ULISSE Algorithm in Multidisciplinary Research

• ULISSE algorithm is a polysemous term describing distinct methods in various fields, including latent class-based selection in healthcare and one-shot retrieval in astronomy.
• Each variant employs domain-specific techniques—such as EM with BIC, fixed CNN features, and randomized truncation—to optimize performance and reduce computational complexity.
• The method’s effectiveness and limitations vary by context, making precise disciplinary interpretation essential for its practical application in research.

Searching arXiv for “ULISSE” to ground the article in the relevant literature. The term ULISSE algorithm does not denote a single, universally standardized method across the arXiv literature. Instead, it refers to several distinct algorithms and methodological frameworks developed in different domains, including latent class–based item selection for geriatric assessment, one-shot astronomical image retrieval, Gaussian-process inference, variable-length subsequence indexing, linear unfolding in experimental physics, and ultrasonic surface reconstruction [(Bartolucci et al., 2014); (Doorenbos et al., 2022); (Filippone et al., 2015); (Linardi et al., 2020); (Laszlo, 2011); (Giannoccaro et al., 2014)]. In some cases ULISSE is an explicit acronym, such as aUtomatic Lightweight Intelligent System for Sky Exploration in astronomy and Unbiased LInear System SolvEr in Gaussian-process inference (Doorenbos et al., 2022, Filippone et al., 2015). In other cases, the underlying paper does not define the acronym and the label is used descriptively or project-specifically [(Laszlo, 2011); (Giannoccaro et al., 2014)]. This multiplicity makes “ULISSE algorithm” a polysemous research term whose meaning depends entirely on disciplinary context.

1. Statistical item selection in the ULISSE nursing-home study

In the statistical literature, ULISSE refers to the nursing-home quality-of-life project analyzed in “Item selection by Latent Class-based methods” (Bartolucci et al., 2014). The paper studies questionnaires made of a large number of polytomous items and applies a latent class (LC) model to cluster subjects into homogeneous groups corresponding to different degrees of impairment of the health conditions (Bartolucci et al., 2014). The data comprise $n = 1739$ subjects in the first wave and $J = 75$ polytomous items grouped into 8 sections: Cognitive Conditions, Auditory and View Fields, Humor and Behavioral Disorders, Activities of Daily Living, Incontinence, Nutritional Field, Dental Disorders, and Skin Conditions (Bartolucci et al., 2014).

The algorithm is closely related to that proposed by Dean and Raftery in 2010 and is aimed at finding the subset of items that provides the best clustering according to the Bayesian Information Criterion while also selecting the optimal number of latent classes (Bartolucci et al., 2014). The LC model assumes a discrete latent variable $U$ with $K$ classes, conditional independence of item responses given class, and Missing at Random (MAR) missingness, with estimation performed by the EM algorithm (Bartolucci et al., 2014). For polytomous items, the observed-data likelihood is written as

$$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$

with the corresponding BIC

$\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$

where

$p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$

for the polytomous LC parameter count (Bartolucci et al., 2014).

The item-selection procedure alternates inclusion and exclusion steps. For a candidate item, the comparison is between a class-dependent model $M_1$, in which the item participates in clustering, and a class-independent/noise model $M_0$, in which it is modeled under a single-class LC when excluded from the clustering set (Bartolucci et al., 2014). The search is stepwise forward–backward, updates $K$ by minimizing BIC over $J = 75$0, and terminates when no inclusion yields $J = 75$1 and no exclusion yields $J = 75$2 (Bartolucci et al., 2014). A notable modification is the “random check” step: after each accepted inclusion or exclusion, EM is rerun from many random starts to mitigate local maxima (Bartolucci et al., 2014).

For the ULISSE dataset, the full-set model with all $J = 75$3 items and $J = 75$4 selected $J = 75$5 by BIC (Bartolucci et al., 2014). Across initial subset sizes $J = 75$6, the best overall solution by $J = 75$7 came from $J = 75$8 and yielded $J = 75$9 of 50 items and $U$0, with

$U$1

(Bartolucci et al., 2014). All items from CC, AVF, and ADL were retained; two items from I were retained; and most excluded items belonged to DD, HBD, NF, and SC (Bartolucci et al., 2014). Validation used both sensitivity to the initial subset and a bootstrap-like resampling with $U$2 samples; 45 items were selected in all final solutions across both initial-subset experiments and resamples, and more than three-quarters of resamples reproduced the original selection, indicating good stability (Bartolucci et al., 2014).

