{"id":"20b09efd-9a43-4f13-b4d7-cf8d100c22d2","arxiv_id":"2602.05155","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"A network of agents can minimize total post-sharing loss variance with a signed linear risk-sharing rule; a compact formula emerges when friends share equally.","lead":"This paper works out the best way for people connected in a network to share each other's insurance losses, assuming only friends can share risk. It gives exact formulas for the optimal sharing rule, including a simple one when friends split losses equally, and connects the result to the graph Laplacian.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2.1 overstates uniqueness of Γ: the linear system (6) can be singular (e.g., a 3-node graph with only edge {1,2}), so Γ is not uniquely 'determined'; A* itself remains unique.","rationale":"I agree with the reader that the paper's main results are mathematically sound and close to acceptance, but I locate the load-bearing issue differently. The reader's weakest assumption — that signed risk-sharing is permitted — is a scope statement the paper makes explicitly in Section 1.2; it is not a hidden assumption. Restricting A to nonnegative entries would indeed change the problem, but the paper never claims to solve that constrained problem, so the reader's concern does not undermine the stated theorem. The sharper technical defect is in Theorem 2.1's uniqueness claim for Γ. The proof establishes that the optimal A* can be represented in the form (5) with some Γ, but it does not show that the linear system (6) has a unique solution. The explicit 3-node disconnected counterexample shows the system can be singular and admits multiple Γ yielding the same A*. Since A* is unique by strict convexity, the central conclusion survives, but the theorem's wording should be qualified. This is a conditional-acceptance-level issue, consistent with the reader's verdict, so I do not change the verdict.","tokens_in":17182,"tokens_out":40192,"duration_ms":370561,"concrete_test":"For the connected 4-node star with μ=1 and Σ=I, construct the m=6 linear system (6) and compute the rank of its coefficient matrix. If the rank is less than 6, the Γ non-uniqueness persists even in a connected graph, and Theorem 2.1 should be amended to say 'any Γ in the affine solution set of (6)' and to discuss when Γ is unique. If the rank is 6, add a connectivity assumption to Theorem 2.1 and note the disconnected counterexample. Independently verify the 3-node example above by substituting both Γ solutions into (5) and confirming they yield the same A*.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the closed-form characterization of the unique optimum via (5) and the linear system (6). Step 3.3 never proves that the m×m coefficient matrix for Γ is nonsingular, and it need not be. For the disconnected graph on {1,2,3} with edge set {{1,2}}, μ=1, Σ=I, the equations (6) have a one-parameter family of solutions: both γ13=γ23=0, γ31=γ32=3/2 and γ13=γ23=1, γ31=γ32=1/2 satisfy (6), and both produce the same matrix A* = [[1/2,1/2,0],[1/2,1/2,0],[0,0,1]]. Thus the phrasing 'Γ ... determined by the linear system (6)' is false as written for graphs outside the connected case, and the parametrization is overcomplete. This does not threaten the uniqueness of A* — strict convexity of the objective guarantees that — nor the algebra of the KKT derivation; it is a precision defect in Theorem 2.1's statement. The reader's concern about signed risk-sharing is a modeling choice explicitly announced in Section 1.2, not a correctness issue; the nonnegativity-constrained problem is a different optimization that the paper does not claim to solve.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies optimal linear risk-sharing on an undirected graph, where each agent's post-sharing loss may depend only on their own loss and the losses of their neighbors. Agents minimize half the total variance of post-sharing losses subject to full allocation, actuarial fairness, and network constraints. Theorem 2.1 gives a closed-form characterization of the unique solution A* via a matrix Γ that solves the linear system (6). Theorem 2.2 specializes to the case where each agent's risk is split equally among friends and derives a Laplacian-based formula. Section 2.5 states nonnegativity conditions, and the paper provides several worked examples, including a barbell