Pith. sign in

REVIEW 1 major objections 4 minor 33 references

In ruin models with unknown risk laws, replacing the adjustment coefficient by its empirical estimate makes the data-driven capital buffer fail to reach the prescribed safety exponent c.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 17:05 UTC pith:L2IMTTSC

load-bearing objection A solid paper with a real negative result and a clean profile framework, but the nonparametric optimality claim goes beyond what the stated assumptions support. the 1 major comments →

arxiv 2607.17732 v1 pith:L2IMTTSC submitted 2026-07-20 math.ST math.OCmath.PRstat.TH

Decision-Centric Large Deviations for Data-Driven Capital Buffers in Ruin Models

Yf Henkes , Bart P.G. van Parys , Bert Zwart This is my paper
classification math.ST math.OCmath.PRstat.TH MSC 60F1062F1291B30
keywords ruin probabilitylarge deviationscapital bufferadjustment coefficientdata-driven decisionplug-in estimatorrelative entropyexponential envelope
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies how an insurer should choose a capital buffer when the increment law of the risk process is unknown and the ruin tolerance is exponentially small, δ = e^{-cn}. It shows that the naive plug-in rule, which substitutes the empirical adjustment coefficient into the classical buffer formula, is unsafe: the joint probability of the historical sample and a future ruin event decays at a rate strictly below the target c. The paper then constructs the pointwise minimal safe profile f*(c,q), coupling the statistical cost I_P(q) of observing statistic q with the future ruin exponent γ(P) f induced by the chosen buffer. Lower semicontinuous majorants of f* are safe, while regular continuous rules falling strictly below f* are unsafe, so f* is the tipping point between safe and unsafe data-driven buffers. The framework is worked out in closed form for three parametric models and, in nonparametric classes, yields a safe, tractable conservative dual rule under a two-level exponential envelope.

Core claim

The central discovery is that data-driven capital buffers under rare-event constraints need a decision-centric profile, not just a consistent estimator of the adjustment coefficient. Proposition 3.1 proves that the empirical plug-in rule C_n = cn/γ(P_n) has ex ante ruin probability decaying at a rate strictly below the desired c, because a local likelihood-ratio perturbation of the empirical law changes the adjustment coefficient linearly while the relative entropy cost is only quadratic. The profile f*(c,q) = sup_{P∈P} (c - I_P(q))_+ / γ(P) is then identified as the exact safety threshold: any lower semicontinuous majorant of f* is safe, and any regular continuous rule strictly below f* at

What carries the argument

The central object is the pointwise minimal safe profile f*(c,q) and its associated variational exponent J_P(f,c) = inf_q{ I_P(q) + γ(P) f(c,q) }. The profile is equivalently characterized through the threshold representation f*(c,q) = inf{t ≥ 0 : F_q(t) ≥ c}, where F_q(t) = inf_{P∈P} { I_P(q) + γ(P)t }. The argument runs on two mechanisms: Sanov-type LDPs for the observed statistic Q_n and the Cramér–Lundberg logarithmic asymptotic for the future ruin probability, combined through the Lundberg bound. The plug-in failure is driven by an explicit perturbation h that makes γ(Q_ε) move linearly in ε while D(Q_ε||P) = O(ε²). In the nonparametric part, safety of the dual rule is proven by an exp-

Load-bearing premise

The whole construction rests on the observed statistic Q_n satisfying a large-deviation principle with a good rate function I_P for every law P in the model class, and the impossibility result additionally needs the Cramér–Lundberg logarithmic limit to hold for the law witnessing the violation; in fully nonparametric settings the paper notes it lacks such a Sanov theorem and instead proves safety of the dual rule directly.

