alphaFeedback

WorksMathematics

Morse-Maxwell Theory for Directed Acyclic Graph

AI contributions
Not recorded for this paper
View details

Abstract

The Maxwell graph of a directed acyclic graph is obtained from its iterated basin and hill boundaries. The unweighted counts of extrema at these levels need not satisfy Morse inequalities. We assign multiplicities by the reduced homology of lower and upper order links and retain the vertices with nonzero local homology. The resulting graph preserves both boundary hierarchies, the local groups, and the homology of all sublevel and superlevel order complexes. Its descending and ascending Morse numbers satisfy the weak and strong Morse inequalities over any field. We also construct a Morse chain complex and compute the local groups in several examples. Finally, we consider homology on the original Maxwell graph and prove analogous inequalities for path homology using relative local groups.

Introduction

Let a graph carry a height function with distinct values, and orient each edge from its higher endpoint to its lower endpoint. A vertex may then reach one or several local minima. The vertices that reach at least two minima form the basin boundary. Repeating this operation, and its analogue for maxima, gives two boundary hierarchies. The Maxwell graph retains the extrema of all their levels and the reachability relations between them. We seek Morse inequalities associated with this construction.

Maxwell’s On Hills and Dales [8] describes regions of depression and elevation and the lines separating them. His counting arguments allow several topographic events to occur at one point. The graph construction follows this viewpoint by using all descending paths; related constructions appear in [12, 13]. The number of extrema at each boundary level alone, however, is insufficient for Morse inequalities. The examples below show that local multiplicities are needed and that the original Maxwell reduction can change homology.

Let DD be a finite directed acyclic graph. Reachability gives a partial order P(D)P(D): we write v≤uv \leq u when a directed path leads from uu to vv. Let K(D)=ΔP(D)K(D) = \Delta P(D) be the simplicial complex whose simplices are the nonempty chains of this order. Fix a field F\mathbb{F}. For each vertex vv, its lower order link L−(v)L^{-}(v) is the order complex of its strict descendants. The local contribution in degree rr is

μr−(v)=dim⁡FH~r−1(L−(v);F).\mu^{-}_{r}(v) = \dim_{\mathbb{F}} \widetilde{H}_{r-1}(L^{-}(v);\mathbb{F}).

We use H~−1(∅;F)=F\widetilde{H}_{-1}(\varnothing;\mathbb{F}) = \mathbb{F}. Thus a sink contributes one generator in degree zero. A vertex whose lower link has three contractible components contributes two generators in degree one. Define the Morse numbers and Betti numbers by

mr(D)=∑v∈V(D)μr−(v),βr(D)=dim⁡FHr(K(D);F).(1)m_r(D) = \sum_{v \in V(D)} \mu^{-}_{r}(v), \qquad\beta_r(D) = \dim_{\mathbb{F}} H_r(K(D);\mathbb{F}). \tag*{(1)}

Theorem 1.1 (Morse inequalities). For every finite directed acyclic graph DD, every field F\mathbb{F}, and every q≥0q \ge0,

∑r=0q(−1)q−rmr(D)≥∑r=0q(−1)q−rβr(D),(2)\sum_{r=0}^{q}(-1)^{q-r}m_r(D) \ge\sum_{r=0}^{q}(-1)^{q-r}\beta_r(D), \tag*{(2)}

with equality for all sufficiently large qq. The weak inequalities mr(D)≥βr(D)m_r(D) \ge\beta_r(D) also hold.

Reversing all arrows gives the ascending inequalities. To preserve these local groups under reduction, we retain the extrema of both boundary hierarchies and every vertex with nonzero lower or upper local homology. We call the resulting graph the homological Maxwell graph. It has the same boundary levels, local groups, and Betti numbers as DD. In particular, (1) can be computed entirely on the reduced graph.

The homology in (1) is reachability homology [6]. For a DAG, it is also path homology of the transitive closure [11]. It can differ from path homology of DD [5] and from singular simplicial homology of the original digraph [3].

Forman’s discrete gradient [4] pairs incident cells of consecutive dimensions, with each cell in at most one pair and no nontrivial closed gradient paths. The unpaired cells are critical; counting gradient paths with signs gives the boundary of the Morse complex. Lin, Wang, and Yau [7] adapt this approach to digraphs, using a Morse function that extends to the transitive closure and a gradient preserving the path chain complex of the original digraph. Their reduction then preserves path homology.

Our construction uses all descending paths determined by the given orientation, without choosing a matching or an auxiliary Morse function. The local homology at a vertex determines its contribution, which may have multiplicity greater than one or occur in several degrees. In smooth Morse theory [9], a nondegenerate critical point of index rr contributes just one local generator in degree rr. The proof of (2) uses the standard homological argument behind Morse inequalities; compare [10, 14]. The additional task here is to preserve the two Maxwell hierarchies and the local groups under reduction. The deletion argument is related to reductions of finite posets [2].

§ 5 explains what changes for path homology. Its Betti numbers do not give Morse inequalities for the unweighted Maxwell level counts ∣Bj∣|B_j|. A three-leaf star gives an obstruction for any homology with the usual H0H_0. Nor can path Betti numbers simply replace βr\beta_r in (2): the subdivided diamond considered there has m1=0m_1 = 0 but first path Betti number one. We obtain path-homological Morse inequalities for every finite DAG by assigning each vertex the relative path homology of adding it to its strict descendants. The sums of these local dimensions are the appropriate Morse numbers. For the unchanged Maxwell graph, its reachability homology similarly gives inequalities with multiplicities computed on that graph. These chain constructions do not yet give a general graph reduction preserving both path homology and the Maxwell hierarchies.

Definitions and proofs occupy § 2 and § 3; the latter also gives a Morse complex and computational bounds. § 4 contains the examples, including a six-vertex graph with β2=1\beta_2 = 1.

Definitions

Directed graphs and their order complexes

Throughout, D=(V,A)D = (V, A) is a finite directed acyclic graph (DAG), without loops or parallel arrows. We use u⇝vu \leadsto v for reachability, allowing a path of length zero. A sink has no outgoing arrow and a source has no incoming arrow. The opposite graph DopD^{\mathrm{op}} reverses all arrows. A height function is a function f:V→Rf: V \to\mathbb{R} strictly decreasing on arrows. Every DAG admits a height function. Different choices give the same construction, since only reachability is used.

Definition 2.1. The reachability order P=P(D)P = P(D) is given by v≤uv \leq u if u⇝vu \rightsquigarrow v. Its order complex K(D)=ΔPK(D) = \Delta P has one simplex for each nonempty chain of PP. For U⊆VU \subseteq V, write P∣UP|_U for the restricted order and K(U)=Δ(P∣U)K(U) = \Delta(P|_U).