2. One-shot sky exploration in astronomy

In astronomy, ULISSE is explicitly defined as aUtomatic Lightweight Intelligent System for Sky Exploration (Torbaniuk et al., 27 Jul 2025). The method introduced in “ULISSE: A Tool for One-shot Sky Exploration and its Application to Active Galactic Nuclei Detection” operates in a one-shot retrieval setting: a single prototype image is supplied, and the system returns a ranked list of sosia objects sharing similar morphology and photometric appearance, without training any task-specific model (Doorenbos et al., 2022). The implementation uses EfficientNet-b0 pretrained on ImageNet, discards classifier layers, and extracts a $U$3 feature map whose spatial mean yields a 1280-D image-level embedding $U$4 (Doorenbos et al., 2022).

The astronomical ULISSE pipeline uses SDSS DR16 cutouts built as JPEG color composites from g, r, i bands, with thumbnails of $U$5 pixels, center-cropped to $U$6 pixels and then resized to $U$7 for the CNN (Doorenbos et al., 2022). Similarity is defined by Euclidean distance in feature space,

$U$8

although cosine similarity is discussed as an alternative (Doorenbos et al., 2022). For $U$9k images, a k-d tree is built over the embeddings; the reported build time is ~25–30 s, and queries return in ~1 s on a single GPU workstation after feature extraction (Doorenbos et al., 2022). Feature extraction for 100k images takes 3–4 min (RTX 2060) (Doorenbos et al., 2022).

The main published application is AGN candidate retrieval in a cleaned SDSS sample of 99,991 objects with BPT labels from the DR8 MPA-JHU galSpec catalog (Doorenbos et al., 2022). In this sample the class proportions are AGN 12%, Composite 5.8%, SFG 44.1%, and Unclassified 38.1%, so the random guess baseline is 12% for AGN (Doorenbos et al., 2022). Across 8 X-ray confirmed AGN prototypes, ULISSE achieved average $K$0 and $K$1 among the top-300 neighbors, with prototype-dependent ranges of ~21–53% for AGN alone and up to ~65% when composites are included (Doorenbos et al., 2022). The best prototype, #4, yielded 53.0% AGN and 65.4% AGN+Comp, whereas late-type prototypes such as #6 yielded 21.3% AGN and 38.9% AGN+Comp (Doorenbos et al., 2022). The method is reported to be most effective in retrieving AGN in early-type host galaxies, while spiral or late-type prototypes are more affected by SFG confusion (Doorenbos et al., 2022).

A later astronomical paper extends ULISSE from candidate retrieval to parameter transfer. In “ULISSE: Determination of star-formation rate and stellar mass based on the one-shot galaxy imaging technique,” the same one-shot feature-space retrieval is used to estimate $K$2 and $K$3 from a single g/r/i composite image (Torbaniuk et al., 27 Jul 2025). The method retrieves $K$4 nearest neighbors by Euclidean distance and transfers their physical properties via either an unweighted mean or a normalized mean using weights $K$5, where $K$6 is the population of the SFR–$K$7 bin containing neighbor $K$8 (Torbaniuk et al., 27 Jul 2025). On a Random working subset of 100,000 galaxies, with 290 targets sampled across the $K$9–$$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$0 plane and multiple morphology, AGN, dust, and redshift strata, ULISSE achieved the following joint performance within 1 dex for all targets: $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$1, $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$2, and $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$3, versus random baselines of 14.8%, 60.3%, and 34.8%, respectively (Torbaniuk et al., 27 Jul 2025). Separate-axis retrieval gave 94.1% within $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$4 for stellar mass with the mean estimator and 81.0% for SFR (Torbaniuk et al., 27 Jul 2025).

These astronomical variants share a common structure: training-free transfer learning, fixed ImageNet features, Euclidean nearest-neighbor search, and the use of a single prototype rather than a labeled training set (Doorenbos et al., 2022, Torbaniuk et al., 27 Jul 2025). This suggests a stable methodological identity for ULISSE within astronomy, despite the shift from object detection to regression-like neighbor transfer.

