Arce: Stylometry, Game Theory, and Radar Estimation

• Arce is a term used in diverse domains, referring to a historical candidate in stylometric authorship, a contributor in counter-terrorism game theory, and a radar 3D localization estimator.
• In stylometry, advanced machine learning techniques and Delta dendrograms highlight Juan Arce de Otálora as the closest match for the Lazarillo, though open-set verification leaves attribution uncertain.
• In radar signal processing, ARCE employs a constrained least-squares approach with angular beam constraints to enhance 3D localization in multi-platform networks.

Searching arXiv for the provided topics and papers to ground the article in current metadata and confirm the relevant senses of “Arce.”
“Arce” denotes distinct entities in contemporary research literatures rather than a single concept. In Hispanic stylometry, it primarily refers to Juan Arce de Otálora, who emerges as the strongest candidate in a machine-learning authorship study of Lazarillo de Tormes, although that study explicitly concludes that no certain attribution can be made with the given corpus [1611.05360]. In game-theoretic work on counter-terrorism, Arce M. denotes the co-author of a benchmark non-cooperative model later reanalyzed under cooperative game theory [2101.01079]. In radar signal processing, ARCE is the angular and range constrained estimator, a constrained least-squares method for 3D localization in multi-platform radar networks [2308.06972].

1. Referential scope

The term spans at least three technically unrelated usages.

Referent	Domain	Role
Juan Arce de Otálora	Stylometric authorship attribution	Strongest candidate author for Lazarillo de Tormes
Arce M.	Counter-terrorism game theory	Co-author of the benchmark non-cooperative payoff model
ARCE	Multi-platform radar networks	Angular and range constrained estimator for 3D localization

The first usage is biographical and philological; the second is bibliographic and game-theoretic; the third is an acronym for an estimation algorithm. The literature therefore treats “Arce” as a surname in some contexts and as a technical abbreviation in others [1611.05360] [2101.01079] [2308.06972].

2. Juan Arce de Otálora in the authorship problem of Lazarillo de Tormes

In “The Life of Lazarillo de Tormes and of His Machine Learning Adversities” [1611.05360], the authorship of Lazarillo de Tormes is treated as a machine-learning / stylometric attribution problem, not as a traditional literary-historical argument. The study builds its own corpus because existing digital corpora were insufficient. That corpus spans 1499–1589, includes the main candidate authors and additional “impostors” or control authors, and represents the Lazarillo in multiple forms: a version with interpolations, one without interpolations, and the second part (1555). The candidate set used in the main experiments includes, among others, Juan Arce de Otálora, Alfonso de Valdés, Diego Hurtado de Mendoza, Juan Luis Vives, Lope de Rueda, Francisco Cervantes de Salazar, Sebastián de Horozco, Pedro Mejía, Juan de Valdés, and Cristóbal de Villalón.

The preprocessing pipeline is unusually detailed. It normalizes old Spanish spelling and removes marginalia, footnotes, headers, footers, page numbers, Latin and Greek quotations, split words, duplicate punctuation, numbers, chapter markers, and speaker names in plays; it also expands abbreviations. The project further built a crowdsourced OCR-review tool, Festos, on top of Tesseract and DocumentCloud to clean manuscript-based texts.

Methodologically, the study surveys and tests several families of methods: distance-based clustering and stylometry, dimensionality reduction, and supervised classification. The first phase uses Burrows-style Delta and its variants, with Delta defined as the mean absolute difference between z-scores for a set of word variables in a text-group and a target text, with smaller Delta indicating greater similarity. The paper also tests Eder’s Delta, which it notes performs best in its setup. It then applies normalized compression distance (NCD),
$$
NCD(x, y) = \frac{C(x+y) - \min(C(x),C(y))}{\max(C(x),C(y))},
$$
where (C(x)) is the compressed size of text (x), and (x+y) is concatenation. Dimensionality reduction and visualization are carried out with PCA and LDA. The predictive stage includes Ridge classification, Bernoulli and Multinomial Naive Bayes, nearest centroid, and later linear and nonlinear SVMs, MaxEnt, random forests, SGD, and bagging. The final verification stage uses Koppel and Schler’s unmasking method, treating authorship as an open-set verification problem rather than a closed-set classification problem.