What would settle it

Simulate a random walk with increments N(-1,1), n = 50, and estimate the normalized exponent - (1/n) log P{M > n c / γ(P_n)} / c for the plug-in rule; if it converges to at least 1 for any c > 0, Proposition 3.1 fails. Alternatively, exhibit a regular continuous rule f' with f'(c,q0) < (c - I_P0(q0))/γ(P0) at some q0 with I_P0(q0) < ∞ that is nevertheless safe at exponent c under P0, which would contradict Theorem 4.4.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The plug-in adjustment-coefficient rule is not merely statistically inefficient but asymptotically unsafe on the exponential scale; its ruin probability decays at a rate strictly below the target c.
  • For any model class satisfying the LDP assumption, safe data-driven buffers are exactly those dominating the profile f*; lower semicontinuous majorants of f* are safe and regular rules below it are unsafe.
  • In parametric models, the profile f* has closed-form expressions and exhibits a clean finite/infinite phase transition at the critical exponent ccrit(q), the relative-entropy cost of driving the adjustment coefficient to zero.
  • In nonparametric settings, a single exponential moment bound is insufficient: it forces the rule to the data-independent buffer c/r, so a two-level exponential envelope (with bounds on both E e^{aX} and E e^{bX}) is necessary and sufficient for a genuinely data-dependent safe rule.
  • The ex ante viewpoint implies that a zero buffer at some observed statistic can still be safe, because the statistical rarity of the data alone already supplies the target exponent in the joint probability.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The plug-in failure mechanism is generic: any decision rule that feeds an estimated parameter into a rare-event constraint while the tolerance scales exponentially with sample size is likely to suffer the same first-order optimism, suggesting the profile construction transfers to other rare-event decision problems.
  • The degeneracy of the single exponential envelope highlights a general design principle for model classes under rare-event constraints: uniform integrability of the relevant exponential moments is the crux, and a two-level envelope is a minimal practical way to enforce it.
  • A testable extension would be to compare the conservative dual rule against bootstrap confidence-band rules at moderate n: the theory predicts the dual rule stays at or above the target exponent while plug-in falls below, which could be checked by simulation in non-Gaussian models.
  • The exact characterization of the safe envelope suggests a bridge to distributionally robust optimization: the profile f* can be seen as the minimal cost buffer that is robust over the relative-entropy geometry induced by the LDP, potentially allowing closed-form robust buffers for other actuarial quantities.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

1 major / 4 minor

Summary. The paper studies data-driven capital buffers in a ruin-theoretic random walk model when the increment law is unknown. A decision maker observes a statistic Q_n from n historical observations and sets capital C_n = n f(c,Q_n), with target logarithmic ruin exponent c = log(1/delta)/n. The central object is the pointwise minimal safe profile f*(c,q) = sup_P (c - I_P(q))_+/gamma(P), which balances the LDP cost of observing q with the future ruin exponent. The main results are: (i) the empirical plug-in rule cn/gamma(hat P_n) fails to attain the target exponent (Prop. 3.1); (ii) lower semicontinuous majorants of f* are safe (Thm. 4.2), while regular continuous rules that fall strictly below f* at a witnessed point are unsafe (Thm. 4.4 and Cor. 4.5); (iii) explicit closed-form profiles for three parametric examples (Rademacher, Gaussian, exponential differences); and (iv) in the nonparametric setting, a single exponential envelope degenerates to a data-independent rule, whereas a two-level exponential envelope yields a conservative dual rule f_cd that is proved safe (Prop. 6.5) and is claimed to be optimal among regular rules.

Significance. If the results hold, the paper gives a clean large-deviation characterization of the boundary between safe and unsafe data-driven capital buffers in the ex ante regime, and it shows that naive plug-in calibration is asymptotically too optimistic. The proof of Prop. 3.1 uses an explicit likelihood-ratio perturbation and is a valuable negative result. The parametric examples are fully explicit and the appendix computations for Rademacher, Gaussian, and exponential-difference profiles check out. The nonparametric sections identify a genuine degeneracy of single-moment envelopes and construct a tractable finite-dimensional dual rule. The paper is generally careful and self-aware: it explicitly flags the absence of a suitable Sanov theorem for the empirical measure at the needed generality. The main caveat is that one load-bearing optimality claim in the nonparametric setting is not supported by the stated assumptions. The safety claims are established by direct arguments independent of the missing LDP.

