{"id":"33d95e02-9756-4af3-bdcc-4577c4acc3f5","arxiv_id":"2411.15881","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For sums of iid variables in the domain of normal attraction of an α-stable law with α in (1,2), the paper gives uniform and non-uniform rates for the error in replacing the call function expectation by its stable limit.","lead":"This paper proves explicit error bounds for approximating the expected payoff of a call option when the underlying risk factors follow a heavy-tailed distribution in the domain of a stable law. The bounds work without any second moment assumption, which previous call-function approximations required, so they apply to data with infinite variance.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of the general-δ non-uniform bound depends on an unverified import from [12] and a Lemma 3.1 bound that mixes the δ=0-improved exponent with the general-δ constant; the claimed c_{2,M} does not follow as written.","rationale":"I read the paper in good faith. The uniform bound Theorem 1.3 is a plausible corollary of [12] applied to the Lipschitz function g_M, and the M-decay structure of Theorem 1.4 is plausible. The strongest claim, however, is the non-uniform general-δ bound, whose proof is almost entirely delegated: Lemma 3.1 is stated with a mixture of the general-δ constant η3 and the δ=0-improved exponent, and the final theorem is asserted by 'the same argument' without displaying the optimization over the truncation parameter a. This is exactly the kind of load-bearing step that needs independent verification. The reader flagged the imported Stein solution representation as the weakest assumption; I agree partially, but I would sharpen the concern: the more concrete problem is that Lemma 3.1, as typeset, asserts a stronger M-decay than Lemma 2.5 provides for δ≠0. If that is only a typo and the intended exponent is 2(α−1)/(3α−1), the theorem may survive; if the improved exponent is genuinely needed, the general-δ claim is unsupported. Because the issue is specific and testable, and because the final theorem might still be correct after a modest revision, I keep the reader's CONDITIONAL verdict unchanged rather than moving to reject.","tokens_in":15626,"tokens_out":28431,"duration_ms":250680,"concrete_test":"Independently re-derive Theorem 1.4 for γ=2−α and δ≠0, using only Lemma 2.5's bound ||f''_{g_M}||∞ ≤ η3,α,δ M^{-2(α−1)/(3α−1)} and the exact truncation argument of [12, Theorem 1.4], tracking the choice of the truncation parameter a and any resulting n-dependence. If the displayed c_{2,M} with the log M term emerges with no extra n factors, the concern is resolved. If the improved exponent M^{-(α^2−1)/(α^2+2α−1)} is required at any point for δ≠0, the general-δ theorem is unproved. In parallel, check the stated hypotheses of [12, Lemma 2.3] against the unbounded function g_M(x)=(x−M)_+.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 1.4 is the central claim, but its proof is a single sentence: it says the result follows from Lemma 2.5 and Lemma 3.1 by the same argument as [12, Theorem 1.4]. Lemma 3.1 is load-bearing, and as stated it contains an internal mismatch. In the γ=2−α and γ=0 cases, the first error term is bounded using the δ=0-improved decay M^{-(α^2−1)/(α^2+2α−1)} and the constant η4,α. But for general δ ≠ 0, the only available non-uniform f'' estimate is Lemma 2.5, which gives the worse rate M^{-2(α−1)/(3α−1)} with η3,α,δ. The improved exponent is not a harmless typo: it claims a stronger decay than the proved f'' bound supports. Moreover, the statement of Theorem 1.4 for general δ uses the unimproved exponent in the final c_{2,M}, so the paper never shows how the general-δ bound follows from Lemma 3.1 as written. Separately, g_M is unbounded, and the hypotheses of [12, Lemma 2.3] and [12, Theorem 1.2], which supply the Stein solution representation and the uniform f' and f'' bounds, are not checked against this unbounded Lipschitz target function. Since the non-uniform constant c_{2,M} and the claimed first non-uniform stable call-function bound rest on this chain, the central claim is not fully supported by the proof as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies stable approximation for expectations of the call function g_M(x)=(x-M)_+ for sums of i.i.d. heavy-tailed random variables in the domain of normal attraction of an α-stable law with α∈(1,2). Under the distributional assumption (1.1) and the tail-bias condition (1.2), it claims a uniform bound in Theorem 1.3 and a non-uniform bound in Theorem 1.4. The non-uniform constant c_{2,M} decays like a power of M in general, with an improved decay c_{3,M} in the symmetric case δ=0. The proof strategy follows the Stein method developed by Chen, Nourdin and Xu [12]: it imports the Stein equation and solution representation, derives new heat-kernel estimates for stable densities in Appendix A, obtains non-uniform bounds for the second derivative of the Stein solution in Lemma 2.5 and Lemma 2.6, and then applies a zero-bias based Taylor-like extension in Lemma 3.1. The claimed n-rates match the optimal Kolmogorov rates of [31], and a Pareto example is provided with numerical support.","tokens_in":15924,"tokens_out":14790,"duration_ms":125307,"significance":"If the results are correct, the paper would provide the first non-uniform stable-approximation bound for the call function without a second-moment assumption, which is a genuine extension of the existing CDO-oriented normal and Poisson approximation results and is relevant for heavy-tailed financial applications. The explicit constants, the treatment of the asymmetric case, and the improved symmetric-case rate are valuable. The paper also contains original heat-kernel estimates in Appendix A. The main caveat is that the currently written proof does not fully support the central non-uniform theorem: several load-bearing estimates are either imported without checking their hypotheses or stated with an internal exponent/constant mismatch. These are fixable but require substantial revision of Section 3.","major_comments":[{"comment":"Lemma 3.1 is stated for general δ, but its first error term in case (i) is bounded by a constant times a/M^{(α^2-1)/(α^2+2α-1)}. This is the improved exponent from Lemma 2.6, which is proved only for δ=0 and has constant η_{4,α}; the only general-δ non-uniform estimate available, Lemma 2.5, gives M^{-2(α-1)/(3α-1)} with constant η_{3,α,δ}. Moreover, case (iii) uses η_{4,α} in its first term even though no δ=0 restriction is present. Since Theorem 1.4 is stated to follow from Lemma 2.5 and Lemma 3.1, the general-δ constant c_{2,M} does not follow as written. The authors must either restrict the improved terms to the symmetric case with the correct constants or redo the general-δ estimates using Lemma 2.5.","section":"Section 3.2, Lemma 3.1(i), (iii)"},{"comment":"The solution representation (2.3) and the uniform bounds ‖f'_g‖_∞≤α and ‖f''_g‖_∞≤4η_{2,α,δ} are imported from [12, Lemma 2.3 and Theorem 1.2], but the hypotheses of those results are not checked for g_M(x)=(x-M)_+. This function is unbounded and only Lipschitz, not C^2_b, so differentiating the representation under the integral in the proof of Lemma 2.5 needs justification. Since every non-uniform constant in Theorems 1.4 relies on estimates for f''_{g_M}, the authors should either state the exact hypotheses of the imported results and verify them for unbounded Lipschitz test functions, or give a truncation/approximation argument.","section":"Section 2, Eq. (2.3) and Lemma 2.4"},{"comment":"Lemma 3.1 is proved only for γ∈[0,2−α], while Theorem 1.4 also contains the branch γ∈(2−α,∞). The proof of Theorem 1.4 says it follows from Lemma 2.5 and Lemma 3.1, but no argument is supplied for γ>2−α. This branch appears with its own M-dependence in c_{2,M} and c_{3,M}, so an explicit treatment or an extension of Lemma 3.1 to this range is needed.","section":"Section 3.2, Lemma 3.1 vs Theorem 1.4"},{"comment":"As printed, the statement of Lemma 2.5 has a sign inconsistency in the exponent. The denominator M^{2(1-α)/(3α-1)} would make the displayed upper bound for f''_{g_M} grow like M^{2(α-1)/(3α-1)}, which contradicts the intended decay and the bound used in Theorem 1.4. The proof-ending display uses the reciprocal exponent M^{-2(α-1)/(3α-1)}. This must be corrected in the statement; the current wording makes the central non-uniform estimate ambiguous.","section":"Lemma 2.5"}],"minor_comments":[{"comment":"In the bound for R, direct computation gives