For example, w<v<uw < v < u determines a triangle in K(D)K(D), even if u→wu \to w is not an arrow of DD. Thus K(D)K(D) is not generally the one-dimensional underlying graph. Throughout the paper homology has coefficients in the fixed field F\mathbb{F}. We use ordinary homology for K(D)K(D) and augmented reduced homology for local links. The empty complex has H~−1=F\widetilde{H}_{-1} = \mathbb{F} and no other reduced homology. A complex is acyclic over F\mathbb{F} if all its augmented reduced homology groups vanish; in particular it is nonempty.

Basins, hills, and the Maxwell graph

For U⊆VU \subseteq V, let D[U]D[U] denote the induced directed subgraph.

Definition 2.2. The basin and hill of a vertex vv are

BD(v)=D[{u:u⇝v}],HD(v)=D[{u:v⇝u}].B_D(v) = D[\{u : u \rightsquigarrow v\}], \qquad H_D(v) = D[\{u : v \rightsquigarrow u\}].

The sets of reachable sinks and preceding sources are

ZD(v)={s∈Sink⁡(D):v⇝s},TD(v)={t∈Source⁡(D):t⇝v}.Z_D(v) = \{s \in\operatorname{Sink}(D) : v \rightsquigarrow s\}, \qquad T_D(v) = \{t \in\operatorname{Source}(D) : t \rightsquigarrow v\}.

The basin boundary and hill boundary are

ΓB(D)=D[{v:∣ZD(v)∣≥2}],ΓH(D)=D[{v:∣TD(v)∣≥2}].\Gamma_B(D) = D[\{v : |Z_D(v)| \geq2\}], \qquad\Gamma_H(D) = D[\{v : |T_D(v)| \geq2\}].

Thus the basin boundary consists of vertices from which descending paths lead to different minima. Basins may overlap. The word “boundary” refers to this combinatorial definition.

Definition 2.3. The descending and ascending boundary hierarchies are

DB0=D,DBj+1=ΓB(DBj),Bj=Sink⁡(DBj),D_B^0 = D,\qquad D_B^{j+1} = \Gamma_B(D_B^j),\qquad B_j = \operatorname{Sink}(D_B^j),
DH0=D,DHj+1=ΓH(DHj),Hj=Source⁡(DHj).D_H^0 = D,\qquad D_H^{j+1} = \Gamma_H(D_H^j),\qquad H_j = \operatorname{Source}(D_H^j).

The boundary of the empty graph is empty. The topographically essential set is

W(D)=⋃j≥0Bj∪⋃j≥0Hj.W(D) = \bigcup_{j \geq0} B_j \cup\bigcup_{j \geq0} H_j.

At each step we use the extrema of the current stage. Since a nonempty stage loses its extrema at the next step, both hierarchies terminate. A vertex may belong to both a descending and an ascending level.

Definition 2.4. For U⊆VU \subseteq V, let DUD_U be the downward Hasse diagram of P(D)∣UP(D)|_U. There is an arrow u→vu \to v exactly when v<uv < u and no element of UU lies strictly between them. The Maxwell graph is

M(D)=DW(D).\mathcal{M}(D) = D_{W(D)}.

An arrow of DUD_U can represent a path through omitted vertices. The Hasse diagram is the transitive reduction of the restricted comparability graph [1].

Local Morse data

For v∈Pv \in P, let P<vP_{<v} and P>vP_{>v} denote its strict lower and upper sets.

Definition 2.5. The lower and upper order links, local groups, and multiplicities are

L−(v)=Δ(P<v),L+(v)=Δ(P>v),L^-(v)=\Delta(P_{<v}), \qquad L^+(v)=\Delta(P_{>v}),
Er±(v)=H~r−1(L±(v);F),μr±(v)=dim⁡FEr±(v)(r≥0).E_r^\pm(v)=\widetilde{H}_{r-1}(L^\pm(v);\mathbb{F}), \qquad\mu_r^\pm(v)=\dim_{\mathbb{F}} E_r^\pm(v) \quad(r\geq0).

A vertex is descending critical if some μr−(v)\mu_r^-(v) is nonzero, and ascending critical if some μr+(v)\mu_r^+(v) is nonzero. Otherwise it is regular in the corresponding direction. Write C−(D)C^-(D) and C+(D)C^+(D) for these two critical sets.

The links contain all strict descendants or ancestors. Regularity requires every local group to vanish; a zero alternating sum of their dimensions is insufficient. We call the degree rr of a local generator its Morse degree. It need not agree with the boundary level of the vertex.

The descending Morse numbers are mr(D)m_r(D) from (1). The ascending ones are

mr+(D)=∑v∈Vμr+(v)=mr(Dop).m_r^+(D)=\sum_{v\in V}\mu_r^+(v)=m_r(D^{\mathrm{op}}).

The order complexes of PP and its opposite have the same simplices, so they have the same Betti numbers.

The homological Maxwell graph

Definition 2.6. The homological essential set and homological Maxwell graph are

SF(D)=W(D)∪C−(D)∪C+(D),MFh(D)=DSF(D).S_{\mathbb{F}}(D)=W(D)\cup C^-(D)\cup C^+(D), \qquad\mathcal{M}_{\mathbb{F}}^h(D)=D_{S_{\mathbb{F}}(D)}.

The graph retains the labels BjB_j, HjH_j and the local groups Er±(v)E_r^\pm(v) at its vertices.

Some vertices are retained only because they belong to W(D)W(D) and may have zero local groups in both directions. The set SF(D)S_{\mathbb{F}}(D) can depend on the coefficient field.

Vertices outside SF(D)S_{\mathbb{F}}(D) contribute zero to mrm_r and mr+m_r^+. Proposition 3.5 shows that the local groups at retained vertices can also be computed in the restricted order P(D)∣SF(D)P(D)|_{S_{\mathbb{F}}(D)}.

Proofs

The Morse inequalities

Choose a linear extension v1,…,vnv_1,\ldots,v_n of P(D)P(D) in which strict lower elements appear first. Set

Ki=K({v1,…,vi}),K0=∅.K_i=K(\{v_1,\ldots,v_i\}), \qquad K_0=\varnothing.

At step ii we attach all chains whose greatest element is viv_i.

Lemma 3.1 (Local attachment). *For every r≥0r\geq0,

Hr(Ki,Ki−1)≅Er−(vi).H_r(K_i,K_{i-1})\cong E_r^-(v_i).

Proof. The new closed star is the cone vi∗L−(vi)v_i*L^-(v_i), and its intersection with Ki−1K_{i-1} is L−(vi)L^-(v_i). Therefore its relative chains are the augmented chains of L−(vi)L^-(v_i) shifted up by one degree. This also covers a sink: the empty lower link supplies one relative generator in degree zero.