network.","tokens_in":17539,"tokens_out":26436,"duration_ms":237221,"significance":"If the main results hold, the paper extends the complete-graph result of Feng, Liu, and Taylor to arbitrary connected networks, and the graph-Laplacian connection for equal-sharing rules is a useful structural insight. The proofs are self-contained and use standard convex-optimization arguments; no parameters are fitted and the examples reproduce the stated objective values. The signed risk-sharing assumption is explicitly stated and is a modeling choice rather than a technical flaw. However, the statement of Theorem 2.1 overclaims uniqueness of the auxiliary matrix Γ, and Corollary 2.2 is misstated as written. These are fixable, but they are substantive enough to require revision.","major_comments":[{"comment":"The statement that Γ is 'determined by' the linear system (6) is not correct for graphs that are not connected. For the 3-vertex graph with edge set {{1,2}}, μ=1, Σ=I, the system (6) is singular: both γ13=γ23=0, γ31=γ32=3/2 and γ13=γ23=1, γ31=γ32=1/2 solve it, and both give the same A* in (5). The proof in §3 constructs a Lagrange-multiplier matrix Γ but never proves the coefficient matrix of (6) is nonsingular. Since the paper's stated scope is connected networks, Theorem 2.1 should either explicitly assume G is connected and prove uniqueness of Γ there, or reformulate (6) as 'let Γ be any solution' and note that all solutions yield the same A*. The uniqueness of A* is not affected, as it follows from strict convexity.","section":"Theorem 2.1, Eq. (6)"},{"comment":"The statement is false as written: 'the constant ĉ is nonnegative if and only if ĉ ≤ μ_i for i={1,2} and Cov(...)≤...' is an ill-posed biconditional, since a nonnegative ĉ can exceed μ_i. The correct statement should combine Lemma 2.2 and Corollary 2.1: for a 2-agent network, Â has all nonnegative entries iff 0 ≤ ĉ ≤ μ_i (i=1,2) and the covariance inequality holds. As printed, it is a mathematical error in a stated result and is used in §2.6.4, so it needs correction.","section":"Corollary 2.2"}],"minor_comments":[{"comment":"The introduction refers to 'Proposition 2.2' for nonnegativity of Laplacian-based rules, but no Proposition 2.2 appears in Section 2.5; the relevant results are Lemma 2.2 and Corollaries 2.1–2.2. Please relabel or correct the reference.","section":"Section 1.1"},{"comment":"The proof is headed 'Proof of Lemma 2.1' and contains several occurrences of μ^TΣμ in the denominator where the statement requires μ^TΣ^{-1}μ. The final condition is correct, but the proof should be cleaned up.","section":"Proposition 2.1 proof"},{"comment":"The displayed numerical check of Lemma 2.1 appears to be inconsistent: with the printed μ and μ_* the stated product ∥μ−μ_*∥1∥μ∥∞=18 does not match the displayed vector. Please verify the arithmetic.","section":"Section 2.6.1"},{"comment":"The equality constraints in (18) are always linearly dependent: the sum of the n equations Aμ=μ equals the μ-weighted sum of the n equations 1^T A=1^T. Thus the Lagrange multiplier ν* is not unique. This does not affect the derivation, since x* is unique and any ν* works, but the proof should acknowledge the non-uniqueness to avoid implying that (18) has a unique solution.","section":"Section 3.2, KKT system"}],"recommendation":"major_revision","confidential_remarks":"The central derivation is sound and the paper is self-contained, with no circularity. The main obstacle to acceptance is the overstatement in Theorem 2.1 regarding uniqueness of Γ and the clearly misstated Corollary 2.2. Both are local and fixable; I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, self-contained convex optimization paper. It takes the known complete-graph result of Feng-Liu-Taylor and solves the same problem when the allocation matrix is forced to be zero on non-edges. Theorem 2.1 gives a closed-form expression for the unique optimal A* via Lagrange multipliers, and Theorem 2.2 handles the equal-share-of-friends case with a graph Laplacian formula. The nonnegativity conditions in Section 2.5 are useful and correctly derived. I checked the worked examples; the reported A* matrices and objective values check out.\n\nWhat's genuinely new is the support constraint on A. The KKT technique is standard, but nobody had written down this network version, and the Laplacian connection is a nice observation. The paper is also honest: it says up front that it allows signed risk-sharing, and it doesn't pretend that nonnegativity is guaranteed.