Case Study 2 — Which Scientific Papers Matter? Citation Networks and Eigenvector Influence
Field: bibliometrics and the science of science. This case study applies the chapter's dominant-eigenvector idea to the network of scholarly citations, where it underlies real influence metrics used to rank journals and papers — and exposes, once again, why counting is not the same as measuring importance.
The problem: a citation count is a popularity contest
Science runs on citations. When one paper cites another, it acknowledges a debt — the cited work was useful, foundational, or worth arguing with. It is tempting to measure a paper's importance simply by counting its citations: the more papers cite you, the more important you are. This is the bibliometric version of the link-counting we dismantled in §29.1.1, and it has the identical flaw. A citation from a landmark paper in a top journal surely signals more importance than a citation from an obscure, never-cited preprint — yet raw counting weighs them equally. Worse, citation counts can be inflated by self-citation rings, by review articles that cite hundreds of papers indiscriminately, and by the sheer size of a field (a paper in a huge, fast-moving area accrues citations a paper in a small field never could).
The fix is the chapter's recursive idea, and it has a name in this setting: eigenvector-based influence, the principle behind the Eigenfactor metric for journals and related paper-level scores. [verify] A paper is influential if influential papers cite it — importance defined recursively, resolved as the dominant eigenvector of the citation network. The citation graph is just another directed graph; importance flows along citations toward the cited works; and the stable distribution of that flow is, once more, the eigenvalue-$1$ eigenvector found by power iteration.
Step 1: the citation network as a stochastic matrix
Model each paper as a node and each citation as an arrow from the citing paper to the cited paper — the citing paper "votes for" the work it builds on, just as a web page links to the pages it endorses. Importance should flow to cited papers, so we want the citing paper to pass rank along its outgoing citations. Column $j$ of the link matrix therefore distributes paper $j$'s rank equally among the papers it cites, each receiving $1/(\text{number of references in } j)$.
A wrinkle specific to citations makes the dangling-node machinery essential. Citations point backward in time: a paper can only cite work that already existed. So the oldest, most foundational papers in a network cite nothing within the network — they are dangling nodes by construction, exactly the dead ends of §29.5. Without the patch, the foundational papers would leak away all the rank that flows into them, and the ranking would collapse. The dangling-node fix (replace their empty columns with a uniform jump) is not an edge case here; it is the normal situation for precisely the most important papers.
Consider five papers, $P_1$ through $P_5$, ordered oldest to newest, citing as follows:
- $P_2$ cites $P_1$;
- $P_3$ cites $P_1, P_2$;
- $P_4$ cites $P_1, P_2, P_3$;
- $P_5$ cites $P_1, P_2, P_3, P_4$;
- $P_1$ cites nothing (it is the oldest — a dangling node).
This is a textbook "foundational paper" structure: $P_1$ is cited by everyone, $P_5$ is the recent survey that cites everyone but is itself cited by no one yet.
Step 2: rank by the dominant eigenvector
We build the link matrix, patch $P_1$'s empty column to uniform, form the Google matrix with $d = 0.85$, and compute the dominant eigenvector — then compare it to the raw citation counts.
