motivated through hierarchies, clusters and risk budgets, and deals with estimation error inside the construction itself, by using the covariance piecemeal. The question that separates them is where to handle estimation uncertainty: at the estimation level, at the construction level, or both.

The Schur complement connects them. Augment each cluster's covariance with the Schur complement of the others before allocating, and the hierarchical construction reproduces the global optimization exactly, when each split is budgeted by the child's conditioned variance (the taxonomy says which encodings do this and which only approach it). A single dial $\gamma \in [0, 1]$ slides smoothly between the two worlds, a construction-level answer that sits on top of whatever estimation-level answer is in use. The best out-of-sample portfolios usually live somewhere in between.

The bridge first appeared in a blog post in November 2022, on one tree, and was formalized in the 2024 paper. The 2026 notes extend it to a family: nested clustered optimization, hierarchical equal risk contribution, hierarchical risk parity and plain inverse variance are each the near end of a bridge whose far end is the same optimizer, and the notes prove when the best place to stand is strictly inside. The taxonomy says which published “Schur” is which, and one picture shows all five roads descending to the same floor.

The intuition. When people build heuristic allocators they chunk: the assets are cut into blocks, by sector, by a dendrogram, by a clustering, and each block is handled on its own, with the covariance between blocks thrown away at the cut. The cut is a hard decision made on soft evidence. The Schur complement is a softer way to cleave. Conditioning a block on the others keeps what the cross-block covariance says while still solving one block at a time, and the dial says how much of it to keep. The cleaving can follow structure that is presumed, sectors or a prior, or structure that is measured, an estimated correlation. In either case, at one end of the dial it is the chunking everyone already does, at the other it is the global optimizer, and in between it is the same recursion trusting the cross-block evidence by degrees.

Scope. The scope here is to demonstrate the connection to the optimization school rather than to compete with it. Optimization done properly already handles estimation uncertainty at the estimation level: under quadratic loss the rule that minimizes out-of-sample variance is the optimizer applied to the posterior-mean covariance, and nonlinear shrinkage solves that problem exactly over rotation-invariant rules. The heuristic constructions belong to the same school. Each is the near end of a path whose far end is that optimizer, the path is a Schur complement damped by degrees, and the damping is the same reliability ratio a Bayesian would put on the coupling. Where to stand on the path is then a question one can answer, and the answer is often inside.

Heuristic Inverse variance diagonal only Hierarchical risk parity bisection, inverse-variance splits Hierarchical minimum variance bisection, min-variance splits HERC clusters, inverse cluster variance NCO clusters, an optimizer across Global optimization Minimum variance Σ⁻¹1, or Σ⁻¹σ, or Σ⁻¹μ each path is a Schur complement, damped by a dial dial = 0: the method as published inside: where an estimated covariance says to stand dial = 1: the optimizer, exactly

Five near ends, one far end. Inverse variance, hierarchical risk parity, hierarchical minimum variance, HERC and NCO each reach the same optimizer when their dial is turned to one, with the split rule the taxonomy page names for each, and the interior is where an estimated covariance says to stand. The 3D view shows the same picture with heights.

The block-inverse identity, the $\gamma$ interpolation, and the recursion at arbitrary hierarchy depth are derived in the introduction and the paper.

Implementations

Three maintained libraries cover the batch, streaming, and estimation layers: skfolio (fixed universe, sklearn idiom), allocation (online, evolving universe, Fiedler seriation), and precise (online covariance and the Schur pseudo-likelihood). Install commands, repositories, and tutorials are on the implementations page.

Further reading

Cite

Cotton, P. (2024). “Schur Complementary Allocation: A Unification of Hierarchical Risk Parity and Minimum Variance Portfolios.” arXiv preprint arXiv:2411.05807. 
@article{cotton2024schur,
  author  = {Cotton, Peter},
  title   = {Schur Complementary Allocation: A Unification of Hierarchical
             Risk Parity and Minimum Variance Portfolios},
  journal = {arXiv preprint arXiv:2411.05807},
  year    = {2024},
  url     = {https://arxiv.org/abs/2411.05807}
}

The bibliography lists works on, citing, or extending Schur complementary allocation, together with the spatial-statistics side of the same idea. Every entry also appears on the literature map or the timeline.