esc

Works

2026

Manuscript cs.LO

Towards a Quantitative Operational Semantics

L. Melodia · Manuscript.

Over quasi-pseudometric spaces, a GSOS law gives every model a least distance for which its coalgebra is nonexpansive, and every context is then nonexpansive with the modulus of its symbols. No final coalgebra is needed, so bisimilarity is a congruence even for unbounded branching, and similarity a precongruence for positive rules. For probabilistic specifications, Kantorovich envelopes compute the moduli from the rules, and for the standard process algebra no modulus does better.

Manuscript math.CT

A Homology Theory for Programming Languages

L. Melodia · Manuscript.

The free models of a GSOS law resolve every model and give it a comonad homology. For polynomial behaviour, the Beck modules are the representations of a linearisation, generated by partial derivatives and transitions under a chain rule. With coefficients in the differentials, a complex of length two drawn on the model computes the homology, and its first group embeds into the relations of a presentation modulo their consequences. Constant coefficients exist exactly for linear rules, and they bound from below the number of relations of every presentation.

Preprint math.AT v2

A Note on Laminar Presentations and Finite Chain Models for Ample Groupoid Homology

L. Melodia · hexagon:2610.00218 · with a Lean 4 formalisation.

A clopen partition of each nerve space cuts a finite chain model out of the Matui complex of an ample groupoid G. Helly axioms on the covers constrain nothing, while laminar presentations bound the rank of a model by the size of the cover and, for second countable G, give a complete system of models. No single model suffices in general: H0 over Z is not finitely generated for the Cantor set.

Manuscript math.CT

Turi–Plotkin Homology: Notes for the Talk

L. Melodia · Manuscript.

For a simple distributive law λ:ΣB⇒BΣ, the comonad G=FU of free bialgebras resolves every model, and a coefficient functor turns the resolution into a chain complex. For linear rules with observational coefficients its homology reduces to two groups, which one Smith normal form computes. In degree one it finds laws that hold in a model but not in the free one, such as ρ(ρ(x))=x.

Preprint math.AT v2

Persistent Magnitude Homology for Quantitative Equational Theories

L. Melodia · arXiv:2608.21479.

A quantitative equational theory U gives every metric space A a free algebra TUA. Filtering its length nerve by length yields a persistence module whose associated graded is the magnitude complex, and a long exact sequence joins the two. The barcode in degree n moves by at most (n+1)δ under a perturbation of size δ, and the factor n+1 cannot be improved.

Preprint cs.LG v4

Learning Transfers: Kan Extensions of Topological Invariants of Learned Representations

L. Melodia · arXiv:2606.07627.

A change of task is a functor J, and the invariants of a learned representation form a functor F. The left Kan extension LanJF is the invariant that the source forces on the target, and its largest distance to the observed invariant is the transfer discrepancy. For persistent homology a functorial version is 1-Lipschitz under interleavings, and on finite samples the discrepancy is a bottleneck distance that is computed exactly.

Preprint math.AT v2

Discrete Coefficients and Open Invariant Covers in the Homology of Ample Groupoids

L. Melodia · arXiv:2603.20861.

The map C∙(G;Z)⊗ZA→C∙(G;A) is always injective, and surjective exactly when every compactly supported continuous map into A is locally constant. The universal coefficient theorem is therefore about discrete coefficients, and it fails for R on the Cantor set. Clopen invariant covers make the Mayer–Vietoris sequence trivial, while open invariant covers give one for every coefficient group.

Peer-Reviewed math.AT v3

Groupoid Homology and Classifying-Space Homology Are Not Isomorphic

L. Melodia · Extracta Mathematicae, to appear · arXiv:2602.13375.

Matui-type homology is built from compactly supported, locally constant chains on the nerve. For a discrete group it is the homology of the classifying space, but for a space seen as a groupoid of units it computes compactly supported cohomology. For the groupoid G of units of the Cantor set X, the two differ already in degree zero: H0(G;Z)≅C(X,Z) is countable, while H0sing(BG;Z) has cardinality 2ℵ0. In every positive degree they agree.

Thesis math.AT v4

Universal Coefficients and Mayer–Vietoris Sequence for Groupoid Homology

L. Melodia · Master’s thesis · arXiv:2602.08998.

Homology of ample groupoids with coefficients in a topological abelian group A is functorial and invariant under Kakutani equivalence. On the nerve spaces Gn, the image of the map Cc(Gn,Z)⊗ZA→Cc(Gn,A) consists of the compactly supported maps with finite image. For discrete A the map is an isomorphism, and the universal coefficient sequence follows. Clopen saturated covers of the unit space give a Mayer–Vietoris sequence.