3. Unbiased linear solves for Gaussian-process inference

In machine learning, ULISSE is the Unbiased LInear System SolvEr proposed in “Enabling scalable stochastic gradient-based inference for Gaussian processes by employing the Unbiased LInear System SolvEr” (Filippone et al., 2015). The context is Gaussian-process regression with covariance parameters $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$5, where posterior sampling over $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$6 requires repeated access to gradients of the marginal log-likelihood,

$$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$7

but where direct Cholesky-based inference is computationally prohibitive for large $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$8 (Filippone et al., 2015).

The key objective is to estimate $$L(\pi,\lambda) = \prod_{i=1}^n \sum_{c=1}^K \pi_c \prod_{j=1}^J \prod_{r=1}^{R_j} \lambda_{jc}(r)^{\mathbf{1}(Y_{ij}=r)},$$9 using only covariance matrix–vector products, without storing or factorizing $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$0, and without bias (Filippone et al., 2015). ULISSE starts from Conjugate Gradients (CG) and uses randomized truncation of the telescoping sum of CG increments. After an early stop at threshold $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$1, the omitted tail is stochastically completed by a Russian roulette scheme with weights $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$2 (Filippone et al., 2015). The estimator $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$3 is unbiased in the sense that

$\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$4

with the parameter $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$5 controlling the variance–time trade-off (Filippone et al., 2015). The paper uses an early-stop threshold of the form $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$6 and experiments with $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$7 and $\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$8 (Filippone et al., 2015).

This unbiased solver is combined with Hutchinson trace estimation and Stochastic Gradient Langevin Dynamics (SGLD). The SGLD update in the paper is

$\mathrm{BIC} = -2\,\ell(\hat\theta) + p\,\log n,$9

with asymptotic correctness under the standard step-size conditions

$p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$0

(Filippone et al., 2015). The quadratic term in the gradient is estimated using two independent ULISSE estimates of $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$1, and the trace term is estimated with $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$2 Hutchinson probes in the experiments (Filippone et al., 2015).

Empirically, the paper reports results on the Concrete dataset with $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$3 and the Census 8L dataset with $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$4 (Filippone et al., 2015). On Concrete, SGLD with ULISSE matched Metropolis–Hastings with exact marginal likelihood in running means and $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$5 standard deviation bands over 40k posterior samples (Filippone et al., 2015). On Census 8L, the implementation achieved ≈ 10,000 posterior samples per day, with effective sample size ≈ 0.1% and one independent sample every ~2.4 hours, using a desktop with an 8-core CPU and NVIDIA GeForce GTX 590 (two GPUs) for covariance matrix–vector products (Filippone et al., 2015). The paper states that ULISSE reduced the number of CG iterations by orders of magnitude at the cost of a modest increase in variance (Filippone et al., 2015).

Within this literature, ULISSE is neither a retrieval system nor an indexing scheme but a debiasing device for iterative linear algebra. The defining property is unbiasedness of the solve, which then propagates to unbiased stochastic gradients.

4. Variable-length subsequence indexing in data-series search

In data management, ULISSE denotes the ULtra compact Index for variable-length Similarity SEarch in data series introduced in “Scalable Data Series Subsequence Matching with ULISSE” (Linardi et al., 2020). The paper presents ULISSE as the first single-index approach that answers similarity queries over subsequences of variable length within a user-specified range $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$6 (Linardi et al., 2020). It supports both non-normalized and Z-normalized series, both Euclidean distance and Dynamic Time Warping (DTW), and both k-NN and $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$7-range queries (Linardi et al., 2020).

The central representation is an envelope that summarizes many contiguous, overlapping subsequences of different lengths using a single set of min–max PAA coefficients, later discretized into iSAX symbols (Linardi et al., 2020). The parameter $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$8 determines how many starting offsets are summarized together: increasing $p = (K-1) + \sum_{j=1}^J K\,(R_j-1)$9 yields fewer, larger envelopes and a more compact index but looser bounds, while decreasing $M_1$0 yields tighter envelopes and a larger index (Linardi et al., 2020). The tree structure is akin to iSAX, with internal nodes storing symbolic summaries $M_1$1, leaves storing several envelopes and pointers to raw series on disk, and an in-memory list of all envelopes at maximum iSAX cardinality for exact search (Linardi et al., 2020).

