betaFeedback

WorksMathematics

Syzygy matrices and higher extensions over string algebras

Abstract

Over a finite-dimensional string algebra, higher extension dimensions may eventually vanish, become periodic, or grow without bound. We show how their behaviour can be read from a finite directed graph. Second syzygies split into uniserial modules from a finite family; their syzygies define a nonnegative integer matrix, so higher extension dimensions become weighted counts of walks. Only the part of the graph relevant to the chosen pair of modules contributes. This gives four growth regimes: eventual vanishing, periodicity, quasi-polynomial growth, and exponential growth. We derive the construction from projective covers, give explicit extension formulas over truncated cycles, and illustrate the growth regimes with examples.

Introduction

For gentle algebras, higher extensions have a rather simple long-term behaviour. Baur and Schroll showed that, over an algebraically closed field, the extension dimensions for any two finite-dimensional modules eventually either vanish or repeat periodically [BS21]. Their description also has a geometric interpretation on the surfaces associated with gentle algebras. Extensions of degree one had already been described rather explicitly, in terms of strings and their overlaps [2], and, for algebras from triangulations of unpunctured marked surfaces, in terms of crossings of curves [3]. It is natural to ask what remains of this picture for the larger class of string algebras.

The answer is not simply more complicated periodicity. Polynomial and exponential growth can occur, sometimes over very small quivers. For example, a radical-square-zero algebra with two loops at one vertex has dim⁡kExt⁡An(S,S)=2n\dim_{\mathbb{k}}\operatorname{Ext}_{A}^{n}(S,S)=2^{n}, whereas two looped vertices joined by one arrow already give dim⁡kExt⁡An(M,N)=n+1\dim_{\mathbb{k}}\operatorname{Ext}_{A}^{n}(M,N)=n+1. We will see both examples below. What is perhaps less obvious is that the behaviour depends on the target as well as on the source: a projective resolution may grow, while its extension spaces with a particular module stay bounded or even vanish.

One reason to study these numbers is that they measure how a projective resolution grows. For T=A/rad⁡AT=A/\operatorname{rad} A, the number dim⁡kExt⁡An(M,T)\dim_{\mathbb{k}}\operatorname{Ext}_{A}^{n}(M,T) is exactly the number of indecomposable projectives in its nn-th term. A general target NN sees only part of this resolution. In Example 4.6, for instance, a single source gives the three sequences 11, nn, and (n2)\binom{n}{2}, depending on the target. In the bounded derived category these are the dimensions of the morphism spaces from MM to the shifts N[n]N[n]. Thus even when the modules themselves have a concrete description, their higher extensions can have quite different sizes.

Computing the first few terms may suggest a pattern, but it does not say when that pattern begins, whether it persists, or which part of the algebra is responsible for it. A finite graph answers these questions without requiring us to construct an infinite resolution. It also separates two sources of growth that will recur in the examples: a syzygy can split repeatedly within one family, or it can keep sending new summands into another.

For string algebras, the work of Huisgen-Zimmermann and Huisgen-Zimmermann–Smalø provides the key observation: after two syzygies one meets only uniserial string modules, and there are finitely many possibilities [8 HZS05]. We take these modules as the states of a finite system. Their syzygies give a nonnegative integer matrix BB; two vectors, one coming from MM and one from NN, then give

dim⁡kExt⁡Ai(M,N)=vMTBi−3hN(i≥3).(1)\dim_{\mathbb{k}}\operatorname{Ext}_{A}^{i}(M,N)=v_{M}^{\mathsf{T}}B^{i-3}h_{N}\qquad(i\geq3). \tag*{(1)}

Thus the extension dimensions count walks in a directed graph. When the cyclic components are directed cycles, staying in one gives periodic behaviour and passing through several can give polynomial growth. Branching within a strongly connected component gives exponential growth. Only the part of the graph which connects the two vectors matters.

There is a related connection with automata in the representation theory of string algebras. Rees constructs finite automata whose accepted words describe strings [11]. In the graph used here, a step instead records a syzygy summand. The extension formula is a weighted-automaton calculation, in the sense of [BR11, §1.6].

The use of a syzygy quiver goes back to Cibils and was developed further by Howard, especially in the study of complexity and growth of projective resolutions [Cib93; How10; How15]. We recall the projective-cover calculation behind (1.A), give the growth alternatives, and then return to gentle algebras, truncated cycles, and linked cycles. This article presents partial results from the author’s already submitted dissertation [10], with the truncated-cycle computations written in a more general form. The surface interpretation is treated there.

Use of AI. OpenAI’s GPT-6 was used to improve the writing of this paper.

Preliminaries

Throughout, k\mathbb{k} is a field, QQ is a finite quiver and A=kQ/IA=\mathbb{k}Q/I, where Jm⊆I⊆J2J^{m}\subseteq I\subseteq J^{2} for some m≥2m\geq2 and JJ is the arrow ideal. Paths are composed from left to right; all modules are finite-dimensional right modules. Write Px=exAP_{x}=e_{x}A, Sx=Px/rad⁡PxS_{x}=P_{x}/\operatorname{rad}P_{x}, and T=A/rad⁡A=⨁x∈Q0SxT=A/\operatorname{rad}A=\bigoplus_{x\in Q_{0}}S_{x}. In particular, each vertex simple is one-dimensional over k\mathbb{k}; no assumption that k\mathbb{k} is algebraically closed is needed.

A projective cover is a surjection from a projective whose kernel lies in its radical. Every module has a projective cover, unique up to isomorphism, and a decomposition into indecomposables, unique up to order and isomorphism. We use these facts in the form of [ARS95, I, Theorem 4.2, II, Theorem 2.2]. Write the cover sequence as

0⟶ΩX⟶P(X)⟶X⟶0.0\longrightarrow\Omega X\longrightarrow P(X)\longrightarrow X\longrightarrow0.

Iterating gives the minimal projective resolution of XX. The following facts explain how we will read extension dimensions from it.

Lemma 2.1 (Projective covers and extensions). For finite-dimensional AA-modules XX, YY the following hold.

  1. Syzygies are additive: Ω(X⊕Y)≅ΩX⊕ΩY\Omega(X \oplus Y) \cong\Omega X \oplus\Omega Y; projective summands disappear. In particular, Ext⁡Ai(P,Y)=0\operatorname{Ext}^{i}_{A}(P,Y)=0 for i≥1∗if∗i \ge1 *if*P$ is projective.

  1. For i≥1i \ge1, dimension shift gives

Ext⁡Ai+1(X,Y)≅Ext⁡Ai(ΩX,Y).\operatorname{Ext}^{i+1}_{A}(X,Y) \cong\operatorname{Ext}^{i}_{A}(\Omega X,Y).

Moreover,

Ext⁡A1(X,Y)≅coker⁡(Hom⁡A(P(X),Y)⟶Hom⁡A(ΩX,Y)).(2)\operatorname{Ext}^{1}_{A}(X,Y) \cong\operatorname{coker}\bigl(\operatorname{Hom}_{A}(P(X),Y) \longrightarrow\operatorname{Hom}_{A}(\Omega X,Y)\bigr). \tag*{(2)}

Consequently,

dim⁡kExt⁡A1(X,Y)=dim⁡kHom⁡A(ΩX,Y)−dim⁡kHom⁡A(P(X),Y)+dim⁡kHom⁡A(X,Y).\dim_{k}\operatorname{Ext}^{1}_{A}(X,Y)=\dim_{k}\operatorname{Hom}_{A}(\Omega X,Y)-\dim_{k}\operatorname{Hom}_{A}(P(X),Y)+\dim_{k}\operatorname{Hom}_{A}(X,Y).
  1. If ⋯⟶P1⟶P0⟶X\cdots\longrightarrow P_{1}\longrightarrow P_{0}\longrightarrow X is a minimal projective resolution, then

Ext⁡Ai(X,T)≅Hom⁡A(Pi,T)(i≥0).\operatorname{Ext}^{i}_{A}(X,T)\cong\operatorname{Hom}_{A}(P_{i},T)\qquad(i\ge0).

