Allow any copula, and a long-only tail-risk optimizer buys one asset.
The reason is direct. Hold the marginal loss distributions fixed while the joint law ranges over every coupling consistent with them. For any law-invariant, subadditive, comonotonically additive, positively homogeneous functional, the comonotonic coupling is worst: every asset suffers its bad draw together. Portfolio risk under that coupling is exactly the weighted sum of the marginal risks. The weights therefore enter through an affine objective, whose optimum on a simplex lies at a vertex. An ambiguity set this broad assigns no value to diversification. The paper's abstract leads instead with its quasi-convexity characterization, leaving the concentration paradox second.
Liu and Liu make no empirical claim. There is no sample period and no return data. Their evidence consists of theorems and constructed examples: a three-asset synthetic illustration in Section 7, a separate three-asset normal case in Section 8, and a two-asset normal-versus-Laplace example. Unless identified as ours, every figure below comes from those exercises.
The allocation results that matter
Proposition 5 contains the SA-CA-PH concentration statement, which the authors describe as adapted from Pflug and Pohl (2018). The paper claims four contributions. Three bear directly on allocation.
First comes Theorem 1. A risk measure is weakly consistent with the majorization order if and only if it is quasi-convex. Weak consistency takes a more diversified weight vector, obtained from the original through doubly stochastic smoothing, and compares it with the worst permutation of those original weights. The diversified vector cannot carry greater risk than that permutation. Theorem 1 gives a genuine characterization and differs from the convex-order condition in Chen et al. (2022). Quasi-convex yet non-convex functionals can still reward diversification in this weak sense. Proposition 1 shows why the weak formulation matters. Requiring majorization to lower risk for every collection of assets sharing a common mean forces rho(X) = f(E[X]), which responds only to the mean.
The practitioner result follows in Theorems 3 and 4, where the vertex argument extends beyond subadditivity. VaR generally lacks subadditivity, while RVaR is grouped with functionals that fail at least one of the three properties. Under stated conditions, both still concentrate. VaR+ at level alpha requires marginals in the decreasing-density class beyond their alpha-quantile, the increasing one, or n = 2. RVaR requires the decreasing-density class at level 1 - alpha - beta. The proofs use convolution bounds from Blanchet et al. (2025) and Fadina et al. (2025), followed by the same passage from linearity to an extreme point. Worst-case coupling produces the concentration. In the paper's account, this follows from the structure of the problem rather than the particular properties of the chosen risk measure.
Proposition 3 then reverses the familiar 1/n intuition for distributions with finite means. With identically distributed marginals, the equally weighted portfolio becomes the riskiest under the worst-case measure, while a single-asset portfolio is safest. Three settings deliver the reversal: rho is law-invariant, SA, CA and PH; rho = VaR_alpha with F in M^alpha_D or M^alpha_I; or rho = RVaR with F in M^{1-alpha-beta}_D. The authors say part (ii), covering VaR, agrees with Proposition 7.1 of Chen et al. (2022). The SA-CA-PH and RVaR cases supply the new material. Earlier work they cite on diversification failure, Chen et al. (2025a) and Chen et al. (2025b), depended on infinite-mean Pareto tails. Dependence ambiguity is sufficient here.
Risk aversion chooses a name
The three assets are arranged so each loses on either mean loss or tail risk. E[X1] = 4.0000 and its VaR+ at 0.99 is 10.0485. For E[X2] = 4.8000, the corresponding value is 6.8937; for E[X3] = 5.3000, it is 6.0951. The authors explicitly say the parameters were "deliberately chosen so that no asset simultaneously has the smallest mean loss and the smallest risk." The candor also confirms that this is a constructed illustration rather than a calibration.
Under VaR+ at 0.99, the optimizer invests fully in Asset 1 when kappa is below 0.2536. It switches to Asset 2 between 0.2536 and 0.6261, then to Asset 3 above 0.6261. RVaR(0.01, 0.10) produces the same three-region pattern, with cutoffs of 0.4116 and 0.8067. Greater risk aversion simply moves the portfolio through a menu of single names. At either threshold, the neighboring vertices tie, making every mixture of that pair optimal under the linear objective. Everywhere else, the solution contains no mixture.
Tail shape changes the switching point. The experiment fixes E[X3] = 5.3000 and SD(X3) = 0.2000 while varying the lognormal shape across [0.2, 1.5]. The Asset-2-to-Asset-3 cutoff ranges over [0.4815, 0.6261] for VaR+ and [0.6719, 0.8663] for RVaR. As the paper observes, dependence on eta is non-monotone. At fixed mean and variance, added skewness does not consistently increase a fixed 0.99 quantile. The three-region form survives throughout. The vertex remains stable; its identity moves.
The Laplace example carries the objection
Theorem 5 sharpens the issue because standard deviation is subadditive without being generally comonotonically additive. Comonotonic dependence still supplies the worst case. When every marginal belongs to one location-scale family, Theorem 5 makes worst-case SD linear in the weights and concentration follows. Beyond that family, concentration may fail, and the paper claims no converse.
Example 1 shows the failure with n = 2. One asset is standard normal; the other is a mean-zero, variance-one Laplace. Each vertex has worst-case SD of exactly 1. The equal mix instead has variance (1 + c)/2 with c < 1 because its increasing, standardized quantile functions are not affinely related. The value is strictly below 1. No vertex solves the problem.