Proof of Theorem 1.1. Let ai,ra_{i,r} be the rank of the connecting map

Hr(Ki,Ki−1)⟶Hr−1(Ki−1).H_r(K_i,K_{i-1}) \longrightarrow H_{r-1}(K_{i-1}).

Here ai,0=0a_{i,0}=0 (ordinary H−1=0H_{-1}=0). Set br(Ki)=dim⁡Hr(Ki)b_r(K_i)=\dim H_r(K_i). Exactness gives

br(Ki)=br(Ki−1)+μr(vi)−ai,r−ai,r+1.b_r(K_i)=b_r(K_{i-1})+\mu_r(v_i)-a_{i,r}-a_{i,r+1}.

Summing over the vertices, with Ar=∑iai,rA_r=\sum_i a_{i,r}, gives

mr(D)−βr(D)=Ar+Ar+1,A0=0.(3)m_r(D)-\beta_r(D)=A_r+A_{r+1}, \qquad A_0=0. \tag*{(3)}

Since each ArA_r is a sum of ranks, the weak inequalities follow. Alternating sums cancel A1,…,AqA_1,\ldots,A_q, while A0=0A_0=0, giving

∑r=0q(−1)q−r(mr(D)−βr(D))=Aq+1≥0.\sum_{r=0}^{q}(-1)^{q-r}\bigl(m_r(D)-\beta_r(D)\bigr)=A_{q+1}\geq0.

All groups vanish in high degrees, giving eventual equality.

For the polynomial form of the inequalities, write mD(t)=∑rmr(D)trm_D(t)=\sum_r m_r(D)t^r and PD(t)=∑rβr(D)trP_D(t)=\sum_r\beta_r(D)t^r.

Corollary 3.2. There is a polynomial QD(t)Q_D(t) with nonnegative integer coefficients such that

mD(t)−PD(t)=(1+t)QD(t).(4)m_D(t)-P_D(t)=(1+t)Q_D(t). \tag*{(4)}

In particular ∑r(−1)rmr(D)=χ(K(D))\sum_r(-1)^r m_r(D)=\chi(K(D)).

Proof. Take QD(t)=∑r≥0Ar+1trQ_D(t)=\sum_{r\geq0}A_{r+1}t^r in (3), and then evaluate at t=−1t=-1.

Preservation under the Maxwell reduction

We first prove that retaining W(D)W(D) preserves both boundary hierarchies, and then show that regular vertices can be deleted without changing the local homology groups.

Lemma 3.3 (The boundary hierarchies). If W(D)⊆U⊆V(D)W(D)\subseteq U\subseteq V(D), then

Bj(DU)=Bj(D),Hj(DU)=Hj(D)B_j(D_U)=B_j(D), \qquad H_j(D_U)=H_j(D)

for every jj. The stage vertex sets are U∩V(DBj)U\cap V(D^j_B) and U∩V(DHj)U\cap V(D^j_H), respectively.

Proof. The basin boundary is predecessor-closed: if u⇝vu\rightsquigarrow v and vv reaches two sinks, then uu reaches them as well. Repetition shows that each DBjD^j_B is predecessor-closed in DD. Consequently every directed path ending in that stage remains in it, so its internal reachability is the restricted original order.

Every comparable pair in a finite poset is joined by a chain of covers. Hence DUD_U realizes exactly the order P(D)∣UP(D)|_U. Induct on jj. At stage zero the asserted vertex set is immediate. Suppose it is known at stage jj. Every vertex of the original stage reaches a sink in BjB_j, and all those sinks belong to W⊆UW\subseteq U. The restricted stage therefore has exactly the same sinks. Each retained vertex reaches the same stage sinks as before, so the condition of reaching at least two gives the required vertex set at stage j+1j+1. Reverse arrows for the ascending statement.

Lemma 3.4 (Deletion of a regular vertex). Suppose L−(z)L^{-}(z) is acyclic over F\mathbb{F}. Deleting zz from a finite poset PP preserves the homology of ΔP\Delta P and the reduced homology of every remaining lower order link. The upper statement holds when L+(z)L^{+}(z) is acyclic over F\mathbb{F}.

Proof. For a simplicial complex KK and vertex zz,

Hr(K,K−z)≅H~r−1(lk⁡Kz).(5)H_r(K,K-z)\cong\widetilde{H}_{r-1}(\operatorname{lk}_K z). \tag*{(5)}

This follows by taking the cone on the link modulo its base. For an order complex,

lk⁡ΔPz=L−(z)∗L+(z).\operatorname{lk}_{\Delta P}z=L^{-}(z)*L^{+}(z).

Over a field, a join with an acyclic factor is acyclic: the augmented join formula is

H~k(A∗B)≅⨁i+j=k−1H~i(A;F)⊗FH~j(B;F).\widetilde{H}_k(A*B)\cong\bigoplus_{i+j=k-1}\widetilde{H}_i(A;\mathbb{F})\otimes_{\mathbb{F}}\widetilde{H}_j(B;\mathbb{F}).

The indices include −1-1. With our augmented convention, A∗∅=AA*\varnothing=A, and the term j=−1j=-1 gives H~k(A∗∅)≅H~k(A)\widetilde{H}_k(A*\varnothing)\cong\widetilde{H}_k(A). The relative groups in (5) vanish, proving preservation of the homology of the full complex.

The lower link at a remaining vertex vv changes only if z<vz<v. Inside L−(v)L^{-}(v), the link of zz is

L−(z)∗Δ({u:z<u<v}).L^{-}(z)*\Delta(\{u:z<u<v\}).

Its first factor is acyclic. The same relative argument proves preservation of the reduced homology of L−(v)L^{-}(v). Reversing the order proves the upper statement.

Proposition 3.5 (Preservation of the data). Let S=SF(D)S=S_{\mathbb{F}}(D) and Q=P(D)∣SQ=P(D)|_S. Inclusion ΔQ↪K(D)\Delta Q\hookrightarrow K(D) induces homology isomorphisms. At every retained vertex it also induces isomorphisms on all lower and upper local groups. The levels Bj,HjB_j,H_j are unchanged.

Proof. Delete the vertices outside SS one at a time. They are initially regular in both directions. At each step, Lemma 3.4 preserves the global homology and all remaining local groups. Thus every vertex still to be deleted is regular in the current restricted order, and induction gives the homology isomorphisms. The boundary levels are preserved by Lemma 3.3, since W⊆SW\subseteq S.

