{"id":"01a50830-cdba-4839-8cbd-e9f0f83008fa","arxiv_id":"1908.01633","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Using payoffs-beliefs duality, the paper proves necessary and sufficient conditions for positive value of information and classifies marginal value near no information as zero, finite, or infinite.","lead":"The paper uses convex geometry to analyze when information is valuable to a decision maker, giving bounds on the value of information from the geometry of available actions and beliefs. It shows that information matters only when it changes the set of optimal decisions, and classifies small amounts of information as having zero, finite, or infinite marginal value.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3's upper bound is proved with an invalid constant: CA=inf_{a∈A} sup||a−a'|| fails on the paper's own Example 1 (VoI=0.75 vs CA·E[d]=√122/15≈0.736). The proof needs CA over a∈A*(pbar), not all of A; fixable, but the printed bound is wrong.","rationale":"I read the paper as establishing a duality framework and using it to bound the value of information. Proposition 2 and Theorems 5–7 appear mathematically sound; the main theorems are existential with constants that can depend on global geometry, and the proofs are broadly correct. The weakest point I found is in the proof of Theorem 3's upper bound, where the constant CA is computed by an infimum over all of A while the argument's key identity (46b) only applies to actions in A^*(pbar). This is not a mere cosmetic slip: in the paper's own Example 1, the printed constant gives CA·E[d] ≈ 0.736 while the true VoI is 0.75, so the displayed inequality fails. Because Theorem 3 only asserts existence of constants, the theorem can be repaired by taking the infimum over A^*(pbar) or by using the diameter of A, so the central claims survive. This strengthens the case for a conditional verdict but does not require rejection. The reader's stated weakest assumption was about expected-utility foundations; my concern is different, hence I disagree on the identification of the single most load-bearing issue, though I agree the appropriate verdict remains conditional.","tokens_in":24291,"tokens_out":45744,"duration_ms":425958,"concrete_test":"Recompute Theorem 3's upper bound for Example 1 with prior pbar = 1/2 and the full-information structure q = (1,0) or (0,1) with equal probability. Using the printed formula CA = inf_{a∈A} sup_{a'∈A} ||a−a'|| gives CA = √61/4 and E[d(q,∆^c_A(pbar))] = 4√2/15, yielding product ≈ 0.736 < VoI = 0.75. Then re-run the same inequality with the infimum restricted to a ∈ A^*(pbar), giving CA = √5 ≈ 2.236 and product ≈ 0.843 > 0.75, confirming that the proof needs this restriction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 3 contains a concrete error in the upper-bound constant. It defines CA = inf_{a∈A} sup_{a'∈A} ||a−a'|| and claims φ_a(q) ≤ (sup_{a'∈A}||a−a'||) d(q, ∆^c_A(pbar)). The displayed inequality is justified by invoking (46b), which holds only for a ∈ A^*(pbar), not for arbitrary a ∈ A. The constant as written is false. In the paper's Example 1, with A = conv{(3,0),(2,2),(0,5/2),(0,0)}, prior pbar = (1/2,1/2), and full information q = (1,0) or (0,1) with probability 1/2, we have VoI_A(q) = (3 + 5/2)/2 − 2 = 3/4. The confidence set is [1/3,4/5], so E d(q, ∆^c_A(pbar)) = (√2/3 + √2/5)/2 = 4√2/15. The proposed CA is the circumradius of the triangle formed by (3,0),(0,5/2),(0,0), i.e. √61/4 ≈ 1.9526. The product CA·E[d] = √122/15 ≈ 0.7363, which is strictly less than VoI = 0.75. Thus the upper bound as proved is not true with this constant. The theorem is salvageable: restricting the infimum to a ∈ A^*(pbar) (or replacing CA by diam A) makes the inequality valid, so this is a proof gap rather than a fatal flaw in the central result.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the value of information in finite-state decision problems through convex duality. The decision problem is encoded by a compact convex set A of state-indexed payoff vectors; the value function v_A is the support function of A, optimal actions are exposed faces, and beliefs supporting an action form normal cones. An information structure is modeled as a random posterior q with expectation equal to the prior. The paper gives a necessary and sufficient condition for zero value of information (Proposition 2), global upper and lower bounds based on the confidence set, the indifference kernel, and the smoothness of the value function (Theorems 3, 5, and 7), and applies these bounds to the marginal value of information near no information in Section 5, including