{"id":"8b06c2d9-25d6-4a57-a3aa-d330df177505","arxiv_id":"1908.08924","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"LIRGOMAX computes the exact first-order response of PageRank to weak perturbations and applies it to 5.416 million Wikipedia articles to find nodes most sensitive to injection at one node and absorption at another.","lead":"The authors develop LIRGOMAX, a method that computes how a large network's PageRank rankings respond to a small perturbation, and they apply it to the English Wikipedia. The method identifies the articles most affected when probability is injected at one article and absorbed at another, which they interpret as the effective pathway between the two.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'effective pathway' interpretation is the weak point: Section 2.5 admits no mathematical basis, the implemented subset is 20+20 by sign rather than top-40 |P1|, and no flow-based or ground-truth validation supports the claim.","rationale":"The mathematics of the pump-model derivation (Eqs. 9–16) is sound: iteration (12) converges in the subspace orthogonal to E^T because G0 is a positive stochastic matrix and V0 has zero sum; the projection (13) controls roundoff. The reader correctly identified the unvalidated interpretive step from linear response magnitude to pathway membership, and the paper itself flags the missing argument in Section 2.5. My reading strengthens that concern by noting that the implemented subset selection is not the top 40 by |P1| but a 20-negative/20-positive split, so the displayed REGOMAX structure is partly a consequence of that arbitrary choice. I therefore keep the CONDITIONAL verdict: the paper is a useful sensitivity algorithm with a correct core derivation, but the 'effective pathway' application needs either a formal justification or a benchmark against a flow-based ground truth. (I also note that Eq. (17) appears to omit a factor (G0)_kl in the sensitivity expansion; this does not affect the pump-model examples but should be corrected.)","tokens_in":14363,"tokens_out":14756,"duration_ms":148825,"concrete_test":"Recompute the subset for the Cambridge→Harvard case as the strict 40 largest entries of |P1| (and likewise for Napoleon→Alexander I), then rerun REGOMAX on this subset. If the selected nodes and the Grr+Gqr^(nd) friend network differ materially from Tables 1–2 and Figures 2–4 and 6–8, the reported 'pathway' is an artifact of the arbitrary 20/20 sign-balanced split rather than the stated |P1| ranking.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central application claim — that the nodes with the largest |P1| form the 'effective pathway' between the injection and absorption nodes — is unsupported. Section 2.5 explicitly states that there is 'no simple mathematical argument' that the two initial nodes even belong to the top-response set, and the paper offers no external ground truth or flow-based benchmark. A further concrete difficulty is that the algorithm as implemented does not select the top 40 entries by |P1| at all: it selects the 20 most negative and the 20 most positive entries. In Table 1, for instance, negative entries with KL ranks 24–68 are omitted, while positive entries with ranks as low as 278 are included; Table 2 similarly includes positive entries down to KL=209 while skipping intervening ranks. The REGOMAX subset, and therefore Figures 2–4 and 6–8, is hence dependent on an arbitrary sign-balanced split, not on the stated magnitude criterion. Since P1 is a derivative of the stationary PageRank, a large |P1| can arise from high PageRank or global probability redistribution (e.g., United States at KL=4) rather than from lying on a directed source-to-target path. Without an independent pathway metric, LIRGOMAX is a sensitivity algorithm; the 'pathway discovery' reading remains a conjecture.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops a linear response formalism for the Google matrix/PageRank problem, deriving the first-order correction P1 to the PageRank vector under a weak perturbation of either the Google matrix or the stochastic process. The central algorithmic contribution is LIRGOMAX, which solves the inhomogeneous PageRank equation P1 = G0 P1 + V0 by iteration with a projection step, avoiding the numerical loss of precision of finite-difference estimates. Two perturbation models are treated: a pump model with probability injection at one node and absorption at another, and a PageRank-sensitivity model for individual matrix-element changes. The method is applied to the 2017 English Wikipedia network (N ≈ 5.4 million) for two injection/absorption pairs (University of Cambridge / Harvard University, and Napoleon / Alexander I of Russia). In each case the authors select 40 nodes from the extremes of P1, apply the REGOMAX reduction to that subset, and present reduced Google matrices and 'friend' networks, which they interpret as revealing the effective pathway between the initial pair of nodes.","tokens_in":14684,"tokens_out":5541,"duration_ms":58407,"significance":"If assessed as a computational method for the exact ε→0 limit of PageRank linear response, the contribution is solid and potentially useful. The derivation in Section 2.4 is internally consistent; the iteration (12) with projection (13) is a standard and efficient way to solve the singular inhomogeneous equation, and the reported scalability to a 5-million-node network