{"id":"2508e969-2118-4a30-a9ff-b69a38b29530","arxiv_id":"2411.12323","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Mirror descent with a tamed gradient provably computes risk budgeting portfolios for sub-additive positive homogeneous risk measures, with a.s. convergence and explicit averaged-iterate rates.","lead":"This paper designs an optimization routine that computes portfolios where each asset's share of total risk matches a target budget, for many different ways of measuring risk. The routine comes with mathematical convergence guarantees and beats an older stochastic method in numerical tests, especially for portfolios with many assets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3.2 is false as stated: for a two-asset volatility model with nonzero mean returns, the tamed expected gradient has negative inner product at an interior point, so the supermartingale step in the proof of Theorem 3.2 fails and the claimed a.s. convergence/rate is unproved.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: Lemma 3.2 is asserted without proof and is false in general, and the stochastic convergence theorem depends on it. My stress test makes this concrete with an interior counterexample for volatility with nonzero mean returns, the very risk measure highlighted in the paper. This is not a stylistic or consensus objection: the falsity of the lemma invalidates the supermartingale inequality (A.8) and the liminf argument in Step 1 of the proof of Theorem 3.2. The paper contains no machine-checked proof, no released code, and no numerical experiment with nonzero mean returns for volatility, so the invalid step is not rescued by independent evidence. The deterministic MD analysis, the ES case satisfying (3.11), and the extensive SMD-versus-t-SGD comparisons remain valuable and appear sound; hence the appropriate action is a major revision conditional on repairing Lemma 3.2, proving the claimed adaptation for volatility and variantiles, and ideally releasing code. I therefore keep the reader's CONDITIONAL verdict rather than escalating to REJECT, because the numerical evidence suggests the algorithms work in practice and the flaw may be repairable.","tokens_in":27679,"tokens_out":17935,"duration_ms":180974,"concrete_test":"Verify Lemma 3.2 against the d=2 volatility example above. With mu=(10,0), Sigma=I_2, b=(1,1), r=1/sqrt(2), T=3/2, kappa=r, a=(1-r)*10*(T-r)/2, xi=a-10T, z=(xi,(T,r)), z*=(-10r,(r,r)), compute F=(2a, kappa(2T+20a)-kappa/T, kappa*2r-1) and evaluate <F, z-z*>. The numerical value is about -1.4, contradicting the lemma. A second independent check: re-derive Step 1 of Appendix A.3 without Lemma 3.2; if no alternative argument is provided, Theorem 3.2 should be weakened by adding a monotonicity condition on the tamed operator or by removing the volatility and variantile claims.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 3.2 rests on Lemma 3.2, which asserts that the inner product of z-z* with the expected tamed gradient is strictly positive for every z != z*, with the proof omitted because it allegedly follows from convexity of h. Convexity alone is insufficient: the taming factor kappa(y)=min_i y_i multiplies every coordinate of the y-gradient by the same scalar, which can destroy strict monotonicity when the mean return vector is nonzero. A concrete d=2 volatility counterexample: take mu=(10,0), Sigma=I_2, b=(1,1), g(x)=x^2, L(xi,x)=(x-xi)^2. Then z*=(xi*,y*) with xi*=-10/sqrt(2) and y*=(1/sqrt(2),1/sqrt(2)). Choose y=(3/2,1/sqrt(2)), so kappa=1/sqrt(2), set a=(1-kappa)*10*(3/2-1/sqrt(2))/2 and xi=a-15. The tamed expected gradient is F=(2a, kappa(3+20a)-kappa/(3/2), 0), and the inner product <F, z-z*> is approximately -1.4 < 0. This point is interior and lies in B_m for m=2.5 >= ||y*||_1, so it is within the algorithm's domain. Because inequality (A.8) needs the inner product to be positive to yield a Robbins-Siegmund supermartingale, and Step 1 of the proof uses liminf of that inner product to conclude z_infinity=z*, a negative value invalidates both the a.s. convergence and the rate bound (3.12). The paper also concedes that condition (3.11) fails for volatility and variantiles and only states that the proof can be adapted; even if (3.11) were repaired, the false lemma would still leave Theorem 3.2 unsupported. The deterministic MD results and the ES numerics are not affected, but the advertised general-risk-measure stochastic convergence theorem is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes deterministic and stochastic mirror descent algorithms for computing risk budgeting portfolios under positive homogeneous and sub-additive risk