# Eigenvector influence on a tiny citation network; compare to raw citation counts.
import numpy as np
np.set_printoptions(precision=4, suppress=True)
papers = ['P1', 'P2', 'P3', 'P4', 'P5']
cites = {'P1': [], 'P2': ['P1'], 'P3': ['P1', 'P2'],
'P4': ['P1', 'P2', 'P3'], 'P5': ['P1', 'P2', 'P3', 'P4']}
idx = {p: i for i, p in enumerate(papers)}; n = 5
M = np.zeros((n, n))
for citer, refs in cites.items():
j = idx[citer]
if refs:
for c in refs:
M[idx[c], j] += 1.0 / len(refs)
else:
M[:, j] = 1.0 / n # patch dangling foundational paper P1
G = 0.85 * M + 0.15 / n * np.ones((n, n))
w, V = np.linalg.eig(G)
r = np.abs(V[:, np.argmax(w.real)].real); r /= r.sum()
counts = {p: sum(p in refs for refs in cites.values()) for p in papers}
for p in papers:
print(f"{p}: influence {r[idx[p]]:.4f} raw citations {counts[p]}")
print("influence ranking:", [papers[i] for i in np.argsort(-r)])
# P1: influence 0.4066 raw citations 4
# P2: influence 0.2198 raw citations 3
# P3: influence 0.1542 raw citations 2
# P4: influence 0.1202 raw citations 1
# P5: influence 0.0991 raw citations 0
# influence ranking: ['P1', 'P2', 'P3', 'P4', 'P5']
Here the eigenvector ranking and the raw citation ranking agree — $P_1 > P_2 > P_3 > P_4 > P_5$ either way — because this network is so cleanly hierarchical that "most cited" and "cited by the most important" point the same direction. (Notice, incidentally, that this is numerically the same matrix and the same rank vector as the chapter's five-page web in §29.6: a citation network and a web are the same kind of object, and we deliberately reused the structure to make the point.) The agreement is reassuring but also a little misleading, because it hides the cases where the two methods diverge — and those divergences are exactly where eigenvector influence earns its keep.
Step 3: where counting and influence disagree
To see the methods part ways, imagine a different network. Paper $X$ is cited fifty times, but all fifty citations come from obscure papers that are themselves never cited — a citation farm, or simply a flurry of forgettable follow-ups in a crowded subfield. Paper $Y$ is cited only five times, but those five citations come from genuine landmark papers, each of which is cited thousands of times. Raw counting declares $X$ the winner, fifty to five. Eigenvector influence declares $Y$ the winner, because $Y$'s five endorsers each carry enormous rank, while $X$'s fifty endorsers carry almost none (they are cited by no one, so they hold little rank to pass along). This is precisely the spam-resistance of §29.1.1 transplanted into science: you cannot inflate your eigenvector influence by collecting citations from unimportant papers, because the recursion discounts them to near-nothing. Influence is not how many cite you; it is how much rank your citers had to give.
This is the substantive reason metrics like Eigenfactor were developed to supplement raw citation counts and the journal impact factor. They reward being cited by important venues and papers, they naturally down-weight the citation-stuffing review articles (a paper that cites five hundred others passes only $1/500$ of its rank to each, so its individual votes are cheap), and they allow more honest comparison across fields of different sizes. [verify] The construction is the chapter's, line for line: a directed graph of citations, a column-stochastic matrix with the dangling foundational papers patched, a damping factor for uniqueness and convergence, and the dominant eigenvector as the answer, computed by power iteration over a graph far too large for direct eigen-decomposition.
Why the eigenvector is the right tool, not a gimmick
It would be fair to ask whether dressing citation-counting in eigenvector clothing is real progress or mathematical theater. The answer is that the recursion captures something counting genuinely cannot: the transitive nature of intellectual influence. A foundational paper's importance is not just that many cite it, but that the papers citing it are themselves important, and the papers citing those are important, and so on — influence propagates through the network, and only a fixed-point computation can account for the full cascade. Truncating the recursion at one level (counting direct citations) throws away all of that structure. The eigenvector sums the entire infinite cascade exactly, in one self-consistent vector, which is why $\mathbf{r} = M\mathbf{r}$ is the right equation and not merely a fancier tally.
The same caveats from the sports case study apply, and bibliometricians take them seriously. Any single metric encodes a definition of "important" that is partly a values choice; citation practices differ across disciplines in ways no normalization fully erases; and reducing a body of scholarship to one number invites exactly the gaming and misuse that periodically scandalizes academia. What the linear algebra contributes is a transparent, manipulation-resistant core — recursive influence solved as a dominant eigenvector — that is demonstrably harder to fake than raw counts. It is the identical dominant eigenvector that ranks the web and ranks sports teams, applied to the citation graph, and it closes the chapter's argument: across web search, athletics, and the literature of science, importance in a network of endorsements is the eigenvalue-$1$ eigenvector, and power iteration is how you find it.