How Far Can a Sparse Polygenic Score Go in Each Ancestry? Dual Bounds That Separate Fitting Bias from the Discovery Gap

A polygenic risk score is a weighted sum of a person's genetic variants that predicts disease risk. Scores built from European-dominated studies predict much worse in people of African or other non-European ancestry. When a new multi-ancestry method improves African-ancestry accuracy a little, nobody can say how much room is left. One possibility is that the fitting procedure is still pulled toward the majority population, which better optimization or reweighting would fix. The other is that the chosen variants simply do not carry the signal in that population, which only new genetic discovery can fix. Today these two explanations cannot be told apart, because every method returns a single point estimate and no statement about what was achievable.

We propose to compute that statement. We pose sparse multi-ancestry scoring as a constrained problem: fit well overall, use at most $k$ variants, and hit a target accuracy in each ancestry group. We then use Lagrangian duality to obtain a lower bound on the error that any score from that variant set could reach in each group. Because the variant budget makes the problem non-convex, the dual bound is generally not tight. The size of that "duality gap" is itself informative. It is exactly the set of trade-offs that group reweighting can never reach. Where the bound is close to the best score we can actually find, the remaining African-ancestry shortfall is certified as a discovery gap rather than an optimization failure.

Core Claim

Three claims are combined here, and to our knowledge none appears in the work we surveyed. First, for summary-statistic polygenic scores with squared loss, each group's loss is a convex quadratic in the effect sizes. All non-convexity therefore comes from the variant budget or non-convex sparsity penalty. That makes a weak-duality lower bound computable at genome scale. Second, the Lagrangian of the constrained problem is precisely a group-reweighted penalized regression. So the duality gap measures how much accuracy reweighting forfeits compared with solving the constrained problem directly. Third, if the bound is evaluated on the held-out ancestry's own data, it becomes a conservative certificate of the form "no $k$-variant score from these variants could do better here." That is an operational definition of the discovery gap that Ancestry Calibration of Polygenic Risk Scores Improves Risk Stratification and Effect Estimation in African American Adults names but cannot measure.

Formally, let $\beta$ be the shared effect sizes plus per-ancestry offsets. Let $L_g(\beta) = \beta^\top R_g \beta - 2\beta^\top r_g + 1$ be the standardized squared-error loss in ancestry group $g$. Here $R_g$ is that group's correlation matrix among variants (its linkage disequilibrium, or LD) and $r_g$ is its vector of marginal variant–trait associations. The problem is

$$ p^\star = \min_{\beta:\ \|\beta\|_0 \le k} \; L_{\mathrm{EUR}}(\beta) + \rho\|\beta\|_2^2 \quad \text{s.t.} \quad L_g(\beta) \le c_g \ \text{for each non-European group } g, $$

where $\|\beta\|_0$ counts nonzero effects, $\rho$ is a small ridge weight, and $c_g$ is the target loss for group $g$. Its dual function is

$$ d(\lambda) = \min_{\|\beta\|_0\le k} \; L_{\mathrm{EUR}}(\beta) + \sum_g \lambda_g \big(L_g(\beta) - c_g\big) + \rho\|\beta\|_2^2, $$

with multipliers $\lambda_g \ge 0$. Evaluating $d(\lambda)$ at fixed $\lambda$ is a group-reweighted fit. Any lower bound on that inner minimum still lower-bounds $p^\star$. We obtain such a bound from the convex relaxation of the variant-selection indicators, which the ridge term keeps informative. As [Lasserre's measure-theoretic view, summarized in the file] predicts, maximizing $d$ amounts to optimizing over randomized mixtures of sparse scores. The gap $p^\star - \max_\lambda d(\lambda)$ is the price of insisting on a single deterministic score.

Relation to Prior Work

Two papers motivate the duality machinery. The empirical duality gap of constrained statistical learning bounds the gap between empirical and statistical duals for constrained neural classifiers. On UCI Adult it raised gender insensitivity to 99.7% at under 0.1% accuracy cost. Learning with Statistical Equality Constraints shows that overparameterization heals the gap and enforces demographic parity on COMPAS. Both use duality to train. Neither uses the dual value as a reported bound. Both concern a binary demographic attribute and dense networks, where the gap is argued to vanish. We study the opposite regime: deliberately sparse, interpretable scores, where the gap is expected to be non-zero and is the object of interest. Their result constrains us. If hidden strong duality holds for polygenic scores too, our second claim reduces to "reweighting is enough," and we will report it as such. Bridging the Duality Gap in Non-Convex Optimization supplies the warning we act on: without a duality-gap certificate, a non-convex solver can stall at a poor point with no signal. Here, that means a minority-dominated objective silently stalls at a majority-dominated stationary point.

