A mathematical formulation and computational exploration of Yayoi Kusama’s tentacle artworks Authors: Daniella Iliatanov, Mary Ostrovsky, Michal Pustějovský, Shai Gul Published: 2025 ABSTRACT AbstractYayoi Kusama, renowned for her three-dimensional modern artworks, is particularly famous for her tentacle-themed pieces. Each of these works features a room with multiple surfaces, resembling tentacles, where each of these surfaces may be viewed as an incomplete monotonically decreasing sequence (where it has been cut before the convergence). While Kusama’s approach is purely artistic, it raises an intriguing mathematical question: Given a surface defined by a monotonically decreasing sequence of j (sufficiently dense) layers, can we predict the surface continuation (by layers) for , and is a convergence being obtained? We aim to provide a mathematical perspective on abstract art and enable its analysis from a computational perspective. To achieve our purpose, we first made a 3D computer model of each of the tentacles which was created by the designer in the team (for a given 2D image). Second, the computational team received this modeling, sampled it; and formulated a prediction algorithm for a given tentacle. Third and last, each of the tentacle predictions has been sent back to the designer, which smoothly reconstructs each tentacle by the algorithm continuation, locates each continuation in a similar way to the original room in the 2D image and wraps it with the respective texture. In this fashion, the viewer can compare the original artistic artwork of the tentacle room to the mathematical analysis predicted room.Subject terms: Computational science, Scientific data Introduction Numerous studies have focused on reconstructing three-dimensional objects, from given multiple projections-two-dimensional images, located in different positions; see1–6, where the depth, color and more need to be approximated. The case of 3D reconstruction from a single 2D image is much more involved and is usually based on acquaintance with the given scene and its properties. Yayoi Kusama famous is contemporary sculpture and installation artist, who is also active in painting, and performance (see7). Her works have been exhibited worldwide, showcasing various motifs and materials, such as mirrors and textured surfaces. All are characterized by a range of geometrical patterns/ideas which catch our eyes; even though Kusama’s does not have any mathematical background or intends to transmit any mathematical concepts (as well as we know). Each exhibition usually has its own character in different museums, where some of the artworks have similar ideas but different implementations. This manuscript refers to her ‘tentacle’ artwork which have different versions in different exhibitions. It is characterized by tentacles that are attached to the floor, as in Fig. 1, each of which is wrapped with a texture of a decreasing sequence of white disks (or other textures), where it seems that the end of each tentacle has been cut and smeared. While Kusama’s approach is purely artistic, for us these surfaces raises an intriguing mathematical question: Given a surface defined by a monotonically decreasing sequence of n (sufficiently dense) layers, can we predict the surface continuation (by layers) for , and is convergence obtained? We aim to give a mathematical formulation and exploration to the above and to visualize a continuation of each of the tentacles by these monotonic rules. Finally, we would like to approximate and visualize the “tentacle room” by the obtained continuation. Ohio 2012 tentacles artwork, photo credit to Tim Perdue. In the right side is the original image with a respective numbering for each tentacle. The left side, if two tentacles have the same numbering, then from our point of view, they are isometrically symmetrical up to translation, reflection, and rotation. Notice that difficulties arise in characterizing tentacle ‘G’ by the given orientation, which may conceal possible twisting of the tentacle. We explore the artistic artwork as a theoretical mathematical data, and present the artistic community a theoretical computational point of view, which they have never been aware of. To the best of our knowledge, this is the first time surface prediction has been applied in an artistic context but plays a major rule in various fields of science. In the engineering materials field, it can help define new materials or the growth of crystals, see8. In the field of manufacturing, there is great interest in predicting surface topography characteristics, see9. From a geographical point of view, the main purpose is to forecast changes in the earth surface as landscape, see10. In11 control algorithms are applied to classify surfaces to the anti-lock brake system in the vehicle industry, by trying to predict the surface that touches the tire and as a result to control the wheel slip to achieve maximum friction. It is also an important concept in astrophysics where a size, shape (and motion) prediction of the asteroid is needed, see4,12. It can also be applied to archaeology, where art as sculptures, may be sorted by defining similarities between surfaces, as in13, which will lead to the predicted surface. All these models, with the right adjustment to our