Brownian, Poisson, and binary-signal examples. The main technical tool is standard convex analysis, with proofs collected in an appendix.","tokens_in":24689,"tokens_out":9194,"duration_ms":94875,"significance":"The framework is elegant and potentially useful: representing the value function as a support function gives a transparent geometric interpretation of when information matters, and the paper separates conditions on the decision problem from conditions on the information structure in a way that the Radner-Stiglitz literature often does not. Proposition 2 is a clean characterization, and the asymptotic classification in Section 5 (zero, finite, or infinite marginal value) is a nice synthesis of existing results, including a comparison with De Lara and Gilotte (2007). The paper is also honest about the scope of its model: it assumes expected utility with randomization, so actions are identified with payoff vectors in a convex compact set. However, two substantive issues need attention before the paper can be accepted: the proof of the upper bound in Theorem 3 uses an invalid constant, and Section 3.2 misidentifies the undecided priors in Example 1. These are repairable, but they affect the reliability of the main results as written.","major_comments":[{"comment":"The proof of the upper bound in Theorem 3 is incorrect as written. The proof fixes an arbitrary a in A and asserts φ_a(q) ≤ (sup_{a' in A} ||a−a'||) d(q, Δ^c_A(pbar)), using φ_a(p)=0 for p in the confidence set via (46b). But (46b) is only established for a in F_A(pbar)=A^*(pbar), not for arbitrary a in A. Consequently the derived constant C_A = inf_{a in A} sup_{a' in A} ||a−a'|| is not valid. This is not merely a cosmetic issue: in Example 1, with prior pbar=(1/2,1/2) and binary full information, VoI_A(q)=3/4, while the claimed constant gives C_A·E[d(q,Δ^c_A(pbar))]=√122/15≈0.736, so the displayed upper bound fails. The theorem is salvageable by taking C_A to be the diameter of A or by restricting the infimum to a in A^*(pbar), but the corrected proof is needed because this bound is used in Proposition 11 and in the Section 5 examples.","section":"Appendix A.3.1, Eq. (49), Theorem 3"},{"comment":"The manuscript states that in Example 1 the agent is undecided at pbar=1/2 and pbar=3/4, with several optimal actions and a nondifferentiable value function. This is incorrect: at both of these beliefs the unique optimal action is (2,2) and the value function is differentiable. The actual kinks, where two pieces of the value function meet, are at p=1/3 (between (3,0) and (2,2)) and p=4/5 (between (2,2) and (0,5/2)). This misidentification appears twice, in Section 3.2 and in the subsequent remark on Theorem 5, and it undermines the illustration of the undecided case. The text should either use p=1/3 and p=4/5 for Example 1 or replace Example 1 with a genuinely indifferent example.","section":"Section 3.2 and the remark after Theorem 5"}],"minor_comments":[{"comment":"There is a missing closing parenthesis in the displayed expression for E[d(q_θ, Δ^c_A(pbar))]; it should read E[d(q_θ, Δ^c_A(pbar))]=0 for θ small enough.","section":"Example 14, Section 5.2"},{"comment":"The text says 'a standard Browian process'; this should be 'Brownian process'.","section":"Example 12, Section 5.2"},{"comment":"In the density argument near the end of the proof, the notation for the closure of the set is hard to follow as rendered; consider rewriting this step with an explicit closure operation to improve readability.","section":"Proof of Proposition 2, Appendix A.3.1"}],"recommendation":"major_revision","confidential_remarks":"The Theorem 3 constant error is the main technical issue; it is localized and fixable, but it is load-bearing because the upper bound is advertised as one of the paper's main global estimates. The Example 1 misidentification in Section 3.2 suggests that the example-based discussion should be checked carefully. I do not see a circularity problem with the De Lara and Gilotte comparison; Proposition 11 is shown to imply their result rather than relying on it. With a corrected Theorem 3 proof and a corrected Example 1 discussion, the paper would be a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper does something genuinely useful: it re-expresses the value of information through the support function of the action set and gets a clean division into three regimes. The confidence set characterization (Prop. 2)—zero value iff posteriors stay where all prior-optimal actions remain optimal—is simple and right. The bounds in Theorems 3, 5, and 7 give global control from local geometry at the prior, and Section 5's classification of marginal value as zero, finite, or infinite is a real advance over the Radner-Stiglitz nonconcavity papers. Proposition 11 indeed subsumes the De Lara-Gilotte sufficient condition, so the novelty is not just cosmetic.