Corollary 3.6 (Idempotence). The homological Maxwell graph is canonical for P(D)P(D) and F\mathbb{F}, and

MFh(MFh(D))=MFh(D).\mathcal{M}_{\mathbb{F}}^{\mathrm{h}}\left(\mathcal{M}_{\mathbb{F}}^{\mathrm{h}}(D)\right)=\mathcal{M}_{\mathbb{F}}^{\mathrm{h}}(D).

Its Morse numbers and Betti numbers equal those of DD.

Proof. The vertex set and its Hasse diagram are defined without choices. Proposition 3.5 preserves WW, both critical sets, and their local groups. Their union in the reduced graph is therefore exactly SS. Taking its Hasse diagram again changes nothing. Vertices outside SS contribute zero to the Morse sums.

An order ideal contains all lower elements of each of its vertices; an order filter contains all upper elements.

Proposition 3.7 (Sublevel and superlevel homology). For every order ideal or order filter U⊆P(D)U\subseteq P(D), inclusion

K(S∩U)↪K(U)K(S\cap U)\hookrightarrow K(U)

induces homology isomorphisms. These isomorphisms commute with inclusions between nested such sets.

Proof. For an ideal, delete U∖SU\setminus S with larger vertices first. When zz is deleted, all of its original lower elements are still present, so its lower link is the original acyclic L−(z)L^{-}(z). Apply Lemma 3.4 at each step. For a filter use smaller vertices first and upper links. For U1⊆U2U_1\subseteq U_2, both composites K(S∩U1)→K(U2)K(S\cap U_1)\to K(U_2) are the same inclusion. This proves naturality.

In particular, for a height function ff decreasing on arrows, the sets {f≤t}\{f \le t\} are ideals and {f≥t}\{f \ge t\} are filters. Hence the reduced and original order complexes have isomorphic sublevel and superlevel persistent homology. The conclusion is a homology isomorphism; acyclic links need not give homotopy equivalences.

A Morse chain complex

The local groups form the chain groups of a Morse complex.

Proposition 3.8. There is a differential on

Cr(D)=⨁v∈SF(D)Er−(v)C_r(D) = \bigoplus_{v \in\mathrm{SF}(D)} E_r^-(v)

making C∗(D)C_*(D) chain homotopy equivalent to C∗(K(D);F)C_*(K(D); \mathbb{F}). A nonzero component of the differential from the summand at vv to the summand at uu requires u<vu < v.

Proof. Group simplices according to their greatest vertex. Their boundary matrix is triangular with respect to the poset labels: a face has the same greatest vertex or a smaller one. The diagonal block at vv is the relative complex of (v∗L−(v),L−(v))(v * L^-(v), L^-(v)).

Over a field, choose bases in each diagonal block separating its homology representatives from elementary acyclic pairs. A pair a,ba,b has consecutive degrees, the same vertex label, and a nonzero coefficient cc of bb in dada. It can be cancelled in the full complex. Replace bb by c−1dac^{-1}da and subtract suitable multiples of aa from the other generators of its degree to remove their new bb coefficient. The equation d2=0d^2=0 then splits off this two-term acyclic summand. This is a chain homotopy equivalence.

Cancellation preserves the order condition. A new matrix entry from label ww to label uu can only factor through the cancelled label vv, so u≤v≤wu \le v \le w. In particular no other diagonal block changes. Cancel all the acyclic pairs. The surviving graded space at vv is E∗−(v)E_*^-(v), and the remaining differential strictly lowers its vertex label. Every vertex outside SF(D)\mathrm{SF}(D) has zero local groups, giving the displayed direct sum.

The basis choices affect the differential but not the numbers mrm_r. Writing dr:Cr→Cr−1d_r:C_r \to C_{r-1} gives

mr(D)−βr(D)=rank⁡dr+rank⁡dr+1.m_r(D) - \beta_r(D) = \operatorname{rank} d_r + \operatorname{rank} d_{r+1}.

Coefficients may connect vertices separated by several cover arrows. This differential is therefore not generally the incidence matrix of the reduced graph viewed as a one-dimensional complex. Reversing the order gives an ascending Morse complex.

First boundary levels and subdivision

Proposition 3.9. A nonsink reaching exactly one sink has zero descending local groups. If v∈B1v \in B_1 reaches kk sinks of DD, then

μ1−(v)=k−1,μr−(v)=0(r≠1).\mu^-_1(v) = k - 1, \qquad\mu^-_r(v) = 0 \quad(r \ne1).

Proof. In the first case the unique sink is a least element of the lower order, so its complex is a cone. In the second case every strict descendant of vv lies outside the first basin boundary: otherwise predecessor closure would contradict v∈Sink⁡(ΓB(D))v \in\operatorname{Sink}(\Gamma_B(D)). Each descendant therefore reaches exactly one sink. Partition them by this sink. There are kk nonempty parts, each with a least element, and elements of different parts are incomparable. The lower link is a disjoint union of kk cones.

The proposition applies within each boundary stage. The groups computed there may differ from the local groups in the original order.

Proposition 3.10. Adding or removing transitive arrows does not change MFh(D)\mathcal{M}_{F}^{h}(D). Subdividing an arrow into a directed path through new vertices of indegree and outdegree one preserves the graph, its levels, and its local multiplicities, with the natural identification of old vertices.

Proof. Transitive arrows do not change reachability. For a single subdivision u→x→vu \to x \to v, the strict lower set of xx has greatest element vv, and its strict upper set has least element uu. Both order links are cones. Deleting xx recovers the original order, and Lemma 3.4 preserves all old local groups.

The vertex xx reaches the same sinks as vv. Inductively, in every descending stage containing xx, the vertex vv is present and their membership at the next step agrees. Thus xx is never a stage sink. The dual statement uses uu and the ascending stages. All old levels are unchanged, so xx belongs to neither WW nor C−∪C+C^{-} \cup C^{+}. Iterate for a longer subdivision.

Computing the construction

Let n=∣V∣n = |V|, m=∣A∣m = |A|, and let RR be the number of nonempty chains of P(D)P(D).

The boundary hierarchy can be computed by propagating at most two distinct reachable-sink identifiers per vertex in reverse topological order. One identifier means a unique reachable sink; two suffice to detect ambiguity. Repeat on the retained boundary, and reverse arrows for the hill hierarchy. Each stage costs O(nj+mj)O(n_j + m_j), and there are at most nn stages in each direction. The total is O(n(n+m))O(n(n + m)) time with O(n+m)O(n + m) working space when stages are stored by exit labels.

