Skip to content

Tameness and scope: when finite descriptions exist

The chapter on indicator presentations described the square module using two regions and a coefficient matrix. The chapter on finite encodings recovered the same module from a finite poset, its spaces and maps, and a classifier on the parameter plane. When can a module have descriptions like these?

The square suggests that we should look for regions on which the same spaces and maps can be used throughout. To understand what this requires, we will describe its regions more carefully and then compare it with a module whose spaces all have dimension one. That second module keeps changing its maps as we move along the real line, and those changes will prevent any finite encoding.

Our finite constructions already supply the condition we need. The next chapter, why finite computations stay tame, explains that practical argument and the role of abelian categories. Here we establish exactly what the condition means and which descriptions it guarantees.

You need the preceding chapters' definitions of modules, finite encodings, and fringe presentations. We use a fixed coefficient field $\mathbb{k}$; take $\mathbb{k}=\mathbb{Q}$ in the examples. The general statements apply to an arbitrary parameter poset $Q$.

What should it mean for a module to be constant on regions?

Recall the square $S=[0,2]^2$ in $\mathbb{R}^2$. Its module $M$ has $M(q)=\mathbb{k}$ inside $S$ and $M(q)=0$ outside. A structure map between comparable points inside is the identity; every other structure map is the uniquely determined zero map of the appropriate shape.

There are two kinds of spaces here. Partition the plane into

\[R_1=S,\qquad R_0=\mathbb{R}^2\setminus S.\]

Use $\mathbb{k}$ as the common space for $R_1$ and $0$ for $R_0$. Once these spaces are identified with the stalks, the maps depend only on which two regions contain their endpoints:

Source regionTarget regionMap, whenever the parameters are comparable
$R_1$$R_1$Identity on $\mathbb{k}$
$R_1$$R_0$Zero map $\mathbb{k}\to0$
$R_0$$R_1$Zero map $0\to\mathbb{k}$
$R_0$$R_0$Unique map $0\to0$, also its identity

The table tells us how to recover each map once we know the endpoint regions. A partition by stalk dimensions would leave those maps unspecified.

For a general module, choose a partition $Q=\bigsqcup_i R_i$, a vector space $V_i$ for each region, and isomorphisms $\alpha_q:V_i\to M(q)$ for $q\in R_i$. It is a constant subdivision if, for every ordered pair $(R_i,R_j)$, the map

\[\alpha_r^{-1}\,M(q\leq r)\,\alpha_q:V_i\longrightarrow V_j\]

is independent of the chosen comparable pair $q\in R_i$, $r\in R_j$. We require this even when $i=j$, including the comparison of a point with itself. This is the compatibility condition in Ezra Miller's Definition 2.6.

To read the equation, start with a vector in $V_i$ and follow it through $M(q)$ and $M(r)$. The last isomorphism expresses the answer in the common space $V_j$. We use the same identifications for every comparison, so the answer must agree even when we choose different endpoints in the two regions. Choosing a fresh pair of bases for each arrow would evade that requirement. The cited definition calls this compatibility no monodromy.

For points in the same region, comparison with a self-map forces the displayed map to be the identity on $V_i$. Thus any structure map between comparable points of that region must be an isomorphism. For incomparable points there is no structure map to require. Regions need not be connected, convex, or bounded; the square's entire exterior is already one region in this subdivision.

The notion of tameness used here

A module is pointwise finite-dimensional if each stalk $M(q)$ has finite dimension. Ezra Miller also calls this $Q$-finite; it does not mean that the parameter poset $Q$ is finite.

In this chapter, tame means pointwise finite-dimensional and admitting a constant subdivision with finitely many regions. Both requirements belong to Definition 2.11.

The square is tame because the two regions above work and their common spaces have dimensions one and zero. This describes the module on all of $\mathbb{R}^2$, including its unbounded exterior. The word “finite” counts the regions and the dimensions of their spaces; it places no bound on the parameter domain.

The two requirements control different things. Finite-dimensional stalks limit the size of each individual space. A finite constant subdivision limits how many descriptions of spaces and maps are needed across the domain. Together they give a uniform bound on stalk dimensions: take the largest of the finitely many dimensions $\dim V_i$. The counterexample below will show that even a uniform bound of one does not suffice by itself.