surfaces, have been considered for the “tentacles room” predictions. To achieve our goal, as a first step, we had to convert a 2D image of a ‘tentacles room’ to 3D computer-aided design1. As mentioned above, usually reconstructing the 3D ‘scene’ from a single image is not easy, but since many versions exist to the “tentacles room” and it respective 2D images, we believe our approximated computer modeling to each of the tentacles is reasonable (all images versions from different exhibitions define the tentacles, without holes, smooth, and without radical changes from different angle). Using this computerized representation, we can break down the room into its tentacle-like components, which each can be explored geometrically. Notice, that each of the tentacles needs to be sampled carefully, such that it will best approximate the original surfaces and our geometrical model. We choose to represent each tentacle as a finite sequence of ellipses, where our algorithm predicts the next ellipses in the sequence until a convergence to area zero is obtained. After finishing all predictions, the designer in the team smeared each of the sequences onto a smooth surface, which will represent each of the tentacles’ continuation. For comparing our prediction to the original artwork, each tentacle is located as in the original artwork (and orientation) and wrapped with the respective decreasing texture of white disks. We attempt to show that abstract art can sometimes embody a purely mathematical perspective-even if the artist was unaware of it-and that creativity is a concept that belongs not only to art, but also to mathematics. Results Our reference for this manuscript is a single 2D image by the photographer Tim Perdue2 of the“tentacles room” by Yayoi Kusama, see Fig. 1, which is composed of nine tentacles with a white disks texture. We observe that the space defined by these tentacles is open, lacking a roof or ceiling, which may suggest an infinite vertical extent. The perspective is oblique, revealing only two vertical walls of the room; the corner edges are barely discernible, and the boundaries of the wall surfaces remain indeterminate. Consequently, Yayoi Kusama’s space can be interpreted as potentially infinite in three dimensions, thereby providing the desired freedom for continuation. We intend to predict the continuation in 3D to each of the tentacles surfaces. Notice that the texture of the monotonically decreasing white disks along each tentacle reinforces this assumption of convergence. Before diving into the computational methods for the prediction, each of the tentacles in the source image needs to be computer-modeled by the designer in the team in 3D software, such that locating the objects (tentacles) naked-without the red and white disks texture, in the right orientation and location will be faithful to the original image. Our computer model, devoid of artistic texture, will serve as our source for generating the data used in our computational methods. By examining other images of the tentacle artworks in different exhibitions, from different positions, we assumed that each of the tentacles without holes, smooth or radical changes from different angles, this assumption helps us to define a 3D representation from a single 2D image. Our strategy that each tentacle in the 3D representation (and respective to the 2D image) will be explored and estimated separately by our geometrical approach; and at the end of the process, with the help of the designer, each of the tentacle’s continuation will be located in the same position and texture respective to the original image. Before processing these 3D objects/tentacles, from our respective, not all nine tentacles need to be explored, see Fig. 1. In the right image, tentacles ‘A’, ‘B’ and ‘F’, may be considered identical up to translation, reflection and rotation, and in a similar way tentacles ‘D’ and ‘H’. The position of tentacle ‘G’ does not give the required data to understand the curvature of the surface so we decided to concede it exploration. This observation reduced the number of tentacles requiring exploration and prediction from nine to five. Slicing each tentacle to a two-dimensional sequence In Fig. 2 is our approximated computer design modeling of each of these five tentacles. Since we are interested in preserving the tentacle structure, the sampling of the computered model, needs to be done according to our geometrical construction and represent best each of the tentacles. Imagining each of the five tentacles in Fig. 1 in 3D. In our approach, this data sampling will be obtained by slicing in a certain rule each tentacle, which will lead to layers of 2D geometrical objects (where, when it is dense enough tends to the tentacle). Since we have five different tentacles, we need to decide which geometrical sequence approximation fits all best, and how many sampling points will be taken from every layer. In this slicing process as a sequence of 2D layers that build each tentacle, we will do our best to avoid 2D non-convex shapes, where the most challenging tentacle to represent by a sequence of 2D geometrical shapes3 in Tentacle (3), see Fig. 2, which the respective medial axis4 is not monotonically increasing and may lead in the upper layers to approximation with the respective