\n\nThe soft spots are minor. The statement of Theorem 2.1 says the off-edge entries of Γ 'are determined by' the linear system (6). That is not literally true for disconnected graphs. Take three nodes with a single edge {1,2}, μ=1, Σ=I; the system (6) has a one-parameter family of Γ solutions, all giving the same A*. Strict convexity still gives unique A*, so the main theorem survives. The fix is either to state Theorem 2.1 for connected graphs, as the summary suggests, or to say that some Γ solving (6) yields the unique A*. I'd also like the authors to flag that the linear system may be singular in general, so numerical solvers should be aware.\n\nThe signed-allocation issue is a modeling choice, not an error. The paper explicitly says it makes no nonnegativity assumption, and Section 2.5/2.6 give conditions and examples for when the signed optimum is nonnegative and how network or equal-share constraints can restore nonnegativity. If someone needs nonnegative A, they're solving a different problem. The authors don't claim to solve it.\n\nCitation pattern is clean: proper credit to [22], no self-citation, no fitted constants or selective data. The proof of Proposition 2.1 has a mislabeled 'Proof of Lemma 2.1' heading, but that's typographical.\n\nBottom line: this is a solid paper for the P2P insurance / decentralized risk-sharing subfield. It deserves a serious referee. I'd send it to peer review and ask for the Theorem 2.1 wording fix; after that it should be close to acceptance.","headline":"Solid KKT-based generalization of complete-graph risk-sharing to general networks; Theorem 2.1 needs a wording fix on uniqueness of Γ, but the central results hold.","tokens_in":17983,"tokens_out":4098,"would_cite":true,"duration_ms":43148,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91B30","90C20","05C50"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that on any connected friendship network, the optimal signed linear actuarially fair risk-sharing rule is unique and given by an explicit closed-form matrix; with equal shares among friends, the rule reduces to a graph-Lap","keywords":["decentralized insurance","peer-to-peer insurance","risk-sharing","network","graph Laplacian","variance minimization","actuarially fair","signed linear risk-sharing"],"falsifier":"Take a small connected graph (e.g., four agents with one missing edge), pick a positive-definite Σ and positive µ, solve the quadratic program (4) with a generic convex QP solver, and compare every entry of the numerical minimizer with the formula from Theorem 2.1; any mismatch—or any violation of the KKT system (6)—would refute the characterization.","tokens_in":17130,"feed_emoji":"🤝","tokens_out":5402,"duration_ms":48434,"temperature":0.7,"pith_summary":"The paper studies how a group of agents connected by a friendship network should share their losses when only directly connected agents may exchange risk. It shows that among all linear, actuarially fair risk-sharing rules, there is exactly one that minimizes the sum of post-sharing variances, and it gives a closed-form expression for that rule in terms of the loss means, covariance, and network structure. When friends are required to take equal shares, the optimal rule takes the especially simple form A = I - c L M^{-1}, where L is the graph Laplacian. A sympathetic reader would care because this turns network-based insurance design into a tractable optimization problem and unifies earlier work that only covered fully connected networks.","feed_headline":"Friends-only risk sharing has one optimal rule","feed_subtitle":"A closed-form formula minimizes total variance on any connected network; equal shares reduce to the graph Laplacian.","key_machinery":"The load-bearing object is the linear risk-sharing rule H(X) = AX. The 'only friends share risk' constraint is encoded by requiring A to vanish outside the graph's adjacency-plus-diagonal support. The proof vectorizes A and rewrites the problem as a strictly convex quadratic program with equality constraints, so the KKT conditions are necessary and sufficient; the Lagrange multipliers reduce to a matrix Γ supported on non-edges, and Γ is the only quantity that must be computed by solving a linear system. In the equal-share case, the constraint forces A = I - L S, and