Two papers supply the primal method. Moreau Envelope Augmented Lagrangian Method for Nonconvex Optimization with Linear Constraints stabilizes augmented Lagrangian multipliers for weakly convex non-smooth objectives. Variable Smoothing for Weakly Convex Composite Functions gives an $O(\epsilon^{-3})$ method for minimum-concave-penalty composites, versus $O(\epsilon^{-4})$ for plain subgradients. A Moreau envelope is a smoothed copy of a kinked function obtained by blending it with a quadratic. Both papers were tested only on synthetic quadratic programs and image denoising. We use them to find good feasible scores, which give the upper side of the sandwich. Our constraints are convex quadratics rather than the linear constraints the first paper assumes, so we either linearize them per outer step or treat that extension as a contribution.

On the genetics side, Multi-ancestry polygenic risk scores using phylogenetic regularization (DendroPRS) shares effects along a population tree with lasso and group-lasso penalties. PRISM feeds functional annotations and ancestry-matched fine-mapping into penalized regression. Both return one score with no statement of optimality. They are the strongest primal competitors and also the inputs whose variant sets we certify. Ancestry Calibration recalibrates scores against continuous African ancestry fraction. It improves stratification, but by construction it cannot change who ranks above whom within an ancestry bin.

Research Program

The first component is the bound itself. For each trait, ancestry grouping, and variant budget $k$, we compute $\max_\lambda d(\lambda)$. The dual function is concave but non-smooth, because the active variant set switches as $\lambda$ moves. We therefore maximize it with a bundle-type supergradient method and record the multipliers. Each evaluation of $d$ needs only $R_g$ and $r_g$. These come from GWAS summary statistics and LD reference panels: UK Biobank ancestry subsets and the 1000 Genomes continental panels described in the file. We work block-wise along the genome, so memory scales with LD blocks rather than millions of variants.

The second component is the primal side. We solve the constrained problem directly with the Moreau-envelope augmented Lagrangian, using a non-convex sparsity penalty as a continuous stand-in for the hard budget, followed by hard thresholding to exactly $k$ variants. We also solve it starting from DendroPRS and PRISM solutions, and keep the best feasible score as the upper bound. Sweeping the targets $c_g$ traces two curves. The certified lower frontier comes from the dual. The achieved upper frontier comes from the primal. The weighted-sum frontier from reweighting lies between them and is supported only on the dual's convex hull.

The third component turns the sandwich into a diagnosis. We define a discovery shortfall for group $g$: the gap between the dual lower bound on $L_g$ under the European-derived variant set and the same bound after adding ancestry-specific fine-mapped variants. We define an optimization shortfall as the gap between a published method's achieved $L_g$ and the dual bound. Making the bound conservative is a design choice. When we compute it from the target group's own held-out sufficient statistics, the bound reflects an oracle that tunes on evaluation data. It can therefore only be optimistic about achievable accuracy, so an "infeasible" verdict errs on the safe side. Continuous admixture enters by defining the groups $g$ as bins of genetic similarity to Yoruba, following Ancestry Calibration. Each bin then gets its own constraint, which avoids the discrete branch boundaries that Gap 4 criticizes in tree regularization.

Baselines

The naive reference point is a European-trained score applied unchanged. Alongside it we run a simple pooled score fit on all individuals. DendroPRS reports a correlation of 0.03 for the pooled model on BMI in aggregated non-White UK Biobank participants, against 0.10 for single-population and 0.14 for DendroPRS (Multi-ancestry polygenic risk scores using phylogenetic regularization). These numbers anchor the scale on which any bound should sit.

The strongest existing methods are DendroPRS and PRISM. PRISM reports an average 13.10% $R^2$ improvement for African ancestry over annotation-agnostic models, and 36.5% for the FEV1/FVC ratio (PRISM). We rerun both on identical UK Biobank splits, with identical LD panels and their recommended tuning grids. To keep the comparison fair, both are then truncated to the same budget $k$. Beating them has two meanings. On the primal side, our constrained solve should reach a lower European loss at equal non-European loss, or the reverse. On the diagnostic side, we report how far each method sits above the certified bound. A method within a few percent of the bound has nothing left to gain from better fitting.