convex hull5, see a closeup in Fig. 3. This observation indicates that while Tentacle 1,2,4,5 are defined by a monotonically decreasing sequence of areas, which we can assume will lead to linear prediction of the tentacles as a monotonic decreasing sequence, the prediction for Tentacle 3 may not lead to a linear result. A detailed view of the third tentacle illustrates two slicing strategies: circular cross-sections and slices parallel to the z-axis. While slicing parallel to the z-axis generally provides a more accurate approximation model for all tentacles, in this tentacle, the 2D layers slicing leads that at some of the layers, non-convex 2D domains are obtained. This geometrical property may lead that in our geometrical approximation the continuation will not be linear. Several methods for the approximation of each tentacle by plane layers were considered, but in the end we narrowed to the following two geometrical slicing ideas. First, we consider approximating each layer as a circle, see Fig. 3. In this approach, each circle is defined by translation, rotation and the radius, all in 3D, which makes difficulties in the multiple prediction models due to a high number of variables, and even leads to contradictions. So, reducing the number of variables was necessary. It led us to define a parallel slicing of the z-axis respective to the (x, y) plane, in a way similar to14, where in this fashion only one rotation matrix of the geometrical planar shape is considered. This leads us to the next step of sampling each layer and approximating it as a sequence of geometrical shapes. A geometrical perspective to the data set Our approach is to slice each tentacle such that the height is constant and fit a geometrical shape that fits to each sampled points to a given layer. After test and trail with approximation of different geometrical shapes and number of sampled points in each layer, we choose the geometrical layers sequence can be considered as ellipse’s sequence6 approximation, where in each layer ten points have been sampled (randomly), since choosing the minimal number of points to define the ellipse did not lead to a good approximation, as in Fig. 4. In Fig. 5, is the approximation for each of the tentacles by the sequence of ellipses from the sampling data (ten points in a layer). The five tentacles have been sampled. In this image, we choose 34 layers (that the viewer will notice the differences) parallel to the z-axis, where in each layer, ten points have been sampled. With this uniform approach of parallel planes sampling to all tentacles, we had to compromise on the input of the third tentacle created data set. Given the data set of five tentacles in Fig. 4, each tentacle has been approximated by a sequence of ellipses. The dots are the ellipses’ centers sequence, which a respective curve needs to be fitted. The geometric approximation of these ellipses for each tentacle will be our input to the continuation prediction model. This fitting process of each ten points to an ellipse will be obtained by determining the center of each ellipse, the two axes , and the angle of rotation (and the respective lifting). We will denote that data set by the sequence , and of course the height of the ellipse which is constant. Fig. 5 illustrates the curving of the ellipses centers in 3D for the given data set. So, we aim to give prediction to the data set , for , where, in our assumption, some of the variables in this vector are independent. Prediction algorithm for tentacle continuation For this purpose, we examined several solutions, including numerical methods, machine learning, and deep learning. Ultimately, we chose the strategy described below: prediction using generalizations of least squares and regressions. We selected this approach because it provided the most satisfactory results. The other methods required significantly more data for the models to learn effectively. However, our data set is limited and contains problematic entries that could lead to errors in learning-based models. The approach we chose is simple, easy to apply to our data, and successfully led us to a viable solution. First, we will predict the 3D curve of the ellipses centers . Second, we will determine the corresponding values of the canonical ellipse- where from our point of view the area (which needs to be evaluated by a prediction model) of the ellipses determines best the relation between ai and bi. Third, each ellipse will be rotated in the right direction; and last, each ellipse will be lifted to the right height. By these vectors predictions we will construct, with the help of the designer, each of the continued tentacles and wrap it with the same texture as the original artwork, such that a comparison can be obtained. Having established the process for achieving our objectives, we can now define it in mathematical terms. Given the center’s data set by layers , which needs to be approximated by a respective 3D curve efficiently, especially in the points where the curvature increases. We found that the following polynomial model approximated best the center’s curve: (i)Define the data as and predict using a least square of order three , where is the prediction value and is the number of layers. Define the data as and predict using a least square of