Thus dim⁡kExt⁡Ai(X,T)\dim_{k}\operatorname{Ext}^{i}_{A}(X,T) counts the indecomposable summands of PiP_i, with multiplicity.

Proof. Direct sums of projective covers are projective covers [ARS95, I, Proposition 4.3]. Applying Hom⁡A(−,Y)\operatorname{Hom}_{A}(-,Y) to the cover sequence gives dimension shift and the exact sequence

0⟶Hom⁡A(X,Y)⟶Hom⁡A(P(X),Y)⟶Hom⁡A(ΩX,Y)⟶Ext⁡A1(X,Y)⟶0;0\longrightarrow\operatorname{Hom}_{A}(X,Y)\longrightarrow\operatorname{Hom}_{A}(P(X),Y)\longrightarrow\operatorname{Hom}_{A}(\Omega X,Y)\longrightarrow\operatorname{Ext}^{1}_{A}(X,Y)\longrightarrow0;

see also [ARS95, I, §5]. The cokernel and dimension formulas follow.

For the last assertion, minimality means that the image of each differential lies in the radical of the preceding projective. Every homomorphism from a projective to TT kills this radical. Hence the differentials vanish after applying Hom⁡A(−,T)\operatorname{Hom}_{A}(-,T). Here Hom⁡A(Px,T)≅Tex≅k\operatorname{Hom}_{A}(P_x,T)\cong T e_x\cong k, so each indecomposable projective contributes one to the dimension. □

For string modules we follow the construction in [BR87, §3]. The classification over an arbitrary field is also given in [CB18, Theorems 1.1–1.2]. We recall the part of the construction needed for the cover calculations.

The syzygy matrix

Strings and projective arms

Write s(α)s(\alpha) and t(α)t(\alpha) for the source and target of an arrow α\alpha. A string algebra has monomial relations, at most two arrows entering and at most two leaving each vertex, and at most one nonzero successor and one nonzero predecessor for each arrow. More explicitly, for each α\alpha there is at most one composable β\beta with αβ∉I\alpha\beta\notin I and at most one composable γ\gamma with γα∉I\gamma\alpha\notin I. It is gentle if its relations are quadratic and each arrow also has at most one composable zero successor and one composable zero predecessor. Throughout this section, AA is a string algebra.

A basis path is a path with nonzero image in AA. A string is a composable word in arrows and formal inverses, without immediate cancellation or a subword belonging to II or its inverse. Trivial strings are allowed. The module M(w)M(w) has one basis vector at each successive occurrence of a vertex in ww; an arrow in the word acts between the corresponding vectors in its own direction. All other arrow actions are zero. Occurrences remain distinct even when their vertex labels agree, and M(w)≅M(w−1)M(w)\cong M(w^{-1}).

A peak is a basis occurrence that survives in the top; a valley is one in the socle. The projective cover of a string module has one summand for each peak. A string with one peak is cyclic. A direct string pp gives a uniserial module M(p)M(p), whose submodules are totally ordered. Every uniserial string module has a direct representative after reversing the word if necessary.

For each arrow β\beta from xx, there is a unique maximal basis path beginning with β\beta. Call it the β\beta-arm of PxP_x. The path basis of PxP_x consists of exe_x and the nontrivial prefixes of its at most two arms. Distinct arms have disjoint path bases, even if they visit the same quiver vertices.

Lemma 3.1 (the uniserial kernel). Let pp be a direct basis path starting at xx. The cover Px↠M(p)P_x \twoheadrightarrow M(p) retains the initial segment specified by pp on its arm and kills every other arm. Its kernel is the direct sum of the remaining tails. In particular, ΩM(p)\Omega M(p) is a sum of at most two uniserial string modules.

Proof. The cover sends exe_x to the top generator. The prefixes of pp map to the successive basis vectors of M(p)M(p); all other basis paths are killed. On each arm, the killed paths form a final segment, generated by its first path. Unique continuation makes this segment uniserial. The segments have disjoint path bases and together span the kernel. For a trivial pp, both arms are killed and the same argument applies. □

In Figure 1, filled dots survive in the quotient and open dots form the kernel. An exhausted arm contributes zero, not a trivial string; a one-vector tail contributes a simple. The kernel vanishes exactly when pp exhausts the sole arm (or pp is trivial at a sink).

Figure 1. A uniserial cover and its kernel tails.

Example 3.2 (why arrow ideals alone need not suffice). Consider the bound quiver

Image not extracted

Figure 6.

The ideal aA≅M(b)aA \cong M(b) has cover Pt(a)P_{t(a)} with path basis et(a),b,bce_{t(a)}, b, bc. Its kernel is the simple kbckbc. Every arrow ideal has dimension two, so this kernel is not an arrow ideal. The uniserial state set must also contain simples.

Reduction after two syzygies

Set Ω0M=M\Omega^0 M=M, and write XX for the non-projective part of a module XX. The reason a finite matrix suffices is a theorem of Huisgen-Zimmermann for monomial algebras. We use its following consequence for string algebras.

Proposition 3.3 (second-syzygy reduction). Every second syzygy is a direct sum of ideals pApA, where pp is a positive-length basis path. In particular, it is a direct sum of uniserial string modules from a fixed finite family.

Proof. The monomial second-syzygy theorem [8 Theorem I], in the form [ZH92, Theorem A(1)], says that the syzygy of a submodule of a projective is a direct sum of principal ideals generated by positive-length paths. Apply it to ΩM⊆P(M)\Omega M \subseteq P(M), passing to the opposite algebra to obtain the right-module statement. The theorem holds over an arbitrary field.

For a string algebra, a positive-length path has at most one nonzero continuation at each step. The path basis of pApA therefore forms a single chain, so pApA is a uniserial string module. There are only finitely many basis paths. □

For a string module, we can see the first kernel directly in its cover. The description is that of [HZS05, Proposition 1(1)(a)]; the path-basis calculation below also works over an arbitrary field.

There is one projective for each peak of the string. Its basis paths follow the string until they reach an end or a valley; beyond that point they map to zero. At an interior valley, two paths pp, qq reach the same basis vector, so their difference ζ=(p,−q)\zeta=(p,-q) belongs to the kernel. Their final arrows are distinct. An outgoing arrow can therefore extend at most one of them, and every nonzero extension of either path belongs to ζA\zeta A.

The kernel is consequently the direct sum of the unused endpoint tails and these valley modules. For a trivial string, the two possible arms are its endpoint tails. They exhaust the killed paths and the differences, and have disjoint supports in the path basis of the cover. Here, as in the string itself, repeated vertex labels still represent distinct basis occurrences. Each summand has one peak, so its own cover has only arm tails in its kernel. This gives the second-syzygy reduction for strings. A valley module need not itself be projective.