Post-hoc calibration is run on REGARDS, as in Ancestry Calibration. That paper reports height correlation rising from 0.24 to 0.32, and the type 2 diabetes top-versus-bottom-decile odds ratio rising from 7.97 to 10.77. We apply it on top of every score, including ours, because it is complementary rather than competing.

The ablation that isolates the key ingredient is a reweighting sweep: the inner problem of $d(\lambda)$ solved with the same Moreau solver on a dense grid of $\lambda$, at the same compute budget as the constrained solve. Any accuracy the constrained solve attains beyond the best reweighted fit is attributable to the non-convex gap and nothing else. A second ablation removes the convex relaxation of the inner minimum and uses the primal solver's local minimum instead. We expect this "bound" to be violated on small problems, which shows why the relaxation is necessary for validity.

Why not just…?

Why not just reweight the ancestry groups in the loss? Reweighting is our Lagrangian. At a fixed weight it is identical to evaluating $d(\lambda)$, so it can only reach trade-offs on the convex hull of the achievable frontier. This is a real threat. If the hidden strong duality found for overparameterized models in Learning with Statistical Equality Constraints also holds for sparse scores, the constrained solve adds nothing. The reweighting ablation settles it directly. We will publish the measured gap even if it is near zero, because a certified zero gap is itself useful: it means reweighting provably suffices.

Why not just fit an ancestry-specific score on the minority cohort and look at its held-out $R^2$? That shows what one estimator achieves on a few thousand people. It says nothing about the best attainable score, and it ignores the joint constraint on European performance. It also confounds the small sample size with missing signal. DendroPRS already shows single-population models losing to shared models on BMI (0.10 versus 0.14). We include it as a primal candidate. It cannot bound anything.

Why not just calibrate post hoc? Calibration fixes location and scale shifts across ancestry, and REGARDS shows that this matters clinically. But it is a monotone transformation within each ancestry bin, so it leaves within-bin ranking and discrimination untouched. Its own authors note that it cannot address the discovery gap. Our bound is computed on the quantity calibration cannot move.

Why not just solve the sparse problem exactly with mixed-integer optimization? That would remove the gap, but it does not scale to genome-wide variant sets. We use exact solves only where they are feasible, on small Geuvadis cis-windows, to validate the bound.

Proposed Experiments

The first decisive check is validity and tightness on problems small enough to solve exactly. For Geuvadis gene-expression prediction in European and African samples, cis-windows hold tens to hundreds of variants. There we enumerate or exactly solve the $k$-sparse constrained problem and confirm that the dual bound never exceeds the true optimum. We also record how much of the gap the convex relaxation of the inner minimum loses. DendroPRS reports improvements for 67 genes there, but no reported per-gene $R^2$ values were found on Lacuna for Geuvadis. We will therefore report against our own reruns.

The second check is planted simulation. We simulate traits on 1000 Genomes haplotypes with African-specific causal variants, then delete those variants from the candidate set. The African-ancestry lower bound should rise by roughly the deleted heritability while the optimization shortfall stays flat. Restoring the variants should reverse this. If the bound does not respond, it is not measuring the discovery gap, and the diagnostic claim fails.

The third experiment is the main real-data study, on UK Biobank. It covers BMI (where DendroPRS reports 0.14), FEV1/FVC (where PRISM reports its 36.5% gain), height, and type 2 diabetes, for African, South Asian, and East Asian ancestry subsets. For several budgets from hundreds to tens of thousands of variants, we report the certified and achieved frontiers, the reweighting frontier, and each published method's distance above the bound. We then repeat with PRISM's ancestry-matched fine-mapped variants added. This splits each trait's African-ancestry shortfall into an optimization share and a discovery share.

Finally, we transfer the certified scores to REGARDS African American adults. Bounds there are computed from REGARDS held-out statistics, and ancestry bins come from genetic similarity to Yoruba. We then ask whether calibration on top of our scores preserves the reported top-decile gains.

The proposal fails cleanly in three ways. If bounds are routinely violated on Geuvadis, the method is invalid. If the gap is negligible everywhere, the conclusion is that reweighting suffices. If the bounds are too loose to distinguish any published method from the frontier, the bound is uninformative. Each outcome would still settle a question the field currently cannot answer.