major comments (1)
  1. [§6.2, fcd optimality claim] The paragraph after the equality fcd(c,Q_n)=f*_{P_{a,b,B}}(c,Q_n) states: 'With the pointwise lower bound of Section 4, it is then optimal among regular rules.' This invokes Theorem 4.4/Corollary 4.5, which require Assumption 2.1: an LDP with good rate function for the statistic Q_n. Here Q_n is the empirical measure. But the opening paragraph of Section 6 explicitly says 'we are not aware of a suitable version of Sanov's theorem' at the required level of generality and calls f*_C 'only a candidate solution.' No such LDP is supplied. Therefore the optimality claim for fcd among regular rules is not established; only the direct safety proof of Proposition 6.5 is established. Please either supply a suitable LDP (e.g., in a topology that controls the exponential moments entering fcd) or qualify the optimality claim as conditional on such an LDP. This is load-bearing because the nonparametri
minor comments (4)
  1. [Abstract] The abstract says 'regular continuous rules that fall strictly below f* are, under mild conditions, unsafe.' The precise statement (Cor. 4.5) requires the violation to be witnessed at a point by a law P0 with finite I_P0(q0) and with the Cramér–Lundberg asymptotic holding. Suggest adding 'at a witnessed point' for accuracy.
  2. [§5, before Proposition 5.2] The sentence 'Because ccrit is continuous, the set {q : f*(c,q)=+∞} is open, so f*(c,·) is lower semicontinuous' could be misread as claiming openness of the infinite set alone implies lower semicontinuity. The lsc conclusion also uses the explicit continuous finite branch; consider rephrasing to make that clear.
  3. [§7] The performance measure -log B^rho_{n,c}/(cn) is based on the expected Cramér–Lundberg upper bound, not the true ruin probability. The paper does say 'on this Cramér–Lundberg scale,' but it would help to state explicitly that a value below one for a rule means the Lundberg bound does not certify the target exponent, not that the true probability necessarily decays slower than e^{-cn}.
  4. [Throughout] Minor typographical issues: 'ex ante' appears unhyphenated in several places; accents in 'Cramér' are inconsistently rendered; in Section 6.2, 'e(n+1)^4' would be clearer with a space or parentheses. These do not affect the mathematics.

Circularity Check

1 steps flagged

Definitional envelope only; plug-in failure, lower-bound necessity, and nonparametric safety results are independently derived.

specific steps
  1. self definitional [Section 4, Definition 4.1 and Theorem 4.2]
    "This pointwise family of inequalities leads to the following envelope f∗ which we will refer to as the pointwise minimal safe profile. ... f∗(c,q)=sup_{P∈P}(c−I_P(q))+/γ(P). ... For every P∈P and q∈Q, g(c,q)≥f∗(c,q)≥(c−I_P(q))+/γ(P). Hence I_P(q)+γ(P)g(c,q)≥c for all q. Proposition 2.2 and Lundberg's inequality give the result."

    The profile f* is defined as the pointwise envelope of the exact safety inequalities I_P(q)+γ(P)f≥c. Theorem 4.2 proves safety of any l.s.c. majorant by reading off this defining inequality, so the 'safe majorants' property is true by construction rather than an independently derived prediction. This is benign: the substantive content lies in Proposition 3.1 (plug-in failure), Theorem 4.4 (necessity below f*), and the explicit parametric/nonparametric profiles, none of which reduce to the definition.

full rationale

The core derivation is self-contained under Assumption 2.1 and standard Lundberg/LDP facts. The empirical plug-in failure (Proposition 3.1) is established by a likelihood-ratio perturbation argument with independent content. The pointwise lower bound (Theorem 4.4) is a genuine LDP-localization argument, not a restatement of f*'s definition. The parametric closed forms are computed from the threshold representation, and the nonparametric safety proof (Proposition 6.5) uses the finite-dimensional dual and an external aggregation inequality [1], not fitted data. Self-citations [31,32,33] provide context and inspiration but are not load-bearing for the central results. The only definitional feature is f* itself, which is constructed to satisfy the pointwise safety inequalities; this is intrinsic to the 'tipping point' formulation and does not invalidate the derived results. One self-flagged gap deserves note: Section 6 states it is 'not aware of a suitable version of Sanov's theorem' for the empirical-measure statistic, so the nonparametric 'optimal among regular rules' claim relies on an LDP the paper disclaims. That is a correctness/support gap, not circularity, and does not raise the circularity score beyond the minor definitional aspect.

