{"id":"6f8b2b8e-4b3c-49e5-9ac0-ec3ff6c52dd6","arxiv_id":"2501.06871","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors introduce an extended Banzhaf index and claim it makes exact power calculations possible for any hierarchical voting game in near-linear time.","lead":"This paper studies how to measure the voting power of individual voters inside large hierarchical systems, such as elections with regional levels. The authors propose a new formula that they claim computes the standard Banzhaf power index exactly, and much faster, even when lower levels are not perfectly balanced.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5 is not well-posed as printed: Eq. (3) mixes the global voter set N with local child sets E_{t(i)}, and the missing normalization choice changes MEBPI by powers of 2; on a 3-leaf unbalanced tree the global reading gives 1/16 instead of the exact BPI 1/4.","rationale":"The reader's verdict (REJECT, low confidence) is directionally right, but the specific stated reason is not: the flat-game check in the reader's weakest_assumption is wrong, because for leaves |L(j)|=1 makes the denominator product in Eq. (3) equal to 1, so EBPI does reduce to BPI in flat games. The real load-bearing defect is that Eq. (3) is underspecified across the recursion in MEBPI: the same symbols N and n are used for the global voter set in p^{EBPI}_i(G) and for the player set of the local subgame U_{t(j)} in the factors of p^{MEBPI}_i. Depending on the reading, the product in Theorem 5 can differ from BPI by powers of 2 at every internal level. The 3-leaf unbalanced example provides a concrete falsification of the global reading and shows that the theorem is not merely unproved but not well-posed as printed. Since all proofs are in a missing supplementary file and no machine-checked or reproducible verification is supplied, the central claim cannot be accepted. The construction may be salvageable with an explicit local normalization, which is why the concern is about correctness risk rather than fraud; the correct remedy is to correct Eq. (3) and provide the proof, but for this version rejection is appropriate.","tokens_in":13790,"tokens_out":21123,"duration_ms":212840,"concrete_test":"Compute the 3-leaf example exactly: leaves a1,a2 under internal node A with local unanimity, leaf b, root characteristic v_root(S)=v_A(S) ∨ v_b(S). Exact BPI for a1 is 1/4. Then evaluate MEBPI_{a1} using Eq. (3) under (i) global N and n, (ii) local n but global denominator product, and (iii) local N, n, and denominator. Only reading (iii) returns 1/4. This check decides whether the printed formula supports Theorem 5; the authors should also state which reading is intended and supply the deferred proof.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim, Theorem 5, rests entirely on Eq. (3), but Eq. (3) is not a well-defined function of the local subgames used in the MEBPI recursion. The numerator sums over S ⊆ E_{t(i)}\\{i}, the children of i's parent, while the denominator multiplies 2^{|L(j)|-1} over j ∈ N\\{i}; if N is the global leaf set, the denominator is always 1 (since leaves have |L(j)|=1), whereas if N is the player set of the local subgame, it is a product over that level only. The paper never disambiguates. This is not the flat-game issue the reader identified: for flat games the denominator is 1 and EBPI does reduce to BPI. The problem is recursive normalization. Consider the tree whose root is the OR of an internal node A and a leaf b, where A is the unanimity game on leaves a1 and a2. Exact BPI of a1 is 1/4. Applying Eq. (3) literally with global N={a1,a2,b} and n=3 gives pEBPI_A(U_root)=1/4 and pEBPI_{a1}(U_A)=1/4, hence MEBPI_{a1}=1/16. If n is taken locally in each subgame while the denominator keeps the global leaf product, one gets 0.75. Only the unstated reading in which both n and the denominator product are local to the child set gives the correct 1/4. Theorem 5 therefore holds only under a corrected definition that the paper does not state, and all proofs are deferred to a supplementary file not present in this version.