A portfolio manager has no evidence here for choosing stochastic gradient Langevin dynamics (SGLD) over a convex CVaR program. Ding, Sabanis and Zhang carefully prove that SGLD can minimize a problem their reformulation has made harder. The theory earns its place. But the headline claim of constants independent of the number of assets requires holding fixed a moment that grows as assets are added.
The paper reports no trading results, returns or risk-adjusted performance. All its numerical evidence comes from simulated, independent Gaussian losses.
Why abandon the convex formulation?
Under the Rockafellar and Uryasev representation, CVaR at level 0.95 is the minimum over a threshold η of η + E[(portfolio loss - η)+]/0.05. The minimizing η is the VaR. With long-only weights on the simplex, the objective is convex; historical scenarios turn it into a linear program.
The authors instead remove the constraints by expressing weights as a softmax of free logits w. This makes the weights positive and ensures they sum to one. They also add a penalty γ|θ|² on the threshold and logits together. The resulting objective is nonconvex in w, and its stochastic gradient jumps because it contains the indicator 1{portfolio loss ≥ η}. This gives a Langevin sampler an unconstrained problem to work on. As β increases, the Gibbs measure proportional to exp(-βV) concentrates on global minimizers.
That choice changes the solution. Setting the expected η-gradient to zero produces a tail probability of 0.05(1 + 2γη), rather than 0.05, so the fitted threshold targets a slightly different quantile. For γ = 5×10^-5 and η near 1.9, the shift is about two parts in ten thousand. The penalty pulls logits toward zero and thus weights toward equality. In the second three-asset example, the third asset is a near-certain loss of 3 with variance 10^-4. Even the reference solution assigns it 0.0004 rather than zero; SGLD assigns it 0.0068.
Elsewhere, the small examples agree closely. SGLD's first three-asset solution has weights (0.4631, 0.2102, 0.3267), VaR 1.9042 and CVaR 2.2310. The reference values are (0.4637, 0.2102, 0.3261), 1.9056 and 2.2304. At N = 15, the top three weights match to the second decimal, and CVaR is 1.3038 against 1.3001. Those references minimize the regularized objective. They check the solver while leaving the regularization bias unmeasured.
Linear in N, empty numerically at β = 10^7
Theorem 1 bounds the Wasserstein-2 distance from the iterate's law to the Gibbs measure by 2C5·λ^{1/4} + e·C6·exp(-nλ/T0). The step-size contribution shrinks slowly. Halving its error requires a step sixteen times smaller.
Theorem 2 turns that result into an excess-risk bound, multiplying the convergence term by C_V and adding a Gibbs term of order (N+1)/(2β)·log(...). With C5, C6 and C_V all O(√N), and T0 independent of N, the combined bound grows roughly linearly in N. The comparison is with the (N+1)^{9/2} dependence of tamed SGHMC in the parameter regime used for that paper's CVaR application. This is a large improvement. It also avoids the density assumption on losses required by the multilevel VaR papers the authors survey.
The "dimension-independent" constants deserve a closer look. The radius beyond which convexity holds is r = 2(1 + E[M(X)])/(γ(1 - q̄)). Both this radius and the semi-convexity constant k contain E[M(X)]; the growth constant m contains E[M(X)²]. The paper defines M(x) = 2·max_i |x_i|. As the asset count rises, so does the expected maximum absolute loss, slowly for Gaussians and faster for fat tails. Calling r, k and m independent of N therefore treats E[max_i |X_i|] as fixed. A portfolio of 160 assets does not have the same expected maximum as a portfolio of 30.
Then there is the size of the constants. The experiments use γ = 5×10^-5 and q̄ = 0.95, giving a denominator γ(1 - q̄) of 2.5×10^-6. This makes r equal to 8×10^5·(1 + E[M]). The bound on the log-Sobolev constant is (2/(βγ))·exp(β(4r(√m + √L·r) + γr²/2)), with β = 10^7. Its exponent enters T0, while C5 contains exp(2kT0). At the experimental settings, the explicit bound provides no quantitative guidance. The scaling argument survives, and that is where the authors put their emphasis.