Axiom & Free-Parameter Ledger

1 free parameters · 6 axioms · 0 invented entities

The key inputs are the large-deviation assumption on the statistic and the exponential tail/ruin asymptotic; the paper proves the rest. Envelope constants a,b,B are user-chosen model-class parameters, not estimated from data.

free parameters (1)
  • two-level envelope parameters (a,b,B) = a=1, b=4, B=2e^4 in Section 7 numerical illustration
    These define the nonparametric model class P_{a,b,B}; they are chosen by the modeller, not estimated from data. The theoretical safety result holds for any 0<a<b, B>1, so they are not fitted to the central claim.
axioms (6)
  • domain assumption The statistic Q_n satisfies an LDP with good rate function I_P for every P in the model class (Assumption 2.1).
    Load-bearing: Proposition 2.2 and Theorem 4.4 derive the decay rate and the necessity result from this.
  • domain assumption The Cramér–Lundberg logarithmic asymptotic (4) holds for the true law and for laws P0 used in lower-bound arguments, requiring Cramér's condition E_P e^{γX}=1 with γ(P) ∈ (0,b).
    Needed for the exact exponential decay of the ruin probability; without it the tipping-point characterization may fail.
  • standard math Lundberg's inequality (1) holds for all laws in the model class.
    Standard random-walk ruin bound, e.g., Asmussen [2].
  • standard math Exponential aggregation inequality (Agrawal–Juneja–Glynn, Lemma E.1 in [1]).
    Used in Proposition 6.5 to control empirical suprema of exp-concave functions over the compact sets N_y and N.
  • standard math Convex duality (Lagrange duality theorem with Slater condition) for infinite-dimensional convex programs used in the dual representation.
    Used in Lemma 6.1 and Lemma 6.4 for strong duality.
  • domain assumption The true data-generating law belongs to the model class P (or P_{a,b,B}).
    The safety guarantee is over the chosen model class; if the true law is outside it, the profile is not valid.

pith-pipeline@v1.3.0-alltime-deepseek · 24764 in / 23041 out tokens · 250023 ms · 2026-08-01T17:05:57.427880+00:00 · methodology

0 comments
read the original abstract

We consider an insurance risk model with a random walk structure in which the underlying probability law is unknown. A decision maker observes a statistic $Q_n$ computed from $n$ historical observations and then needs to choose a capital buffer of the form $C_n=n f(c,Q_n)$, where $c=\log(1/\delta)/n$ balances the amount of data and the tolerated ruin probability $\delta$. The classical safe capital buffer is inversely proportional to the adjustment coefficient $\gamma$, the exponential rate at which the ruin probability decays; when the law is unknown, this coefficient must be inferred from the data. The profile $f$ couples the statistical cost of observing an atypical historical statistic with the future ruin exponent induced by the resulting decision. We study the joint ex ante probability (over both the historical sample and an independent future risk process) that the future maximum exceeds the data-dependent buffer. We show that the naive plug-in rule fails to achieve the prescribed logarithmic decay exponent $c$, illustrating the adverse impact of model uncertainty when making decisions under rare-event constraints. We then identify a profile $f^*=f^*(c,Q_n)$ with the following appealing properties: its lower semicontinuous majorants are safe, while regular continuous rules that fall strictly below $f^*$ are, under mild conditions, unsafe. We illustrate the potential applicability of our framework by developing three parametric examples. In nonparametric settings, we show that a single exponential envelope leads to degenerate capital buffers, whereas a two-level exponential envelope yields a nondegenerate capital buffer which we prove to be safe.

Figures

Figures reproduced from arXiv: 2607.17732 by Bart P.G. van Parys, Bert Zwart, Yf Henkes.

Figure 1
Figure 1. Figure 1: Normalized finite-sample exponent − log B ρ n,c/(cn) for the empirical plug-in rule, the Gaussian LDP rule, and the nonparametric conservative dual rule. The benchmark distribution is P = N(−1, 1), and n = 50. The dashed horizontal line marks the target value one. induced decision must control. Within this framework we showed that the naive plug-in rule fails to attain the prescribed safe logarithmic decay… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 2 linked inside Pith

  1. [1]

    Optimal δ-correct best-arm selection for heavy-tailed distributions

    Shubhada Agrawal, Sandeep Juneja, and Peter Glynn. Optimal δ-correct best-arm selection for heavy-tailed distributions. arXiv:1908.09094, 2023