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies hierarchical voting games, which are simple games structured as trees with voters at the leaves. It identifies a property called balance as the condition under which the standard multiplicative Banzhaf power index (MBPI) decomposition equals the Banzhaf power index (BPI). It then introduces an extended Banzhaf power index (EBPI) and a multiplicative version (MEBPI), and claims in Theorem 5 that MEBPI equals BPI in all hierarchical voting games, balanced or not, yielding an O(n d K 2^K) algorithm for exact BPI. The paper applies this to Slovenian National Council elections and to vocabulary selection in sentiment analysis, using Stanza's sentiment classifier as a characteristic function. The central theoretical claim is that MEBPI reproduces BPI exactly, and the experiments compare MEBPI and MBPI against naive BPI on small sentences.","tokens_in":14246,"tokens_out":13634,"duration_ms":119021,"significance":"If Theorem 5 is correct, the paper makes a useful contribution: it would give a polynomial-ish decomposition for exact BPI in hierarchical games that are not balanced, with an explicit complexity improvement over the naive exponential enumeration. The identification of balance as the key assumption behind the classical MBPI decomposition is also a clarifying observation, and the Slovenian National Council example is a nice illustration of the potential speedup. However, the current text does not support the central claim: the definition of EBPI in Eq. (3) is internally inconsistent under the paper's own flat-game convention, the proof of Theorem 5 is deferred to a supplementary file not present in this version, and the sentiment experiments apply the theory to a non-monotone classifier. The paper does not include machine-checked proofs or reproducible code, and as submitted it does not provide enough evidence for the main theorem.","major_comments":[{"comment":"Eq. (3) as printed contradicts Lemma 3. In a flat voting game, the denominator \\prod_{j\\in N\\setminus\\{i\\}} 2^{|L(j)|-1} equals 2^{n-1}, while the numerator contains an additional factor 1/2^{n-1} inside the summation. Consequently p_i^{EBPI}(G) = p_i^{BPI}(G)/2^{n-1}, not p_i^{BPI}(G). This is not a cosmetic issue: Lemma 4 and Theorem 5 both rely on EBPI reducing to BPI in balanced cases, so the central equality is false as the definition is printed.","section":"Section 4, Eq. (3), Lemma 3"},{"comment":"Eq. (3) is not a well-defined function of the local subgame used in the MEBPI recursion. It mixes the local child set E_{t(i)} with the global set N in the denominator, and the expression v(S\\cup\\{i\\}) is undefined when S\\cup\\{i\\} contains internal nodes under the semantics of Section 2.1, where v is defined on subsets of the leaf set N. On the three-leaf unbalanced tree described in the stress-test note (root OR of a unanimity game on a1,a2 and a leaf b), reading N as the global leaf set gives MEBPI_{a1}=1/16 instead of the exact BPI 1/4; only an unstated local reading gives the correct value. Theorem 5 therefore holds only under a corrected and explicitly stated definition that the paper does not provide.","section":"Section 4, Eq. (3), Theorem 5"},{"comment":"All proofs are deferred to a supplementary file that is not present in this version; the text states 'All proofs appear in the supplementary material' in footnote 4 and again before Theorem 5. The central claim, Theorem 5, is therefore not proven in the submitted manuscript. I cannot verify the equality MEBPI=BPI or the claimed complexity bounds without that proof, and the proof is load-bearing for the paper's main contribution.","section":"Section 4, Theorems 2 and 5"},{"comment":"The experiments use Stanza's sentiment classifier as the characteristic function of a voting game. The paper itself acknowledges that sentences are not monotonic in their clauses and gives a concrete non-monotonic example, yet the theory of Sections 2-4 is restricted to monotone voting games with v(\\emptyset)=0 and v(N)=1. A trained neural classifier does not automatically satisfy monotonicity, so the comparison of MEBPI to naive BPI on these sentences is outside the theorem's scope and does not validate Theorem 5. The paper should either construct a monotone characteristic function or explicitly frame the sentiment application as a heuristic use of the formalism rather than a consequence of the theorem.","section":"Section 5, Experiments"}],"minor_comments":[{"comment":"The domain of EBPI is unclear: Section 2 defines voters as leaf nodes N, but MEBPI applies p^{EBPI} to local subgames U_{t(j)} whose voter set E_{t(j)} may contain internal nodes, and Lemma 4 and Theorem 5 write p^{BPI}_i for i\\in M even though BPI was defined only for voters i\\in N. The paper should specify whether 'voting game' refers to the flat local game or the full hierarchical game in Eq. (3).","section":"Section 2.1, Definitions"},{"comment":"Reference [10] is cited as 'Seth J. Chandler. Yelp Dataset,' but the Yelp dataset is not authored by Chandler; this appears to be a citation error that should be