Example 3.4 (a valley kernel need not be projective). Consider the two bound quivers in Figure 2. In both, the cover of M(ab−1)M(ab^{-1}) is P1⊕P3P_1\oplus P_3, and the two paths into the valley give the kernel generator ζ=(a,−b)\zeta=(a,-b). In (a), all arrows kill ζ\zeta, so

Figure 2. Valley kernels: (a) S2S_2; (b) M(d)M(d).

ΩM(ab−1)≅S2,dim⁡P2=2.\Omega M(ab^{-1})\cong S_2,\qquad\dim P_2=2.

In (b), ζd=(ad,0)≠0\zeta d=(ad,0)\ne0, and the kernel is M(d)M(d). The extra killed path adad is already part of the valley module. Both algebras are string but not gentle: cc has two zero predecessors.

The matrix and the two endpoint vectors

Let U\mathcal{U} contain one representative of each isomorphism class of non-projective uniserial string modules. It is finite because there are only finitely many basis paths. Lemma 3.1 shows that taking syzygies and deleting projective summands stays within direct sums of this family.

For an indecomposable VV, write (X:V)(X:V) for its multiplicity in XX. Define

BUV=(ΩU:V),B_{UV}=(\Omega U:V),
vM=((Ω2M:U))U∈U,hN=(dim⁡kExt⁡A1(U,N))U∈U.v_M=\bigl((\Omega^2 M:U)\bigr)_{U\in\mathcal{U}},\qquad h_N=\bigl(\dim_k\operatorname{Ext}^1_A(U,N)\bigr)_{U\in\mathcal{U}}.

Rows index source modules; columns index their syzygy summands. We regard vM,hNv_M,h_N as column vectors. The matrix depends on AA, while the two vectors record the chosen pair.

Theorem 3.5 (the extension formula). Every row of BB has sum at most two. For all M,NM,N,

dim⁡kExt⁡Ai(M,N)=vMTBi−3hN(i≥3).\dim_k\operatorname{Ext}^i_A(M,N)=v_M^{\mathsf T}B^{i-3}h_N\qquad(i\ge3).

Proof. Lemma 3.1 gives the row bound. By Proposition 3.3, vMv_M records all non-projective summands after two syzygies. Syzygy is additive and kills projective summands, so one further step multiplies the multiplicity row by BB. Induction gives vMTBnv_M^{\mathsf T}B^n for Ωn+2M\Omega^{n+2}M. Now apply dimension shift and additivity from Lemma 2.1.

The exponent begins at zero in degree three. The first two positive degrees require their own cover calculations. For hNh_N, use the cokernel or the dimension formula in Lemma 2.1; maps ΩU→N\Omega U \to N that extend to the cover do not contribute.

Example 3.6 (a complete calculation). Let A=k[x]/(x3)A = k[x]/(x^3), S=A/(x)S = A/(x) and U=A/(x2)U = A/(x^2). The two covers give ΩS≅U\Omega S \cong U and ΩU≅S\Omega U \cong S. In the order (S,U)(S,U),

B=(0110).B = \begin{pmatrix} 0 & 1 \\ 1 & 0 \end{pmatrix}.

Take M=S⊕U⊕2⊕AM = S \oplus U^{\oplus2} \oplus A and N=SN = S. Then Ω2M≅S⊕U⊕2\Omega^2 M \cong S \oplus U^{\oplus2}, so vM=(1,2)Tv_M = (1,2)^{\mathsf T}. Every map A→SA \to S kills rad⁡A\operatorname{rad} A; both restriction maps in {(2.A)} are zero. Since Hom⁡A(S,S)\operatorname{Hom}_A(S,S) and Hom⁡A(U,S)\operatorname{Hom}_A(U,S) each have dimension one, hS=(1,1)Th_S = (1,1)^{\mathsf T}. Hence

dim⁡kExt⁡Ai(M,S)=(1,2)Bi−3(1,1)T=3(i≥3).\dim_k \operatorname{Ext}^i_A(M,S) = (1,2)B^{i-3}(1,1)^{\mathsf T} = 3 \qquad(i \ge3).

The same covers give 3 in degrees one and two. Degree zero is 4: the projective summand contributes to Hom, but not to positive Ext.

Once BB is known, a new pair of modules requires only the two vectors vMv_M and hNh_N. For gentle string pairs over an algebraically closed field, the arrow and overlap basis of [2 Theorem A] is another way to compute hNh_N.

Corollary 3.7 (rational series). The truncated Poincaré series, regarded as a formal power series over Q\mathbb{Q}, is

∑i≥3dim⁡kExt⁡Ai(M,N)ti=t3vMT(I−tB)−1hN.\sum_{i\ge3}\dim_k \operatorname{Ext}^i_A(M,N)t^i = t^3v_M^{\mathsf T}(I-tB)^{-1}h_N.

It is rational, with denominator dividing det⁡(I−tB)\det(I-tB). The higher extension dimensions satisfy a linear recurrence.

Proof. Sum the matrix powers as a formal geometric series and use the adjugate formula. A recurrence follows by applying the characteristic polynomial of BB to its powers. □

The denominator can be smaller for a particular pair. In Example 3.6, det⁡(I−tB)=1−t2\det(I-tB)=1-t^2, but the truncated series is 3t3/(1−t)3t^3/(1-t). The syzygies alternate while the target gives both states the same weight. The constant tail 3 needs only a one-state matrix description, 3⋅(1)n⋅13\cdot(1)^n\cdot1. In general, linear minimization over Q\mathbb{Q} can reduce the number of states further [BR11, §2.3], though the resulting states need not represent modules or have nonnegative integral transition weights.

Growth from the syzygy graph

View a nonnegative integer matrix BB as the adjacency matrix of a directed multigraph, with BUVB_{UV} edges from UU to VV. Thus (Bn)UV(B^n)_{UV} counts walks of length nn, and

an=vTBnha_n = v^{\mathsf T}B^n h

counts them with initial and terminal weights. This is a weighted automaton over N\mathbb{N} with one letter: each reading of that letter applies BB, and the word of length nn has weight ana_n. We use the same walk-counting methods that appear in the growth theory of regular languages [Gaw+10].

A vertex is accessible if it can be reached from supp⁡v\operatorname{supp} v, and coaccessible if it can reach supp⁡h\operatorname{supp} h. We keep only vertices with both properties. Borrowing the terminology of automata theory [13], we call the resulting graph the trimmed graph for (v,h)(v,h), or the trimmed syzygy graph for (M,N)(M,N) when v=vMv=v_M and h=hNh=h_N. Restricting to it does not change ana_n.

We use the usual strongly connected components. A component is cyclic if it contains a cycle, and it is a directed cycle when its internal edges, counted with multiplicity, form a single cycle (possibly a loop). The component graph is acyclic.

The four alternatives

A sequence is eventually quasi-polynomial of degree dd if, on each residue class modulo some L>0L > 0, its sufficiently large terms agree with a polynomial of degree at most dd, and the degree is exactly dd on at least one class. Degree zero means eventual periodicity with a nonzero tail.

Lemma 4.1. Let CC be the adjacency matrix of a cyclic strongly connected directed multigraph. Then ρ(C)≥1\rho(C) \ge1, with equality precisely when the graph is a directed cycle.