For Euclidean search, ULISSE defines a lower bound $M_1$2 between the query PAA and an envelope, and for DTW it defines $M_1$3, a PAA–iSAX lower bound between a DTW envelope and a ULISSE envelope (Linardi et al., 2020). Query processing uses a best-first approximate search over the tree to obtain an initial best-so-far, followed by a cache-friendly in-memory sequential scan and raw verification guarded by lower bounds and early abandoning (Linardi et al., 2020). The disk access pattern is explicitly sequential and coalesced, and each verification reads at most $M_1$4 points from disk (Linardi et al., 2020).

The reported implementation is in C, with bulk loading via iSAX 2.0, on datasets including synthetic random-walk series and real collections such as electric power GAP, EEG/EMG/ECG, ASTRO, and SEISMIC (Linardi et al., 2020). The paper reports that, versus the UCR Suite, ULISSE is up to ~12× faster for Euclidean exact k-NN on 5 GB datasets for $M_1$5–$M_1$6 and up to ~5× faster on 100–750 GB datasets for the same length regime (Linardi et al., 2020). For DTW exact k-NN it is up to ~10× faster on real datasets for moderate $M_1$7, and versus Index Interpolation its index is ~100× smaller and answers k-NN ~10× faster across ranges 256–4096 (Linardi et al., 2020). Against CMRI, ULISSE is up to ~15× faster overall for approximate search on non-normalized data, and against KV-Match it is up to ~10× faster for DTW $M_1$8-range search on large datasets (Linardi et al., 2020).

In this setting, the meaning of ULISSE is entirely different from the machine-learning and astronomical uses. The commonality is only nominal: the method is an index structure, not a solver or retrieval model.

5. Linear unfolding and ultrasonic reconstruction

Two additional bodies of work use ULISSE in yet different senses.

In experimental physics, “A Linear Iterative Unfolding Method” presents a linear iterative algorithm for recovering an unknown spectrum $M_1$9 from a smeared measurement $M_0$0 under a folding operator $M_0$1 (Laszlo, 2011). The paper explicitly states that it does not define the acronym “ULISSE”, and in this context ULISSE refers to the linear iterative method introduced there (Laszlo, 2011). The iteration is

$M_0$2

or, in discrete matrix form,

$M_0$3

(Laszlo, 2011). Its sole regularization parameter is the stopping order of the iteration, and the paper provides explicit propagation formulae for bias error, statistical error, and systematic error (Laszlo, 2011). The method is proved to converge, in the noise-free case, to the true distribution modulo the kernel of the response operator under the practical condition $M_0$4, which the paper states holds for convolutions, calorimeter response functions, and momentum reconstruction response functions based on tracking in magnetic field (Laszlo, 2011).

In ultrasonic sensing, “Least Entropy-Like Approach for Reconstructing L-Shaped Surfaces Using a Rotating Array of Ultrasonic Sensors” likewise states that the paper does not explicitly define the acronym ULISSE; the summary uses it descriptively for the method (Giannoccaro et al., 2014). The approach reconstructs two smooth orthogonal planes forming an L-shaped corner from data acquired by four in-air ultrasonic transducers mounted in a rotating linear array (Giannoccaro et al., 2014). It first computes the waveform-energy indicator

$M_0$5

and identifies the critical angle $M_0$6, which marks the corner direction dominated by multiple reflections (Giannoccaro et al., 2014). After removing the four points at $M_0$7 and splitting the remaining scans into two plane-specific subsets, each plane is fit by minimizing the least entropy-like objective

$M_0$8

where $M_0$9 and $K$0 (Giannoccaro et al., 2014). Optimization uses Nelder–Mead from six initializations, and the paper reports that cross-terms in the recovered plane equations drop by roughly one to two orders of magnitude relative to least squares (Giannoccaro et al., 2014).

These two uses illustrate a recurring feature of the term ULISSE outside astronomy, Gaussian processes, and time-series indexing: the name may denote a method associated with a specific project or explanatory shorthand rather than a stable acronym defined in the paper itself [(Laszlo, 2011); (Giannoccaro et al., 2014)].

