The global minimum variance portfolio minimizes risk without any return target — it asks only “what combination of assets has the least volatility?” Without constraints, the answer involves shorting overexposed assets to hedge factor risk. With a long-only constraint — no short positions — the answer changes structurally, not just quantitatively.
Note (2026-11-10): I checked the abstract and searched the full text of arXiv 2603.07692. The paper's result is a threshold rule. In the one-factor case the assets that get weight are those whose factor exposure (beta) is below a threshold index, and the paper says the solution does not depend on the assets above it. In the multi-factor case it describes the included assets as those whose exposure vectors lie below a threshold set by a hyperplane in the space of exposures. The essay says the included assets are “the ones whose factor exposures happen to cancel through positive-weight combinations” and that the solution lives on “facets of the weight simplex”. I did not find either of those in the text. The paper's picture is low exposure under a threshold, not cancellation, and its geometry is the hyperplane in exposure space. The essay also leaves out the paper's empirical part (about 1,000 US stocks, with shrinkage estimators of the covariance).
For covariance matrices arising from factor models, the relationship between the long-only minimum variance portfolio and the factor exposures has an explicit characterization. In the single-factor case, the solution can be written directly in terms of covariance parameters — closed-form expressions for which assets receive weight and which are excluded. In the multi-factor case, the structure is geometric: the constraint boundary shapes the portfolio through the interaction of factor exposures, and the solution lives on specific facets of the weight simplex determined by the factor structure.
The through-claim: the long-only constraint doesn't just truncate the unconstrained solution at zero — it changes what the portfolio is doing. The unconstrained minimum variance portfolio hedges across all factors by shorting assets with high exposure and buying assets with low exposure. The long-only version cannot hedge — it can only select. The assets it includes are not the ones with the lowest variance but the ones whose factor exposures happen to cancel through positive-weight combinations. The constraint transforms an optimization over all positions into a selection problem among assets whose exposures naturally diversify. The geometry of the factor model determines which assets are even candidates, before any optimization begins.