Proof. Every row sum is at least one. If y>0y > 0 is a left Perron eigenvector, then yTC1≥yT1y^{\mathsf{T}}C\mathbf{1} \ge y^{\mathsf{T}}\mathbf{1}, with strict inequality whenever a row sum exceeds one. Thus ρ(C)≥1\rho(C) \ge1, with equality only when every row sum is one. Strong connectivity then gives a directed cycle [14 Chapter 1].

Theorem 4.2 (growth alternatives). Let an=vTBnha_n = v^{\mathsf{T}}B^n h for a nonnegative integer matrix BB and nonnegative integer vectors v,hv, h. Exactly one of the following holds.

  1. The trimmed graph is acyclic. Then ana_n is eventually zero.

  2. The trimmed graph contains a cycle, every cyclic component is a directed cycle, and no walk meets two cyclic components. Then ana_n is eventually periodic and not eventually zero.

  3. Every cyclic component is a directed cycle, and the greatest number met by a walk is r≥2r \ge2. Then ana_n is eventually quasi-polynomial of degree r−1r - 1.

  4. Some cyclic component of the trimmed graph has spectral radius greater than one. If ρ\rho is the largest such radius, then

lim sup⁡n→∞an1/n=ρ>1.\limsup_{n\to\infty} a_n^{1/n} = \rho> 1.

In (2) and (3), the least common multiple of the cycle lengths is a quasiperiod. The same conclusions hold for higher extension dimensions, with the index shift in Theorem 3.5.

Proof. An acyclic graph has bounded walk length. Every cycle in the trimmed graph lies on a walk from the initial to the terminal support, so it gives arbitrarily long counted walks. Suppose now that all cyclic components are directed cycles. Fix the route through the component graph, its connecting edges, and its initial and terminal support vertices. In each cycle, a walk follows the unique path from entry to exit of length less than one circuit, with any number of complete circuits inserted. There are finitely many such routes γ\gamma, and each contributes

cγtdγ∏j=1sγ(1−tℓγ,j),cγ>0,\frac{c_\gamma t^{d_\gamma}}{\prod_{j=1}^{s_\gamma}(1-t^{\ell_{\gamma,j}})}, \qquad c_\gamma> 0,

to ∑n≥0antn\sum_{n\ge0} a_n t^n, where sγs_\gamma is the number of cycles on the route. Routes without cycles contribute a polynomial.

Let LL be the least common multiple of the cycle lengths. All poles of the series are LL-th roots of unity, so its coefficients are eventually quasi-polynomial with quasiperiod LL [15 Proposition 4.4.1]. To find the degree, put r=max⁡γsγr = \max_\gamma s_\gamma. The pole at t=1t=1 has order exactly rr, since the positive route contributions cannot cancel there, and no other pole has larger order. In a partial fraction expansion, a pole of order jj contributes a polynomial of degree j−1j - 1 times an LL-periodic sequence. The degree is therefore r−1r - 1. The cases r=1r = 1 and r≥2r \ge2 give (2) and (3).

Finally, let DD be the adjacency matrix of a cyclic component of maximal radius ρ>1\rho> 1. The trimmed matrix is block triangular with the component matrices on its diagonal, so its spectral radius is ρ\rho. The spectral-radius formula for matrix powers gives lim sup⁡an1/n≤ρ\limsup a_n^{1/n} \le\rho.

For the reverse inequality, fix entry and exit walks through a vertex zz of DD. They give an+c≥K(Dn)zza_{n+c} \ge K(D^n)_{zz} for some cc and K>0K > 0. If dd is the period of DD, the block of DdD^d containing zz is primitive and has spectral radius ρd\rho^d. Perron–Frobenius theory therefore gives (Ddm)zz∼czρdm(D^{dm})_{zz} \sim c_z\rho^{dm}, with cz>0c_z > 0. This subsequence proves the reverse inequality and hence (4); Lemma 4.1 exhausts the possibilities.

The row bound gives dim⁡kExt⁡Ai(M,N)≤∥vM∥1∥hN∥∞2i−3\dim_k \operatorname{Ext}_A^i(M,N) \le\|v_M\|_1\|h_N\|_\infty2^{i-3} for i≥3i \ge3, so an exponential rate is at most two. Trimming matters: the full matrix may grow faster than the coefficient for a chosen pair. Even in the exponential case some residue classes can vanish.

Remark 4.3 (finer asymptotics). A limit need not exist: residue classes may vanish or have different rates. Finer asymptotics follow from periodic-class decomposition and Perron–Frobenius theory [12, 14]. For example, B=(0210)B = \begin{pmatrix}0 & 2 \\ 1 & 0\end{pmatrix} and v=h=(1,0)Tv = h = (1,0)^{\mathsf{T}} give a2m=2ma_{2m} = 2^m and a2m+1=0a_{2m+1} = 0. Thus lim sup⁡an1/n=2\limsup a_n^{1/n} = \sqrt{2}, although every odd term vanishes.

Small examples of the four behaviours

For the following examples we impose all paths of length two as relations. Their extensions can be computed directly from the quiver.

Lemma 4.4 (Radical-square-zero simples). For A=kQ/J2A = kQ/J^2, let DxyD_{xy} be the number of arrows from xx to yy. Then

ΩSx≅⨁ySy⊕Dxy,dim⁡kExt⁡Ai(Sx,Sy)=(Di)xy(i≥0).\Omega S_x \cong\bigoplus_y S_y^{\oplus D_{xy}}, \qquad\dim_k \operatorname{Ext}_A^i(S_x,S_y) = (D^i)_{xy} \quad(i \ge0).

Proof. The radical of PxP_x is the direct sum of the simples at the targets of arrows from xx. Iterate and apply Lemma 2.1. □

In Figure 3, each vertex has at most two incoming and two outgoing arrows, so these radical-square-zero algebras are string. A walk records successive syzygies; it need not survive as a nonzero path in AA.

Figure 3. Four growth behaviours for kQ/J2kQ/J^2.

Example 4.5 (vanishing and periodicity). In Figure 3a, the only nonzero extension between S1S_1 and S3S_3 occurs in degree two, where it has dimension one. Although the length-two path is zero in AA, it still counts as a walk. In (b),

dim⁡kExt⁡An(S1,S1)={1n even,0n odd.\dim_k \operatorname{Ext}_A^n(S_1,S_1) = \begin{cases} 1 & n\ \text{even},\\ 0 & n\ \text{odd}. \end{cases}

Example 4.6 (one or two transfers). In Figure 3c, a walk from 0 to 1 chooses one of its nn steps at which to cross the connecting arrow. A walk from 0 to 2 chooses two crossing steps. Thus

dim⁡kExt⁡An(S0,S0)=1,\dim_k \operatorname{Ext}_A^n(S_0,S_0) = 1,
dim⁡kExt⁡An(S0,S1)=n,dim⁡kExt⁡An(S0,S2)=(n2).\dim_k \operatorname{Ext}_A^n(S_0,S_1) = n, \qquad\dim_k \operatorname{Ext}_A^n(S_0,S_2) = \binom{n}{2}.

The last two series are t/(1−t)2t/(1-t)^2 and t2/(1−t)3t^2/(1-t)^3, respectively. The source is the same in all three cases; changing the target changes the growth.

Example 4.7 (Fibonacci growth). In Figure 3(d), the adjacency matrix is

F=(1110).F=\begin{pmatrix}1&1\\1&0\end{pmatrix}.