Across these procedures, Arce de Otálora repeatedly appears near the top. In the Delta dendrograms, the Lazarillo is placed among a group that includes Juan Arce de Otálora and Alfonso de Valdés; the authors state that “among the ones more closely related to the author of the Lazarillo, with or without interpolaciones, we find Juan Arce de Otálora and Alfonso de Valdés.” In the PCA of stopwords, Arce is one of the authors whose chunks show the most similar use of function words to the Lazarillo. In the LDA analysis, the authors state that the only viable candidacies might be Pedro Mejía, Alfonso de Valdés, Juan Arce de Otálora, and to a lesser extent Juan Luis Vives and Cristóbal de Villalón. The study’s summary judgment is correspondingly explicit: “the most likely author seems to be Juan Arce de Otálora, closely followed by Alfonso de Valdés” [1611.05360].

3. Evidential status and limits of the Lazarillo attribution

The strongest support for Arce de Otálora comes from stylometric distances and supervised prediction. In the results table for supervised classification, Ridge with the “total” feature set achieves the best scores overall: precision 0.9718, recall 0.9696, F-score 0.9701. When those models are used to assign Lazarillo chunks, Juan Arce de Otálora is the winner in both the max-wins and average-chunk strategies. Table 6 reports Arce with the highest chunk average in that phase, 42.50, far above Alfonso de Valdés (8.40) and others. In the later supervised setup, Table 8 again places Arce first: he receives the most chunks in the max-wins strategy and the highest average assignment, with an average of 36.82 chunks, followed by Pedro Mejía (11.00), Alfonso de Valdés (7.64), Gaspar Gil Polo (5.09), Juan Luis Vives (3.82), Cristóbal de Villalón (1.18), and Lope de Rueda (1.00). The authors emphasize that in the best-performing models, “almost 37 out of the 73 chunks of the Lazarillo are always assigned to Otálora regardless of the method” and that in some models more than 86% of the chunks go to Arce.

The open-set verification stage tempers that result rather than overturning it. In the reduced candidate pool, Figure 11 is described as showing that the degradation curve for the Lazarillo matches Arce’s profile best, followed by Alfonso de Valdés. Yet the unmasking classifier ultimately labels the Lazarillo as “different-author” for all candidates. The paper states that the method “did not assign a clear winner.” Arce therefore emerges as the closest stylistic match, not as a proven author.

This distinction is central to the paper’s interpretation. The authors repeatedly warn that the case is an open-set authorship problem: the true author may not be in the candidate corpus at all. They also stress that the corpus is incomplete and imbalanced, that class imbalance can distort distance and classification results, and that some methods are sensitive to text length and to the imbalance in the number of chunks per author. They add that NCD groupings may partly reflect length rather than style. The conclusion accordingly remains deliberately limited: the statistical evidence “seem[s] to point out in the direction of Arce de Otálora by a wider margin with regards to Valdés,” but no certain attribution can be made with the given corpus [1611.05360].

4. Arce and Sandler in counter-terrorism game theory

In “Counter-terrorism analysis using cooperative game theory” [2101.01079], Sung Chan Choi explicitly builds on Arce M. and Sandler (2005). In that benchmark formulation, counter-terrorism is modeled as a non-cooperative game: communication among players is not allowed, or, if it is allowed, there is no mechanism to enforce any agreement they may make. The policy set consists of Preemption, Deterrence, and Status quo / no action. Choi summarizes the same incentive structure used by Arce and Sandler: Preemption generates a public benefit for both players but is privately costly to the preemptor, whereas Deterrence can impose a public cost on the other player by deflecting attacks while giving private benefit to the deterrer. This induces a prisoner’s dilemma structure.

The reproduced symmetric (3\times 3) payoff matrix is
$$
\begin{array}{c|ccc}
& \text{Preempt} & \text{Status Quo} & \text{Deter} \
\hline
\text{Preempt} & (2,2) & (-2,4) & (-6,6) \
\text{Status Quo} & (4,-2) & (0,0) & (-4,2) \
\text{Deter} & (6,-6) & (2,-4) & (-2,-2)
\end{array}
$$
with public benefit from preemption (=4), private cost of preemption (=6), public cost of deterrence (=4), and private benefit from deterrence (=6).