2024

Thesis math.AT v1

Algebraic and Topological Persistence

L. Melodia · Bachelor’s thesis · arXiv:2410.08323.

Chain complexes, simplicial and singular homology, exact sequences and relative homology lead up to excision, which shows that simplicial and singular homology agree on triangulable spaces. On this basis the thesis develops persistent homology and persistent cohomology of filtrations of point clouds, with many proofs written anew to make the connections between the techniques visible.

2021

Peer-Reviewed cs.LG v6

Homological Time Series Analysis of Sensor Signals from Power Plants

L. Melodia, R. Lenz · ECML PKDD 2021 Workshops, CCIS 1524, Springer, pp. 283–299.

Sliding windows turn sensor signals into point clouds, and their Vietoris–Rips persistence gives Betti curves in degrees 0 and 1. A network with three branches of convolutions and LSTMs reads the raw signal and both curves and assigns each signal its designation in the plant. On 18,000 signals from four power plants it reaches an accuracy of 63.8% on the full designation.

Peer-Reviewed stat.ML v13

Estimate of the Neural Network Dimension Using Algebraic Topology and Lie Theory

L. Melodia, R. Lenz · ICPR 2020 Workshops, LNCS 12665, Springer, pp. 15–29.

If the data live on a connected abelian Lie group, that group is Tq×Rp, and its Betti numbers (qk) determine q. Reading them off persistence landscapes estimates the dimension of an embedding that keeps the topology of the data, and with it a lower bound for the number of neurons that a network layer needs to learn this topology.

2020

Peer-Reviewed cs.CG v16

Persistent Homology as Stopping-Criterion for Voronoi Interpolation

L. Melodia, R. Lenz · IWCIA 2020, LNCS 12148, Springer, pp. 29–44.

Voronoi interpolation densifies a sample by natural neighbours, and sooner or later it changes the topology of the data. Persistent homology is computed after every step, and the interpolation stops once the bottleneck and Wasserstein distances to the original persistence diagram exceed a threshold. Experiments on handwritten signatures from the MOBISIG database compare three filtrations.

2018

Thesis stat.ML v12

Deep Learning Estimation of Absorbed Dose for Nuclear Medicine Diagnostics

L. Melodia · Master’s thesis · arXiv:1805.09108.

In radionuclide therapy with 177Lu, a single dose kernel ignores differences between tissues, and Monte Carlo transport is slow. A U-residual network learns the map from local density to dose kernel instead, so that density-adapted kernels need no transport simulation for each patient. On unseen patients it reaches a continuous intersection over union of 0.86.

2015

Peer-Reviewed linguistics

brauchen mit und ohne zu + Infinitiv vor dem Hintergrund des sozialen Kräftefeldes

L. Melodia · In K. Šichová, R. Krapp, P. Rössler and V. Dovalil (eds.), Standardvarietät des Deutschen, Logos, Berlin, pp. 83–96.

A sociolinguistic case study, in German, of brauchen with and without zu before an infinitive. The variation between the two forms is read against Ammon’s social field of forces of a standard variety, in which codifiers, model speakers and writers, language experts and norm authorities together decide what counts as standard German.

Code

GSOSQPMet Towards a Quantitative Operational Semantics. TeX GSOSHomology A Homology Theory for Programming Languages. TeX TuriPlotHom Turi-Plotkin Homology. TeX pages My private website. HellyHomology A note on laminar presentations and finite chain models. TeX PersHomAlg Persistent Homology for Quantitative Algebras. TeX CV This is my personal résumé. TeX ANKI A collection of my ANKI decks, mostly mathematics. CatTrans A categorical framework for transfer learning. TeX TwirlFlake Homological Time Series Analysis of Sensor Signals from Power Plants. Python SIML Persistent Homology as Stopping-Criterion for Voronoi Interpolation. TeX NTOPL Estimation of Neural Network Dimension using Algebraic Topology and Lie Theory. Python MADVK Deep Learning Estimation of Absorbed Dose for Nuclear Medicine Diagnostics. TeX BSc Algebraic and Topological Persistence. TeX MSc Universal Coefficients and Mayer-Vietoris Sequence for Groupoid Homology. TeX WengChun Weng-Chun Erlangen website. CSS UCT-Mayer-Vietoris Presentation for UCT and Mayer-Vietoris Sequence for Groupoid Homology. TeX Pomodoro A Rust implementation of the ToMATo algorithm. Rust Homolipop Persistent homology for rings and persistent k-theory modulo finite fields. Python Novikovbon Exact persistent homology over finite fields from nested matrices, implemented rigorously and efficiently in Rust. Rust ToroidalLearn A toroidal neural network layer for Keras. Operate on an n-torus, the natural framework for all types of signals. Python Talks Talks I have given at various occasions. KnowDisc Knowledge Discovery in Databases. TeX