    Pith/arXiv arXiv 1908

  2. [2]

    Springer, New York, second edition, 2003

    Søren Asmussen.Applied Probability and Queues. Springer, New York, second edition, 2003

    2003

  3. [3]

    World Scientific, Singapore, second edition, 2010

    Søren Asmussen and Hansj¨ org Albrecher.Ruin Probabilities. World Scientific, Singapore, second edition, 2010

    2010

  4. [4]

    Robust bounds on risk-sensitive functionals via R´ enyi divergence.SIAM/ASA Journal on Uncertainty Quantification, 3(1):18–33, 2015

    Rami Atar, Kenny Chowdhary, and Paul Dupuis. Robust bounds on risk-sensitive functionals via R´ enyi divergence.SIAM/ASA Journal on Uncertainty Quantification, 3(1):18–33, 2015

    2015

  5. [5]

    Robust solutions of optimization problems affected by uncertain probabilities

    Aharon Ben-Tal, Dick den Hertog, Anja De Waegenaere, Bertrand Melenberg, and Gijs Rennen. Robust solutions of optimization problems affected by uncertain probabilities. Management Science, 59(2):341–357, 2013

    2013

  6. [6]

    S´ andor Cs¨ org˝ o and Jozef L. Teugels. Empirical Laplace transform and approximation of compound distributions.Journal of Applied Probability, 27(1):88–101, 1990

    1990

  7. [7]

    Springer, New York, second edition, 1998

    Amir Dembo and Ofer Zeitouni.Large Deviations Techniques and Applications. Springer, New York, second edition, 1998

    1998

  8. [8]

    Duchi, Peter W

    John C. Duchi, Peter W. Glynn, and Hongseok Namkoong. Statistics of robust optimization: a generalized empirical likelihood approach.Mathematics of Operations Research, 46(3):946– 969, 2021

    2021

  9. [9]

    Duffy and Anthony P

    Ken R. Duffy and Anthony P. Metcalfe. How to estimate the rate function of a cumulative process.Journal of Applied Probability, 42(4):1044–1052, 2005

    2005

  10. [10]

    Duffy and Anthony P

    Ken R. Duffy and Anthony P. Metcalfe. The large deviations of estimating rate functions. Journal of Applied Probability, 42(1):267–274, 2005

    2005

  11. [11]

    Duffy and Sean P

    Ken R. Duffy and Sean P. Meyn. Estimating Loynes’ exponent.Queueing Systems, 68(3–4):285–293, 2011

    2011

  12. [12]

    Ellis.A Weak Convergence Approach to the Theory of Large Deviations

    Paul Dupuis and Richard S. Ellis.A Weak Convergence Approach to the Theory of Large Deviations. Wiley, New York, 1997

    1997

  13. [13]

    Katsoulakis, Yannis Pantazis, and Petr Plech´ aˇ c

    Paul Dupuis, Markos A. Katsoulakis, Yannis Pantazis, and Petr Plech´ aˇ c. Path-space information bounds for uncertainty quantification and sensitivity analysis of stochastic dynamics.SIAM/ASA Journal on Uncertainty Quantification, 4(1):80–111, 2016

    2016

  14. [14]

    Katsoulakis, Yannis Pantazis, and Luc Rey-Bellet

    Paul Dupuis, Markos A. Katsoulakis, Yannis Pantazis, and Luc Rey-Bellet. Sensitivity analysis for rare events based on R´ enyi divergence.The Annals of Applied Probability, 30(4):1507–1533, 2020

    2020

  15. [15]

    A bootstrap procedure for estimating the adjustment coefficient.Insurance: Mathematics and Economics, 10(3):181–190, 1991

    Paul Embrechts and Thomas Mikosch. A bootstrap procedure for estimating the adjustment coefficient.Insurance: Mathematics and Economics, 10(3):181–190, 1991

    1991

  16. [16]

    Ganesh and Neil O’Connell

    Ayalvadi J. Ganesh and Neil O’Connell. An inverse of Sanov’s theorem.Statistics & Probability Letters, 42(2):201–206, 1999

    1999

  17. [17]

