Multiparameter persistence in Julia
TamerOp
TamerOp is the Toolkit for Algebraic Module Encodings over $\mathbb{R}^n$ and Other Posets, a Julia library for studying how mathematical objects change with one or more parameters.
A persistence module records a vector space at each parameter and linear maps that follow its vectors as the parameters change. We describe it through a finite encoding: spaces and maps on finitely many labels, together with an assignment from parameters to labels. These labels form a partially ordered set, or poset, recording which ones can be compared. The finite description lets us recover the original spaces and maps.
If you already have an input in mind, construct its module and explore the maps .
Computing with a finite encoding
Different inputs lead to spaces and maps that we can study together.Begin with
Build from filtered data Points, images, graphs, or cells with grades Define a module directly Sloped polyhedral regions, presentations, or finite posetsStart from a filtration
As we increase a threshold, we can add points, edges or cells to a shape. This growing family is a filtration. Its homology records connected components, loops around holes and their higher dimensional counterparts, together with the maps that follow them through the growth. A finite encoding keeps the resulting persistence module available for further study.
For example, a compatible family of linear maps between two modules is called a module morphism. We can examine its kernel, the vectors it sends to zero, or its image, the vectors it reaches. A resolution describes a module through generators, relations and relations among those relations. These algebraic questions apply to modules obtained from filtered data too.
When the question concerns the birth and death of classes along a single parameter, the direct persistence routines compute their intervals from the filtered complex. The ring lesson begins with this calculation.
Describe a module directly
We can also specify the spaces and maps directly, using geometric regions or a finite poset. Consider a module whose space is one dimensional on the diagonal strip $0 \leq x+y \leq 1$ and zero elsewhere, with identity maps between comparable parameters inside the strip. Its exact region encoding retains the sloped boundaries and describes the spaces and maps throughout the plane.
Once this module is encoded, the same operations on the finite model let us study it. The starting description can therefore come from an algebraic or geometric question about modules themselves.
Try both routes in From inputs to computed objects. The explanation of finite encodings develops the connection to Ezra Miller's theory of modules over posets.
Performance you can examine
The benchmark reports compare matched mathematical requests, with independently checked answers, timings, memory measurements, and downloadable results.
The QPA comparison measures algebra on finite modules, where TamerOp takes substantially less time for the matched requests. The PHAT comparison measures complete ordinary barcodes over the field F₂, whose arithmetic is modulo two. Its aggregate difference is smaller, and individual cases favor each tool.
The Ripser.py comparison covers ordinary Rips persistence, including computations on selected subsets of points, called landmarks, and retained cocycles, which evaluate homology classes. On the 1,024-point circle, the query takes 7.01 seconds with TamerOp and 30.4 seconds with Ripser.py. Larger planar inputs also favor TamerOp; landmark results are mixed, and some sparse and small requests favor Ripser.py.
The separate original C++ Ripser comparison studies larger requests for barcodes. TamerOp is about 1.8 times as fast on the two circle inputs; most dense planar cases and the larger sphere perform similarly. C++ Ripser is 1.6–1.8 times as fast on the sparse spheres. These observations were made on a shared machine, with substantial variation in absolute times.
The CubicalRipser comparison studies large images and volumes. TamerOp is 2.18 times as fast in the aggregate for the fixed set of requests. For a smooth 256³ array whose values specify when voxels appear, the request takes 6.67 seconds with TamerOp and 14.2 seconds with CubicalRipser. Binary volumes whose values are assigned to vertices favor CubicalRipser, while smooth 2D cases are much closer. The report explains these input conventions and how the returned barcodes were compared.
The two Ripser studies use different workloads and timing boundaries. Each report states its inputs, versions, timing conditions, and limits, so you can judge what the evidence means for your work.
For the algorithms behind the results, read the implementation accounts.
From theory to application
The documentation moves from understanding the mathematics, through exploring the library, to completing a particular task. These are independent entrances: start with the kind of answer you need.
Find a route through the ideas in the learning map, browse a subject in the topic map, or look up an operation in API reference.