The walk counts satisfy a0=a1=1a_0=a_1=1 and an=an−1+an−2a_n=a_{n-1}+a_{n-2} for n≥2n\ge2. Thus dim⁡kExt⁡An(S1,S1)\dim_k\operatorname{Ext}_A^n(S_1,S_1) is 1,1,2,3,5,8,…1,1,2,3,5,8,\ldots, with generating function 1/(1−t−t2)1/(1-t-t^2) and exponential rate (1+5)/2(1+\sqrt{5})/2. Here branching occurs within one recurrent component.

Example 4.8 (two loops). For the bound quiver

x⇄0⇄yx \rightleftarrows0 \rightleftarrows y

I=(x2,xy,yx,y2)I=(x^2,xy,yx,y^2),

each step of a walk has two choices. Thus A=k⟨x,y⟩/(x,y)2A=k\langle x,y\rangle/(x,y)^2 has dim⁡kExt⁡An(S,S)=2n\dim_k\operatorname{Ext}_A^n(S,S)=2^n. The full non-projective uniserial string family is (S,M(x),M(y))(S,M(x),M(y)), with matrix

B=(200100100).B=\begin{pmatrix}2&0&0\\1&0&0\\1&0&0\end{pmatrix}.

The projective AA has radical S⊕2S^{\oplus2}; the covers of M(x)M(x) and M(y)M(y) each have simple kernel. Only SS is recurrent; the two other states feed into it after one step. Quadratic relations alone therefore allow exponential growth.

Periodic examples

The gentle case

Write R(α)=αAR(\alpha)=\alpha A for an arrow ideal. In the small gentle quiver below, the dotted curve marks the relation ab=0ab=0:

Small gentle quiver with vertices 1, 2, and 3, arrows a and b, and dotted relation curve

Figure 1 (PDF p. 9).

The gentle conditions simplify both parts of the kernel calculation. Quadraticity also makes every positive-length path ideal isomorphic to an arrow ideal. Together with Proposition 3.3, this gives periodicity over an arbitrary field.

Lemma 5.1 (gentle kernels). For a gentle algebra, ΩR(α)\Omega R(\alpha) is R(β)R(\beta) if αβ∈I\alpha\beta\in I for an arrow β\beta starting at t(α)t(\alpha), and is zero if no such arrow exists. For a string module,

ΩM(w)≅L0⊕P⊕L1,\Omega M(w)\cong L_0\oplus P\oplus L_1,

where PP is projective and each LjL_j is zero or an arrow ideal.

Proof. The cover Pt(α)→αAP_{t(\alpha)}\to\alpha A sends uu to αu\alpha u. A basis path is killed precisely when it begins with an arrow β\beta such that αβ∈I\alpha\beta\in I: quadraticity excludes a first obstruction farther along the path. There is at most one such β\beta, so the kernel is βA\beta A or zero.

For the second assertion, use the string-kernel description following Proposition 3.3. At a valley xx, the final arrows of p,qp,q are distinct. Each outgoing arrow has exactly one nonzero predecessor among them, by the two gentle predecessor conditions. Quadraticity implies that every positive-length basis path from xx extends exactly one of p,qp,q. Distinct paths have disjoint images in the path basis of the cover, so Px→(p−q)AP_x\to(p-q)A, ex↦p−qe_x\mapsto p-q, is injective as well as surjective.

For any positive-length basis path u=vαu=v\alpha, quadraticity gives uw≠0uw\ne0 if and only if αw≠0\alpha w\ne0 for every basis path ww starting at t(α)t(\alpha). Thus αw↦uw\alpha w\mapsto uw defines an isomorphism αA≅uA\alpha A\cong uA. In particular, each endpoint tail is an arrow ideal up to isomorphism. This gives the decomposition of [BS21, Proposition 3.1]. □

Corollary 5.2 (gentle periodicity). Higher extension dimensions over a gentle algebra over any field are eventually periodic, including the possibility of eventual zero.

Proof. By Proposition 3.3 and the path-ideal isomorphism just proved, Ω2M\Omega^{2}M is a direct sum of arrow ideals. The first part of Lemma 5.1 sends each arrow ideal to another arrow ideal or to zero. There are finitely many arrows, so each such sequence terminates or becomes periodic. Taking a common period for the finitely many summands makes the higher syzygies of MM periodic. Dimension shift gives eventual periodicity of dim⁡kExt⁡Ai(M,N)\dim_k \operatorname{Ext}^{i}_{A}(M,N) for every target NN. □

Over an algebraically closed field this recovers the vanishing-or-periodicity conclusion of [BS21, Theorem 3.4], which also gives a finer endpoint criterion for a nonzero periodic tail. In that setting, bands satisfy a stronger conclusion: their projective and injective dimensions are at most one. This is [BS21, Corollary 3.6] for quasi-simple bands; a filtration by quasi-simple bands and the long exact sequences for Ext give the same bounds in higher rank. For the radical-square-zero cycle kCm/J2kC_m/J^2 one sees the period directly: ΩSj=Sj+1\Omega S_j=S_{j+1} and

dim⁡kExt⁡Ai(Sj,Sl)=[l≡j+i(modm)].\dim_k \operatorname{Ext}^{i}_{A}(S_j,S_l)=[l\equiv j+i\pmod m].

Brackets denote the indicator of the enclosed condition.

Truncated cycles

Let

n≥1,r≥2,Λn,r=kCn/Jr.n\ge1,\qquad r\ge2,\qquad\Lambda_{n,r}=kC_n/J^r.

where CnC_n has arrows aj:j→j+1a_j:j\to j+1, with indices modulo nn. Let U(j,ℓ)U(j,\ell) be the uniserial module with top SjS_j and length ℓ\ell, for 1≤ℓ≤r1\le\ell\le r. Its successive composition factors are Sj,…,Sj+ℓ−1S_j,\ldots,S_{j+\ell-1}, and U(j,r)=PjU(j,r)=P_j. We allow r>nr>n, so vertex labels can repeat in a composition series. The cover calculation is unchanged; the Hom calculation must count each eligible occurrence separately.

Proposition 5.3 (complementary tails). For 1≤ℓ<r1\le\ell<r,

ΩU(j,ℓ)≅U(j+ℓ,r−ℓ),Ω2U(j,ℓ)≅U(j+r,ℓ).\Omega U(j,\ell)\cong U(j+\ell,r-\ell),\qquad\Omega^{2}U(j,\ell)\cong U(j+r,\ell).

The least syzygy period of U(j,ℓ)U(j,\ell) is

per⁡ΩU(j,ℓ)={n/gcd⁡(n,ℓ),r=2ℓ,2n/gcd⁡(n,r),r≠2ℓ.\operatorname{per}_{\Omega}U(j,\ell)= \begin{cases} n/\gcd(n,\ell), & r=2\ell,\\ 2n/\gcd(n,r), & r\ne2\ell. \end{cases}

The least positive even return time is always 2n/gcd⁡(n,r)2n/\gcd(n,r). The stable syzygy matrix is a permutation matrix.