Aα∫_{-(2A)^{1/α}}^{(2A)^{1/α}} (1+δ)1_{(0,∞)}(u)+(1−δ)1_{(-∞,0)}(u) over |u|^{α-1} du equal to α/(2−α)(2A)^{2/α}‖f''_{g_M}‖_∞ a, not 2α/(2−α)(2A)^{2/α}‖f''_{g_M}‖_∞ a. The displayed constant appears to be too large by a factor of 2, so the explicit constants in the following display should be rechecked.","section":"Section 3.2, proof of Lemma 3.1"},{"comment":"The notation ‖f''_{g_M}(y)‖_∞ in the statements is not standard: the left side contains a free variable y while the norm is in y. The statements should read ‖f''_{g_M}‖_∞.","section":"Lemma 2.5 and Lemma 2.6"},{"comment":"The phrase 'Rn is defined by 1.6' should read 'Rn is defined by (1.6)'; the equation number is missing parentheses in this and several other references.","section":"Theorem 1.4"},{"comment":"Figure 1 labels one curve 'Density of Z', but the random variable Z is not defined in the text; it should be S_α(1,0) or another explicitly defined quantity.","section":"Example 1.6 and Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on [12] and [13], but those are published works with independent derivations, so I do not see a circularity problem. The novelty of the non-uniform estimates is real. The main concern is proof completeness rather than correctness of the overall strategy: the internal mismatch in Lemma 3.1, the uncovered γ>2−α branch, and the unchecked hypotheses for the imported Stein solution are all fixable, but they affect the central theorem. The paper fits the scope of math.PR and would be a reasonable contribution after a careful revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chen, Qi, and Zhang extend the Stein's-method stable CLT of [12] to the call function g_M(x) = (x-M)_+ without finite second moments. The genuinely new piece is Theorem 1.4, a non-uniform bound whose constant decays in M, plus a better M-rate in the symmetric case using improved heat-kernel bounds. That fills a real gap: all earlier call-function approximations assumed finite variance. Lemma 2.2's heat-kernel estimates and Lemma 2.5's non-uniform f'' bound are derived carefully, and the constants are internally consistent; those parts give me some confidence in the overall approach.\n\nBut the stress-test note is right, and the problem is load-bearing. The proof of Theorem 1.4 is a single sentence: 'by the same argument' as [12], from Lemmas 2.5 and 3.1. Lemma 3.1, the workhorse, states error terms for γ=2−α and γ=0 whose first term contains the improved δ=0 exponent M^{-(α^2−1)/(α^2+2α−1)} multiplied by the general-δ constant η3,α,δ. That exponent is proved only in Lemma 2.6, for δ=0, using η4,α. For general δ, Lemma 2.5 gives the worse rate M^{-2(α−1)/(3α−1)}. So as written, Lemma 3.1 claims a stronger decay than the available estimates support. The theorem may still be true, but the chain does not follow.\n\nTwo other gaps, minor but real: the Stein solution representation (2.3) is imported from [12, Lemma 2.3] without checking it applies to the unbounded, non-smooth call function, and the differentiation under the integral in (2.8) is not justified. Both are likely fixable with a few lines, but they need to be stated.\n\nThe uniform bound Theorem 1.3 is essentially a corollary of [12, Theorem 1.4] with a Lipschitz function, and the paper only half-acknowledges this. The simulation is not a direct test of the theorem and lacks error bars; treat it as illustrative.\n\nOverall: the core idea is sound and the topic is well-motivated for heavy-tailed finance, but the central theorem's proof is not in a citable state. It deserves a serious referee, and the authors should be asked to repair Lemma 3.1 and fill the technical gaps before publication. I would not cite it in its current form.","headline":"Plausible non-uniform stable call-function bound, but the key lemma's M-rate for general δ is not supported by the proved f'' estimate.","tokens_in":16503,"tokens_out":5046,"would_cite":false,"duration_ms":40174,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60F05","60E07"],"pacs":[],"model":"deepseek-v4-flash","headline":"For iid sums in the domain of normal attraction of an α-stable law with $1 < \\alpha < 2$, the expected call payoff can be replaced by the stable-law call price with an explicit error bound, requiring no second moment.","keywords":["stable