Louvre: Cultural Heritage and Scientific Analysis

• Louvre is a historic museum and research infrastructure that preserves artefacts such as Elamite tablets and Islamic ceramics while supporting multidisciplinary inquiry.
• Research at the Louvre integrates traditional analysis with modern techniques, exemplified by studies on ancient mathematics and computational visitor behavior using sensor and random-walk methods.
• Advanced analytical tools like PIXE, RBS, and deep learning power non-destructive evaluation of artworks and inspire development of museum-oriented AI systems.

Searching arXiv for papers on the Louvre and related museum research. The Louvre is a museum and cultural landmark that appears in recent research as a preserved artefact repository, a site of scientific analysis, a monitored visitor environment, and a recurring reference point in computational models of cultural heritage and urban visual narratives. Within the arXiv literature, the Louvre is associated with Old Elamite mathematical tablets from Susa, large-scale studies of visitor movement and congestion, non-destructive materials analysis of Islamic ceramics, and museum-oriented AI systems for question answering and visual reasoning (Heydari et al., 2022, Yoshimura et al., 2016, Chabanne et al., 2011, Bongini et al., 2020).

1. Institutional role as collection, archive, and research infrastructure

Several studies treat the Louvre not merely as an exhibition venue but as a research infrastructure. Susa Mathematical Text No. 2 and SMT No. 26 are part of a set of 26 mathematical clay tablets excavated from Susa in 1933 by French archaeologists and now held in the Louvre Museum; the museum also provides photographs used in scholarly analysis (Heydari et al., 2022, Heydari et al., 2022). In the materials-science literature, research on Islamic lustre ceramics was enabled by access to more than 110 preserved objects or fragments supplied by French national museums, notably the Louvre’s Department of Islamic Art, and by ion-beam analyses performed at the AGLAE accelerator at C2RMF in the Louvre’s Palais (Chabanne et al., 2011).

This combination of preservation, cataloguing, imaging, and analytical access is central to the Louvre’s scholarly function. The mathematical papers depend on museum-held tablets and updated photographs; the ceramics paper depends on museum-held reference objects and non-destructive instrumentation. This suggests an institutional role that joins curatorship with laboratory-based research rather than separating art-historical and scientific inquiry.

2. Elamite mathematics in the Louvre’s Susa tablets

Two Louvre tablets are especially important for the history of ancient mathematics. The first, Susa Mathematical Text No. 2, dates from between 1894 and 1595 BC and contains on its reverse a regular heptagon with numerical inscriptions and an explicit area instruction: “A (regular) heptagon. You multiply (the square of a side) by 4 and you subtract one-twelfth (of the result from the result itself), and (you see) the area” (Heydari et al., 2022). In modern notation, for side length $a$,

$$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$

The paper compares this Elamite approximation with the exact area and with Babylonian and Heronian formulas:

Tradition	Area formula for a regular heptagon with side $a$	Error vs. exact
Exact	$$\frac{7}{4}\cot\left(\frac{180^\circ}{7}\right)a^2 \approx (3;38,2,5,5)\times a^2$$	0
Elamite	$\frac{11}{3}a^2 = (3;40)\times a^2$	0.9%
Babylonian	$(3;41)\times a^2$	1.36%
Heron	$\frac{43}{12}a^2 = (3;35)\times a^2$	1.39%

The geometric explanation proposed in the paper is a “cut and paste” construction based on an area $4a^2$ partitioned into 12 equal parts and reduced by one part; the residual error is less than 1% (Heydari et al., 2022). The historical significance is twofold. First, the formula is more accurate than other contemporaneous Babylonian formulas and also more accurate than Heron’s much later expression. Second, the stepwise inscription is used to argue for genuine Elamite innovation, while still acknowledging broader debate about transmission and independence.

A second Louvre tablet, SMT No. 26, is studied as an Elamite problem on the bisection of trapezoids by a transversal strip rather than by a single transversal line (Heydari et al., 2022). The problem imagines a trapezoidal plot divided equally between two brothers, with the partition taking the form of a finite-thickness “party wall.” For trapezoid bases $a>b$ and height $h$, with the height divided into $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$0 equal segments, the $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$1-th transversal line has length

$$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$2

and the area to the left of the $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$3-th strip is

$$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$4

The condition for a bisecting strip yields a quadratic equation in $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$5, and the tablet’s scribe chooses $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$6 and $$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$7, values likely selected because they produce an exact integer solution in the sexagesimal system (Heydari et al., 2022). The contrast with standard Babylonian practice is explicit: the Babylonian line-bisector formula

$$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$8

is independent of height, whereas the Elamite strip construction requires a more elaborate discrete and quadratic analysis. The Louvre tablets therefore document a mathematically sophisticated corpus distinct from the better-known Babylonian material.