order three , where is the prediction value and is the number of layers. (ii)Define another data and predict , as has been done in (i). Define another data and predict , as has been done in (i). (iii)Construct the prediction curve . Construct the prediction curve . Fig. 6, illustrates the curve prediction of the ellipses centers (compared to the original data set of each tentacle). The black triangles represent the given data set for the ellipses’ centers. The red curve is the obtained prediction curve by a degree three polynomial, which led to the best approximation. An Algorithm for Tentacle Prediction Our technique for predicting , the largest and smallest diameters respectively, is based on the area of each ellipse, which we believe needs to converge to zero. The following procedure has been formulated for this prediction: (i)Given the axis of the ellipses data set , . Compute the respective area of the ellipse in the layer i, which is the sequence . Given the axis of the ellipses data set , . Compute the respective area of the ellipse in the layer i, which is the sequence . (ii)By linear regression from degree two, predict , where . By linear regression from degree two, predict , where . (iii)Define the data set sequence as () and by linear regression from degree two, predict . Define the data set sequence as () and by linear regression from degree two, predict . (iv)Compute Compute Notice that the prediction can be obtained by different methods, but since and together define the area, from our mathematical perspective, the error is reduced. Lastly, for the angle , we define the data set as , which is estimated using a degree-two linear regression. Combining all results produced the vector of ellipses , representing the continuation of each tentacle. For the respective pseudo-code for the complete procedure, which plots a prediction vector , which represents the construction of the new tentacle and its continuation, see Algorithm 1. In this algorithm, we decided to stop at the point where the diameters diverge, where each tentacle behaved differently; some really converged, where the sequence tends to infinity (as we expected), while others seemed to diverge shortly after the initial convergence. Smearing this sequence will lead to the desired result. Lastly, wrapping each tentacle in the exact same texture, color, shading, etc., as the original room is challenging, since one has been obtained manually by 3D software modeling, compared to a photo of a real visual art. Fig. 7, gives a comparison of our modeling with the same number of layers to the original artwork (Fig. 1), all tentacles in the same location and the same height (without adding the additional layers). The reader can compare our computerized approximations until the n layer and the original image. To obtain the described result, we made relaxation to all layers, which leads to the left image. In the right image, we applied a similar texture to the original artwork. Notice, that our model (as other prediction models), does not plot exactly the original data set since it is determined by the continuation. Fig. 8a, is our continuation approximation to each of the explored tentacles, positioned in the same orientation and location as the original artwork. Fig. 8b, is our final result, where each of which is wrapped by the texture similarly to the original artwork with same rule of decreasing sequence of red circles and a similar lightning. In the first two images our approximation of the “tentacles room” after the continuation (where we did not expand the room), first naked and second with the texture (and a similar lightning to the original). Notice that we had to expand the room due to the continuation. In the third image, an example of one of the tentacles printed in 3D. Conclusions Predicting surfaces in unsupervised learning poses inherent challenges due to the absence of direct error estimation relative to an original surface. First, our study demonstrates that incorporating an external parameter can effectively guide and control such predictions. Drawing on Kusama’s artwork, we identified that the auxiliary information embedded in the texture of white disks along each tentacle provides a natural error-regulating mechanism, as the sequence of disks decreases monotonically from the floor upward. Under the theoretical assumption that disk size converges to zero at a finite height, the continuation of each tentacle remains both mathematically bounded and practically constrained. This integration of external texture-based data thus transforms an unsupervised problem into a partially supervised framework, improving prediction reliability. Defining an appropriate data set for predicting tentacle-like surfaces required addressing the challenge of generating meaningful data from computer-aided models. Unlike typical machine learning settings where sampled data is pre-existing, our data originated from model-based sampling and thus required transformation into geometric representations suited to surface prediction. By leveraging the convergent texture pattern-modeled as a sequence of two-dimensional geometric shapes progressing from the floor upward, we ensured that the continuation of this sequence would theoretically converge to zero area. Maintaining as we can