\n\nThe convex-analysis machinery is standard, proofs are mostly traceable, and the insurance example gives a nice concrete illustration. No circularity, and the one self-citation is used as a comparison that the paper actually generalizes.\n\nNow the soft spots, in order of importance.\n\nFirst, the proof of Theorem 3's upper bound has a genuine gap. It defines phi_a for arbitrary a in A, then invokes (46b) which only holds for a in A^*(pbar). As a result the displayed constant C_A = inf_{a in A} sup_{a' in A} ||a-a'|| is invalid; on the paper's own Example 1 with full-information q, VoI = 3/4 while C_A E d = sqrt(122)/15 ≈ 0.736, so the bound as constructed is false. The theorem itself is very likely true—the fix is to restrict the infimum to A^*(pbar) or use the diameter of A—but a referee should require the correction. This is a proof error, not a dead end.\n\nSecond, the running example in Section 3.2 misidentifies the undecided priors: the kinks are at 1/3 and 4/5, not 1/2 and 3/4. That mistake also shows up in the discussion after Theorem 5 and should be corrected.\n\nThird, the abstract says \"global estimates ... from local properties ... only,\" but the constants depend on global geometry. The authors acknowledge this in the caveat after Theorem 3, so it's an overstatement in the abstract, not a mathematical error.\n\nWho is this for? Decision theorists and math-economists working on the value of information, especially the Radner-Stiglitz agenda. It deserves serious peer review; conditional on fixing the proof gap and the example, I'd expect it to be a solid contribution. I would accept it for review and lean towards acceptance after revision.","headline":"A genuinely useful convex-analysis treatment of the value of information, with a fixable proof gap in Theorem 3 and a wrong running example; worth refereeing once repaired.","tokens_in":25232,"tokens_out":6733,"would_cite":true,"duration_ms":63288,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46N10","91B06"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper makes the value of information a geometric fact: information is worthless exactly when posteriors stay in the set where prior-optimal actions remain optimal, and worth a measurable amount otherwise.","keywords":["value of information","convex analysis","payoffs-beliefs duality","support function","confidence set","marginal value of information","information structures","decision theory"],"falsifier":"Compute the value of information in the paper's four-action example (Table 1) at prior $p=1/2$ with a signal that sends the posterior to $p=0$ or $p=1$ equiprobably: Proposition 2 predicts $\\mathrm{VoI}=\\frac12 v(0)+\\frac12 v(1)-v(\\tfrac12)=0.75>0$ because both posteriors lie outside the confidence set $[1/3,4/5]$. If such a direct calculation ever returned zero, or if a numerical search over compact convex polytope action sets produced a posterior distribution that escapes the confidence set with positive probability yet has zero expected gain, the central characterization would be refuted.","tokens_in":24080,"feed_emoji":"📈","tokens_out":8278,"duration_ms":82990,"temperature":0.7,"pith_summary":"This paper gives a convex-analysis account of the value of information in decision problems. It treats each action as a payoff vector indexed by states of nature and each belief as a probability vector, pairing them by the scalar product that defines expected payoff; the value function is then the support function of the set of available actions. The central result characterizes zero information value: information is worthless exactly when, almost surely, the posterior belief falls in the confidence set of beliefs at which every action optimal at the prior remains optimal. When posteriors can leave that set with positive probability, information has positive value, and the paper supplies global upper and lower bounds on that value in terms of how far posteriors move. At the margin, near no information, the agent's local behavior at the prior, whether confident, undecided, or flexible, determines whether the marginal value of information is zero, positive and finite, or infinite.","feed_headline":"Information is worthless unless it can change the action","feed_subtitle":"New bounds express a signal's worth by how far its posteriors travel from the prior-optimal