The key non-cooperative conclusion is that ((\text{Deter}, \text{Deter})) is a pure Nash equilibrium because deterrence is dominant, even though ((\text{Preempt}, \text{Preempt})) gives both players higher payoffs. Choi’s contribution is not to deny this model but to reinterpret the same strategic environment cooperatively [2101.01079].

5. Cooperative reformulation of the Arce–Sandler model

Choi studies three cooperative frameworks: TU solution, NTU Nash bargaining solution, and NTU lambda-transfer approach. In the TU setting, for payoff bimatrix ((A,B)), the players first maximize total surplus,
$$
o = \max_{i,j}(a_{ij}+b_{ij}),
$$
and then split the surplus subject to disagreement payoffs
$$
D_1 = pAq,\qquad D_2 = pBq.
$$
The TU solution is
$$
\left(\frac{o + D_1 - D_2}{2},\ \frac{o - (D_1 - D_2)}{2}\right),
$$
or, using (\delta=\operatorname{Val}(A-B)),
$$
\phi^* = \left(\frac{o+\delta}{2},\frac{o-\delta}{2}\right).
$$
For the Arce–Sandler matrix, the difference matrix (A-B) has a saddle point, the threat strategies are (p^*=(0,0,1)^T), (q^*=(0,0,1)^T), the disagreement point is ((D_1,D_2)=(-2,-2)), the total surplus is (o=4), and therefore
$$
\phi^* = (2,2).
$$
This corresponds to ((\text{Preempt}, \text{Preempt})) and requires no side payment.

Choi then treats the same game as an NTU game. The Nash bargaining solution maximizes
$$
f(u,v) = (u-u^)(v-v^),
$$
and the paper includes a uniqueness theorem for that maximizer. Using the disagreement point ((u^,v^)=(-2,-2)), Choi finds that the maximum again occurs at
$$
(\bar u,\bar v)=(2,2).
$$
The lambda-transfer analysis introduces a conversion factor (\lambda>0), transforming the game into ((\lambda A,B)) with solution
$$
\phi(\lambda)=\left(\frac{o(\lambda)+\delta(\lambda)}{2\lambda},\ \frac{o(\lambda)-\delta(\lambda)}{2}\right),
$$
where
$$
o(\lambda)=\max_{i,j}(\lambda a_{ij}+b_{ij}),\qquad \delta(\lambda)=\operatorname{Val}(\lambda A-B).
$$
For the example, the unique relevant conversion factor is
$$
\lambda^* = 1,
$$
and this again yields
$$
\phi(\lambda^*) = (2,2).
$$

The same conclusion persists in the generalized model. Choi parameterizes the Arce–Sandler structure with (B) for the public benefit from preemption, (c) for the private cost of preemption, (b) for the private benefit of deterrence, and (C) for the public cost of deterrence, under the assumptions
$$
B < c < 2B,\qquad C < b < 2C,
$$
and then imposes
$$
B=C,\qquad c=\alpha B,\qquad b=\beta B,
$$
with
$$
1<\alpha,\qquad 1<\beta<2.
$$
In normalized notation, the TU solution is
$$
(2-a,\,2-a),
$$
which translates back to
$$
(2B-c,\ 2B-c).
$$
The generalized NTU and lambda-transfer analyses again select the same symmetric preemption-preemption outcome. The main contrast is therefore sharp: Arce and Sandler (2005) yield ((\text{Deter}, \text{Deter})) under non-cooperative equilibrium logic, whereas Choi’s cooperative reformulation yields ((\text{Preempt}, \text{Preempt})) under all three cooperative solution concepts [2101.01079].

6. ARCE in multi-platform radar networks

In “3D Localization and Tracking Methods for Multi-Platform Radar Networks” [2308.06972], ARCE denotes the angular and range constrained estimator. The setting is a multi-platform radar network (MPRN) with one transmitter—with one receiver co-located with it, forming a monostatic active radar—and multiple spatially separated receivers. For a target at 3D position (\mathbf{x}_k), the bistatic range measurement at receiver (i) is
$$
\rho_k^{(i)} = | \mathbf{x}_k - \mathbf{t}_k | + | \mathbf{x}_k - \mathbf{r}_k^{(i)} | + w_k^{(i)}, \qquad i=1,\ldots,S,
$$
where (\mathbf{t}_k) is the transmitter position, (\mathbf{r}_k^{(i)}) is the position of receiver (i), and (w_k^{(i)}) is zero-mean measurement noise. Geometrically, each bistatic range defines an ellipsoid in 3D. With noise, the ellipsoids no longer meet at a single point, so direct intersection is not robust.