For the dimension experiment, N takes values in {30, 38, 48, 62, 78, 99, 126, 160}. The authors average 100 chains of 2×10^6 steps at λ = 5×10^-4 and estimate a log-log slope of about one. They describe this as consistent with Theorem 2; the abstract goes further and says the experiments exhibit the predicted dimension dependence. Yet the experiment does not distinguish the terms in the bound. Any error accumulating coordinate by coordinate can produce a slope of one, including the additional variance from stochastic gradients at λ = 5×10^-4. The fitted slope therefore cannot distinguish the theorem's constants from generic per-coordinate scaling. All eight points also use one fixed draw of means and volatilities.
5 seconds for SGLD, 2,748 for PSGLA
The timing table supplies the only computational comparison. It covers a single N = 10 instance, averages each checkpoint over 100 trajectories, and sets tolerance at 5×10^-3. The reported iterations to tolerance, final excess risk and elapsed time are:
- SGLD: 41,000, 4.8607×10^-3, 5 s
- tSGHMC: 143,000, 4.7941×10^-3, 22 s
- Proximal stochastic subgradient: 43,000, 4.2746×10^-3, 1,190 s
- PSGLA: 29,000, 4.3068×10^-3, 2,748 s
PSGLA takes the fewest iterations and the longest time. The authors explain that the proximal methods "incur substantially higher computational costs due to the numerical proximal updates used in our implementations." On their account, the 238x elapsed-time gap (1,190 s against 5 s) largely measures the cost of their numerical proximal step. The proximal methods also end further below the 5×10^-3 target: 4.27×10^-3 and 4.31×10^-3, compared with 4.86×10^-3 for SGLD. A checking grid spaced at 10^3 iterations allows those differences.
We did not find a run solving the original convex problem directly, either through a scenario linear program or through the closed-form Gaussian objective already used for the references. A desk would try that first.
Every method in the table tackles the same nonconvex softmax formulation. The authors themselves restrict their claim to that setting: the abstract claims efficiency "for the optimization problem under consideration", and the comparison section closes with "for the CVaR portfolio optimization problem considered here." The table therefore ranks solvers within the authors' chosen formulation. Why restrict the benchmark to softmax when the convex program in their own equation (1) is available? Nor does the table establish portfolio quality, since all four methods stop at the same excess-risk target.
Our ETF adaptation became an equal-weight book
We built an adaptation of our own. The paper trades nothing, and the figures below belong to our construction. We replaced simulated Gaussians with historical daily losses on SPY, QQQ, IWM, TLT and GLD. At each month-end, we used the trailing 504 daily joint losses and selected γ, λ and β from a 27-point grid over 12 monthly validation intervals. We then refit from zero using 100,000 SGLD updates, resampling whole rows from the scenario history. Our backtest covers 2020-01-01 to 2024-07-01. Commissions are $0.0040 a share, subject to a $1.00 order minimum.
Our execution imposed a 20% position cap. Across five assets, that forces every holding to 20%, so the SGLD portfolio, the convex CVaR benchmark and equal weight all reduced to the same monthly-rebalanced equal-weight book. From January 2020 to July 2024, that book returned 42.69%, with a Sharpe of 0.64 (Sortino 0.84) and volatility of 14.82%. Maximum drawdown was -28.62%, and Calmar was 0.29. These figures tell us nothing about SGLD's allocation ability. We also filled at the close instead of the intended next session's open. Our implementation is at fault.
Removing the cap would still leave a separate gap between the theory and our data. Daily returns have serial dependence, and the equity sleeves are correlated. The paper assumes i.i.d. observations, and its diagonal-covariance Gaussians never test those conditions.
The probability result earns its place
The structural lemmas are a genuine advance in probability theory. They establish linear growth, convexity at infinity and a one-sided Lipschitz condition for a discontinuous gradient without a density assumption. Extending the argument to SCAD-penalized least squares shows that it reaches beyond this application. The portfolio case remains incomplete. I would change my view if SGLD beat a scenario LP on held-out CVaR at N in the hundreds, with E[max_i |X_i|] reported alongside N.
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.