measures, building on the variational characterization of risk budgeting portfolios in [15]. The algorithms are based on a tamed gradient kappa(y) grad Gamma_g(y) with kappa(y) = min_i y_i ^ 1, which is bounded on the domain and intended to vanish only at the minimizer. The deterministic MD scheme is proved to converge with a non-asymptotic rate on the averaged sequence (Theorem 3.1), and a stochastic MD scheme is claimed to converge almost surely with a similar rate (Theorem 3.2) under condition (3.11). Numerical experiments compare the methods with projected SGD for Expected Shortfall, volatility, and deviation measures across portfolio sizes up to 250 assets.","tokens_in":28167,"tokens_out":9399,"duration_ms":100446,"significance":"If the stochastic convergence theorem were correct, the paper would make a useful contribution: it would provide the first a.s. convergence guarantee with a quantitative rate for risk budgeting portfolios across general risk measures, going beyond the unanalyzed projected SGD of [15]. The deterministic convergence proof is coherent, the explicit proximal updates in (3.4) and (3.10) are a practical strength, and the numerical study is extensive, including stability comparisons with classical and tamed SGD. However, the central stochastic claim rests on a false monotonicity lemma, and one of the paper's headline examples, volatility, does not satisfy the square-integrability condition required by Theorem 3.2. The advertised main result is therefore not established as it stands.","major_comments":[{"comment":"Lemma 3.2 is false as stated, and Theorem 3.2 relies on it in an essential way. The proof is omitted with the explanation that the result follows from convexity of h, but the tamed expected gradient is not grad h(z): the y-components are multiplied by the common scalar kappa(y) = min_i y_i ^ 1. Convexity gives <z - z*, grad h(z)> > 0, which does not imply the displayed inequality for the tamed vector. A concrete counterexample is the two-asset volatility model with X ~ N((10,0), I_2), b = (1,1), g(x) = x^2, and L(xi,x) = (x - xi)^2. Here z* = (xi*, y*) with xi* = -10/sqrt(2) and y* = (1/sqrt(2), 1/sqrt(2)). Take y = (3/2, 1/sqrt(2)), which is interior and lies in B_m for m = 2.5 >= ||y*||_1, and set kappa = 1/sqrt(2), a = (1-kappa)*10*(3/2 - 1/sqrt(2))/2, and xi = a - 15. A direct computation gives <E[(d_xi H, kappa grad_y H)](z), z - z*> approximately -1.4 < 0. This is not a boundary artifact. Inequality (A.8) in Appendix A.3 requires this inner product to be nonnegative for the Robbins-Siegmund argument, and Step 1 of the proof uses the liminf of this inner product to conclude that z_infinity = z*. A negative value invalidates both the a.s. convergence claim and the rate bound (3.12).","section":"Appendix A.3, Step 1"},{"comment":"Theorem 3.2 does not cover two of the paper's headline examples, and the asserted adaptation is not provided. The text explicitly states that condition (3.11) is not met for volatility (g(x) = x^2) and for deviation measures with p > 1, including variantiles, and then says that the proof of Theorem 3.2 can be readily adapted, without giving details. This is a load-bearing gap: the proof in Appendix A.3 uses (3.11) to control E[Y_{k+1} | F_k] <= N_*^2 and to prove sup_n E[M_n^2] < infinity. For the quadratic and polynomial losses, the squared derivatives grow in xi, so those controls fail. Since the abstract and introduction advertise volatility and variantiles among the covered risk measures, the paper must either prove a theorem that covers these cases or substantially restrict the claims.","section":"Section 3.3.3"}],"minor_comments":[{"comment":"In the Notations paragraph, 'weletlet' should read 'we let'.","section":"Notations"},{"comment":"In Remark 5, 'upper-boudn' should be 'upper bound'.","section":"Remark 5"},{"comment":"In the sentence introducing the proximal mapping, 'proxymal' should be 'proximal'.","section":"Section 3.3.2"},{"comment":"The footnote references 'Theorem 3.5', but no Theorem 3.5 appears in the paper; the intended reference is presumably Theorem 3.2.","section":"Section 4.2.2, footnote 3"}],"recommendation":"reject","confidential_remarks":"The deterministic part and the Expected Shortfall numerical study are the solid core of the paper, but the advertised stochastic convergence result for general risk measures is unsupported by a false lemma and an unproved adaptation for volatility and variantile. A routine revision cannot repair this without either imposing restrictive assumptions that contradict the 'general risk measures' claim or redesigning the taming mechanism. I would encourage the authors to resubmit a version with a corrected stochastic proof or with claims restricted to the cases that are actually covered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this paper's real contribution is the deterministic mirror descent treatment of risk budgeting with a tamed gradient. The stochastic convergence theorem as stated is not established, because Lemma 3.2 is false as stated and the paper itself concedes that condition (3.11) fails for volatility and p>1 deviation measures. It still deserves a serious referee, but the advertised generality needs repair.