Proof. The kernel of U(j,r)↠U(j,ℓ)U(j,r)\twoheadrightarrow U(j,\ell) is its complementary tail, of length r−ℓr-\ell and top j+ℓj+\ell. Applying this twice restores the length and advances the top by rr. If r≠2ℓr\ne2\ell, odd returns are impossible because they change the length. Even returns advance the top by a multiple of rr, giving period 2n/gcd⁡(n,r)2n/\gcd(n,r). If r=2ℓr=2\ell, every syzygy preserves length and advances the top by ℓ\ell, so the least period is n/gcd⁡(n,ℓ)n/\gcd(n,\ell). Finally the map (j,ℓ)↦(j+ℓ,r−ℓ)(j,\ell)\mapsto(j+\ell,r-\ell) is a permutation of the finite state set. □

Example 5.4 (even and odd periods). For (n,r)=(5,3)(n,r)=(5,3), the orbit of S0S_0 begins

U(0,1)⟼U(1,2)⟼U(3,1)⟼U(4,2)⟼U(1,1).U(0,1)\longmapsto U(1,2)\longmapsto U(3,1)\longmapsto U(4,2)\longmapsto U(1,1).

Its length is ten: every second step advances the top by three, and an odd return cannot restore the length. For (n,r)=(6,4)(n,r)=(6,4), the module U(0,2)U(0,2) instead has period three:

U(0,2)→ΩU(2,2)→ΩU(4,2)→ΩU(0,2).U(0,2)\xrightarrow{\Omega}U(2,2)\xrightarrow{\Omega}U(4,2)\xrightarrow{\Omega}U(0,2).

Its least positive even return time is six.

Maps and extensions from eligible depths

A map between two uniserial modules is determined by the image of a top generator. In the target, the allowed path-basis vectors give a basis of maps. We number their depths from zero at the top. The vertex label gives a congruence, and the annihilation condition gives an interval of allowed depths. To compute Ext, we remove the maps that extend to the preceding projective in the resolution.

Theorem 5.5 (extensions over a truncated cycle). Let X=U(j,ℓ)X=U(j,\ell) and Y=U(k,m)Y=U(k,m), where 1≤ℓ<r1\leq\ell<r and 1≤m≤r1\leq m\leq r. Write ΩiX=U(ai,bi)\Omega^{i}X=U(a_i,b_i); explicitly,

(a2t,b2t)=(j+tr,ℓ),(a2t+1,b2t+1)=(j+tr+ℓ,r−ℓ).(a_{2t},b_{2t})=(j+tr,\ell),\qquad(a_{2t+1},b_{2t+1})=(j+tr+\ell,r-\ell).

For every i≥1i\geq1,

dim⁡kExt⁡Λn,ri(X,Y)=#{q∈Z:max⁡(0,m−bi)≤q<min⁡(m,r−bi),q≡ai−k(modn)}.\dim_k\operatorname{Ext}^{i}_{\Lambda_{n,r}}(X,Y)=\#\left\{q\in\mathbb{Z}: \begin{aligned} &\max(0,m-b_i)\leq q<\min(m,r-b_i),\\ &q\equiv a_i-k\pmod n \end{aligned} \right\}.

For arbitrary 1≤ℓ,m≤r1\leq\ell,m\leq r, degree zero is given by

dim⁡kHom⁡(U(j,ℓ),U(k,m))=#{q∈Z:max⁡(0,m−ℓ)≤q<m, q≡j−k(modn)}.\dim_k\operatorname{Hom}(U(j,\ell),U(k,m)) =\#\left\{q\in\mathbb{Z}:\max(0,m-\ell)\leq q<m,\ q\equiv j-k\pmod n\right\}.

For a projective source, positive-degree extensions vanish.

Proof. Let y0,…,ym−1y_0,\ldots,y_{m-1} be the usual path-basis vectors of YY, with yqy_q at depth qq and vertex k+qk+q. A homomorphism from U(a,b)U(a,b) is determined by the image of its top generator. Such an image must lie at vertex aa, and it must be annihilated by the path of length bb. Thus a basis for the possible images is given by

yqwithq≡a−k(modn),max⁡(0,m−b)≤q<m.y_q\quad\text{with}\quad q\equiv a-k\pmod n,\qquad\max(0,m-b)\leq q<m.

These vectors give independent homomorphisms, even when the same vertex occurs several times in YY. Taking a=ja=j and b=ℓb=\ell proves the Hom formula.

To calculate extensions, put di=r−bid_i=r-b_i. The embedding of ΩiX=U(ai,bi)\Omega^iX=U(a_i,b_i) into the preceding projective Pai−diP_{a_i-d_i} begins at depth did_i. A map from that projective is determined by an arbitrary vector of YY at vertex ai−dia_i-d_i. On a basis vector ysy_s, restriction multiplies by the path of length did_i, giving ys+diy_{s+d_i} if s+di<ms+d_i<m, and zero otherwise. Thus the image of restriction is spanned by exactly the eligible basis maps with q≥diq\geq d_i. Conversely, each such map lifts by sending the projective’s top generator to yq−diy_{q-d_i}. The cokernel in (2.A) therefore retains precisely the eligible depths q<diq<d_i. Combining this with the annihilation bound and applying dimension shift gives the formula. □

The dimensions can exceed one: over kC2/J7kC_2/J^7,

dim⁡kExt⁡1(U(0,3),U(1,3))=2.\dim_k\operatorname{Ext}^{1}(U(0,3),U(1,3))=2.

Here ΩU(0,3)=U(1,4)\Omega U(0,3)=U(1,4), whose generator can map to either circled occurrence in the target:

U(1,3):\qquad\circled{1}\xrightarrow{\phantom{000}}0\xrightarrow{\phantom{000}}\circled{1}
q=0q=1q=2q=0\qquad\qquad q=1\qquad\qquad q=2

Both maps survive in Ext: the preceding projective contains the syzygy at depth three, so its restriction to this length-three target is zero.

Every positive-degree sequence has period dividing 2n/gcd⁡(n,r)2n/\gcd(n,r), but its least period can be shorter than the source’s syzygy period. We now determine the least periods for the arrow-ideal pairs considered in this dissertation.

Least periods for arrow ideals

Corollary 5.6 (arrow-ideal dimensions and least periods). Assume n≥3n \ge3 and 2≤r≤n2 \le r \le n, and write R(aj)=U(j+1,r−1)R(a_j)=U(j+1,r-1). Put φ(i)=⌊i/2⌋r+(i mod 2)\varphi(i)=\lfloor i/2\rfloor r+(i\bmod2). For i≥1i\ge1,

dim⁡kExt⁡Λn,ri(R(aj),R(al))=[l−j≡φ(i)(modn)].\dim_k\operatorname{Ext}^{i}_{\Lambda_{n,r}}(R(a_j),R(a_l))=[l-j\equiv\varphi(i)\pmod n].

Let g=gcd⁡(n,r)g=\gcd(n,r). If r=2r=2, the least period is nn. If r≥3r\ge3 and l−jl-j is congruent to 00 or 11 modulo gg, the least period is 2n/g2n/g. Otherwise all positive-degree terms are zero. In degree zero,

dim⁡kHom⁡(R(aj),R(al))=[(j−l) mod n∈{0,…,r−2}].\dim_k\operatorname{Hom}(R(a_j),R(a_l))=[(j-l)\bmod n\in\{0,\ldots,r-2\}].

Proof. Proposition 5.3 gives

Ω2mR(aj)=R(aj+mr),Ω2m+1R(aj)=Sj+(m+1)r.\Omega^{2m}R(a_j)=R(a_{j+mr}),\qquad\Omega^{2m+1}R(a_j)=S_{j+(m+1)r}.

Substituting into Theorem 5.5 leaves one possible depth on each parity and gives the displayed congruence. The even degrees attain exactly the class 00 modulo gg, and the odd degrees attain class $1.