Reachability itself can be computed in O(n(n+m))O(n(n + m)) time by a search from each vertex. Once the order complex is listed, dense boundary-rank computations give a coarse bound of O(nR3)O(nR^3) field operations for all lower and upper local groups, treating them one at a time, with O(R2+n2)O(R^2 + n^2) storage. The cancellation procedure in Proposition 3.8 takes at most R/2R/2 steps, each with O(R2)O(R^2) matrix updates, hence O(R3)O(R^3) field operations.

After SF(D)S_F(D) is known, the Hasse diagram on it can be obtained in O(∣SF(D)∣(n+m))O(|S_F(D)|(n + m)) time. For each retained uu, scan its retained descendants with larger elements first. Select an uncovered vv as a cover of uu, then mark vv and all its descendants. Every noncover is marked by a preceding cover, and a cover cannot be marked by a different retained element without an intermediate vertex. Sharing the marking set among these searches gives O(n+m)O(n + m) work per uu.

These bounds describe direct implementations. Since a total order has R=2n−1R = 2^n - 1, enumerating the order complex can already require exponential time. Recognizing conical links and using sparse matrices or cancellations can reduce the computation. Over Q\mathbb{Q}, the bounds above count field operations rather than bit operations.

Examples

We compute the descending Morse numbers; reversing arrows gives the ascending ones. In each example the identity mD(t)−PD(t)=(1+t)QD(t)m_D(t) - P_D(t) = (1 + t)Q_D(t), with QD(t)Q_D(t) having nonnegative coefficients, verifies all the Morse inequalities.

A three-leaf star

Let DD have arrows v→a,v→b,v→cv \to a, v \to b, v \to c; see Figure 1(a). The sinks a,b,ca,b,c each contribute μ0=1\mu_0 = 1. The lower link of vv is three isolated points, so μ1(v)=2\mu_1(v) = 2. All vertices are retained:

Three-leaf star and bipartite four-cycle

Figure 1. Two different roles for degree-one generators: component mergers in (a), and a merger together with a surviving cycle in (b). Crossings of arrows are not vertices.

(m0,m1)=(3,2),(β0,β1)=(1,0).(m_0,m_1)=(3,2), \qquad(\beta_0,\beta_1)=(1,0).

The inequalities read

3≥1,2−3=−1=0−1.3 \ge1,\qquad2-3=-1=0-1.

Equivalently, mD(t)−PD(t)=2+2t=2(1+t)m_D(t)-P_D(t)=2+2t=2(1+t).

Take C0=F⟨a,b,c⟩C_0=\mathbb{F}\langle a,b,c\rangle and two generators e1,e2e_1,e_2 at vv in degree one, with

de1=b−a,de2=c−a.de_1=b-a,\qquad de_2=c-a.

The boundary has rank two. Adding vv joins three components into one, and the multiplicity two counts the two independent mergers.

A directed bipartite graph

Take sources u,vu,v, sinks a,ba,b, and all four arrows from a source to a sink, as in Figure 1(b). There are no directed paths of length two, so the order complex is the four-cycle itself. Each source has a two-point lower link. Thus

(m0,m1)=(2,2),(β0,β1)=(1,1).(m_0,m_1)=(2,2),\qquad(\beta_0,\beta_1)=(1,1).

The strong inequalities are 2≥12\ge1 and 2−2=1−12-2=1-1. The polynomial difference is 1+t1+t.

A Morse complex has degree-one generators eu,eve_u,e_v with deu=dev=b−ade_u=de_v=b-a. Its boundary has rank one, and eu−eve_u-e_v generates H1H_1. In the basis ev,eu−eve_v,e_u-e_v, the first generator joins components and the second represents the cycle.

More generally, orient Kp,qK_{p,q} from its pp sources to its qq sinks, where p,q≥2p,q\ge2. All vertices are essential and retained. Each source has a lower link consisting of qq points, giving

m0=q,m1=p(q−1),β0=1,β1=(p−1)(q−1).m_0=q,\qquad m_1=p(q-1),\qquad\beta_0=1,\qquad\beta_1=(p-1)(q-1).

Hence

mD(t)−PD(t)=(q−1)(1+t).m_D(t)-P_D(t)=(q-1)(1+t).

For K3,4K_{3,4} this gives (m0,m1)=(4,9)(m_0,m_1)=(4,9) and β=(1,6)\beta=(1,6). There are only three vertices in B1B_1, so unweighted level counts would fail even the weak degree-one inequality.

The diamond

Let the arrows be a→ba\to b, a→ca\to c, b→db\to d, c→dc\to d. There is one source and one sink. The first basin and hill boundaries are empty, and W={a,d}W=\{a,d\}. Every nonsink has a conical lower link with least element dd; every nonsource has a conical upper link with greatest element aa. Thus SF(D)={a,d}S_{\mathrm{F}}(D)=\{a,d\} as well.

The full order complex consists of the two triangles abdabd and acdacd joined along their common edge adad. This edge is present in the order complex because aa reaches dd. The complex is a disc, although the underlying one-dimensional graph is a circle. Only dd contributes to the descending Morse numbers:

The diamond reduces to an arrow

Figure 2. The diamond reduces to an arrow. Both reachability order complexes are contractible.

mD(t)=PD(t)=1,QD(t)=0.m_D(t) = P_D(t) = 1,\qquad Q_D(t) = 0.

All Morse inequalities are equalities.

A six-vertex graph with second homology

Arrange six vertices in three pairs {a1,a2}\{a_1,a_2\}, {b1,b2}\{b_1,b_2\}, {c1,c2}\{c_1,c_2\}. Include all four arrows from the aa-pair to the bb-pair and all four from the bb-pair to the cc-pair; see Figure 3. The descending levels are the cc-, bb-, and aa-pairs, respectively, and every vertex is retained.

A DAG with vertices in three pairs and arrows from the a-pair to the b-pair and from the b-pair to the c-pair

Figure 3. A DAG whose order complex is the boundary of an octahedron. The annotations give the single nonzero descending multiplicity at each vertex.

Each cic_i is a sink. The lower link of each bib_i is the two-point set {c1,c2}\{c_1,c_2\}. The lower link of each aia_i is the four-cycle on the bb- and cc-pairs. Consequently

(m0,m1,m2)=(2,2,2).(m_0,m_1,m_2) = (2,2,2).

A chain contains at most one vertex from each pair, and any such selection is a chain. Therefore

K(D)=S0∗S0∗S0≅S2,(β0,β1,β2)=(1,0,1).K(D) = S^0 * S^0 * S^0 \cong S^2,\qquad(\beta_0,\beta_1,\beta_2) = (1,0,1).

The strong inequalities are

2≥1,2−2=0≥−1,2−2+2=2=1−0+1.2 \ge1,\qquad2 - 2 = 0 \ge-1,\qquad2 - 2 + 2 = 2 = 1 - 0 + 1.