A constant subdivision is not yet an encoding

The square's two regions describe its constant data, but we cannot use those same two pieces as the fibers of a map that preserves order. To see the obstruction, follow the three comparable points

\[(-1,1)\leq(1,1)\leq(3,1).\]

The first and last points are exterior, while the middle point is inside. If $z$ labels the exterior and $a$ labels the square, monotonicity would force $z\leq a\leq z$. Antisymmetry then gives $z=a$, incompatible with the different stalk dimensions.

The constant subdivision allowed us to record the map for each pair of region names without ordering the names themselves. An encoding requires that order, because its labels must support a module whose structure maps recover the original ones. We therefore need to separate some parts of the exterior.

For this square we already have a suitable refinement. Record

\[u(x,y)=\mathbf{1}_{\{x\geq0,\ y\geq0\}},\qquad c(x,y)=\mathbf{1}_{\{x>2\ \text{or}\ y>2\}}.\]

Both bits increase with the parameters, so $\pi(x,y)=(u(x,y),c(x,y))$ is order-preserving into $\{0<1\}^2$. Its four fibers split the exterior into three pieces and retain the square as the fiber labeled $(1,0)$. Give that label the space $\mathbb{k}$ and the other three labels zero. This is the encoding on four labels developed earlier.

Description of the same square moduleWhat is finite?
Constant subdivisionTwo regions with compatible identifications of spaces and maps
Signature encodingFour labels, their ordered module, and the classifier $\pi$
Encoding by nine regionsNine labels, their ordered module, and the band classifier
Fringe presentationOne upset, one downset, and the coefficient matrix $[1]$

Each count refers to a different part of the description, so comparing two regions with four labels tells us nothing about the number of persistent classes. We have constructed useful descriptions without claiming that any of them is the smallest possible one.

Three equivalent ways to have finite data

For a fixed field and a pointwise finite-dimensional module over any poset, the following are equivalent:

  1. It admits a finite constant subdivision.
  2. It admits a finite encoding: $M\cong\pi^*N$ with $\pi:Q\to P$ order-preserving, $P$ finite, and every $N(p)$ finite-dimensional.
  3. It admits a finite fringe presentation: an image of a map from a finite sum of upset indicators to a finite sum of downset indicators, with one scalar on each component's intersection.

The first two conditions are related by Theorem 4.22; the third is included in Ezra Miller's syzygy theorem, Theorem 6.12. The pointwise finite-dimensional hypothesis is explicit here: a finite number of constant regions alone places no dimension bound on their common spaces.

These are existence statements about the same module. They do not identify the regions in one description with the labels or summands in another. Here is how the constructions fit together.

From an encoding to a constant subdivision

Take the nonempty fibers $\pi^{-1}(p)$. Identify each stalk in that fiber with $N(p)$ using the given module isomorphism. A map between fibers is then $N(p\leq p')$, independent of which comparable original parameters were chosen. There are only finitely many fibers, so they give a finite constant subdivision.

From a fringe presentation to an encoding

For each upset, record membership; for each downset, record membership in its complement. This produces finitely many increasing bits. Their patterns form a finite ordered set. At each pattern, select the active rows and columns of the coefficient matrix and take its image. Between comparable patterns, restrict the appropriate projection onto the remaining target coordinates to the image.

The chapter on indicator presentations explained why these spaces and maps recover the presented module. The construction uses the fixed scalar coefficients as well as the regions; membership patterns alone would lose the linear algebra.

From a constant subdivision to an encoding

Here we use the finite encoding theorem to organize the compatible regional data into a finite ordered model. The construction forms upsets from the constant regions and uses membership in those upsets to obtain labels that respect the original order.

The conclusion concerns an encoding of the module. Its fibers may differ from the given pieces, so we should not assume that an arbitrary chosen subdivision will survive the construction as a refinement. For our square, we can see directly that the four fibers above do refine the original two regions.