is a genuine practical advantage. The paper honestly acknowledges some limitations, notably in Section 2.5 where it concedes that there is no mathematical argument guaranteeing the initial nodes belong to the top-response set. However, the paper's central application claim—that the selected nodes constitute the 'effective pathway' between the injection and absorption nodes—is an interpretive step that is not validated by any independent ground truth, flow-based metric, or comparison with alternative pathway-finding methods. As written, the paper demonstrates a sensitivity-analysis algorithm; the pathway-discovery language overstates what is established.","major_comments":[{"comment":"The identification of the nodes with the largest linear-response values as an 'effective pathway' between injection and absorption nodes is unsupported. Section 2.5 explicitly states that there is 'no simple mathematical argument' that the two initial nodes even belong to the top-response set, and Sections 3.1–3.2 provide no external validation: no comparison with explicit hyperlink paths, random-walk hitting or commute probabilities, betweenness/flow centrality, or any other pathway benchmark. Because the abstract and Discussion present pathway identification as the main application, this is a load-bearing issue. The authors should either provide such validation or explicitly reframe the contribution as a sensitivity-analysis tool rather than a pathway-discovery method.","section":"Section 2.5 and Section 3"},{"comment":"The implemented node-selection rule is not the top-40 subset by |P1| implied by the surrounding narrative; it is the union of the 20 most negative and 20 most positive entries of P1. Table 1 omits negative entries with KL ranks between 24 and 68 while including positive entries with KL ranks as low as 278, and Table 2 shows the same pattern (e.g., positive entries down to KL=209 while intervening negative entries are skipped). This sign-balanced selection changes the REGOMAX subset and therefore changes all subsequent reduced-matrix results in Figures 2–4 and 6–8. The authors should justify this unstated modeling choice or report results for the literal top-40 magnitude criterion, since the current rule is an additional arbitrary element in the pipeline.","section":"Section 2.5, Tables 1 and 2"}],"minor_comments":[{"comment":"There are small language errors: 'matricies' in Section 4 should be 'matrices', and Section 3.2 contains the duplicated article in 'In contrast to the the previous example'.","section":"Section 4 and Section 3.2"},{"comment":"The statement that results are insensitive to α in the range 0.5 ≤ α ≤ 0.95 is cited to previous work; since P1 is a new derived quantity, a one-line check that the reported top-response tables are stable under this α variation would make the algorithmic claim more self-contained.","section":"Section 2.1"},{"comment":"The caption explains that friends are chosen by the largest matrix elements in the same column of Grr+Gqr^(nd), but the direction convention of the drawn arrows should be stated explicitly in the captions, because the column-stochastic convention may make the link direction non-obvious to readers.","section":"Figures 4 and 8"},{"comment":"Reference [10] is cited as an arXiv preprint; if a peer-reviewed version exists, it should be cited instead, and reference [16] is a project URL without a persistent identifier, so the exact data and code snapshot used for the figures should be archived with a DOI.","section":"References"},{"comment":"The term 'pathway' is used before it is operationally defined; the manuscript would be clearer if it stated in Section 2.5 that 'pathway' is a heuristic label for the set of nodes with large |P1|, rather than a mathematically defined object.","section":"Section 2.5"}],"recommendation":"major_revision","confidential_remarks":"The technical core of Section 2 is sound and likely publishable; the main concern is that the application claims go beyond what is demonstrated. I would not reject the paper on mathematical grounds. A revision that addresses the unsupported 'pathway' interpretation and the arbitrary sign-balanced selection rule, or that clearly reframes the contribution as sensitivity analysis, would bring the manuscript in line with its evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is solid. The derivation in Section 2 is correct: expanding G and F to first order, getting the inhomogeneous equation P1 = G0 P1 + V0, and solving by power iteration with a projection step is standard but cleanly done. The algorithm genuinely scales to the 5.4M-node Wikipedia graph, and the exact ε→0 limit is a real improvement over finite-difference sensitivity. The pump model normalization Di = 1/P0(i), Dj = −1/P0(j) is a nice trick that makes the first-order source vector sparse and interpretable. Credit where due: this is a useful tool for PageRank sensitivity analysis, and the authors are transparent about the machinery and provide code and data links.