Mapping Firms' Locations in Technological Space: A Topological Analysis of Patent Statistics

Emerson G. Escolar; Mitsuru Igami; Yasin Ozcan; Yasuaki Hiraoka

REVIEW 3 major objections 4 minor 5 references

Mapping Firms' Locations in Technological Space: A Topological Analysis of Patent Statistics

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Firms whose patent portfolios branch off into unique trajectories tend to post higher revenue, EBIT, and market value: one extra unit of flare length is associated with about 40%, 39%, and 31% higher performance, respectively, even after…

desk verdict The first economics application of TDA/Mapper is a genuinely useful measurement paper, but the performance regression is over-sold and the unexamined single-linkage clustering choice leaves the central correlation less secure than it looks. read the letter →

arxiv 1909.00257 v7 pith:7HOWN3W5 submitted 2019-08-31 econ.EM cs.DMmath.ATmath.CO

Emerson G. Escolar , Yasuaki Hiraoka , Mitsuru Igami , Yasin Ozcan This is my paper · ORCID

classification econ.EMcs.DMmath.ATmath.CO

keywords topologicaldataanalysisMapperflarelengthpatentstatisticstechnologicalspaceinnovationfirmperformanceproductdifferentiation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that the direction of a firm's inventive activity can be described by the shape of its patent portfolio as it moves through a 430-dimensional technology space. It builds a graph in which each node is a cluster of firm-years, using an algorithm that preserves global structure through local clustering, and defines a firm's "flare length" as the longest finite branch its nodes form before rejoining the main map. The central empirical claim is that flare length is positively associated with 2005 revenue, EBIT, and market value, with one extra unit of flare length corresponding to roughly 40% higher revenue, 39% higher EBIT, and 31% higher market value even after controlling for total patent count. The association survives sector controls, balanced-panel checks, and alternative distance metrics, which matters because it suggests the shape of a portfolio carries information about firm outcomes that raw portfolio size does not.

What carries the argument

The machinery is the Mapper shape graph together with the graph-theoretic flare measure. Mapper projects the 430-dimensional point cloud of logged patent counts by firm-year onto a two-dimensional PCA filter, covers the image with a 20-by-20 grid of overlapping squares, clusters the points in each square using single linkage with cosine distance, and links clusters that share points. The resulting graph's nodes contain firm-years; for each firm, one takes the subgraph of nodes containing that firm, marks interior nodes whose neighbors all remain in the subgraph, and computes each interior node's exit distance to the outside. A connected component of the interior is a flare if it connects to the rest of the graph (an island if it does not), and the firm's flare length is the largest finite exit distance across its flares. The regressions use this single number as the key regressor.

What would settle it

Run the same construction with a different filter function (say, a single principal axis or a random projection) or a systematically different cover, and re-estimate column 3 of Table 3; if the 0.34 coefficient on flare length loses significance or changes sign under a small parameter perturbation, the flare-performance link is a tuning artifact. Alternatively, permute patent-class labels across firm-years and recompute flare lengths: if the permuted data still yield positive performance correlations, the measure is not capturing genuine technological uniqueness.

Watch

Extended reading notes

Core claim

The paper's central claim is that a firm's technological uniqueness, measured graph-theoretically as flare length in a Mapper shape graph, is positively and significantly associated with firm performance. In the authors' own terms, the regressions in Table 3 report that an extra unit of flare length is associated with 40%, 39%, and 31% higher revenue, EBIT, and market value, respectively, after controlling for log patent count. The association is not an artifact of industry composition or survivorship: it persists within sectors and industries and in balanced-panel subsamples, and F-tests reject the joint insignificance of the flare and island indicators for all three performance measures. The paper presents this as a predictive relationship, not a causal one, and interprets it through the industrial-organization idea of product differentiation: firms whose patent portfolios move in directions no one else follows tend to be the firms with stronger market outcomes.