Why a finite encoding also yields a fringe presentation

There is a finite construction using linear algebra that explains this direction. Start with the finite module $N$ on $P$. At a label $x$, form

\[F(x)=\bigoplus_{p\leq x}N(p),\qquad E(x)=\bigoplus_{r\geq x}N(r).\]

After choosing a basis for every $N(p)$, $F$ is a finite sum of principal upset indicators and $E$ a finite sum of principal downset indicators. “Principal” means the upset $\{x:p\leq x\}$ or the downset $\{x:x\leq r\}$ generated by one label.

Define maps

\[F(x)\xrightarrow{a_x}N(x)\xrightarrow{b_x}E(x),\]

where $a_x((v_p)_p)=\sum_{p\leq x}N(p\leq x)v_p$, and $b_x(v)=(N(x\leq r)v)_{r\geq x}$. The summand indexed by $x$ makes $a_x$ surjective and $b_x$ injective: in each case that coordinate uses the identity of $N(x)$. The maps commute with structure maps, so the image of $b\,a$ is isomorphic to $N$ as a module.

For a source summand indexed by $p$ and a target summand indexed by $r$, the composite at any $p\leq x\leq r$ is

\[N(x\leq r)\,N(p\leq x)=N(p\leq r).\]

Its matrix is independent of $x$. Thus every scalar component is constant on its upset–downset intersection, as a fringe presentation requires. If $p\not\leq r$, that intersection is empty and the component is zero. Pulling the construction back along $\pi$ replaces these regions by their preimages and gives a finite fringe presentation of $M$. This establishes existence; it need not give the smallest presentation. If a pulled-back intersection is empty, set its unused coefficient entries to zero.

A counterexample with dimension one everywhere

For this example only, use the one-parameter domain $\mathbb{R}$. Keep $\mathbb{k}=\mathbb{Q}$. Put

\[V(t)=\mathbb{k}\quad\text{for every }t\in\mathbb{R},\]

and, for $s\leq t$, define

\[V(s\leq t)= \begin{cases} \operatorname{id}_{\mathbb{k}},&\lfloor s\rfloor=\lfloor t\rfloor,\\ 0,&\lfloor s\rfloor\ne\lfloor t\rfloor. \end{cases}\]

Here $\lfloor t\rfloor$ is the integer $n$ satisfying $n\leq t<n+1$. The map is the identity within one unit interval and zero whenever it crosses an integer boundary. For example,

\[V(\tfrac14\leq\tfrac34)=[1],\qquad V(\tfrac34\leq1)=[0],\qquad V(1\leq\tfrac32)=[1].\]

We can check the module axioms directly. If $s\leq t\leq u$, their floors are nondecreasing. When the endpoint floors agree, the middle one must agree too, so all the maps are identities. When the endpoint floors differ, at least one of the two successive maps is zero, making their composite zero just like the direct map. A point always has the same floor as itself, which gives the identity self-map.

Equivalently, writing $I_n=[n,n+1)$,

\[V\cong\bigoplus_{n\in\mathbb{Z}}\mathbb{k}[I_n].\]

Here $\mathbb{k}[I_n]$ is the indicator module of the half-open interval $[n,n+1)$. Exactly one summand is present at each parameter. At an integer, one summand ends and the next begins, so the stalk is still $\mathbb{k}$.

The dimension is one throughout the displayed interval. Half-open bars show successive unit intervals. The selected parameters one quarter and three quarters both have label zero and structure matrix [1]; three quarters and one have labels zero and one and structure matrix [0].

The upper line records dimensions, while the bars below show which summand supplies the space. At $t=1$ the first bar ends and the next begins, so the stalk remains $\mathbb{k}$ even though the map from $3/4$ is zero. Filled and open endpoints mean included and excluded, respectively. This is a schematic of a finite window; the unit intervals continue in both directions.

Why no finite constant subdivision can work

Consider the infinitely many parameters $t_n=n+\tfrac12$ for integers $n\geq0$. In any partition into finitely many regions, two of them, say $t_n<t_m$, must belong to the same region.

Their structure map is zero, because they lie in different unit intervals. Changing bases cannot turn a zero map into an isomorphism. But two comparable points in one constant region must have an isomorphism between their stalks: their transported map must agree with the identity self-map. This is a contradiction. Hence $V$ is not tame.