connectedness plus Aµ = µ forces S = c M^{-1}, leaving a single scalar c that is found by minimizing a convex quadratic in c.","core_discovery":"For a connected graph G with n agents, losses X with mean µ and positive definite covariance Σ, Theorem 2.1 characterizes the unique solution A* of the quadratic program that minimizes (1/2)tr(AΣA^T) subject to full allocation (1^T A = 1^T), actuarial fairness (Aµ = µ), and the support constraint a_ij ≠ 0 only when i = j or {i,j} is an edge. The solution is A* = (1/n)11^T + (I - (1/n)11^T)[(1/a)µµ^T + Γ((1/a)Σ^{-1}µµ^T - I)]Σ^{-1}, where a = µ^T Σ^{-1} µ and Γ has zero diagonal and zero entries on edges, with its off-edge entries fixed by a linear system. Theorem 2.2 then shows that if friends must take equal shares of each agent's risk, the unique optimum becomes A = I - ĉ L M^{-1}, with ĉ","pith_inferences":["Because the paper explicitly allows signed transfers, its closed forms do not solve the version of the problem in which A must be entrywise nonnegative; a natural follow-up is to characterize the optimum under that stricter, more insurance-realistic constraint, where the feasible set is no longer an affine subspace and the KKT system changes.","The graph-Laplacian formula suggests that the variance reduction achievable by equal-share sharing is tied to the Laplacian's spectrum; one could test whether networks with larger algebraic connectivity systematically admit lower post-sharing variance under Theorem 2.2.","The barbell example implicitly proposes a network-design heuristic—connect agents whose mean losses lie within a bounded ratio—and a quantitative extension would derive the optimal such ratio from µ and Σ rather than choosing it by hand.","The single-period model could be iterated: applying the same optimal rule each period yields a multi-period scheme whose long-run variance could be compared with the multi-period P2P models cited in the introduction."],"forward_implications":["The earlier complete-graph solution (all agents may share risk) is recovered exactly when Γ = 0, so the network result is a strict generalization.","Imposing a network structure weakly increases the minimized total variance: sharing with fewer counterparties cannot improve the optimum, as the examples show.","In the equal-share setting, the entire rule is determined by graph degrees, entrywise means, and covariances; for a d-regular graph with i.i.d. losses it becomes A = I - L/(d+1), where each agent keeps 1/(d+1) of its own loss and gives 1/(d+1) to each friend.","The nonnegativity conditions provide simple checks for when the signed optimum happens to involve no negative transfers, and the barbell example shows that choosing whom to connect can remove negative entries.","These results give a quantitative basis for comparing network designs by their post-sharing variance and for deciding whether equal sharing among friends is worth its cost."],"fun_headline_variants":["Friends-only risk sharing: one optimal rule","Closed-form rule for risk sharing on social networks","Unique optimal rule when only friends share risk","Graph Laplacian governs optimal friends-only risk sharing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The theorems assume the risk-sharing matrix may contain negative entries, meaning some agents can profit from others' losses; if that is forbidden, the optimization problem changes and the stated formulas no longer describe its optimum.","fun_headline_variants_meta":{"raw":{"variants":["Friends-only risk sharing: one optimal rule","Closed-form rule for risk sharing on social networks","Unique optimal rule when only friends share risk","Graph Laplacian governs optimal friends-only risk sharing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1365,"prompt_tokens":693,"completion_tokens":672,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":437,"tokens_out":672,"duration_ms":7692,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T04:20:00.749321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small connected graph (e.g., four agents with one missing edge), pick a positive-definite Σ and positive µ, solve the quadratic program (4) with a generic convex QP solver, and compare every entry of the numerical minimizer with the formula from Theorem 2.1; any mismatch—or any violation of the KKT system (6)—would refute the characterization.","supporting_citations":[],"review_version":1}