On each active parity, the indicator has least period n/gn/g in the half-degree. Hence the smallest even period is 2n/g2n/g. An odd period 2t+12t+1 would require both parities to be active. Shifting a nonzero even term by this period changes φ\varphi by tr+1tr+1; shifting a nonzero odd term changes it by (t+1)r−1(t+1)r-1. Since both shifted terms must still be nonzero, we obtain

tr+1≡0,(t+1)r−1≡0(modn).tr+1\equiv0,\qquad(t+1)r-1\equiv0\pmod n.

Subtracting forces r≡2(modn)r\equiv2\pmod n, hence r=2r=2 under 2≤r≤n2\le r\le n. In that case φ(i)=i\varphi(i)=i. The Hom formula follows from the same theorem in degree zero. □\square

Example 5.7 (reading the congruence). Over Λ5,3\Lambda_{5,3}, the self-pair R(a0),R(a0)R(a_0),R(a_0) has nonzero positive extensions exactly in degrees 00 and 77 modulo ten. The pair R(a0),R(a1)R(a_0),R(a_1) has them in degrees 11 and 44 modulo ten:

ii12345678910
dim⁡Ext⁡i(R(a0),R(a0))\dim\operatorname{Ext}^{i}(R(a_0),R(a_0))0000001001
dim⁡Ext⁡i(R(a0),R(a1))\dim\operatorname{Ext}^{i}(R(a_0),R(a_1))1001000000

Table 1.

For Λ6,3\Lambda_{6,3}, g=3g=3. The pair R(a0),R(a2)R(a_0),R(a_2) has no positive extensions because 22 is neither 00 nor 11 modulo 33. The source still has periodic syzygies: the target detects none of the relevant first extensions.

For r≥3r\ge3 these algebras are not gentle, but their one-arm covers still force periodic syzygies. For split Nakayama algebras, Erdmann, Klász and Marczinzik show that a nonzero self-extension in degree one of an indecomposable module forces nonzero self-extensions in every positive degree [4]. In the restricted range of Corollary 5.6, Ext⁡1(R(aj),R(aj))=0\operatorname{Ext}^{1}(R(a_j),R(a_j))=0. Outside that range, take kC3/J6kC_3/J^6 and X=U(0,3)X=U(0,3). Here ΩX≅X\Omega X\cong X, and Theorem 5.5 gives dim⁡kExt⁡Ai(X,X)=1\dim_k\operatorname{Ext}^{i}_{A}(X,X)=1 for every i≥1i\ge1, in agreement with the cited result.

Polynomial growth from linked cycles

The growth mechanism is already visible in a radical-square-zero algebra. Take the quiver

Figure 7.

The loops keep the simples where they are, while the arrow bb produces a new summand:

ΩS0≅S0⊕S1,ΩS1≅S1.\Omega S_0\cong S_0\oplus S_1,\qquad\Omega S_1\cong S_1.

Hence ΩiS0≅S0⊕S1⊕i\Omega^iS_0\cong S_0\oplus S_1^{\oplus i}, and for T=S0⊕S1T=S_0\oplus S_1,

dim⁡kExt⁡Ai(S0,T)=i+1(i≥0).\dim_k\operatorname{Ext}^{i}_{A}(S_0,T)=i+1\qquad(i\ge0).

This is a two-vertex version of Example 4.6. The same mechanism works when the persistent summands are longer uniserial modules rather than simples.

Two cycles

Let CnC_n be an oriented cycle with vertex set Z/nZ\mathbb{Z}/n\mathbb{Z}. We take two copies, with vertices pp and p′p' and arrows

ap:p→p+1,ap′:p′→(p+1)′,bp:p→p′(p∈Z/nZ).a_p:p\to p+1,\qquad a'_p:p'\to(p+1)',\qquad b_p:p\to p'\quad(p\in\mathbb{Z}/n\mathbb{Z}).

All indices are modulo nn; Figure 4 shows the complete quiver for n=3n=3.

Figure 4. Two linked 3-cycles; the dotted arcs mark a0b1=b1a1′=0a_0b_1=b_1a'_1=0.

Kill all paths of length rr along the upper cycle and of length ss along the lower cycle, where 3≤r,s≤n3\le r,s\le n. In addition impose, for every pp,

ap−1bp=0,bpap′=0.a_{p-1}b_p=0,\qquad b_pa'_p=0.

We denote the resulting string algebra by Λn,r,s\Lambda_{n,r,s}. The mixed relations ensure that bpA≅Sp′b_pA\cong S_{p'}.

Write U(p,ℓ)U(p,\ell) for the uniserial module of length 1≤ℓ<r1\le\ell<r with top SpS_p, supported on the upper cycle, and U(p′,ℓ)U(p',\ell) for its lower counterpart, with 1≤ℓ<s1\le\ell<s. All links act as zero. The upper projective cover has an extra arm bpb_p, so

ΩU(p,ℓ)≅U(p+ℓ,r−ℓ)⊕Sp′,\Omega U(p,\ell)\cong U(p+\ell,r-\ell)\oplus S_{p'},
ΩU(p′,ℓ)≅U((p+ℓ)′,s−ℓ).\Omega U(p',\ell)\cong U((p+\ell)',s-\ell).

Retaining only the complementary tail permutes the modules on each row. On this family the syzygy matrix therefore has the form

Bcyc=(PC0Q),(Bcycm)12=∑a=0m−1PaCQm−1−a(m≥1),B_{\mathrm{cyc}}=\begin{pmatrix}P&C\\0&Q\end{pmatrix},\qquad(B_{\mathrm{cyc}}^m)_{12}=\sum_{a=0}^{m-1}P^aCQ^{m-1-a}\quad(m\ge1),

where P,QP,Q are permutation matrices and CC records the link simples. The orbits of PP and QQ are the recurrent components of the syzygy graph. The summation counts the possible times of transfer from an upper orbit to a lower one.

After two syzygies these are all the modules we need. By Proposition 3.3, the summands are positive-length path ideals: a path along either cycle leaves a tail shorter than its truncation length, and a link gives a lower simple.

A walk meets at most two orbits, one on each row. By Theorem 4.2, every extension sequence is eventually zero or eventually quasi-polynomial of degree at most one. If all summands of Ω2M\Omega^2 M lie on the lower cycle, the sequence is bounded. A linear term occurs on some residue class precisely when an upper state in supp⁡vM\operatorname{supp} v_M can reach a lower state in supp⁡hN\operatorname{supp} h_N.

Example 6.1 (one new summand at every step). Take n=r=s=3n=r=s=3. The first syzygies of the upper simple S0S_0 are

ΩS0≅U(1,2)⊕S0′,Ω2S0≅S0⊕S1⊕U(1′,2),Ω3S0≅U(1,2)⊕S0′′⊕U(2′,2).\begin{aligned} \Omega S_0 &\cong U(1,2) \oplus S'_0,\\ \Omega^2 S_0 &\cong S_0 \oplus S_1 \oplus U(1',2),\\ \Omega^3 S_0 &\cong U(1,2) \oplus S''_0 \oplus U(2',2). \end{aligned}

At each step the upper summand produces one new lower simple, while each existing lower summand has just one successor. Hence ΩiS0\Omega^i S_0 has i+1i+1 indecomposable summands, all with simple top. Taking T=A/rad⁡AT=A/\operatorname{rad} A, Lemma 2.1 gives

dim⁡kExt⁡Ai(S0,T)=i+1(i≥0).\dim_k \operatorname{Ext}^i_A(S_0,T)=i+1 \qquad(i\ge0).