corrected.","section":"References"},{"comment":"The sentence 'The EBPI formula, which generalizes BPI, is a factor of K slower than the naive BPI formula' is confusing: the formula in Eq. (3) contains normalization terms beyond the standard 2^{n-1}, and the complexity comparison should be stated in terms of the algorithms rather than the index formula.","section":"Section 4, Complexity discussion"},{"comment":"Each data point in Figure 1 is an average over only 10 sentences; it would be helpful to report error bars or standard deviations, since the mean squared error comparisons are likely to be noisy and the visual difference between MBPI and MEBPI is a central experimental claim.","section":"Section 5.1, Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to be an early-stage submission of an AAMAS paper, and the arXiv version does not contain the supplementary proofs that the text repeatedly references. Given the ambiguity in Eq. (3) and the missing proof of Theorem 5, I would not consider acceptance until the definition is corrected and the central theorem is proven in the paper or in an accessible supplement. The novelty relative to existing multi-level power-index decompositions (e.g., Owen-style recursive formulations) should also be clarified, since EBPI may be related to known weighted-Banzhaf constructions. The sentiment application is potentially appealing but is currently not a valid test of the theorem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper identifies something real — balance is the condition that makes the multiplicative Banzhaf decomposition work — and the ambition to extend BPI to unbalanced hierarchies is sensible. But as submitted, the central definition (Eq. 3) is not well-formed, the main theorem is unproven in the text, and the experiments don't validate the theory. I cannot support acceptance.\n\nWhat is genuinely useful: the explicit statement of balance as the key hypothesis is a clarification that deserves to be recorded. The Slovenian example is a nice, concrete use of the balanced decomposition. The proposed extension, if it can be made precise, would give an interesting O(n d K 2^K) algorithm.\n\nThe soft spots are load-bearing. First, Eq. (3) is ambiguous: the numerator ranges over S ⊆ E_{t(i)} \\ {i} (local siblings) while the denominator is a product over j ∈ N \\ {i} with N never specified as global leaves or local players. Under the literal global reading, the normalization is wrong; on a small unbalanced tree the product along the hierarchy gives the wrong BPI. Second, even in the flat case the printed formula does not reduce to BPI: with leaves, the denominator is 1 and the numerator carries a 2^{n-1} factor that makes EBPI equal 2^{2n-2} times BPI, contradicting Lemma 3. That is an internal inconsistency, not a typo to wave away. Third, all proofs are deferred to a supplementary file that is not present, so Theorem 5 is unverifiable from the text. Fourth, the sentiment experiments use a non-monotone classifier as the characteristic function, and they measure MSE against BPI rather than testing the exact equality; the experiment is a heuristic evaluation, not a test of the theorem. Finally, the paper doesn't connect to the compound-game / Banzhaf-Owen literature, where related decompositions and the balance condition have been discussed.\n\nWho is this for: researchers working on power indices in hierarchical or compound games might find the balance observation useful, but they should not rely on Theorem 5 until the definition is repaired and a proof appears.\n\nRecommendation: send to a serious referee if the authors are expected to fix the definition and provide the proof; otherwise desk reject. As it stands, it is not a citable research result.","headline":"The balance insight is real, but the garbled EBPI definition and missing proof make the central theorem unsupportable as written.","tokens_in":14684,"tokens_out":12433,"would_cite":false,"duration_ms":99939,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A12","91B12"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a generalized Banzhaf index, multiplied level by level through a voting hierarchy, recovers the exact Banzhaf power index for all hierarchical voting games, balanced or not, and the resulting algorithm runs in $O(n d…","keywords":["Banzhaf power index","hierarchical voting games","extended Banzhaf power index","multiplicative decomposition","balanced voting games","sentiment analysis","vocabulary selection","social choice"],"falsifier":"Take a flat weighted majority game with three voters of equal weight and quota two (the simplest