3. Conservation science and the Louvre’s Islamic ceramics

The Louvre’s collections also underpin materials-science research on Islamic lustre ceramics from the 9th century to the Renaissance. The study on metallic lustre decoration analyzes more than 110 objects or fragments from the Louvre, the Musée National du Moyen-Âge, the Musée National de la Céramique de Sèvres, and private lenders, with the aim of characterizing the composition and structure of ceramic bodies, glazes, and lustre layers through non-destructive or minimally invasive methods (Chabanne et al., 2011).

The principal analytical techniques are PIXE and RBS. PIXE uses a 3 MeV proton beam and the GUPIX software package to determine major, minor, and trace elements. RBS uses a 3 MeV alpha particle beam and SIMNRA fitting to reconstruct in-depth composition profiles, including the thickness and metal content of the lustre surface layers (Chabanne et al., 2011). Complementary methods include SEM, SEM-EDS, grazing-incidence X-ray diffraction, AFM, TEM, optical microscopy, and X-ray diffraction. For nanoparticle size estimation, the study cites the Scherrer formula

$$S^E_{\Gamma_7(a)} = 4a^2 - \frac{1}{12}(4a^2) = \frac{11}{3}a^2 = (3;40)\times a^2.$$9

The findings describe a long technological evolution rather than a single stable recipe. Two main substrata are distinguished: marly clays and highly siliceous pastes. Two glaze classes are also identified: alkaline glazes and lead glazes, with the latter often opacified with $a$0 (Chabanne et al., 2011). Early Abbasid wares exhibit multilayered lustre structures with a top glaze region free of copper and silver particles, a gradient zone, a principal lustre layer with maximum $a$1 content, and lower transition layers. Later productions vary in the presence of a particle-free surface layer, the total lustre thickness, the relative proportions of copper and silver, and the distribution of those metals through depth.

The Louvre specimens are important in several specific respects. Abbasid samples from Susa and Samara serve as technological reference points for early lustre; Fatimid examples document adaptation to lead-bearing glazes; late Hispano-Moresque pieces show noticeable silver even in copper-colored lustres; and Safavid examples are described as uniquely copper-based lustres over transparent tin-free glazes (Chabanne et al., 2011). One corrected assumption is therefore explicit: later red Spanish lustres were not simply silver-free. The broader significance is methodological as much as historical, since the study presents a reference base for attribution, authentication, and conservation grounded in museum-held material.

4. Visitor mobility, congestion, and length of stay

The Louvre is also a major empirical site for computational visitor studies. One line of work uses anonymized Bluetooth monitoring in the Denon wing over five months in 2010, with eight sensors placed at Entrance Hall (E), Gallery Daru (D), Venus de Milo (V), Salle des Caryatides (C), Great Gallery (B), Winged Victory of Samothrace (S), Salle des Verres (G), and Sphinx (P). After cleaning, 81,498 unique devices were analyzed, and comparison with ticket sales indicated that on average 8.2% of visitors had Bluetooth enabled (Yoshimura et al., 2016). Length of stay at a node was defined as

$a$2

A key result is that the number of unique nodes visited has almost no correlation with total length of stay: Spearman’s $a$3 with $a$4 (Yoshimura et al., 2016). Longer visits therefore do not imply proportionally broader spatial coverage. Earlier entry is associated with longer stays, and the relation between occupancy density and stay duration is nonlinear: at low to moderate occupancy, average stay time increases; then it plateaus; at high occupancy, average stay time drops sharply (Yoshimura et al., 2016). Median stay times are longest at Entrance Hall and Winged Victory, while exhibit-specific medians are longer at Venus de Milo, Salle des Caryatides, and Great Gallery than at Gallery Daru, Sphinx, and Salle des Verres.