Load-bearing premise

The load-bearing premise is that the shape graph is built with the right choices—the PCA filter, 20-by-20 cover with 50% overlap, single-linkage clustering heuristic, and cosine distance—so that the resulting flare lengths are a faithful signature of uniqueness rather than an artifact of those settings.

Editorial extensions

If this is right

A one-unit longer flare is associated with roughly 40% higher revenue, 39% higher EBIT, and 31% higher market value in the baseline specification that controls for log patent count.
The flare and island indicators add statistically significant predictive power beyond patent count: the F-test restricting both coefficients to zero is rejected at the 1% level or better for all three performance measures.
Positive flare-performance associations persist within sectors and industries and in balanced-panel subsamples, so the result is not solely an industry-composition or survivorship artifact.
Alternative distance metrics (Euclidean, correlation, and min-complement) yield similar regression coefficients, so the association is not tied to the cosine-distance choice alone.
Because flare length is relative to the whole graph, a firm's measured uniqueness depends on the trajectories of all other firms in the sample, matching the idea of product differentiation in industrial organization.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The relative nature of flare length suggests a natural extension: use a firm's flare length as an outcome in studies of mergers, entry, or R&D competition, since it shifts when rivals change their patenting.
The same Mapper-plus-flare pipeline could be exported to other high-dimensional economic objects, such as product characteristics or trade baskets, to quantify uniqueness wherever a point cloud has meaningful continuity.
A testable extension is a bootstrap or subsample analysis that resamples firms and re-estimates flare lengths, producing confidence intervals for the 0.27–0.34 coefficients and checking how much of the association is driven by the graph's dependence on the full sample.
One could test the proposed mechanism directly: if long flares proxy for sustained growth and differentiation, flare length should also predict intermediate outcomes like patent citations or R&D productivity years before financial performance is observed.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

I read this as a methods-and-stylized-facts paper, not an econometric contribution per se. The genuinely new piece is the formal graph-theoretic definition of "flare length" as a firm-specific measure of portfolio uniqueness computed from a Mapper graph, applied here to 333 firms' patent portfolios. The math is correct: the boundary lemma and the exit-distance proposition are clean, and the algorithm for computing flare length via a multi-source Dijkstra on the induced subgraph is sensible. The visualizations are compelling, and the comparison with Jaffe's k-means clustering is honest and illuminating. This is the first serious TDA/Mapper application in economics, and the authors know the computational topology literature well.

The main soft spot is the gap between what the abstract claims and what the regressions actually do. The abstract in front of me says the association holds after controlling for portfolio size, survivorship, industry classification, and firm fixed effects. In the full text, the baseline regressions in Table 3 control only for log patent count. Industry checks appear as scatter plots (Appendix Figures 10–11), survivorship checks as balanced-panel subsamples (Tables 7–8), and there are no firm fixed effects anywhere. That is not fatal to the measurement contribution, but it is an overstatement that needs fixing.

The other soft spot, which I think is more load-bearing, is the fixed choice of single-linkage clustering in the Mapper construction. The paper varies the cover resolution, overlap, and distance metric, but never the clustering algorithm or the PCA filter function. Single-linkage is prone to chaining, and chaining artifacts could plausibly inflate flare lengths for firms sitting on thin bridges in the graph. Since the performance regression is computed on those flare lengths, the correlation could partly be an artifact of the clustering choice. Without released code or data, I can't rule that out. The authors should either show robustness to other clustering methods (e.g., DBSCAN or average-linkage) or at least disclose this as a limitation.

Also minor: flare length is a discrete, heavily zero-inflated variable, and treating each integer step as a linear "extra length" with a 40% performance premium is a strong interpretation. The coefficient is what it is, but the economic significance language would deserve more caution.

Who should read this? Anyone working on measuring technological position or thinking about applying topological methods to economic data. It is a serious method paper that deserves a referee, but I would ask for code/data plus robustness on the clustering step and a revised abstract before publication.

Yes, send it to peer review.