actions.","key_machinery":"The load-bearing object is the support function $\\sigma_A(s)=\\sup_{a\\in A}\\langle s,a\\rangle$ of the action set $A\\subset\\mathbb{R}^K$, restricted to beliefs; this is the value function. Convex analysis supplies the dictionary: the subdifferential of the value function at a belief is the exposed face of optimal actions, and the normal cone at an action collects the beliefs that reveal that action as optimal. From these, the paper defines the confidence set of a prior, the posteriors at which every prior-optimal action stays optimal, and the indifference kernel, the directions that break none of the prior's ties. The confidence set carries Proposition 2 and Theorem 3; the indifference kernel and the curvature of the boundary of the action set, via the spherical image and Weingarten maps, carry the undecided and flexible bounds of Theorems 5 and 7.","core_discovery":"The discovery is that the economic question of how much a piece of information is worth is governed by the geometry of the set of available payoff vectors near the prior. Writing $A$ for the compact convex set of actions and $v_A(p)=\\max_{a\\in A}\\langle p,a\\rangle$ for the value function, the set of optimal actions at belief $p$ is the subdifferential $\\partial v_A(p)$, and the set of beliefs at which an action $a$ is optimal is the normal cone $N_A(a)$ intersected with the simplex. Against this backdrop, Proposition 2 states that $\\mathrm{VoI}_A(q)=\\mathbb{E}[v_A(q)]-v_A(\\bar p)$ vanishes if and only if the random posterior $q$ lies almost surely in the confidence set $\\Delta^c_A(\\bar p)=\\bigcap_{a\\in A^\\star(\\bar p)}\\Delta^\\star_A(a)$. Theorems 3, 5, and 7 then bound the value of any information structure by, respectively, the expected distance from the posterior to the confidence set, the expected seminorm distance from prior to posterior when the agent is undecided, and the expected squared distance when the agent is flexible. These local-to-global estimates turn the value of information into a quantity that can be read off from the shape of the action set at the prior alone.","pith_inferences":["The three regimes can be read as a local Taylor expansion of the value function, flat, kinked, or quadratically curved; this suggests that for any information structure, the asymptotic value is fixed by the lowest-order nonzero term in that expansion along the signal's belief displacements.","Because the confidence set is computed only from prior-optimal actions, the bounds offer a robustness tool: an analyst who knows only the local face of the action set at the prior can bound the worst-case value of any information structure without knowing the distribution of signals.","The same payoff-beliefs duality may carry over to infinite state spaces through support functions on dual pairs of locally convex spaces, but the normal-cone and curvature arguments would need functional-analytic reworking; testing the quadratic bound in a Gaussian belief model would be a natural first step."],"forward_implications":["If a signal never moves the posterior outside the confidence set of the prior, it has zero value, no matter how informative it is by other criteria.","If a signal moves the posterior outside that set with positive probability, the value is positive, with upper and lower bounds in terms of expected distance from the posterior to the confidence set and of the probability of leaving an epsilon-neighborhood of it.","At priors where several actions are optimal, a small signal that breaks a tie has first-order value: the value grows like the expected distance between prior and posterior, so the marginal value can be infinite for signals whose belief displacement is of order the square root of the information parameter.","At priors where the optimal action varies smoothly with belief, the value of information grows like the expected squared distance from prior to posterior, making small information second-order and giving diffusion-type signals a positive finite marginal value.","Near no information, the local regime of the value function at the prior, confident, undecided, or flexible, together with the speed at which posteriors spread, decides among three marginal values: zero, a positive finite number, or infinite."],"supporting_citations":[{"why":"Supplies the classical comparison-of-experiments framework and the convex-hull representation of actions that the model builds on.","marker":"[8]"},{"why":"Supplies the representation of information by distributions of posterior beliefs that average out to the prior belief.","marker":"[9]"},{"why":"Provides