    Ganesh and Neil O’Connell

    Ayalvadi J. Ganesh and Neil O’Connell. A large-deviation principle for Dirichlet posteriors. Bernoulli, 6(6):1021–1034, 2000

    2000

  18. [18]

    Empirical bounds for ruin probabilities.Stochastic Processes and their Applications, 8:243–255, 1979

    Jan Grandell. Empirical bounds for ruin probabilities.Stochastic Processes and their Applications, 8:243–255, 1979. 23

    1979

  19. [19]

    Springer, New York, 1991

    Jan Grandell.Aspects of Risk Theory. Springer, New York, 1991

    1991

  20. [20]

    Teugels, and Ann Vanmarcke

    Peter Hall, Jozef L. Teugels, and Ann Vanmarcke. The abscissa of convergence of the Laplace transform.Journal of Applied Probability, 29(2):353–362, 1992

    1992

  21. [21]

    Estimators and bootstrap confidence intervals for ruin probabilities.ASTIN Bulletin, 19(1):57–70, 1989

    Christian Hipp. Estimators and bootstrap confidence intervals for ruin probabilities.ASTIN Bulletin, 19(1):57–70, 1989

    1989

  22. [22]

    Springer, London, 2 edition, 2014

    Achim Klenke.Probability Theory: A Comprehensive Course. Springer, London, 2 edition, 2014

    2014

  23. [23]

    John Wiley & Sons, New York, 1959

    Solomon Kullback.Information Theory and Statistics. John Wiley & Sons, New York, 1959

    1959

  24. [24]

    Robust sensitivity analysis for stochastic systems.Mathematics of Operations Research, 41(4):1248–1275, 2016

    Henry Lam. Robust sensitivity analysis for stochastic systems.Mathematics of Operations Research, 41(4):1248–1275, 2016

    2016

  25. [25]

    Luenberger.Optimization by Vector Space Methods

    David G. Luenberger.Optimization by Vector Space Methods. Wiley, New York, 1969

    1969

  26. [26]

    Extension of some large deviation results for posterior distributions.Journal of the Korean Statistical Society, 43(2):189–200, 2014

    Claudio Macci. Extension of some large deviation results for posterior distributions.Journal of the Korean Statistical Society, 43(2):189–200, 2014

    2014

  27. [27]

    An inverse Sanov theorem for exponential families

    Claudio Macci and Mauro Piccioni. An inverse Sanov theorem for exponential families. Journal of Statistical Planning and Inference, 224:54–68, 2023

    2023

  28. [28]

    Richard O. Michaud. The Markowitz optimization enigma: Is ‘optimized’ optimal?Financial Analysts Journal, 45(1):31–42, 1989

    1989

  29. [29]

    A large-deviations analysis of the GI/GI/1 SRPT queue

    Misja Nuyens and Bert Zwart. A large-deviations analysis of the GI/GI/1 SRPT queue. Queueing Systems, 54(2):85–97, 2006

    2006

  30. [30]

    Pitts, Rudolf Gr¨ ubel, and Paul Embrechts

    Susan M. Pitts, Rudolf Gr¨ ubel, and Paul Embrechts. Confidence bounds for the adjustment coefficient.Advances in Applied Probability, 28(3):802–827, 1996

    1996

  31. [31]

    Tobias Sutter, Bart P. G. van Parys, and Daniel Kuhn. A Pareto dominance principle for data-driven optimization.Operations Research, 72(5):1976–1999, 2024

    1976

  32. [32]

    Bart P. G. van Parys, Peyman Mohajerin Esfahani, and Daniel Kuhn. From data to decisions: Distributionally robust optimization is optimal.Management Science, 67(6):3387–3402, 2021

    2021

  33. [33]

    Bart P. G. van Parys and Bert Zwart. Robust mean estimation for optimization: the impact of heavy tails. arXiv:2503.21421, 2025. 24 A Proofs of the parametric example profiles Proof of Proposition 5.2. Assume first q∈ Qs = [0 , 1/2), the stable region. For y > 0, set Hq(y) = inf{d(q∥p) : 0 < p <1/2, γ(p) ≤y} . Since γ(p) ≤y is equivalent to p≥p y := 1/(1+...

    Pith/arXiv arXiv 2025