convexity in the representation of each layer proved essential, as deviations from convexity could lead to errors when applying the convex hull. This careful construction of the data set forms a critical foundation for achieving reliable predictive modeling of such surfaces. Our investigation demonstrates that tentacle-like surfaces, when modeled from computer-aided designs, can be effectively approximated by sequences of ellipses. By carefully selecting the number of sampled points in each layer to best represent the underlying surface geometry, we achieve a representation that captures essential properties such as convexity, area, rotation, and convergence. This geometric approach shifts the focus from raw point-based sampling to a higher-level description in terms of elliptical layers, each defined by its height, center, axes, etc.. Our results show that this method provides accurate predictions for subsequent layers when the sampled boundaries form convex domains, while non-convex domains may lead to degenerate, thin ellipses. Nevertheless, given that most tentacles in our data set exhibit predominantly convex layers, the approximation remains both mathematically sound and visually faithful, as evidenced by the strong correspondence between predicted and actual structures (e.g., Tentacle 3 compared to Tentacles 1, 2, 4, and 5). Finally, our approximation of the continuation incorporates the assumption that certain variables-such as the axes and area-are interdependent and, consistent with our initial texture-based assumptions, are expected to converge. Identifying and enforcing these dependencies not only aligns the predictions with the underlying geometric constraints but also mitigates potential contradictions within the algorithm across successive prediction steps. This dependency-aware approach thus enhances both the coherence and the stability of the predicted surface continuations. Summary In this research, we seek to push the boundaries by formulating three-dimensional abstract art as a mathematical model governed by a given set of rules, which can then be investigated from a computational perspective. As part of this work, we aim to demonstrate how Yayoi Kusama’s artistic creation of ‘tentacles’ can be interpreted, from the viewpoint of mathematicians and the exact sciences, as a monotonically decreasing sequence of geometric forms which, for artistic reasons, does not reach convergence. To carry out this task, we selected one of Kusama’s images and transformed it into a three-dimensional digital representation, defining it as vector data. An algorithm was then applied to each tentacle to predict its form until convergence (which is a continuation of each tentacle). Finally, the predicted tentacles were positioned in a manner similar to their arrangement in the original artwork, enabling viewers to compare the two and to appreciate how the artwork might appear if each tentacle were extended by a mathematical convergence model to an infinite number of layers. Acknowledgements We express our gratitude to photographer Tim Perdue for granting permission to use his photograph in support of this theoretical research. Author contributions M.P 3D visualization. D.I, M.O and S.G mathematical formulation and codes. Data Availability All data generated and analyzed in this study were derived from a single image of Tim Perdue, for which we obtained explicit permission to explore using mathematical and computational methods. The image was processed and reconstructed into a 3D model by a designer on our team. This 3D representation was subsequently sampled, resulting in a CSV file that numerically represents the original image data. Researchers interested in replicating or extending this work are welcome to use the publicly available dataset and are kindly requested to cite this study accordingly. For data availability please contact the author Shai Gul. Sincerely, Shai Gul Declarations Competing interests The authors declare no competing interests. Footnotes This has been done in Blender software manually. Tim Perdue gave us explicit permission to use and explore the image. A set of points is convex if it contains every line segment between two points in the set. This is the thin version of the shape. the smallest convex set that encloses all the points. While every three points (which are not on the same line) define a circle, in the case of an ellipse, five points define at most one ellipse with unique terms: three non-collinear points, and another two more points are needed to specify the shape and size of the ellipse along those axes. Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement All data generated and analyzed in this study were derived from a single image of Tim Perdue, for which we obtained explicit permission to explore using mathematical and computational methods. The image was processed and reconstructed into a 3D model by a designer on our team. This 3D representation was subsequently sampled, resulting in a CSV file that numerically represents the original image data. Researchers interested in replicating or extending this work are welcome to use the publicly available dataset and are kindly requested to cite this study accordingly. For data availability please contact the author Shai Gul. Sincerely, Shai Gul