the convex-analysis foundations, support functions, normal cones, and subdifferentials used throughout the paper.","marker":"[33]"},{"why":"Supplies the convex-analysis facts on differentiability and subdifferentials used in Proposition 4 and in the flexible-agent proofs.","marker":"[22]"},{"why":"Supplies the geometric convex analysis of smooth bodies, spherical image maps, Weingarten maps, and curvature used in the flexible case.","marker":"[34]"},{"why":"Defines the marginal value of information problem the paper reframes and extends with separate conditions on decision problems and information structures.","marker":"[32]"},{"why":"Gives a tight sufficient condition for zero marginal value of information that Proposition 11 is shown to imply.","marker":"[16]"},{"why":"Provides earlier joint conditions on information and decision problems for null marginal value, which the paper compares with its separate-condition approach.","marker":"[15]"}],"fun_headline_variants":["Payoff geometry sets the price of information","Local shape of payoffs bounds any signal's value","Confidence sets reveal when information pays","Posteriors' distance from prior-optimal sets value","Marginal value of info read from prior's geometry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument depends on representing any decision problem as a compact convex set of payoff vectors and on expected-utility maximization $\\max_{a\\in A}\\langle p,a\\rangle$; if preferences are not expected utility, randomization is forbidden, or actions cannot be reduced to state-indexed payoffs, the duality and every theorem built on it collapse.","fun_headline_variants_meta":{"raw":{"variants":["Payoff geometry sets the price of information","Local shape of payoffs bounds any signal's value","Confidence sets reveal when information pays","Posteriors' distance from prior-optimal sets value","Marginal value of info read from prior's geometry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1459,"prompt_tokens":1000,"completion_tokens":459,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":387}},"tokens_in":616,"tokens_out":459,"duration_ms":5156,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:08:50.239428+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the value of information in the paper's four-action example (Table 1) at prior $p=1/2$ with a signal that sends the posterior to $p=0$ or $p=1$ equiprobably: Proposition 2 predicts $\\mathrm{VoI}=\\frac12 v(0)+\\frac12 v(1)-v(\\tfrac12)=0.75>0$ because both posteriors lie outside the confidence set $[1/3,4/5]$. If such a direct calculation ever returned zero, or if a numerical search over compact convex polytope action sets produced a posterior distribution that escapes the confidence set with positive probability yet has zero expected gain, the central characterization would be refuted.","supporting_citations":[{"cited_title":"Blackwell, Equivalent comparison of experiments , Annals of Mathematical Statistics, 24 (1953), pp","cited_arxiv_id":null,"evidence_quote":"Supplies the classical comparison-of-experiments framework and the convex-hull representation of actions that the model builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the representation of information by distributions of posterior beliefs that average out to the prior belief."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the convex-analysis foundations, support functions, normal cones, and subdifferentials used throughout the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the convex-analysis facts on differentiability and subdifferentials used in Proposition 4 and in the flexible-agent proofs."},{"cited_title":"Schneider, Convex bodies: the Brunn-Minkowski theory , Cambridge University Press, second ed., 2014","cited_arxiv_id":null,"evidence_quote":"Supplies the geometric convex analysis of smooth bodies, spherical image maps, Weingarten maps, and curvature used in the flexible case."},{"cited_title":"Radner and J","cited_arxiv_id":null,"evidence_quote":"Defines the marginal value of information problem the paper reframes and extends with separate conditions on decision problems and information structures."},{"cited_title":"De Lara and L","cited_arxiv_id":null,"evidence_quote":"Gives a tight sufficient condition for zero marginal value of information that Proposition 11 is shown to imply."},{"cited_title":"Chade and E","cited_arxiv_id":null,"evidence_quote":"Provides earlier joint conditions on information and decision problems for null marginal value, which the paper compares with its separate-condition approach."}],"review_version":1}