\n\nThe soft spot is the interpretive leap from P1 to 'effective pathway.' They compute the 20 most negative and 20 most positive entries of P1, not the top 40 by |P1|, despite saying otherwise in the text. That is a concrete mismatch: in Table 1, negative entries with KL up to 68 are skipped while positive entries with KL as low as 278 are included. So the sign-balanced subset is an arbitrary choice, and the claimed magnitude criterion is not what is implemented. More importantly, Section 2.5 explicitly concedes there is 'no simple mathematical argument' that the injection and absorption nodes even belong to the top-response set. Without an independent flow-based or ground-truth metric, the pathway interpretation is a conjecture. Large |P1| can reflect global PageRank redistribution rather than a source-to-target route, as the appearance of United States at KL=4 suggests.\n\nThat said, these are problems with the application framing, not with the mathematics. The linear response computation is correct, and the REGOMAX analysis of the selected subset is a reasonable, if preliminary, illustration. The paper reads as an honest first demonstration. For a reader wanting to compute PageRank sensitivity on huge directed networks, this is a genuinely useful contribution. For a reader expecting validated path discovery, it falls short.\n\nMy recommendation: send it to peer review. A good referee should push the authors to either validate the pathway claim against a flow-based baseline or reframe the paper as sensitivity analysis without the pathway language. The subset-selection discrepancy should also be fixed or explained. This is a worthwhile contribution that needs revision, not rejection.","headline":"Sound linear-response math and a scalable algorithm, but the 'effective pathway' claim rests on an unvalidated interpretive step the authors themselves flag.","tokens_in":15197,"tokens_out":1040,"would_cite":true,"duration_ms":12662,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new algorithm computes the exact linear response of PageRank, locating the pathway between any two nodes.","keywords":["Google matrix","PageRank","linear response theory","LIRGOMAX","REGOMAX","directed networks","injection and absorption","sensitivity analysis"],"falsifier":"Compute $P_1$ with LIRGOMAX on a small directed network where the true pathway is known from an independent flow-based metric, for example maximum flow or random-walk hitting probabilities, and check whether the top-$|P_1|$ subset contains both endpoints and coincides with that pathway. A pair with a low-PageRank absorption node having few incoming links, where the paper itself expects the endpoints to fall outside the top subset, would already falsify the automatic pathway-selection claim.","tokens_in":1776,"feed_emoji":"🕸️","tokens_out":3347,"duration_ms":97933,"temperature":0.7,"pith_summary":"This paper claims that a weak perturbation of a Google matrix, or of the Markov process behind PageRank, produces a first-order correction to the PageRank vector that can be computed exactly, without finite-difference noise, by solving an inhomogeneous equation. It introduces the LIRGOMAX algorithm, which iterates $P_1 = G_0 P_1 + V_0$ to convergence in large directed networks. For the injection/absorption variant, where probability is pumped into one node and removed at another, the entries of $P_1$ with largest absolute value mark the nodes most sensitive to the perturbation, and these form an automatically selected subset of the effective pathway between the two nodes. The paper demonstrates on a five-million-node network that this subset, combined with the REGOMAX reduction, reveals a two-block structure of direct and indirect interactions between the two endpoints. The broader point is that linear response, a standard tool in physics, transfers to PageRank and gives a direct way to probe influence and hidden pathways in directed networks.","feed_headline":"Exact PageRank response maps the pathway between two nodes","feed_subtitle":"Small perturbations become a map of influence between nodes in million-node directed networks.","key_machinery":"The load-bearing object is the inhomogeneous PageRank equation $P_1 = G_0 P_1 + V_0$ together with the iterative solver and the projection $Q(X)=X-(E^TX)P_0$. The equation converts the linear-response problem into a source-driven version of the PageRank fixed-point problem, so it can be solved by the same convergent iteration as PageRank itself. The projection removes the numerical component along $P_0$ that would otherwise accumulate and slow convergence, and it enforces the normalization condition $E^TP_1=0$. For the injection/absorption model, the special choice of the perturbation matrix makes $V_0$ a difference of two columns of $G_0$, which is what ties the response vector to a directed pathway between two nodes.","core_discovery":"The central discovery is that the linear response of PageRank is governed by a single inhomogeneous fixed-point equation. Expanding $G(\\varepsilon)=G_0+\\varepsilon G_1$ and $F(\\varepsilon,P)=P+\\varepsilon F_1(P)$ around $\\varepsilon=0$, the $O(\\varepsilon)$ correction $P_1$ satisfies $P_1 = G_0 P_1 + V_0$ with $V_0 = G_1P_0 + G_0F_1(P_0)$. Because $E^T V_0 = 0$, the apparent singularity of $(1-G_0)^{-1}$ is harmless: $V_0$ has no component along the PageRank eigenvector, and iteration $P^{(n+1)}_1=G_0P^{(n)}_1+V_0$ converges with the same speed as the PageRank power method. For injection at $i$ and absorption at $j$, the special scaling $D_i=1/P_0(i)$, $D_j=-1/P_0(j)$ makes the source term exactly the difference between the $i$-th and $j$-th columns of $G_0$, so the response is driven by the difference of two columns of the Google matrix. Ordering all nodes by $|P_1|$ and keeping the extremes yields the most sensitive