\n\nWhat is new: the tamed-gradient MD construction, the continuous extension of κ∇Γ, a clean deterministic proof (Theorem 3.1) with a non-asymptotic rate, and an approximate stochastic scheme with a claimed a.s. rate. The numerics are extensive: calibrated three-asset mixtures, dimensions up to 250, comparisons against the earlier SGD, and several risk measures. The SMD vs t-SGD comparison is informative. The relationship to the characterization in [15] is transparent, and the deterministic proof does not lean on unstated inputs.\n\nWhere it breaks: Lemma 3.2 asserts strict positivity of the inner product between z−z* and the expected tamed gradient, with the proof omitted because it allegedly follows from convexity of h. Convexity alone cannot deliver that: the taming factor κ(y) multiplies only the y-coordinate of the gradient and can destroy strict monotonicity when the mean return vector is nonzero. I cannot fully verify the arithmetic in the stress-test counterexample from here, but the structural objection is sound, and if the computed negative inner product is right, inequality (A.8) has the wrong sign and the Robbins-Siegmund step fails. That is load-bearing for Theorem 3.2 and the rate (3.12). Separately, the paper states that condition (3.11) is not met for volatility and variantiles and that the proof can be \"readily adapted.\" Not writing that adaptation down leaves the theorem covering only ES and p=1 deviation measures, not the full advertised menu. The deterministic results and the ES numerics stand; the stochastic theorem as printed is unsupported.\n\nWho this is for: quantitative portfolio management researchers and anyone working on risk budgeting algorithms. I would send it to peer review with a clear mandate: fix Lemma 3.2 or provide an interior-limit argument that avoids the false statement, actually prove the adaptation for volatility, and ship code and data. My own verdict is skeptical on Theorem 3.2 but positive on the deterministic part.","headline":"The deterministic MD analysis is genuinely solid, but the advertised stochastic convergence theorem rests on a tamed-gradient monotonicity lemma that is false as stated, so the paper needs repair before its main claim can stand.","tokens_in":28676,"tokens_out":3259,"would_cite":true,"duration_ms":37843,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C05","62L20","62G32","91Gxx"],"pacs":[],"model":"deepseek-v4-flash","headline":"Tamed mirror descent gives provable convergence for risk budgeting portfolios.","keywords":["Risk budgeting","Mirror descent","Risk measures","Expected Shortfall","Stochastic approximation","Tamed gradient","Deviation measures","Portfolio optimization"],"falsifier":"Take the three-asset volatility objective $g(r_\\rho(y))=y^\\top\\Sigma y$ with a nonzero mean-return vector and equal budgets, and compute $\\langle y-y^\\star, \\kappa(y)\\nabla\\Gamma_g(y)\\rangle$ at a boundary point $y$ with one coordinate zero; if any such inner product is negative, Lemma 3.2 is false and the proof's exclusion of boundary limit points collapses. Running the SMD recursion (3.10) on the same example with $\\gamma_n=n^{-0.55}$ and $m=100$ would then directly show whether the sequence still converges.","tokens_in":27521,"feed_emoji":"📉","tokens_out":11144,"duration_ms":109743,"temperature":0.7,"pith_summary":"Risk budgeting portfolios — long-only portfolios whose asset contributions to total risk match preset budgets — are usually found by solving a system of nonlinear equations that is equivalent to minimizing a strictly convex function with a logarithmic barrier. That barrier makes the gradient explode at the boundary of the positive orthant, so standard mirror descent theory does not apply. This paper shows that multiplying the gradient by the taming factor $\\kappa(y)=\\min_i y_i\\wedge 1$ removes the singularity without moving the minimizer, and proves that deterministic and stochastic mirror descent then converge to the risk budgeting weights, with an explicit non-asymptotic rate for the averaged iterates. The stochastic version covers risk measures