Papers
The core write-ups of Schur damping. The bibliography lists these alongside antecedents, independent rediscoveries, and derived works across fields.
Core
- Cotton, P. (2024). Schur Complementary Allocation: A Unification of Hierarchical Risk Parity and Minimum Variance Portfolios. arXiv:2411.05807. Original version: October 29, 2024; latest version: September 19, 2026 (corrected PDF, source). The canonical write-up: the damped Schur complement and the recursion at arbitrary hierarchy depth, with γ interpolating HRP and minimum variance.
- Cotton, P. (2026). Two Sides of Schur Damping: High-Dimensional Pseudo-Likelihoods and Portfolio Allocation. arXiv:2606.14798. The cross-field identity: the same damping, and the same closed-form reliability γ*, in portfolio allocation and in spatial pseudo-likelihoods.
Working papers & notes
- Cotton, P. (2026). When the Out-of-Sample-Optimal Schur Portfolio Lies Between HRP and Minimum Variance. Working paper: PDF. Where to sit on the bridge when the covariance is estimated with error, in four layers: a structural theorem (top-level cross-block noise is invisible at HRP and its expected cost factors as γτ²H, so small noise uniformly separates the optimum from HRP whenever coupling has population value); a local theorem at an exact minimum-variance endpoint (positive marginal noise sensitivity moves the optimum inside by order τ²); and an exactly solved four-asset family whose unique optimizer is interior and slides monotonically from 1 to ½ as the noise grows. A whole-matrix section then noises every entry of the covariance: a sharp second-moment condition Ξ > 0 decides interiority, exact enumeration certifies interior optima under entrywise noise, and under the sample covariance of T Gaussian draws the median-optimal coupling rises with T, from 0.62 at T=15 to 0.96 at T=250. Exact examples bound the general claims: HRP can become locally optimal at a finite noise level (boundary derivative −1/700 + (8/189)τ²), and the surrogate frontier is a clipped monotone inverse of τ² = −V₀′/G′. A closing section shows that every portfolio with a unit budget is the exact minimum-variance portfolio of some positive definite matrix. Separately, the nearest symmetric matrix that rationalizes a chosen allocation differs from the input by a correction of rank at most two, of the same algebraic form as Jagannathan and Ma's, and that nearest matrix can be indefinite; its distance from the data measures how far an allocator's implied beliefs sit from its own input: zero for minimum variance, 0.205 for HRP, 0.385 for equal weight. All exact claims are machine-verified in rational arithmetic.
-
Cotton, P. (2026). Nested Clustered Optimization Is One End of a Schur Bridge, and the Interior Is Sometimes Provably Better.
Working paper: SSRN 7480738,
PDF,
source,
certificate,
demos.
The flat-partition counterpart of the bridge. Block inversion writes the unconstrained
minimum-variance portfolio as one problem per cluster on a Schur-complemented block.
Conditioning each cluster on one knot from every other cluster, a truncation in the manner
of Vecchia approximations, is exact under a rank-one “gateway” model of
cross-cluster dependence. Damping by γ then runs from nested clustered optimization
at γ=0 to the global optimum at γ=1, with no linear solve larger than a cluster
or the number of clusters. The flat and tree
constructions agree at γ=1 and differ in the interior.
Under the gateway model only the knot exposures move along the bridge, in closed form; the member positions are the same at both ends. Under estimation error the optimal γ can be strictly interior, and full coupling can remain locally optimal: a local theorem at the minimum-variance end, with exact rational examples of both. A symmetric family reduces the question to one scalar allocation, gives the optimum in closed form (γ⋆=2/3 in the simplest case), and shows both signs in one model. An exact lost-precision identity makes full coupling the unique noiseless optimum, and a final section treats singular covariances, where damping by γ itself becomes singular. All identities and both examples are machine-verified. -
Cotton, P. (2026). Hierarchical Equal Risk Contribution Is a Corner of a Schur Square, and Two Dials Lead to Minimum Variance.
Working paper: PDF,
source,
certificate.
HERC budgets clusters by inverse cluster variance, which is the rule block inversion gives the
global minimum-variance portfolio. HERC lacks conditioning in two places: each cluster on
the others, and each asset on its cluster mates. Damping the two Schur complements by γ and
η gives a square with HERC at (0,0) and the global optimum at (1,1).
Raffinot's split formula, read with risk contribution as a precision, telescopes down any dendrogram to a flat inverse-fitness rule; the reference implementations sum variances instead and depart from it at three or more clusters. The inner dial has a closed form interpolating the naive and Stevens quantities, a long-only frontier, and on an equicorrelated cluster is the straight segment from inverse variance to cluster minimum variance, so its noise problem is one scalar: an exact interior optimum η⋆=3/5 with gaps 1/45 and 1/20, a sign criterion at full coupling with both signs realized, and a two-dial local theorem at the minimum-variance corner with exact examples where both dials move inside and where one stays. Every claim is machine-verified in rational arithmetic. - Cotton, P. (2026). The Thurstone Tilt Is a Schur Bridge, and Both of Its Ends Are Races. Working paper: PDF, source, certificate. Thurstone polishing dials the estimated correlation into a race with a scalar φ. Taking the reference to be the block-diagonal part of the estimate makes that blend the estimate with its cross-cluster entries scaled by φ, and for two blocks the conditional of that matrix is the damped complement at γ = φ². So the tilt dial is the Schur dial under a square root, the near end races the clusters independently, and the far end races them under the full estimate. Every other bridge here ends at a linear solve; this one ends on the simplex, splits a duplicated position rather than doubling it, and has no pole, the taper being positive semidefinite throughout. The closed-form reliability transfers as φ⋆ = √γ⋆.
- Cotton, P. (2026). The Far End of a Schur Bridge Does Not Care How Its Clusters Are Priced, and the Interior Does. Working paper: PDF, source, certificate. A two-tier allocation prices its clusters with a $k\times k$ matrix, and every construction so far builds that matrix from the same portfolios it holds. Separating the two gives a family with a trial space and a test space. At the far end the $k$ conditioned directions are the blocks of $\Sigma^{-1}u$ and sum to it, so any non-degenerate pairing recovers it and the choice of representative cannot change the far end at all. The interior is not free: an unmatched pairing loses symmetry and can pass through singularity inside the dial, with a positive-definite $6\times6$ example where it does and a rate of about a quarter of estimated markets. Pricing against the cluster portfolios makes the matrix a Gram matrix of $\Sigma$ and removes the possibility.
- Cotton, P. (2026). Covariance Localization Is Schur Damping, and Its Optimal Strength Is a Sampling-Error Correction. Working paper: PDF, source, verification. Ensemble Kalman filters localize by an elementwise taper. A taper that is constant on blocks is Schur damping: the localized covariance is positive semidefinite, its conditional covariance of the unobserved block is A − γBD−1B′, and the localized gain is √γ times the optimal gain. Under sampling error in the cross-covariance the expected analysis error is an exact quadratic in √γ, so the optimal coupling is strictly interior at every noise level, in closed form, and equals the square of a signal-to-noise shrinkage of the gain.
- Cotton, P. (2026). Schur Damping for Perpetual Demand Lending Pools. Working paper: PDF, web note. Proposes the interior of the dial for the single-pool-vs-two-pools decision of Chitra et al.: a partially-merged DeFi lending pool that hedges against $A - \gamma B C^{-1} B^{\mathsf T}$, with γ* set by how well the cross-pool covariance can be estimated. In the undersampled on-chain regime, γ* is pulled toward 0.