approximation","call function","Stein's method","domain of normal attraction","heavy-tailed distributions","non-uniform bounds","CDO pricing","Wasserstein-1 distance"],"falsifier":"For the Pareto example with $\\alpha=1.5$ (Example 1.6), compute $|\\mathbb{E}[(S_n-M)_+]-\\mathbb{E}[(S_{1.5}(1,0)-M)_+]|$ by high-precision simulation across $n=10^2$ to $10^5$; the bound predicts decay like $n^{-1/3}$ jointly with a polynomial improvement in $M$, and an observed slower decay would refute the theorem.","tokens_in":15393,"feed_emoji":"📈","tokens_out":6551,"duration_ms":56403,"temperature":0.7,"pith_summary":"This paper proves that when independent identically distributed summands lie in the domain of normal attraction of an $\\alpha$-stable law with $1<\\alpha<2$, the expected call payoff $\\mathbb{E}[(S_n-M)_+]$ is close to the stable-law expectation $\\mathbb{E}[(S_\\alpha(1,\\delta)-M)_+]$, with an explicit error of order $R_n$. The bounds require only a finite first moment, so heavy-tailed models that previously fell outside Gaussian or Poisson approximation theory are now covered. The non-uniform version shows the error constant decays polynomially in the strike $M$, and in the symmetric case $\\delta=0$ the decay exponent is improved through sharper heat-kernel estimates. The sample-size rates match the optimal rate known for Kolmogorov-distance stable convergence.","feed_headline":"Stable call-payoff error bounds need only finite means","feed_subtitle":"Heavy-tailed sums' expected call payoffs track the α-stable law at rate R_n, with error shrinking as strike M grows.","key_machinery":"The argument runs through the Stein equation for the $\\alpha$-stable generator $\\mathcal{A}_{\\alpha,\\delta}f(y)=d_\\alpha\\int (f(y+u)-f(y)-uf'(y))/(2|u|^{1+\\alpha})$ with skewness weights, together with the solution representation $f_g(y)=-\\int_0^\\infty\\int p_{(1-e^{-t})^{1/\\alpha},\\delta}(u-e^{-t/\\alpha}y)(g(u)-\\nu(g))\\,du\\,dt$. The proof controls the approximation error by a Taylor-like expansion of $\\mathbb{E}[X f'_{g_M}(Y+aX)]-\\mathbb{E}[X]\\mathbb{E}[f'_{g_M}(Y)]$, using a zero-biased coupling and the decay of $B$. The load-bearing estimates are the uniform and non-uniform bounds on the second derivative of the Stein solution: $\\|f''_g\\|_\\infty\\le 4\\eta_{2,\\alpha,\\delta}$ uniformly, and $\\|f''_{g_M}\\|_\\infty\\le \\eta_{3,\\alpha,\\delta}M^{-2(\\alpha-1)/(3\\alpha-1)}$, improved to $\\eta_{4,\\alpha}M^{-(\\alpha^2-1)/(\\alpha^2+2\\alpha-1)}$ when $\\delta=0$.","core_discovery":"The paper's central claim is a pair of error bounds, Theorems 1.3 and 1.4, for the call function $g_M(x)=(x-M)_+$. If $X_1$ has the distribution (1.1) with a bounded tail-bias term $B$ satisfying $|B(x)|\\le L/|x|^\\gamma$, and $S_n$ is the centered, $\\sigma$-normalized sum of $n$ iid copies, then $|\\mathbb{E}[(S_n-M)_+]-\\mathbb{E}[(S_\\alpha(1,\\delta)-M)_+]|\\le c_1 R_n$ uniformly, where $R_n$ is given by (1.6). The non-uniform version replaces $c_1$ by a constant $c_{2,M}$ that decays like $M^{-2(\\alpha-1)/(3\\alpha-1)}$, and for symmetric $\\delta=0$ by a smaller $c_{3,M}$ decaying like $M^{-(\\alpha^2-1)/(\\alpha^2+2\\alpha-1)}$. The point of these bounds is that they require only a finite first moment, not finite variance, so the approximation applies to heavy-tailed summands.","pith_inferences":["Because the proof uses only the Lipschitz property of $g_M$ and derivative regularity of the Stein solution, the same machinery should yield analogous bounds for other Lipschitz or piecewise-linear payoffs, such as put payoffs $(M-x)_+$ or butterfly spreads.","The exponents in $M$ come from a truncation argument that the paper itself flags as possibly suboptimal, so sharper estimates of the Stein solution away from the strike should improve the polynomial decay, particularly for $\\gamma>0$.","When $\\gamma=0$, the sample-size rate contains an integral of $|B(x)|/|x|^{\\alpha-1}$ and a supremum term, so in applications one must quantify the decay of $B$ before the bound becomes a literal power of $n$."],"forward_implications":["Finite first moment suffices: heavy-tailed summands with infinite variance are admitted, provided the tail shape matches (1.1) with $B$ decaying as in (1.2).","The uniform bound gives a call-payoff