Minimum variance in disguise
The generality of the minimum-variance solution, and why every bridge lands on it.
The far end of every bridge is written as minimum variance. The object the bridges reach is the direction $\Sigma^{-1}b$ for a constraint vector $b$, and many portfolio problems, and problems outside portfolios, are exactly that with a different $b$. Whatever is a precision matrix times a vector is reachable by the same recursion, and the block-inversion identity that makes the far end exact is indifferent to which $b$ is in the numerator.
Every symmetric linear system has a financial reading
The 2022 blog post and the 2024 paper build on one observation. Take any symmetric positive-definite $Q$ and any vector $b$, and ask for the solution of $Qx=b$. Read $Q$ as a covariance and $b$ as a vector of exposures, one per asset. The portfolio that minimizes variance $x^\top Qx$ subject to unit exposure $b^\top x=1$ is
$$ w(Q,b) = \frac{Q^{-1}b}{b^\top Q^{-1}b}, \qquad\text{with variance}\qquad \nu(Q,b) = \frac{1}{b^\top Q^{-1}b}, \qquad\text{so}\qquad Q^{-1}b = \frac{w(Q,b)}{\nu(Q,b)}. $$The solution of the linear system is that portfolio scaled by the inverse of its own variance: a holding, and the capital it deserves. Linear algebra does not care what $b$ is, so neither does the reading. Ones give the fully invested minimum-variance portfolio, volatilities give the most diversified one, expected returns give the tangency portfolio, and a hedge ratio, a regression coefficient or a forecast combination are the same object with $Q$ and $b$ renamed.
The reading makes hierarchy possible. Block inversion says the solution splits by blocks, $(\Sigma^{-1}b)_C = (\Sigma^c_C)^{-1}\big(b_C - \Sigma_{C,-C}\Sigma_{-C,-C}^{-1}b_{-C}\big)$: each block is the same problem again on its Schur complement, with its part of $b$ conditioned on the rest as a companion vector. Read financially, each block holds the minimum-variance portfolio of its conditional covariance and is budgeted by the inverse of that portfolio's variance. That is the inverse-variance rule of every hierarchical heuristic, applied to conditioned rather than raw blocks, and it is why the far end of every bridge is exact for every $b$ at once.
Who is in the family
| problem | $Q$ | $b$ | solution |
|---|---|---|---|
| Minimum variance, fully invested | $\Sigma$ | $\mathbf 1$ | $\Sigma^{-1}\mathbf 1$, normalized |
| Maximum diversification (Choueifaty and Coignard) | $\Sigma$ | $\sigma$, the volatilities | $\Sigma^{-1}\sigma$: maximizes $\sigma^\top w/\sqrt{w^\top\Sigma w}$ |
| Tangency, maximum Sharpe ratio | $\Sigma$ | $\mu$, the excess means | $\Sigma^{-1}\mu$, normalized when $\mathbf 1^\top\Sigma^{-1}\mu>0$ |
| Kelly, growth-optimal under Gaussian returns | $\Sigma$ | $\mu$ | $\Sigma^{-1}\mu$, unnormalized: the same direction with its scale kept |
| Maximum decorrelation | $R$, the correlation | $\mathbf 1$ | $R^{-1}\mathbf 1$: minimum variance with the volatilities stripped out |
| Optimal forecast combination (Bates and Granger) | the forecast-error covariance | $\mathbf 1$ | the same formula, on forecasts instead of assets |
| Minimum tracking error with one exposure target | $\Sigma$ | $c$, the exposure | benchmark plus a multiple of $\Sigma^{-1}c$: minimum variance in active weights |
| Hedging one position with several instruments | $\Sigma_{yy}$ | $\Sigma_{yx}$ | $\Sigma_{yy}^{-1}\Sigma_{yx}$, the regression coefficients |
| Generalized least squares, one regressor | $\Sigma$, the error covariance | $x$, the regressor | $\Sigma^{-1}x$ up to scale: the estimator is a portfolio of observations |
| Bayes rule under quadratic loss | $\bar\Sigma$, the posterior mean | $\mathbf 1$ | $\bar\Sigma^{-1}\mathbf 1$: the direct out-of-sample optimum is still in the family |
Two extensions cost nothing. Mean-variance with a budget has solution $\Sigma^{-1}(a\mathbf 1 + c\,\mu)$, two directions from the family combined, and a bridge run with the companion matrix $[\mathbf 1,\ \mu]$ delivers both columns at once, since block inversion applies columnwise. Several affine constraints $B^\top w = c$ give $w = \Sigma^{-1}B\,(B^\top\Sigma^{-1}B)^{-1}c$, which again needs only $\Sigma^{-1}B$.
Who is not
Three families of problems are not a precision matrix times a vector, and no bridge is exact for them at the far end. Long-only and box-constrained problems have an active set, and their solutions are $\Sigma^{-1}b$ only on the assets left free, with $b$ adjusted by multipliers that depend on the data; the recursion still runs, and the HERC square has a closed-form frontier for how far the inner dial can turn before the first short appears, but the far end is not the constrained optimum. Risk parity and its budgeted variants are fixed points of $w_i(\Sigma w)_i = \beta_i$, not linear solves. And objectives that are not quadratic forms, expected shortfall, drawdown, utilities beyond the second moment, have no companion vector at all; hierarchical versions of them exist, HERC admits them, but the far end is exact only through the identity above, which needs a quadratic form.