balanced majority game). Compute each voter's EBPI directly from Equation (3) and compare with the standard BPI value of $\\frac{1}{2}$; the paper's Lemma 3 predicts they are equal. If they differ, then EBPI is not an extension of BPI and Theorem 5 cannot hold as stated, since the base case of the hierarchy already breaks down.","tokens_in":13635,"feed_emoji":"🗳️","tokens_out":9571,"duration_ms":85827,"temperature":0.7,"pith_summary":"The paper is trying to establish that the Banzhaf power index—the standard measure of how often a single voter is decisive—can be computed in hierarchical voting games by multiplying together a generalized local index at each level of the tree, even when the game is not balanced. If true, this makes exact voter power computable in time $O(n d K 2^K)$ instead of the exponential $O(n 2^n)$, which matters for large real systems such as national electorates, ensembled predictors, and word-level sentiment models. The key move is to introduce an extended index, EBPI, that keeps track of the numbers of winning and losing coalitions in sibling subgames, and to claim that a product of these extended indices equals ordinary BPI for every voter. The paper also applies the result to expose a hidden assumption in earlier recursive BPI formulas—balance—and to demonstrate the speedup on sentiment-based vocabulary selection.","feed_headline":"Generalized Banzhaf index recovers exact power in all voting trees","feed_subtitle":"Multiplying each voter's local influence up the tree replaces exponential coalition checks with a fast recursive pass.","key_machinery":"The load-bearing object is the extended Banzhaf power index (EBPI), defined in Equation (3) as a local measure of a voter's criticality that multiplies each critical coalition by the numbers of winning and losing coalitions in all sibling subgames. The companion recursive counters $\\omega(k)$ and $\\lambda(k)$ count winning and losing coalitions in every subgame of the tree, and these counts feed both the EBPI definition and the tree traversal. The multiplicative extended Banzhaf power index (MEBPI) is then the product of EBPI values along the path from the voter to the root. The claim is that this product, not the plain product of BPI values, is what equals BPI in general.","core_discovery":"At the center of the paper is Theorem 5: in every hierarchical voting game, balanced or unbalanced, the multiplicative extended Banzhaf power index (MEBPI) equals the ordinary Banzhaf power index (BPI) for every voter. The argument is that BPI's exponential counting over all coalitions can be factored level by level: the power of a leaf voter is their local decisiveness in their own subgame times the decisiveness of each ancestor in its parent's subgame, provided each local factor is measured by EBPI, which weights each sibling by the number of winning and losing coalitions it can form. In balanced games these weights cancel and EBPI reduces to BPI, so the classical multiplicative formula is recovered as a special case; in unbalanced games the extra weighting is what lets the product stay exact. Because the local factors are computed from a single tree traversal that counts winning and losing coalitions at every node, the claimed equality yields an algorithm whose cost is $O(n d K 2^K)$ for all voters.","pith_inferences":["If Theorem 5 holds, the product-of-local-indices idea could carry over to permutation-based power indices, where the bottleneck is the same exponential enumeration; the paper names this direction as future work.","The EBPI weighting scheme suggests a sampling-based approximation: estimate $\\omega$ and $\\lambda$ by random subset counts, then run the recursive pass, to get an anytime approximate BPI with confidence bounds.","A natural testable extension is to replace the coarse binary sentiment classifier with a three-valued or continuous sentiment function; the paper notes ternary games as a future direction, and the extended index would need further modification because complements can be neutral in a way that breaks the current binary framework.","The vocabulary-selection application carries an implicit claim that grammatical parse trees mirror compositional semantics; if that correspondence degrades on longer or non-compositional sentences, the proxy error would grow—measurable by comparing MEBPI against exact BPI on sentences beyond length 15, where the baseline becomes infeasible."],"forward_implications":["Exact Banzhaf power becomes computable for large hierarchical electorates: the run time $O(n d K 2^K)$ replaces $O(n 2^n)$, so games with millions of voters