Referee Report

3 major / 4 minor

Summary. The paper introduces a topological data analysis approach, based on the Mapper algorithm, to characterize the evolution of 333 large firms' patent portfolios across 430 USPTO classes during 1976–2005. Each firm-year is a point in a 430-dimensional space; the Mapper graph is built using PCA as a filter, a 20×20 overlapping cover, single-linkage clustering, and cosine distance. The authors then define a graph-theoretic notion of a 'flare'—a connected component of a firm's induced subgraph that is not an island—and a 'flare length' measuring how far the firm's unique trajectory extends from the graph's main body. They document that 40.3% of firms exhibit flares and that flare length is positively associated with revenue, EBIT, and market value in 2005, even after controlling for log patent count. They interpret this as evidence that technological uniqueness predicts firm performance. The paper also compares Mapper with Jaffe's (1989) global clustering, arguing that the graph reveals industry connectedness that discrete clusters obscure.

Significance. If the central empirical claim holds, the paper makes a useful methodological contribution by bringing TDA into empirical economics, providing a principled way to summarize high-dimensional patent data without collapsing it into a fixed cluster partition. The formal graph-theoretic definitions and proofs in Sections 2.4–2.5 and Appendix B are rigorous and clearly presented. The authors are careful to describe the analysis as correlational rather than causal. The comparison with Jaffe's clustering is instructive and shows complementary value. The main question is whether the flare-length measure is robust to the analyst-chosen Mapper parameters, especially the clustering algorithm, since the performance correlations in Table 3 are built on the baseline flare lengths.

major comments (3)

[Abstract and Section 5, Eq. (5), Table 3] The abstract states that the flare-performance association 'continues to hold after we control for portfolio size, firm survivorship, industry classification, and firm fixed effects.' The regression specification in Eq. (5) and Table 3 contains only log(patents) as a control; no industry classification or firm fixed effects appear in any regression table. The 'within-sector' evidence in Appendix Figures 10 and 11 is scatter plots, not regressions that control for industry, and a firm fixed effect cannot be estimated in this cross-sectional 2005-outcome design. The full-text v5 abstract, by contrast, only claims control for portfolio size and firm survivorship. The overstatement should be corrected, or the additional controls should actually be implemented and reported.
[Section 2.3 and Appendix Table 6] The sensitivity analysis varies the cover resolution n, overlap o, and distance metric, but never re-estimates the Mapper graph with an alternative clustering algorithm or an alternative filter function. This is a load-bearing gap because single-linkage clustering has a well-known chaining tendency: points connected by thin bridges can be merged into elongated clusters, and in a Mapper cover such chains can produce long graph branches. Since flare length is defined as graph distance from the boundary of the firm's subgraph, chaining artifacts could directly inflate the flare lengths that drive the Table 3 correlations. The paper should either re-estimate flares with e.g., average-linkage or DBSCAN and show that the distribution of flare lengths and the Table 3 coefficients are stable, or explicitly temper the robustness claims.
[Section 5, Table 3] The regression includes a dummy 1{k_i = ∞} for islands-only firms, but it is not stated how k_i is coded in the regression for those firms. If k_i is set to 0 for islands-only observations, the coefficient on flare length measures only finite flares; if set to some large number, the coefficient is not directly interpretable. The paper should state the coding explicitly and discuss how the islands-only coefficient is identified, especially because it changes sign from positive to negative once log(patents) is added.

minor comments (4)

[Appendix Table 6] In the 'Correl' columns, the Log(Patents) coefficient is reported as 0.28 with standard error 0.22 in the revenue regression, which is inconsistent with the 0.04 standard errors in the other distance-metric columns and is likely a typographical error.
[Section 2.3] The text says that n = 15 and 25 and o = 30% produce 'qualitatively similar Mapper graphs,' but no figure or table quantitatively compares flare lengths or downstream regressions under those parameters. A brief quantitative comparison would make this sensitivity claim verifiable.
[Section 3] The sample selection rule—firms that acquired at least four firms with patents—is noted but its implications are not discussed. The results apply to large, M&A-active firms, and the paper would benefit from an explicit statement that the findings may not generalize to smaller or non-M&A firms.
[Section 5] The regressions report ordinary standard errors; because the outcome variables are cross-sectional and likely heteroskedastic, robust standard errors should be reported to confirm the significance claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: flare length is computed from patent data alone, and the performance regressions are a standard in-sample association.