expressible as $g(r_\\rho(y))=\\min_\\xi \\mathbb{E}[L(\\xi,-\\langle y,X\\rangle)]$, including Expected Shortfall and generalized deviation measures, and the numerical section shows it stays stable and often more accurate than projected SGD as the number of assets grows. If the theorems hold, this is the first scheme to give almost-sure convergence with a quantitative rate for risk budgeting across general risk measures.","feed_headline":"Tamed mirror descent provably computes risk budgeting portfolios","feed_subtitle":"A bounded tamed gradient gives almost-sure convergence and explicit rates for ES, volatility, and variantiles.","key_machinery":"The load-bearing object is the tamed gradient $T(y)=\\kappa(y)\\nabla\\Gamma_g(y)$ with $\\kappa(y)=\\min_i y_i\\wedge 1$, extended by continuity to the boundary by setting components to $-b_i$ where the coordinate is zero. It does two jobs: it stays uniformly bounded on every centred ball, which makes the standard mirror-descent Bregman inequality applicable, and it preserves the strict monotonicity $\\langle y-y^\\star,T(y)\\rangle>0$ for $y\\neq y^\\star$, which is what forces any limit point to be $y^\\star$. Around it, the algorithm uses the negative-entropy Bregman divergence $D_F(y,y')=\\sum_i y_i\\log(y_i/y'_i)-\\sum_i y_i+\\sum_i y'_i$ and its explicit proximal map $P^m_y(v)$, the exponential update $y_i e^{-v_i}$ rescaled to the $\\ell^1$-ball of radius $m$. In the stochastic case the state is $z=(\\xi,y)$, the divergence gains a quadratic $\\xi$-term, and the update is given by (3.10), with the same taming factor applied to the $y$-gradient of $H(z,X)=L(\\xi,-\\langle y,X\\rangle)-\\sum_i b_i\\log y_i$.","core_discovery":"The paper's central claim is that the unique minimizer $y^\\star$ of $\\Gamma_g(y)=g(r_\\rho(y))-\\sum_{i=1}^d b_i\\log y_i$ can be computed by deterministic and stochastic Mirror Descent even though $\\nabla\\Gamma_g$ blows up at the boundary of the positive orthant. The resolution is the tamed gradient $\\kappa(y)\\nabla\\Gamma_g(y)$ with $\\kappa(y)=\\min_i y_i\\wedge 1$, which extends continuously to the closed domain, vanishes only at $y^\\star$, and satisfies the strict monotonicity inequality used to rule out boundary limit points. With this tamed gradient and the explicit KL-proximal update, the deterministic scheme converges in the limit and the weighted averages satisfy the rate bound (3.5); the stochastic scheme, built on the representation $g(r_\\rho(y))=\\min_\\xi\\mathbb{E}[L(\\xi,-\\langle y,X\\rangle)]$, converges almost surely to $z^\\star=(\\xi^\\star,y^\\star)$ and its weighted averages satisfy the a.s. bound (3.12). The normalized limit $u^\\star=y^\\star/\\|y^\\star\\|_1$ is the unique risk budgeting portfolio. The theorem applies directly to Expected Shortfall and deviation measures with $p=1$; for volatility and variantiles the stated square-integrability condition (3.11) is violated, and the paper asserts that the proof can be adapted.","pith_inferences":["The taming factor $\\min_i y_i\\wedge 1$ is a general device for convex programs with logarithmic barriers and explicit exponential proximal maps; it could be tested on other entropy-regularized allocation problems outside risk budgeting.","Because the numerics show constant step size converging faster than the provably convergent decreasing-step choices, a natural next conjecture is that simple averaging of later iterates gives the same almost-sure rate under weaker moment assumptions.","If the asserted adaptation to volatility and variantiles is written out fully, the same machinery should yield a central limit theorem for the averaged sequence, an extension the paper already flags as future work.","The rate bound's dependence on $m\\log m$ and dimension $d$, together with the identity $r_\\rho(y^\\star)=1/g'(r_\\rho(y^\\star))$, suggests a practical rule for choosing $m$ from the risk level rather than by blind search."],"forward_implications":["Risk budgeting portfolios for any risk measure of the form $\\rho(Z)=\\min_\\xi \\mathbb{E}[L(\\xi,Z)]$ can be computed by an explicit, projection-free mirror descent recursion that converges almost surely and comes with a non-asymptotic averaged-iterate rate.","The log-barrier singularity no longer blocks stochastic approximation: the factor $\\kappa(y)=\\min_i y_i\\wedge 1$ keeps the update inside the positive orthant while preserving the unique minimizer.","For Expected Shortfall and generalized deviation measures the theoretical guarantees apply directly; numerical tests indicate the SMD method stays stable and beats projected SGD as the number of assets grows.","The user must know or guess a radius $m\\ge\\|y^\\star\\|_1$; choosing $m$ too small provably prevents convergence, while too large $m$ only degrades the explicit rate.","The weighted-average estimate $\\bar z_n$ satisfies an almost-sure bound of order $(\\sum_{k=0}^n\\gamma_{k+1})^{-1}$ up to the inverse of the smallest coordinate, which is where the boundary behavior enters."],"supporting_citations":[{"why":"Supplies the variational characterization (Theorem 2.1) that turns risk budgeting into minimization of $g(r_\\rho(y))-\\sum b_i\\log y_i$, and provides the SGD baseline used for comparison.","marker":"[15]"},{"why":"Supplies the mirror-descent and Bregman-divergence machinery, including the two-point inequality used to derive both convergence and the averaged-iterate rate.","marker":"[35]"},{"why":"Defines the mirror descent algorithm and the classical convergence-rate benchmark that the paper's tamed variant extends.","marker":"[36]"},{"why":"Provides the Bayes-risk and elicitability representation that justifies the stochastic framework $g(r_\\rho(y))=\\min_\\xi \\mathbb{E}[L(\\xi,-\\langle y,X\\rangle)]$ for $g=\\mathrm{Id}$.","marker":"[23]"},{"why":"Defines generalized deviation measures, the family used to cover MAD, volatility, and variantiles as examples in the stochastic framework.","marker":"[42]"},{"why":"Gives the convex-minimization representation of CVaR and Expected Shortfall that the ES example relies on.","marker":"[39]"}],"fun_headline_variants":["Tamed mirror descent: explicit rates for risk budgeting portfolios","Stochastic tamed mirror descent provably computes risk budgets","Tamed gradients make risk budgeting portfolios computable","Tamed gradient tames boundary blow-up in risk budgeting","Tamed mirror descent solves risk budgeting with provable convergence"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything depends on the tamed gradient pushing away from the boundary at every boundary point, and on the noisy gradients having finite second moments; volatility and variantile examples do not satisfy the second condition as stated, so the paper asserts that the proof can be adapted.","fun_headline_variants_meta":{"raw":{"variants":["Tamed mirror descent: explicit rates for risk budgeting portfolios","Stochastic tamed mirror descent provably computes risk budgets","Tamed gradients make risk budgeting portfolios computable","Tamed gradient tames boundary blow-up in risk budgeting","Tamed mirror descent solves risk budgeting with provable convergence"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000484,"raw_usage":{"total_tokens":2380,"prompt_tokens":925,"completion_tokens":1455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1376}},"tokens_in":541,"tokens_out":1455,"duration_ms":11558,"temperature":1.0,"reasoning_tokens":1376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:40:40.423996+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the three-asset volatility objective $g(r_\\rho(y))=y^\\top\\Sigma y$ with a nonzero mean-return vector and equal budgets, and compute $\\langle y-y^\\star, \\kappa(y)\\nabla\\Gamma_g(y)\\rangle$ at a boundary point $y$ with one coordinate zero; if any such inner product is negative, Lemma 3.2 is false and the proof's exclusion of boundary limit points collapses. Running the SMD recursion (3.10) on the same example with $\\gamma_n=n^{-0.55}$ and $m=100$ would then directly show whether the sequence still converges.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the variational characterization (Theorem 2.1) that turns risk budgeting into minimization of $g(r_\\rho(y))-\\sum b_i\\log y_i$, and provides the SGD baseline used for comparison."},{"cited_title":"Nemirovski, A","cited_arxiv_id":null,"evidence_quote":"Supplies the mirror-descent and Bregman-divergence machinery, including the two-point inequality used to derive both convergence and the averaged-iterate rate."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the mirror descent algorithm and the classical convergence-rate benchmark that the paper's tamed variant extends."},{"cited_title":"Embrechts, T","cited_arxiv_id":null,"evidence_quote":"Provides the Bayes-risk and elicitability representation that justifies the stochastic framework $g(r_\\rho(y))=\\min_\\xi \\mathbb{E}[L(\\xi,-\\langle y,X\\rangle)]$ for $g=\\mathrm{Id}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines generalized deviation measures, the family used to cover MAD, volatility, and variantiles as examples in the stochastic framework."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the convex-minimization representation of CVaR and Expected Shortfall that the ES example relies on."}],"review_version":1}