are tractable when the tree is shallow and narrow.","In balanced games the generalized index collapses to the ordinary one, so the classical MBPI decomposition is recovered as a special case, and balance is identified as the precise hidden assumption behind that decomposition.","The sentiment of a sentence can be treated as an unbalanced hierarchical vote over words; on review sentences up to 40 words, MEBPI stays close to exact BPI while being orders of magnitude faster (about 6 seconds versus 700 seconds for a 10-word sentence).","The measured power in the Slovenian National Council is a concrete payoff: a voter in Velenje has roughly three times the power of a voter in Ljubljana, a number that comes from the product formula rather than from enumerating two million voters."],"supporting_citations":[{"why":"Defines the Banzhaf power index, the quantity the paper generalizes and computes.","marker":"[7]"},{"why":"Original Penrose formulation of the same voting-power measure.","marker":"[31]"},{"why":"Establishes the multiplicative BPI decomposition for balanced hierarchical games, the result the paper extends to unbalanced games.","marker":"[29]"},{"why":"Introduces BPI as a vocabulary-selection heuristic, the experimental baseline for word importance.","marker":"[30]"},{"why":"Supplies the sentiment classifier and constituency parser used to define the sentence-level voting game.","marker":"[33]"},{"why":"Source of the review corpus used in the runtime and accuracy experiments.","marker":"[10]"},{"why":"Provides the Slovenian National Council election data used in the voter-power case study.","marker":"[26]"}],"fun_headline_variants":["Extended Banzhaf index makes hierarchical voting exact","Exact power in all voting trees via extended Banzhaf","New index exact on unbalanced hierarchical voting","Multiplicative Banzhaf extends to unbalanced trees","Exact Banzhaf power for any voting hierarchy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The equality theorem inherits all of its force from the claim that EBPI as defined in Equation (3) is a genuine extension of BPI—so in a flat voting game it must reduce exactly to the standard index, and in balanced hierarchical games it must coincide with it; if that base-case reduction fails, the product formula cannot be computing BPI.","fun_headline_variants_meta":{"raw":{"variants":["Extended Banzhaf index makes hierarchical voting exact","Exact power in all voting trees via extended Banzhaf","New index exact on unbalanced hierarchical voting","Multiplicative Banzhaf extends to unbalanced trees","Exact Banzhaf power for any voting hierarchy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1410,"prompt_tokens":1038,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":294}},"tokens_in":654,"tokens_out":372,"duration_ms":3943,"temperature":1.0,"reasoning_tokens":294,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:53:41.107865+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a flat weighted majority game with three voters of equal weight and quota two (the simplest balanced majority game). Compute each voter's EBPI directly from Equation (3) and compare with the standard BPI value of $\\frac{1}{2}$; the paper's Lemma 3 predicts they are equal. If they differ, then EBPI is not an extension of BPI and Theorem 5 cannot hold as stated, since the base case of the hierarchy already breaks down.","supporting_citations":[{"cited_title":"Weighted V oting Doesn’t Work: A Mathe- matical Analysis","cited_arxiv_id":null,"evidence_quote":"Defines the Banzhaf power index, the quantity the paper generalizes and computes."},{"cited_title":"The Elementary Statistics of Majority V oting","cited_arxiv_id":null,"evidence_quote":"Original Penrose formulation of the same voting-power measure."},{"cited_title":"A Priori V oting Power and the US Electoral College","cited_arxiv_id":null,"evidence_quote":"Establishes the multiplicative BPI decomposition for balanced hierarchical games, the result the paper extends to unbalanced games."},{"cited_title":"Game-theoretic V ocabulary Selection via the Shapley Value and Banzhaf Index","cited_arxiv_id":null,"evidence_quote":"Introduces BPI as a vocabulary-selection heuristic, the experimental baseline for word importance."},{"cited_title":"Chandler","cited_arxiv_id":null,"evidence_quote":"Source of the review corpus used in the runtime and accuracy experiments."},{"cited_title":"DRZAVNI SVET–THE NATIONAL COUNCIL OF SLOVENIA","cited_arxiv_id":null,"evidence_quote":"Provides the Slovenian National Council election data used in the voter-power case study."}],"review_version":1}