Several cycles and accessible links

Take finitely many disjoint oriented cycles CuC_u, with nun_u vertices, and kill their paths of length rur_u, where 3≤ru≤nu3\le r_u\le n_u. Join them by arrows, called links, pointing forward in some ordering of the cycles. Each vertex may be the source of at most one link and the target of at most one link. Kill every length-two path containing a link. The resulting algebra is string, and each link b:c→db:c\to d gives bA≅SdbA\cong S_d.

Write Uu(j,ℓ)U_u(j,\ell) for the uniserial module of length 1≤ℓ<ru1\le\ell<r_u with top SjS_j on CuC_u, with all links acting as zero. Its cover gives

ΩUu(j,ℓ)≅Uu(j+ℓ,ru−ℓ)⊕⨁b:j→d a linkSd.(3)\Omega U_u(j,\ell)\cong U_u(j+\ell,r_u-\ell)\oplus\bigoplus_{b:j\to d\ \mathrm{a\ link}}S_d. \tag*{(3)}

All summands are non-projective. Retaining the summand on CuC_u gives the permutation

Fu(j,ℓ)=(j+ℓ,ru−ℓ),Fu2(j,ℓ)=(j+ru,ℓ),F_u(j,\ell)=(j+\ell,r_u-\ell),\qquad F_u^2(j,\ell)=(j+r_u,\ell),

with vertex indices modulo nun_u. Starting from SwS_w, its successive tops are

w,w+1,w+ru,w+ru+1,w+2ru,….w,\quad w+1,\quad w+r_u,\quad w+r_u+1,\quad w+2r_u,\ldots.

Thus, if gu=gcd⁡(nu,ru)g_u=\gcd(n_u,r_u), the source cc of a link occurs exactly when

c≡worc≡w+1(modgu).(4)c\equiv w\quad\text{or}\quad c\equiv w+1\pmod{g_u}. \tag*{(4)}

Every such top recurs periodically. Using that link starts a new persistent summand at its target vertex on the next cycle. Call a chain of links live from vv if (6.B) holds successively, beginning at vv and then at each preceding link target. Let δ∗(v)\delta^*(v) be the largest number of links in such a chain, with value zero if none can be used.

Proposition 6.2 (the degree from a live chain). Let si(v)s_i(v) count the indecomposable summands of ΩiSv\Omega^i S_v, with multiplicity. Then si(v)s_i(v) is eventually quasi-polynomial of degree δ∗(v)\delta^*(v).

Proof. By (6.A), syzygies stay within direct sums of the finitely many modules Uu(j,ℓ)U_u(j,\ell). Since links point forward, the recurrent components of the syzygy graph on these modules are exactly the orbits of the permutations FuF_u. A transfer between reachable orbits uses a live link and enters the next orbit at a simple. The greatest number of orbits met by a walk starting at SvS_v is therefore δ∗(v)+1\delta^*(v)+1. Apply Theorem 4.2 with terminal weight 1.

For a pair (M,N)(M,N), start instead at the summands of Ω2M\Omega^2 M. From Uu(j,ℓ)U_u(j,\ell), the first link is accessible when its source is congruent to jj or j+ℓj+\ell modulo gug_u; all subsequent transfers start from simples and use (6.B). Finally discard states unable to reach supp⁡hN\operatorname{supp} h_N. The last restriction can lower the degree: a target need not detect every summand.

Example 6.3 (a live link and a missed link). Use the quivers in Figure 5, killing all paths of length three along a cycle and all paths of length two containing a link. Starting at S0S_0, the first link produces S0′S_0'. Following only the summand that stays on the middle cycle gives

Figure 5. Growth from S0S_0: (a) quadratic; (b) linear.

S0′⟼U(1′,2)⟼S0′.S_0' \longmapsto U(1',2) \longmapsto S_0'.

The tops are 0′0' and 1′1'. The second link in (a) is therefore used repeatedly, while the link in (b), starting at 2′2', is never used. Although U(1′,2)U(1',2) also has a composition factor at 2′2', only its top produces a link summand in (6.A). The two summand counts therefore have degrees two and one, respectively.

The first few counts make the difference visible:

ii012345678
second link at 0′0'123467101115
second link at 2′2'122334455

For the first row, in degree 2m2m the three cycles contribute 11, mm, and (m+12)\binom{m+1}{2} summands. In degree 2m+12m+1, the middle count increases to m+1m+1 and the other two stay the same. In the second row, the third cycle contributes nothing and the second contributes ⌈i/2⌉\lceil i/2\rceil in degree ii. These numbers also give dim⁡kExt⁡Ai(S0,A/rad⁡A)\dim_k \operatorname{Ext}_A^i(S_0,A/\operatorname{rad} A).

Acknowledgements

I thank Azzurra Ciliberti and Karin Baur for their guidance and supervision during my dissertation.

Rühr-Universität Bochum

Email address: [email protected]

References

  1. [2]İlke Çanakçı, David Pauksztello, and Sibylle Schroll. “On extensions for gentle algebras”. In: Canad. J. Math. 73.1 (2021), pp. 249–292. doi: 10.4153/S0008414X2000005X.DOI
  2. [3]İlke Çanakçı and Sibylle Schroll. “Extensions in Jacobian algebras and cluster categories of marked surfaces”. In: Adv. Math. 313 (2017), pp. 1–49. doi: 10.1016/j.aim.2017.03.016.DOI
  3. [4]Karin Erdmann, Viktória Klász, and René Marczinzik. “Selfextensions of modules over group algebras”. In: J. Algebra 649 (2024). With an appendix by Bernhard Böhmler and René Marczinzik, pp. 319–346. doi: 10.1016/j.jalgebra.2024.03.014.DOI
  4. [8]Birge Huisgen-Zimmermann. “Predicting syzygies over monomial relations algebras”. In: Manuscripta Math. 70 (1991), pp. 157–182. doi: 10.1007/BF02568368.DOI
  5. [10]Martin Irungu. “Higher Syzygies and Ext-Periodicity over String Algebras”. Submitted dissertation. Dissertation. Ruhr-Universität Bochum, Sept. 2026.
  6. [11]Sarah Rees. “The Automata that Define Representations of Monomial Algebras”. In: Algebras and Representation Theory 11.3 (2008), pp. 207–214. doi: 10.1007/s10468-007-9063-4.DOI
  7. [12]Uriel G. Rothblum. “Algebraic eigenspaces of nonnegative matrices”. In: Linear Algebra Appl. 12.3 (1975), pp. 281–292. DOI: 10.1016/0024-3795(75)90050-6.DOI
  8. [13]Jacques Sakarovitch. Elements of Automata Theory. Cambridge University Press, 2009. DOI: 10.1017/CBO9781139195218.DOI
  9. [14]Eugene Seneta. Non-negative Matrices and Markov Chains. Second. New York: Springer, 1981.
  10. [15]Richard P. Stanley. Enumerative Combinatorics. Second. Vol. 1. Cambridge Studies in Advanced Mathematics. Cambridge University Press, 2012. DOI: 10.1017/CBO9781139058520.DOI

Paper details

vlnc:W_c9jamx3a
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Mathematics

Not recorded for this paper

View statement
Looking for related papers…
Contents