full rationale

The paper's central measure, flare length, is defined purely in graph-theoretic terms from the Mapper graph constructed on firms' patent-class vectors (Sections 2.3-2.5, Definitions 5-11). No financial performance variable enters the construction of the Mapper graph or the computation of flare length; the regressions in Section 5 (equation 5 and Table 3) then relate this precomputed graph measure to 2005 revenue, EBIT, and market value while controlling for log patent count. The Mapper tuning choices (n=20, overlap 0.5, cosine distance, single-linkage with the Singh-Memoli-Carlsson heuristic) are explicitly presented as analyst specifications, and the paper varies n, overlap, and distance metric in sensitivity analyses (Appendix Table 6). These choices are not fitted to the outcome variables and are not renamed as predictions. Self-citations in the paper (Ozcan 2015 as data source; Igami and Subrahmanyam 2019 for hard-disk-drive context; Hiraoka et al. 2016 as an illustrative TDA application) are provenance or background references and are not load-bearing for the flare-performance claim. No equation in the paper makes flare length equal to a performance variable by construction, and no fitted parameter is relabeled as a predicted outcome. The closest concern, the analyst-chosen clustering and cover parameters, is a robustness matter rather than circularity, because the parameters are not tuned to the dependent variables and alternative choices yield broadly similar results.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or dimensions are introduced. The 'flare' is a data-derived graph summary, not a postulated hidden cause. The main external assumptions are the representativeness of patent data and the stability of the Mapper parameters.

free parameters (2)

Mapper cover resolution n = 20
Number of cubes in the cover of the PCA-filtered image. Chosen for visual detail (Section 2.3). Sensitivity checks with n=15 and 25 are reported but not exhaustive.
Mapper cover overlap o = 0.5
Overlap ratio of cover intervals. Chosen to balance connectivity and computational burden. Sensitivity with o=0.3 is mentioned.

assumptions (4)

domain assumption Patent class counts are a valid representation of a firm's inventive activities.
The whole analysis treats the vector of patent counts across 430 USPTO classes as the location of the firm in technological space. Stated in Section 1 and Section 3.
ad hoc to paper The Mapper graph with the chosen filter, cover, and clustering preserves the topological features relevant to technological uniqueness.
The flare length depends on the specific Mapper construction in Section 2.3, including PCA filter, n=20, o=0.5, and single-linkage clustering. This is not an established standard for patent data.
domain assumption Cosine distance is an appropriate dissimilarity measure for patent portfolios.
Cosine distance is standard in the innovation literature (Jaffe 1989), so this is a reasonable prior, but it is still a modeling choice that affects the graph.
domain assumption The 5-year moving window and log(+1) transform provide a meaningful smoothing of annual patent counts.
Used in Section 2.3 to reduce volatility and skewness, following Benner and Waldfogel (2008).

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mapping Firms' Locations in Technological Space: A Topological Analysis of Patent Statistics." pith.science (2026). https://pith.science/paper/7HOWN3W5

@misc{pith2026190900257,
  author       = {Pith},
  title        = {Pith review of: Mapping Firms' Locations in Technological Space: A Topological Analysis of Patent Statistics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7HOWN3W5}},
  note         = {Machine review of arXiv:1909.00257}
}

read the original abstract

Where do firms innovate? Mapping their locations and directions in technological space is challenging due to its high dimensionality. We propose a new method to characterize firms' inventive activities via topological data analysis (TDA) that represents high-dimensional data in a shape graph. Applying this method to 333 major firms' patents in 1976--2005 reveals substantial heterogeneity: some firms remain undifferentiated; others develop unique portfolios. Firms with unique trajectories, which we define and measure graph-theoretically as "flares" in the Mapper graph, perform better. This association is statistically and economically significant, and continues to hold after we control for portfolio size, firm survivorship, industry classification, and firm fixed effects. By contrast, existing techniques -- such as principal component analysis (PCA) and Jaffe's (1989) clustering method -- struggle to track these firm-level dynamics.