A second study, using seven Bluetooth sensors over 24 days and 24,452 unique visitors, reaches a closely related conclusion: short- and long-stay visitors tend to visit a similar number of key locations, and Spearman’s correlation between unique nodes and stay duration is 0.072 with $a$5 (Yoshimura et al., 2016). Short stay visitors average about 4.3 nodes, whereas long stay visitors rise only to about 5.5 nodes. The sequence $a$6 is reported as the most common longer path for both groups, 15.2% of visitors visit only one sensor node, and node S is visited by 97% of visitors (Yoshimura et al., 2016). Node G remains comparatively marginal, never exceeding 40% frequency even among longer-stay visitors.

The random-walk analysis of the same 24-day, 24,452-trajectory dataset formalizes selectivity by comparing observed path frequencies with path frequencies under a graph-constrained random walk (Yoshimura et al., 2018). With

$a$7

short-stay visitors exhibit stronger behavioral regularities than long-stay visitors. For example, path $a$8 reaches $a$9 for short stays (Yoshimura et al., 2018). A common misconception is that longer-stay visitors necessarily diversify radically; the Bluetooth and random-walk studies instead show that they often spend more time on broadly similar iconic routes.

5. Computational mediation, visual reasoning, and narrative representation

Museum-oriented AI research provides a further perspective on the Louvre as an environment of image-based interaction. A visual question answering framework for cultural heritage proposes a modular system in which a BERT-based question classifier routes queries either to a visual QA module or to a contextual QA module using external knowledge sources such as museum descriptions or Wikipedia (Bongini et al., 2020). The visual module follows “Bottom-up and Top-down Attention,” using Faster R-CNN features from salient regions, GLoVe embeddings, a GRU question encoder, attention-based multimodal fusion, and final prediction layers. On Artpedia, the question classifier reaches 93.8% accuracy; the visual QA module reaches 52.4% accuracy on visual questions; the contextual QA module reaches 68.4% accuracy and 83.2% F1 on contextual questions; and the full pipeline reaches 57.0% overall accuracy on mixed input (Bongini et al., 2020).

The paper is framed around museum use cases in which visitors photograph works and ask natural-language questions. This suggests a direct applicability to heavily photographed environments associated with the Louvre, where image capture is already a dominant interaction mode. The substantive point is not only automation of labels but adaptive explanation: the system distinguishes questions answerable from image content alone from questions that require structured external knowledge.

A different computational perspective appears in work on visual storyline learning from Flickr albums. For the concept “Paris,” a Skipping Recurrent Neural Network learns storylines in which the Eiffel Tower is prominent early, common landmarks such as the Arc de Triomphe follow, and the Louvre appears late in the sequence (Sigurdsson et al., 2016). The methodological reason is explicit: classic RNNs and LSTMs are dominated by short-term repetition, whereas S-RNN skips through albums to recover longer-term ordered subsets. The Louvre is therefore modeled not simply as a cluster of visually similar interior images but as a late-stage node in a recurrent Paris travel narrative (Sigurdsson et al., 2016). In this literature, the Louvre functions simultaneously as a physical destination and as a structurally identifiable event in learned temporal representations.

6. Homonymous technical usages of “Louvre”

In contemporary arXiv literature, “Louvre” also appears as the name of technical methods unrelated to the museum. “Louvre: Lightweight Ordering Using Versioning for Release Consistency” introduces a microarchitectural mechanism that assigns versions to memory instructions in load/store queues and the write buffer; it reports a 39.6% reduction in ordering instruction latency and an 11% average program-performance improvement over baseline (Kumar et al., 2017). “Louvre: Relaxing Hardware Requirements of Quantum LDPC Codes by Routing with Expanded Quantum Instruction Set” defines a routing framework for generalized bicycle and bivariate bicycle codes; Louvre-7 preserves syndrome-extraction depth while reducing average degree to 4.5 in BB codes, and Louvre-8 further reduces degree to 4 with a slight depth increase (Zhou et al., 28 Aug 2025). “LouvreSAE: Sparse Autoencoders for Interpretable and Controllable Style Transfer” uses an art-specific sparse autoencoder on latent embeddings to build style profiles and reports performance on ArtBench10 while being 1.7–20x faster than prior methods in style evaluations (Panda et al., 22 Dec 2025).

These homonymous usages matter primarily as a terminological caution. In research databases, “Louvre” may designate the museum, artefacts held by the museum, or unrelated computational frameworks. Disambiguation therefore requires attention to disciplinary context, especially in cross-domain search and citation practice.