The polynomial calculation is

(2+2t+2t2)−(1+t2)=(1+t)2,QD(t)=1+t.(2 + 2t + 2t^2) - (1 + t^2) = (1 + t)^2,\qquad Q_D(t) = 1 + t.

Choose generators z1,z2z_1,z_2 in degree two, e1,e2e_1,e_2 in degree one, and c1,c2c_1,c_2 in degree zero. One choice of differential is

dz1=dz2=e2−e1,de1=de2=c2−c1.d z_1 = d z_2 = e_2 - e_1,\qquad d e_1 = d e_2 = c_2 - c_1.

Both maps have rank one and their composition is zero. Thus H0≅FH_0 \cong\mathbb{F}, H1=0H_1 = 0, and H2=F⟨z1−z2⟩H_2 = \mathbb{F}\langle z_1-z_2\rangle, giving the stated Betti numbers.

A vertex missed by the original hierarchy

Consider the nine-vertex directed tree TT with arrows

a→p,b→p,p→x,c→x,x→q,x→d,q→e,q→f.a \to p,\quad b \to p,\quad p \to x,\quad c \to x,\quad x \to q,\quad x \to d,\quad q \to e,\quad q \to f.

Its only nonempty levels are

B0={d,e,f},B1={q},H0={a,b,c},H1={p}.B_0 = \{d,e,f\},\quad B_1 = \{q\},\quad H_0 = \{a,b,c\},\quad H_1 = \{p\}.

Hence the original Maxwell graph deletes xx.

The original directed tree and its Maxwell reduction

Figure 4. The original reduction removes xx and changes reachability homology. The homological Maxwell graph retains xx and equals the tree on the left.

The vertex xx is comparable with every vertex of TT, so K(T)K(T) is a cone. After removing xx, every element of {a,b,p,c}\{a,b,p,c\} is above every element of {q,d,e,f}\{q,d,e,f\}. The first set has order complex equal to the path a−p−ba-p-b and an isolated cc; the second gives the path e−q−fe-q-f and an isolated dd. Each is homotopy equivalent to S0S^0. Their join is the order complex of the original Maxwell graph, hence

K(M(T))≃S0∗S0≅S1.K(\mathcal{M}(T)) \simeq S^0 * S^0 \cong S^1.

The original reduction has changed the first Betti number from zero to one.

Both links at xx have two contractible components, so μ1−(x)=μ1+(x)=1\mu_1^-(x)=\mu_1^+(x)=1. The modified construction retains xx. The nonzero descending contributions are

μ0−(d)=μ0−(e)=μ0−(f)=1,μ1−(q)=μ1−(x)=1,\mu_0^-(d)=\mu_0^-(e)=\mu_0^-(f)=1,\qquad\mu_1^-(q)=\mu_1^-(x)=1,

with all others zero. Thus SF(T)=V(T)S_{\mathbb{F}}(T)=V(T) and

mT(t)=3+2t,PT(t)=1,QT(t)=2.m_T(t)=3+2t,\qquad P_T(t)=1,\qquad Q_T(t)=2.

The inequalities are the same as for the three-leaf star. For example, choose degree-one generators uq,uxu_q,u_x with duq=f−ed u_q=f-e and dux=d−ed u_x=d-e. Their boundary has rank two. The generator at xx is absent from the original Maxwell graph.

DigraphmD(t)m_D(t)PD(t)P_D(t)QD(t)Q_D(t)
Three-leaf star3+2t3+2t1122
Directed K2,2K_{2,2}2+2t2+2t1+t1+t11
Directed K3,4K_{3,4}4+9t4+9t1+6t1+6t33
Diamond111100
Six-vertex example2+2t+2t22+2t+2t^21+t21+t^21+t1+t
Nine-vertex tree3+2t3+2t1122

Table 4.

Computations verify these values over F2\mathbb{F}_2, F3\mathbb{F}_3, and F5\mathbb{F}_5. They also check the reduction and both Morse complexes on all 1,099 DAGs compatible with a fixed topological ordering through five vertices. The computations provide additional checks of the examples and the reduction.

The original Maxwell graph and path homology

An obstruction to unweighted level counts

Write cj(D)=∣Bj(D)∣c_j(D)=|B_j(D)|. The following obstruction applies to any homology theory with the usual zeroth Betti number.

Proposition 5.1. There is no assignment of nonnegative Betti numbers br(D)b_r(D), with b0(D)b_0(D) equal to the number of weakly connected components, for which the strong Morse inequalities with cj(D)=∣Bj(D)∣c_j(D)=|B_j(D)| hold for every DAG.

Proof. Take the star with arrows v→siv \to s_i, 1≤i≤k1 \le i \le k, where k≥3k \ge3. Its only nonempty descending levels have sizes c0=kc_0=k and c1=1c_1=1. It is connected, so the degree-one strong inequality would give

1−k=c1−c0≥b1−b0=b1−1,henceb1≤2−k<0.1-k=c_1-c_0 \ge b_1-b_0=b_1-1,\qquad\text{hence}\qquad b_1\le2-k<0.

This contradicts nonnegativity.

The star is its own Maxwell graph. Passing to cohomology over a field leaves its Betti numbers unchanged and does not remove the obstruction.

Homology on the unchanged Maxwell graph

Put W=W(D)W=W(D), PW=P(D)∣WP_W=P(D)|_W, and define

HrM(D;F)=Hr(ΔPW;F),brM(D)=dim⁡HrM(D;F).(6)H_r^M(D;\mathbb{F})=H_r(\Delta P_W;\mathbb{F}),\qquad b_r^M(D)=\dim H_r^M(D;\mathbb{F}). \tag*{(6)}

Define the local multiplicities in the reachability order of M(D)\mathcal{M}(D) by

μrM(w)=dim⁡H~r−1(Δ{u∈W:u<w};F),mrM(D)=∑w∈WμrM(w).\mu_r^M(w)=\dim\widetilde{H}_{r-1}\left(\Delta\{u\in W:u<w\};\mathbb{F}\right),\qquad m_r^M(D)=\sum_{w\in W}\mu_r^M(w).

Proposition 5.2. The numbers mrM(D)m_r^M(D) and brM(D)b_r^M(D) satisfy the weak and strong Morse inequalities, including eventual equality. The groups HM∗HM_* are invariant under the original Maxwell reduction, under changes of transitive arrows, and under directed subdivision. The original boundary hierarchies are retained.