error of order $R_n$, and in the regimes $\\gamma\\in(2-\\alpha,\\infty)$ and $\\gamma=2-\\alpha$ that rate is $n^{1-2/\\alpha}$ (up to a logarithmic factor), matching the optimal Kolmogorov rate for stable convergence.","The non-uniform bound makes the error smaller as the strike $M$ grows, with polynomial decay of the constant, and the symmetric stable limit ($\\delta=0$) enjoys a better decay exponent from sharper heat-kernel estimates.","The results extend call-function approximation theory to lower-moment settings, so CDO tranche pricing and risk-theory calculations can use stable models where Gaussian or Poisson approximations require second moments."],"supporting_citations":[{"why":"Supplies the Stein equation, the solution representation (2.3), and the stable-CLT error decomposition used in both main theorems.","marker":"[12]"},{"why":"Introduced the $\\alpha$-stable Stein method and the Wasserstein-1 stable CLT whose smoothing and coupling lemmas the proof imports.","marker":"[39]"},{"why":"Provides the symmetric-case heat kernel estimate used in Lemma 2.3, which yields the improved $M$-decay for $\\delta=0$.","marker":"[13]"},{"why":"Established the two-sided optimal rate of convergence to a stable law, used to identify $R_n$ as rate-optimal in $n$.","marker":"[31]"},{"why":"Supplies the simulation representation of $S_\\alpha(1,0)$ used in the Pareto example's numerical check.","marker":"[28]"},{"why":"Gives the smooth-density and scaling facts for stable laws underlying the heat kernel estimates.","marker":"[34]"}],"fun_headline_variants":["Finite means suffice for stable call-payoff bounds","Call-payoff error bounds need only finite first moment","Stein's method yields uniform bounds for call payoffs","Heavy-tailed sums get call-payoff bounds without variance","First moment only: stable call approximation bounds"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on assuming the summands have the exact Pareto-like tail shape (1.1) with a bounded correction $B(x)$ that decays like $|x|^{-\\gamma}$; if the true distribution's tail deviates from that form, or $B$ decays too slowly, the stated sample-size rates degrade or stop being explicit.","fun_headline_variants_meta":{"raw":{"variants":["Finite means suffice for stable call-payoff bounds","Call-payoff error bounds need only finite first moment","Stein's method yields uniform bounds for call payoffs","Heavy-tailed sums get call-payoff bounds without variance","First moment only: stable call approximation bounds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2711,"prompt_tokens":942,"completion_tokens":1769,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":1690}},"tokens_in":558,"tokens_out":1769,"duration_ms":11569,"temperature":1.0,"reasoning_tokens":1690,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:48:02.888148+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the Pareto example with $\\alpha=1.5$ (Example 1.6), compute $|\\mathbb{E}[(S_n-M)_+]-\\mathbb{E}[(S_{1.5}(1,0)-M)_+]|$ by high-precision simulation across $n=10^2$ to $10^5$; the bound predicts decay like $n^{-1/3}$ jointly with a polynomial improvement in $M$, and an observed slower decay would refute the theorem.","supporting_citations":[{"cited_title":"and Xu, L","cited_arxiv_id":null,"evidence_quote":"Supplies the Stein equation, the solution representation (2.3), and the stable-CLT error decomposition used in both main theorems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced the $\\alpha$-stable Stein method and the Wasserstein-1 stable CLT whose smoothing and coupling lemmas the proof imports."},{"cited_title":"and Yang, X","cited_arxiv_id":null,"evidence_quote":"Provides the symmetric-case heat kernel estimate used in Lemma 2.3, which yields the improved $M$-decay for $\\delta=0$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Established the two-sided optimal rate of convergence to a stable law, used to identify $R_n$ as rate-optimal in $n$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the simulation representation of $S_\\alpha(1,0)$ used in the Pareto example's numerical check."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the smooth-density and scaling facts for stable laws underlying the heat kernel estimates."}],"review_version":1}