Figures

Figures reproduced from arXiv: 1909.00257 by the authors.

**Figure 2.** Illustration of the Mapper Procedure [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗

**Figure 3.** illustrates the definitions of interior and boundary. The pink region represents firm i’s subgraph Gi , the green nodes are in the interior Fi , and the purple nodes are in the boundary Gi \ Fi [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 5.** Figure 5: Flares and Financial Performances (a) Revenue 2 4 6 8 10 12 Log(Revenue) 0 2 4 6 8 10 Flare length (b) EBIT 2 4 6 8 10 12 Log(EBIT) 0 2 4 6 8 10 Flare length (c) Market value 2 4 6 8 10 12 Log(Market value) 0 2 4 6 8 10 Flare length Note: The center of each circle repr…

**Figure 8.** Figure 8: Three-Dimensional PCA Note: Red markers are IT firms, green markers are drug makers, and blue markers are all others. A-4 [PITH_FULL_IMAGE:figures/full_fig_p035_8.png]

**Figure 10.** Figure 10: Revenues and Flares by Sector (a) Technology 2 4 6 8 10 12 Log(Revenue) 0 2 4 6 8 10 Flare length (b) Capital Goods 2 4 6 8 10 12 Log(Revenue) 0 2 4 6 8 10 Flare length (c) Health Care 2 4 6 8 10 12 Log(Revenue) 0 2 4 6 8 10 Flare length (d) Consumer Goods 2 4 6 8 10 …

**Figure 11.** Figure 11: Revenues and Flares by SIC Code (a) Computers and Peripherals Cisco Systems Inc Hewlett Packard Co EMC Corp Sun Microsystems Inc Dell Inc 3Com Corp Seagate Technology Inc Juniper Networks Inc Tandberg Data ASA Silicon Graphics International Digi International Inc 2 4 …

**Figure 12.** Figure 12: Raw Data on Selected Technology Firms (a) Hewlett Packard 0 100 200 300 400 500 600 700 800 USPTO Patent Class 1990 1995 2000 2005 Year of Application or Acquisition (b) Dell 0 100 200 300 400 500 600 700 800 USPTO Patent Class 1990 1995 2000 2005 Year of Application …

**Figure 13.** Figure 13: Mapper Graph Based on Jaffe’s Measure (Details) [PITH_FULL_IMAGE:figures/full_fig_p040_13.png]

**Figure 14.** Figure 14: Mapper Graph Based on Jaffe’s Measure and Mahalanobis D [PITH_FULL_IMAGE:figures/full_fig_p041_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

5 extracted references · 4 canonical work pages

[1]

Matching patents to Compustat ﬁrms, 1980-2015: Dynamic reassignment, name changes, and own ership structures

Arora, Ashish, Sharon Belenzon, and Lia Sheer (2019). “Matching patents to Compustat ﬁrms, 1980-2015: Dynamic reassignment, name changes, and own ership structures”. Athey, Susan and Guido W Imbens (2017). “The state of applied eco nometrics: Causality and policy evaluation”. In: Journal of Economic Perspectives 31(2), pp. 3–32. Azoulay, Pierre, Christian...

arXiv 2019
[2]

Topological persis- tence and simpliﬁcation

Elsevier, pp. 129–213. Edelsbrunner, Herbert and John Harer (2010). Computational topology: an introduction . American Mathematical Society. Edelsbrunner, Herbert, David Letscher, and Afra Zomorodian (2 000). “Topological persis- tence and simpliﬁcation”. In: Proceedings 41st Annual Symposium on Foundations of Computer Science. IEEE, pp. 454–463. Epstein,...