Proof. Apply Theorem 1.1 to M(D)\mathcal{M}(D). Lemma 3.3, with U=WU=W, gives the same levels and hence W(M(D))=W(D)W(\mathcal{M}(D))=W(D). Thus M(M(D))=M(D)\mathcal{M}(\mathcal{M}(D))=\mathcal{M}(D). Changing transitive arrows preserves PWP_W. The topographic argument in Proposition 3.10 shows that subdivision introduces no new essential vertex and leaves the restricted order on the old essential vertices unchanged.

For the nine-vertex tree TT of § 4, the Morse and Poincaré polynomials are 3+3t3+3t and 1+t1+t; their difference is 2(1+t)2(1+t), but b1M(T)=1≠0=β1(T)b_1^M(T)=1\ne0=\beta_1(T). Functoriality holds for order-preserving maps carrying essential vertices to essential vertices.

The original counts suffice if each w∈Bjw\in B_j has lower link in PWP_W with the reduced homology of Sj−1S^{j-1}, and all vertices of W∖⋃jBjW\setminus\bigcup_j B_j have acyclic lower links. Then mjM=∣Bj∣m_j^M=|B_j|. This holds in the six-vertex example, but not in the three-leaf star.

Local Morse numbers for path homology

Let Ar(D)A_r(D) be the vector space spanned by allowed elementary paths [v0,…,vr][v_0,\ldots,v_r], with vi→vi+1v_i \to v_{i+1} an arrow for every ii. Following [5], set

Ωr(D)={c∈Ar(D):∂c∈Ar−1(D)},Hrpath(D)=Hr(Ω∗(D),∂).\Omega_r(D)=\{c\in A_r(D):\partial c\in A_{r-1}(D)\},\qquad H_r^{\mathrm{path}}(D)=H_r(\Omega_*(D),\partial).

Here ∂\partial is the alternating vertex-deletion boundary and ∂0=0\partial_0=0. In general Ωr(D)\Omega_r(D) is a space of linear combinations; its basis need not consist of elementary paths.

For a vertex vv, use the induced digraphs

D≤v=D[{u:u≤v}],D<v=D[{u:u<v}],D_{\leq v}=D[\{u:u\leq v\}],\qquad D_{<v}=D[\{u:u<v\}],

with their original arrows, without adding transitive arrows. The inclusion D<v⊆D≤vD_{<v}\subseteq D_{\leq v} embeds their path chain complexes: an allowed chain with allowed boundary in the smaller graph has the same property in the larger graph. Define relative local groups and their dimensions by

Fr−(v)=Hr(Ω∗(D≤v)/Ω∗(D<v)),νr−(v)=dim⁡Fr−(v).(7)F_r^-(v)=H_r(\Omega_*(D_{\leq v})/\Omega_*(D_{<v})),\qquad\nu_r^-(v)=\dim F_r^-(v). \tag*{(7)}

Write pr−(D)=∑vνr−(v)p_r^-(D)=\sum_v\nu_r^-(v) and brpath(D)=dim⁡Hrpath(D)b_r^{\mathrm{path}}(D)=\dim H_r^{\mathrm{path}}(D). The corresponding polynomials are pD(t)=∑rpr−(D)trp_D(t)=\sum_r p_r^-(D)t^r and PDpath(t)=∑rbrpath(D)trP_D^{\mathrm{path}}(t)=\sum_r b_r^{\mathrm{path}}(D)t^r.

Proposition 5.3 (Path-homological Morse inequalities). *For every finite DAG and every q≥0q\geq0,

∑r=0q(−1)q−rpr−(D)≥∑r=0q(−1)q−rbrpath(D),(8)\sum_{r=0}^{q}(-1)^{q-r}p_r^-(D)\geq\sum_{r=0}^{q}(-1)^{q-r}b_r^{\mathrm{path}}(D), \tag*{(8)}

with equality for all sufficiently large qq. Also pr−(D)≥brpath(D)p_r^-(D)\geq b_r^{\mathrm{path}}(D). If DD is transitive, then Fr−(v)≅Er−(v)F_r^-(v)\cong E_r^-(v).*

Proof. Choose a linear extension with descendants first. For its first ii vertices put Di=D[{v1,…,vi}]D_i=D[\{v_1,\ldots,v_i\}]. The inclusions identify the quotient chain complexes:

Ω∗(D≤vi)/Ω∗(D<vi)≅Ω∗(Di)/Ω∗(Di−1).(9)\Omega_*(D_{\leq v_i})/\Omega_*(D_{<v_i})\cong\Omega_*(D_i)/\Omega_*(D_{i-1}). \tag*{(9)}

Indeed, every new allowed path starts at viv_i, and all its remaining vertices are strict descendants. The condition that a boundary be allowed also separates according to the initial vertex: deleting an endpoint always leaves an allowed path, whereas every forbidden face comes from an interior deletion and retains the initial vertex. Consequently forbidden faces cannot cancel between chains with different initial vertices, so each initial-vertex component of an Ω\Omega-chain is itself an Ω\Omega-chain. It follows that

Ω∗(Di)=Ω∗(Di−1)+Ω∗(D≤vi),Ω∗(Di−1)∩Ω∗(D≤vi)=Ω∗(D<vi).\Omega_*(D_i)=\Omega_*(D_{i-1})+\Omega_*(D_{\leq v_i}),\qquad\Omega_*(D_{i-1})\cap\Omega_*(D_{\leq v_i})=\Omega_*(D_{<v_i}).

The subcomplex identities give the chain isomorphism (9).

The exact sequences of this filtration give pr−−brpath=Ar+Ar+1p_r^--b_r^{\mathrm{path}}=A_r+A_{r+1}, where Ar≥0A_r\geq0 and A0=0A_0=0, as in Theorem 1.1. Bounded path lengths give eventual equality. For a transitive DAG, Ω∗(D)\Omega_*(D) is the simplicial chain complex of K(D)K(D), and the local quotient is the cone–base complex of Lemma 3.1.

The local groups are independent of the linear extension. Reversing arrows gives Fr+(v)F_r^+(v) and the ascending inequalities.

Grouping paths by their initial vertex makes the differential triangular, with diagonal homology F∗−(v)F^-_\ast(v). The cancellation proof of Proposition 3.8, or algebraic Morse reduction [14], therefore produces

Pr(D)=⨁v∈V(D)Fr−(v),Hr(P∗(D))≅Hrpath(D).\mathcal{P}_r(D)=\bigoplus_{v\in V(D)}F^-_r(v),\qquad H_r(P_\ast(D))\cong H^{\mathrm{path}}_r(D).

Its differential lowers the reachability label. Set

Spath=W(D)∪{v:F∗−(v)≠0}∪{v:F∗+(v)≠0}.S_{\mathrm{path}}=W(D)\cup\{v:F^-_\ast(v)\ne0\}\cup\{v:F^+_\ast(v)\ne0\}.