nodes, and applying REGOMAX to that subset gives the reduced Google matrix with its direct and indirect interaction components.","pith_inferences":["I read the paper's 'pathway' as a claim about linear sensitivity, not about graph-theoretic paths; a natural test is whether top-$|P_1|$ nodes coincide with high-flow or high-commute-time routes on small networks where those quantities can be computed exactly.","The difference-of-columns form $V_0=(G_0)_{:i}-(G_0)_{:j}$ suggests a Green's-function interpretation: $P_1$ is the response of the Markov chain to a unit source-sink pair, so the same machinery could be applied to node influence, betweenness, or random-walk hitting times.","The paper's own caveat that injection nodes with low PageRank and few incoming links may not appear in the top-response subset implies that the automatic selection is endpoint-blind in those cases; a low-PageRank pair would require a separate criterion for including the endpoints themselves."],"forward_implications":["The PageRank sensitivity $D^{(j\\to i)}(k)=P_1(k)/P_0(k)$ to a matrix-element perturbation can be computed in the exact $\\varepsilon\\to 0$ limit, avoiding the numerical loss of precision of finite-difference formulas.","The method provides an automatic, non-heuristic way to select a small subset of nodes relevant to a perturbation, which can then be studied with REGOMAX; this replaces arbitrary choices of subset.","The same linear-response framework applies to any perturbation of the Google matrix or of the Markov process, not only injection/absorption, so it can quantify how small structural changes propagate through a directed network.","On the tested five-million-node network, the top positive and negative entries of $P_1$ form two coherent blocks around the injection and absorption nodes, showing that the response captures meaningful global structure.","Because the iteration has the same cost as the power method, the computation is feasible for any network where the unperturbed PageRank can be computed."],"supporting_citations":[{"why":"Defines the Google matrix and PageRank eigenproblem that the perturbation theory is built on.","marker":"[6]"},{"why":"Supplies the standard construction, damping factor and iteration method used for the unperturbed PageRank $P_0$.","marker":"[7]"},{"why":"Sets out Google matrix analysis of directed networks and the CheiRank notions used in the response ranking.","marker":"[8]"},{"why":"Provides the English Wikipedia 2017 network with five million nodes used in the numerical demonstrations.","marker":"[9]"},{"why":"Introduces the reduced Google matrix (REGOMAX) that LIRGOMAX feeds with its selected subset.","marker":"[10]"},{"why":"Exemplifies the REGOMAX application and friend-network construction that the paper reuses for the selected subsets.","marker":"[11]"}],"fun_headline_variants":["Linear response maps PageRank's hidden pathways","PageRank response theory reveals node influence paths","LIRGOMAX: fast PageRank sensitivity maps","Weak links expose PageRank pathway in networks"],"cache_read_input_tokens":17280,"weakest_assumption_plain":"The ranking of nodes by the magnitude of the first-order response $|P_1|$ is taken to identify the 'effective pathway' between the injection and absorption nodes; the paper states in Section 2.5 that there is no mathematical argument placing the two endpoints in that top subset, especially for low-PageRank nodes.","fun_headline_variants_meta":{"raw":{"variants":["Linear response maps PageRank's hidden pathways","PageRank response theory reveals node influence paths","LIRGOMAX: fast PageRank sensitivity maps","Weak links expose PageRank pathway in networks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1331,"prompt_tokens":1011,"completion_tokens":320,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":274}},"tokens_in":627,"tokens_out":320,"duration_ms":3614,"temperature":1.0,"reasoning_tokens":274,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:25:19.226169+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $P_1$ with LIRGOMAX on a small directed network where the true pathway is known from an independent flow-based metric, for example maximum flow or random-walk hitting probabilities, and check whether the top-$|P_1|$ subset contains both endpoints and coincides with that pathway. A pair with a low-PageRank absorption node having few incoming links, where the paper itself expects the endpoints to fall outside the top subset, would already falsify the automatic pathway-selection claim.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Google matrix and PageRank eigenproblem that the perturbation theory is built on."},{"cited_title":"Langville and C.D","cited_arxiv_id":null,"evidence_quote":"Supplies the standard construction, damping factor and iteration method used for the unperturbed PageRank $P_0$."},{"cited_title":"Frahm and D.L","cited_arxiv_id":null,"evidence_quote":"Sets out Google matrix analysis of directed networks and the CheiRank notions used in the response ranking."},{"cited_title":"Frahm and D.L","cited_arxiv_id":null,"evidence_quote":"Provides the English Wikipedia 2017 network with five million nodes used in the numerical demonstrations."},{"cited_title":"Frahm, K","cited_arxiv_id":null,"evidence_quote":"Exemplifies the REGOMAX application and friend-network construction that the paper reuses for the selected subsets."}],"review_version":1}