ARCE improves range-only localization by exploiting a priori knowledge of the transmitter antenna beamwidth. The key idea is that the target must lie inside the radar beam, which creates an angular constraint that can be combined with the range equations. The estimator is therefore a constrained LS estimator rather than a generic unconstrained localizer. The paper describes ARCE as solving a constrained least-squares problem of the form
$$
\hat{\mathbf{x}}k^{\text{ARCE}} = \arg\min{\mathbf{x}\in\mathcal{F}} \sum_{i=1}^{S} \left( \rho_k^{(i)} - |\mathbf{x}-\mathbf{t}_k| - |\mathbf{x}-\mathbf{r}_k^{(i)}| \right)^2,
$$
where (\mathcal{F}) is the feasible region defined by the antenna beam constraints. Using Karush–Kuhn–Tucker (KKT) conditions, the feasible domain is partitioned into six subsets; candidate solutions are derived for each subset; at most 26 candidates are considered; and the candidate with smallest cost is selected. The result is described as quasi-closed form.

Within the tracking architecture, the paper distinguishes NAD (non-adaptive) and AD (adaptive) angular constraints. NAD uses the physical transmitter beam directly. AD uses a virtual beam built by intersecting the physical beam with a target-specific beam centered on the predicted target direction, with widths
$$
d_{k,\ell}^{\mathrm{az}} = 2 \tilde{C}\,\sigma_{k,\ell}^{\mathrm{az}}, \qquad
d_{k,\ell}^{\mathrm{el}} = 2 \tilde{C}\,\sigma_{k,\ell}^{\mathrm{el}}.
$$

The paper’s second major contribution is combining ARCE with the particle-based scalable SPA-based MTT method. Each target has state
$$
\mathbf{s}{k,\ell} =
\begin{bmatrix}
\mathbf{x}{k,\ell}^T & \mathbf{v}{k,\ell}^T
\end{bmatrix}^T,
$$
and the tracker maintains beliefs (\tilde{f}(\mathbf{s}{k,\ell}, r_{k,\ell})) for potential targets, where (r_{k,\ell}\in{0,1}) is a Bernoulli existence variable. The combined algorithm proceeds through prediction, initialization, SPA data association, ARCE localization, ARCE-driven resampling, and update. In the resampling step, the least significant fraction (1-\alpha_r) of particles is replaced with particles drawn from a Gaussian centered at the ARCE estimate,
$$
\check{\mathbf{x}}{k|k-1,\ell}^{(p)} \sim \mathcal{N}\left(\mathbf{x}{k,\ell}^{\text{ARCE}}, \sigma_{\text{ARCE}}^2 I\right),
$$
while keeping the velocity component unchanged.

The reported simulations consider a 3D scenario with one monostatic active radar at the origin, four additional receivers, the active antenna pointing along the X-axis, half-beamwidth of (20^\circ), two moving targets, 100 scans and 10 s scan time, and SNRs of (0), (-10), and (-20) dB, under both ideal and non-ideal detection/clutter settings. In the ideal case, stand-alone ARCE is often worse than baseline SPA-MTT at low SNR because ARCE is memoryless and uses only noisy single-snapshot data; however, the combined ARCE+SPA methods outperform baseline SPA-MTT, especially for the target near the antenna beam edge. At 0 dB SNR, ARCE performs better than baseline SPA-MTT during the early part of the run, especially during initialization, while the combined method remains best overall. The paper further reports that when the number of particles is small, for example 500 or 1000, ARCE provides strong gains; as the particle count increases to around 2500, baseline SPA-MTT and ARCE-enhanced SPA become similar. In the non-ideal scenario with missed detections and clutter, the proposed method still outperforms baseline SPA-MTT, especially at (0) dB and (-10) dB [2308.06972].