By Lemma 3.3, the Hasse diagram on SpathS_{\mathrm{path}} retains both hierarchies. The chain model carries extra data: its differential depends on choices and may connect nonadjacent comparable vertices. It need not be the ordinary path complex of this Hasse diagram.

Subdivision and admissible deletions

The diamond provides a simple obstruction to preserving path homology by the original Maxwell reduction. Subdivide its arrow a→ba\to b to obtain the graph JJ:

a→x→b→d,a→c→d.a\to x\to b\to d,\qquad a\to c\to d.

Both original Maxwell graphs are the single arrow a→da\to d. For the diamond, Ω2\Omega_2 is spanned by [a,b,d]−[a,c,d][a,b,d]-[a,c,d], whose boundary fills its only undirected cycle. Thus its first path Betti number is zero. In JJ, the three allowed two-paths have distinct forbidden shortcut faces [a,b][a,b], [x,d][x,d], and [a,d][a,d]. Hence Ω2(J)=0\Omega_2(J)=0. Its five arrows and five vertices form a connected cycle, so

H1path(J;F)≅F.H^{\mathrm{path}}_1(J;\mathbb{F})\cong\mathbb{F}.

Thus the original Maxwell graph does not determine path homology, which is also not subdivision-invariant. Since JJ has a unique sink, all nonsink lower order links are cones, so m1(J)=0m_1(J)=0. This also proves the failure of substituting path Betti numbers in (2) noted in the introduction.

In JJ, only dd contributes in degree zero and only aa contributes in degree one:

pJ(t)=1+t=PJpath(t).p_J(t)=1+t=P^{\mathrm{path}}_J(t).

Removing aa leaves a directed tree, so its relative group records the cycle; all other nonsink local groups vanish. Reversing arrows gives Spath={a,d}S_{\mathrm{path}}=\{a,d\}. The chain model has one generator in each of degrees zero and one and zero differential. The arrow a→da\to d itself has no degree-one path homology.

A useful deletion is possible when zz has exactly one incoming arrow u→zu\to z, one outgoing arrow z→wz\to w, and u→wu\to w is already present. The retraction r:D→D−zr:D\to D-z sending zz to ww and fixing other vertices satisfies id⁡D≃i∘r\operatorname{id}_D\simeq i\circ r via z→wz\to w, where ii is inclusion. Thus deletion preserves path homology; see [15 Section 6]. It preserves reachability on the remaining vertices, and zz belongs to no boundary level by the argument of Proposition 3.10. Both hierarchies therefore survive. Inserting a missing shortcut first is not allowed.

It remains to characterize when Maxwell levels agree with local homological degrees, and to find a canonical graph reduction preserving both path homology and the two hierarchies.

AI use statement. LLMs were used for proof refinement, literature review, and manuscript revision.

The University of Texas Rio Grande Valley, Brownsville, Texas, USA

[email protected]

References

  1. [1]A. V. Aho, M. R. Garey, and J. D. Ullman, The transitive reduction of a directed graph, SIAM J. Comput. 1 (1972), no. 2, 131–137. doi:10.1137/0201008.DOI
  2. [2]J. A. Barmak and E. G. Minian, Simple homotopy types and finite spaces, Adv. Math. 218 (2008), 87–104. doi:10.1016/j.aim.2007.11.019.DOI
  3. [3]M. Burfitt, J. Wu, S. S.-T. Yau, and S.-T. Yau, Computing singular simplicial homologies of digraphs and quivers, arXiv:2509.02065, version 2 (2025). https://arxiv.org/abs/2509.02065.
  4. [4]R. Forman, Morse theory for cell complexes, Adv. Math. 134 (1998), no. 1, 90–145. doi:10.1006/aima.1997.1650.DOI
  5. [5]A. A. Grigor’yan, Y. Lin, Yu. V. Muranov, and S.-T. Yau, Path complexes and their homologies, J. Math. Sci. (N.Y.) 248 (2020), no. 5, 564–599. doi:10.1007/s10958-020-04897-9.DOI
  6. [6]R. Hepworth and E. Roff, The reachability homology of a directed graph, Int. Math. Res. Not. 2025, no. 3, rnae280. doi:10.1093/imrn/rnae280.DOI
  7. [7]Y. Lin, C. Wang, and S.-T. Yau, Discrete Morse theory on digraphs, Pure Appl. Math. Q. 17 (2021), no. 5, 1711–1737. doi:10.4310/PAMQ.2021.v17.n5.a4.DOI
  8. [8]J. C. Maxwell, On hills and dales, Philos. Mag. (4) 40 (1870), no. 269, 421–427; reprinted in The Scientific Papers of James Clerk Maxwell, vol. II, W. D. Niven (ed.), Cambridge University Press, 1890, 233–240. doi:10.1017/CBO9780511710377.018 (collected edition).
  9. [9]J. Milnor, Morse Theory, Annals of Mathematics Studies, vol. 51, Princeton University Press, Princeton, NJ, 1963.
  10. [10]E. G. Minian, Some remarks on Morse theory for posets, homological Morse theory and finite manifolds, arXiv:1007.1930, version 2 (2012). https://arxiv.org/abs/1007.1930.
  11. [11]Yu. V. Muranov, Alexandroff homology and path homology, Math. Notes 112 (2022), 159–162. doi:10.1134/S0001434622070185.DOI
  12. [12]O. R. Musin, On some problems of computational geometry and topology, in Global Analysis—Studies and Applications V, Yu. G. Borisovich, Yu. E. Gliklikh, and A. M. Vershik (eds.), Lecture Notes in Math., vol. 1520, Springer, Berlin, 1992, 57–80. doi:10.1007/BFb0084715.DOI
  13. [13]O. R. Musin, Topographic structure of image, in Computer Analysis of Images and Patterns (CAIP 1993), D. Chetverikov and W. G. Kropatsch (eds.), Lecture Notes in Comput. Sci., vol. 719, Springer, Berlin, 1993, 24–30. doi:10.1007/3-540-57233-3_3.DOI
  14. [14]E. Sköldberg, Algebraic Morse theory and homological perturbation theory, Algebra Discrete Math. 26 (2018), no. 1, 124–129. https://arxiv.org/abs/1311.5803.
  15. [15]C. Wang and S. Ren, A discrete Morse theory for digraphs, arXiv:2007.13425, version 1 (2020). https://arxiv.org/abs/2007.13425.

Paper details

vlnc:W_dp73as6p
v1 · latestHistory (1)

published · Fresh — added to the Hub within the last 30 days

added ·

Usage

views · 1

downloads · 0

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

CC BY 4.0

Original source

  • Mathematics

Not recorded for this paper · View statement →

Looking for related papers…
Contents