work page 2010
[4]

Introduction

Elsevier, pp. 1083–1127. Nelson, Richard (1962). “Introduction”. In: The Rate and Direction of Inventive Activity: Economic and Social Factors . Ed. by Committee on Economic Growth of the Social Sci- ence Research Council Universities-National Bureau Committee fo r Economic Research. Princeton University Press: Princeton, NJ. Nicolau, Monica, Arnold J Lev...

work page 1962
[551]

Towards a new approach to reveal dynamical organization of the brain using topological data analysis

Saggar, Manish et al. (2018). “Towards a new approach to reveal dynamical organization of the brain using topological data analysis”. In: Nature communications 9(1), pp. 1–14. Sampat, Bhaven and Heidi L Williams (2019). “How do patents aﬀect f ollow-on innovation? Evidence from the human genome”. In: American Economic Review 109(1), pp. 203–36. Singh, Gur...

work page 2018
[1236]

Clustering and classiﬁcation of time series using topological data analysis with applications to ﬁnance

Majumdar, Sourav and Arnab Kumar Laha (2020). “Clustering and classiﬁcation of time series using topological data analysis with applications to ﬁnance”. I n: Expert Systems with Applications, p. 113868. Moser, Petra and Shmuel San (2020). Immigration, Science, and Invention. Evidence from the Quota Acts . Tech. rep. Myers, Kyle (2020). “The Elasticity of ...

work page 2020

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.

[1] [1]

Matching patents to Compustat ﬁrms, 1980-2015: Dynamic reassignment, name changes, and own ership structures

Arora, Ashish, Sharon Belenzon, and Lia Sheer (2019). “Matching patents to Compustat ﬁrms, 1980-2015: Dynamic reassignment, name changes, and own ership structures”. Athey, Susan and Guido W Imbens (2017). “The state of applied eco nometrics: Causality and policy evaluation”. In: Journal of Economic Perspectives 31(2), pp. 3–32. Azoulay, Pierre, Christian...

arXiv 2019

[2] [2]

Topological persis- tence and simpliﬁcation

Elsevier, pp. 129–213. Edelsbrunner, Herbert and John Harer (2010). Computational topology: an introduction . American Mathematical Society. Edelsbrunner, Herbert, David Letscher, and Afra Zomorodian (2 000). “Topological persis- tence and simpliﬁcation”. In: Proceedings 41st Annual Symposium on Foundations of Computer Science. IEEE, pp. 454–463. Epstein,...

work page 2010

[3] [4]

Introduction

Elsevier, pp. 1083–1127. Nelson, Richard (1962). “Introduction”. In: The Rate and Direction of Inventive Activity: Economic and Social Factors . Ed. by Committee on Economic Growth of the Social Sci- ence Research Council Universities-National Bureau Committee fo r Economic Research. Princeton University Press: Princeton, NJ. Nicolau, Monica, Arnold J Lev...

work page 1962

[4] [551]

Towards a new approach to reveal dynamical organization of the brain using topological data analysis

Saggar, Manish et al. (2018). “Towards a new approach to reveal dynamical organization of the brain using topological data analysis”. In: Nature communications 9(1), pp. 1–14. Sampat, Bhaven and Heidi L Williams (2019). “How do patents aﬀect f ollow-on innovation? Evidence from the human genome”. In: American Economic Review 109(1), pp. 203–36. Singh, Gur...

work page 2018

[5] [1236]

Clustering and classiﬁcation of time series using topological data analysis with applications to ﬁnance

Majumdar, Sourav and Arnab Kumar Laha (2020). “Clustering and classiﬁcation of time series using topological data analysis with applications to ﬁnance”. I n: Expert Systems with Applications, p. 113868. Moser, Petra and Shmuel San (2020). Immigration, Science, and Invention. Evidence from the Quota Acts . Tech. rep. Myers, Kyle (2020). “The Elasticity of ...

work page 2020