There is also a direct obstruction to a finite encoding. Two of the $t_n$ would have to receive the same finite label $p$. Their map would then be recovered from $N(p\leq p)=\operatorname{id}_{N(p)}$ and would be an isomorphism, again contradicting the zero map. The equivalence theorem also rules out a finite fringe presentation.

The dimension remains one everywhere, but a finite description would have to identify two of these intervals in a way their zero map forbids. The dimension plot hides exactly the distinction used in the proof.

A short rule and a finite window are different claims

The floor formula describes $V$ in a few symbols. It even gives an encoding through the poset $\mathbb{Z}$, with $\mathbb{k}$ at every integer and zero maps between distinct integers. That encoding is infinite. A short program or formula is a different notion of finite description from the finite encodings considered here.

If we restrict the domain to a fixed bounded interval, only finitely many unit intervals meet it. Restricting the floor classifier to that finite set of integer labels gives a finite encoding of the restricted module. The restriction is tame, while the module on all of $\mathbb{R}$ is not. If we widen the window, more integer labels appear while the dimension profile stays flat. Every bounded restriction still uses only finitely many labels. The obstruction on the whole line comes from needing infinitely many of them at once, as the preceding proof shows.

Distinguish this from q-tameness

In one-parameter persistence, q-tame usually means that every map $V(s\leq t)$ with $s<t$ has finite rank; see Frédéric Chazal, William Crawley-Boevey, and Vin de Silva, Section 1.4. Our counterexample has ranks zero or one, so it is q-tame. It is not tame in the sense of finite constant subdivisions used in this chapter. The two terms impose different requirements.

From an existence theorem to an implemented construction

The theorem concerns modules over arbitrary posets and allows abstract upsets, downsets, and classifiers. To compute, we must also provide a representation of those ingredients and methods for evaluating membership and order. Finitely many sets need not come with finite geometric descriptions or an algorithm for testing membership.

The square supplies both kinds of information: its finite presentation has the coefficient $[1]$, and four explicit inequalities describe its regions. The selected encoder can construct the finite model and classify queries. To use the theorem computationally, we need this additional representation of the regions. It does not supply a general procedure that accepts an arbitrary module and decides whether it is tame.

Some concrete TamerOp entry routes are:

Starting informationWhat the package works with
A finite poset, finite-dimensional spaces, and compatible mapsA finite module such as TamerOp.Advanced.PModule, inspected and validated directly. It is already a finite representation.
The square's box regions and coefficient matrixencode(upsets, downsets, coefficients, options) produces an EncodingResult for the selected supported construction.
A supported polyhedral presentationPresentation encoding using the applicable geometric backend, with that backend's generator and arithmetic requirements.
Typed data together with a supported filtrationData ingestion through encode, requesting the completed encoding stage when an EncodingResult is wanted.

In the first row, the identity map of the finite poset already gives an encoding in the mathematical sense, and the software can operate on the finite module directly. There is no need to turn that observation into an encode(module) call.

For geometric presentations, consult the geometry contracts. For data, the ingestion guide explains filtration, degree, grid, truncation, and choices of intermediate stages. These determine which module is represented. A finite encoding recovers that chosen module; it does not undo a change made when constructing the input.

The coefficient field and geometric arithmetic are separate choices. Rational coefficients make the stated linear algebra exact without automatically making every geometric input exact. The exact-grade guide describes guarantees for its specific construction. Likewise, neither the existence theorem nor a successful example asserts that the returned encoding is minimal or inexpensive for every supported input.

The full syzygy theorem also gives finite indicator resolutions, which organize successive relations among simpler modules. To use a resolution for a derived computation, we still need to know which category contains its terms and maps. In particular, Ext and Tor over the finite base require the category and comparison hypotheses described in their guide.

Why do finite constructions meet this condition?

The square has now given us a concrete version of each description in the theorem, and the module on the real line has shown why the maps have to be part of the finiteness condition. We can next ask how much of this reasoning needs to be repeated when we work with a new finite input or perform another calculation on an encoded module. Continue with why finite computations stay tame for the argument using a finite complex, a calculation on the square, and the closure theorems of Ezra Miller and Lukas Waas.