6. Cross-domain characteristics, ambiguities, and limitations

Because the literature contains several unrelated ULISSE methods, any unqualified use of the phrase “the ULISSE algorithm” is ambiguous. A precise interpretation requires the disciplinary context, the associated paper, or both [(Bartolucci et al., 2014); (Doorenbos et al., 2022); (Filippone et al., 2015); (Linardi et al., 2020)]. The ambiguity is not merely nominal: the underlying mathematical objects differ fundamentally. In the nursing-home paper, ULISSE is a latent-class variable-selection procedure driven by BIC and EM (Bartolucci et al., 2014). In astronomy, it is a one-shot nearest-neighbor retrieval framework based on ImageNet embeddings and Euclidean distance (Doorenbos et al., 2022, Torbaniuk et al., 27 Jul 2025). In Gaussian-process inference, it is an unbiased debiasing mechanism for truncated CG and stochastic gradients (Filippone et al., 2015). In time-series databases, it is a variable-length subsequence index built from envelope summaries and iSAX (Linardi et al., 2020). In unfolding and ultrasonic reconstruction, it designates a linear iterative inverse method and a least-entropy robust geometric fitting procedure, respectively [(Laszlo, 2011); (Giannoccaro et al., 2014)].

The limitations reported in each literature are correspondingly domain-specific. The latent-class item-selection method may be affected by violations of local independence, likelihood multimodality, possible violation of MAR, and sensitivity to $K$1 and the initial subset size $K$2, with $K$3 varying between 8 and 10 in sensitivity analyses (Bartolucci et al., 2014). The astronomical ULISSE is sensitive to prototype morphology, domain shift from ImageNet, artifacts and confounders such as bright stars or overlapping sources, and the loss of spatial layout detail due to mean pooling (Doorenbos et al., 2022). The SFR/$K$4 variant additionally notes limitations from thumbnail size, companions, redshift effects, fiber aperture and AGN contamination, dust and inclination, and cross-survey domain shift (Torbaniuk et al., 27 Jul 2025). The GP ULISSE trades computational savings for estimator variance controlled by $K$5, $K$6, and the number of Hutchinson probes $K$7, while SGLD introduces practical bias if step sizes are frozen too early (Filippone et al., 2015). The data-series ULISSE can lose pruning power for extremely wide length ranges, very large $K$8, highly nonstationary series, or large DTW warping windows (Linardi et al., 2020). The unfolding method regularizes only through stopping order and does not strictly enforce positivity or normalization at finite iterations (Laszlo, 2011). The ultrasonic method assumes a structured scene with two planar surfaces and notes scattering in unstructured scenes as a limitation (Giannoccaro et al., 2014).

A plausible implication is that ULISSE has evolved less as a single algorithmic lineage than as a reusable project name or acronym template adopted independently in several subfields. The astronomy and SFR/$K$9 papers form one coherent lineage, but the other uses are methodologically unrelated (Doorenbos et al., 2022, Torbaniuk et al., 27 Jul 2025).

7. Enduring significance of the ULISSE name in research practice

Despite the heterogeneity of meanings, the various ULISSE methods share a recurrent role as computational devices for reducing search or inference burden under practical constraints. The latent-class ULISSE reduces a 75-item questionnaire to a more parsimonious 50-item subset while improving clustering interpretability and stability (Bartolucci et al., 2014). The astronomical ULISSE removes the need for task-specific training and enables rapid candidate retrieval or approximate property estimation directly from images (Doorenbos et al., 2022, Torbaniuk et al., 27 Jul 2025). The GP ULISSE removes the need for marginal-likelihood evaluation and Cholesky factorization while preserving unbiasedness in a Monte Carlo sense (Filippone et al., 2015). The data-series ULISSE replaces multiple fixed-length indexes with a single compact index for variable-length subsequence search (Linardi et al., 2020). The unfolding and ultrasonic variants reduce inverse-problem instability or outlier sensitivity through iterative regularization or entropy-like robustness [(Laszlo, 2011); (Giannoccaro et al., 2014)].

For researchers, the principal editorial caution is therefore terminological rather than conceptual: ULISSE is not a canonical algorithmic family unless the citation makes it so. In astronomy, the phrase generally denotes the EfficientNet-based one-shot similarity-search framework and its neighbor-transfer extensions (Doorenbos et al., 2022, Torbaniuk et al., 27 Jul 2025). In Gaussian processes, it denotes the unbiased Russian-roulette linear solver used inside SGLD (Filippone et al., 2015). In the social-statistical ULISSE project, it denotes the latent class–based item-selection algorithm for nursing-home assessment (Bartolucci et al., 2014). Elsewhere, the name may stand for a project-specific or descriptive method whose acronym is not standardized in the source paper [(Laszlo, 2011); (Giannoccaro et al., 2014)].