The authors state the limitation themselves. The paradox "may be mitigated when assets exhibit substantially different distributional characteristics, thereby potentially restoring the benefit of diversification." This sits uneasily beside the abstract's headline claim for "many widely-used risk functionals." Example 1 already defeats concentration for SD at n = 2. The three-asset SD exercise opens the gap further by combining normal and lognormal marginals. For Z and exp(Z), comonotonic correlation is 1/sqrt(e - 1) = 0.7629. Across 0.6690 < kappa < 1.0276, the worst-case optimum is an interior combination of Assets 2 and 3. At kappa = 0.75, the weights are (0, 0.2035, 0.7965). At kappa = 0.4706, every mixture of Assets 1 and 2 is optimal. The comonotonic-additive measures concentrate exactly under the stated density conditions. Measures outside that class can lose the result quickly.
One unresolved dial
Section 8 blends two objectives: omega weights worst-case dependence risk, and 1 - omega weights the reference model. With a multivariate normal reference and ES, the objective is lambda'mu + c_p times (omega lambda'sigma + (1 - omega) sqrt(lambda'Sigma lambda)). The paper interprets the second component as favoring the asset with the lowest marginal standard deviation. The third rewards diversification when Sigma is positive definite. At omega = 1, the portfolio buys argmin_i ES_p(X_i). The authors carefully describe the resemblance to FRTB as conceptual rather than formulaic. Intermediate omega is characterized only by KKT conditions requiring numerical solution. VaR and RVaR "require separate analysis." No rule is offered for choosing omega.
That open choice contains the investment decision. The dependence-sensitivity table makes the stakes visible. At kappa = 0.50, the independent-reference solution is (0.3280, 0.6720, 0), with HHI 0.5592. Under worst-case comonotonic dependence at the same kappa, it becomes (0, 1, 0), with HHI 1.0000. The premium for worst-case protection is 0.1897. The paper defines this as the optimized increase in the mean-SD criterion and explicitly excludes interpretations as a transaction cost or regulatory capital charge. Its corresponding worst-case objective, V_SD(0.50), equals -5.2500. Using those two reported figures, our arithmetic puts the premium at 3.6% of the objective's magnitude. By kappa = 1.25, it has fallen to 0.0832.
Table 5 compares two fixed dependence models; it does not trace omega and therefore does not establish a discontinuity. The authors say the optimizer need not pass smoothly between regimes because the active set can change. The table still records a large shift, from a two-asset spread to one name, caused entirely by the assumed dependence model.
The simplex does essential work. Short-selling is excluded by construction. Nonnegative weights make comonotonic dependence worst for every allocation at once.
Our ETF blend lost money
We implemented the omega blend using liquid US ETFs. The universe was the top 50 by trailing one-year dollar volume. Rebalancing occurred monthly at the first trading-day close, with signals based on the prior close. The setup used 500 aligned daily returns, empirical 97.5% ES, long-only and fully invested weights, a 10% per-name cap, and omega = 0.50. We charged 5bps one-way plus $0.004 per share. The window ran from 2020-01-01 to 2024-07-01. Our linear program minimizes half the weighted sum of marginal ES plus half the historical portfolio ES across same-date loss scenarios.
The paper reports no figures from such a run. Ours produced total return -29.49%, Sharpe -0.52, maximum drawdown 24.88%, win rate 61.89%, and profit factor 0.31. Average wins were $793.14, against average losses of $4,207.92. A portfolio finishing 29.49% below its starting value cannot also have a 24.88% maximum drawdown, so that figure needs reconciliation before it carries weight.
This weak result speaks first to our construction. Four choices stand out. We minimized risk alone and included no expected-return term. The paper supports that setup only when expected returns are equal or pinned; across leveraged and near-duplicate ETFs, they plainly differ. The 10% cap requires at least ten holdings and prevents a vertex solution, leaving this run unable to test Theorems 3 to 5. With a 500-day window, the 2.5% tail contains roughly 12.5 observations. The marginal ES values treated as known in the paper therefore become noisy plug-in estimates here. Finally, selecting solely on tail risk through a window containing 2022 pushed the portfolio toward the low-volatility end of the universe. Realized volatility was 12.74%, with beta 0.15 to SPY. A long-only, fully invested portfolio had its money precisely where returns were absent. The pattern of many small wins and a few large losses, a 61.89% hit rate beside a 0.31 profit factor, fits a grinding long-only book that is short nothing and predicts nothing.
Would we trade the result?
We checked the theorem statements and their hypotheses rather than reproducing the proofs. The VaR and RVaR extension is a genuine contribution. As a portfolio design rule, we find it close to unusable; this judgment is ours rather than the paper's. The authors introduce the problem through Silicon Valley Bank, where several risk drivers "became highly aligned under stressed market conditions." Pairwise comonotonicity everywhere assumes far more than such alignment. Admitting it consumes the full diversification benefit.
The extension named in the conclusion could change our view: ambiguity sets constrained by partial dependence information, including correlations or other dependence characteristics. Our version would use a Wasserstein ball around a fitted copula. The paper already proves a Wasserstein result in Proposition 6, centered on a reference joint distribution F_0 rather than a copula. Proposition 7 covers the moment case. Both acknowledge that the norm penalty and mean-variance trade-off discourage concentration without ensuring an interior solution. Until the ambiguity set is narrow enough to believe, omega determines the portfolio and remains unset.
Our backtest stops at 2024-07-01, and everything after that date is deliberately left untouched so the same strategy can be checked out of sample later.