The sharp vertex bound
for conditional decomposability
Caius Folkerts
Abstract. We prove that the least number of vertices of a conditionally
Minkowski decomposable -polytope is for every . Equivalently, below this threshold
decomposability is determined by the face lattice. The proof uses a small cover of the
edge graph by rigid subgraphs and studies the homotheties induced on those pieces.
A dimension count in their parameter space forces either a summand of smaller
dimension or a facet satisfying Shephard’s criterion; the remaining three-homothety
case is controlled by concurrence and a forest count. These mechanisms give a
uniform proof without a classification of the combinatorial types.
1. Introduction
A Minkowski decomposition depends on the geometry of a polytope, yet many sufficient
conditions for decomposability depend only on its face lattice. A combinatorial type is
conditionally decomposable, or ambiguous, if it has both decomposable and indecomposable
realizations. The smallest such three-dimensional types have eight vertices. In arbitrary
dimension, Wang and Yost constructed conditionally decomposable -polytopes with
vertices, proved the lower bound , and asked whether their upper bound was sharp [5].
They also established sharpness when the decomposable realization has a segment
summand.
Theorem 1.For every , the least number of vertices of a conditionally
decomposable -polytope is .
The lower bound follows from a stronger assertion. If one realization of a -polytope with
at most vertices is decomposable, then every realization of its face lattice is decomposable.
Wang and Yost’s wedge construction supplies equality.
The numerical threshold comes from the interaction between the graph and the
geometry of parallel maps. Since such a map is homothetic on each rigid part, a cover of
the vertex set by a few rigid subgraphs encodes it by points in the -dimensional space of
homothety parameters. Differences of these parameters locate the edges joining different
parts, giving a dimension bound on the number of vertices in each part; equality then
produces actual faces of the polytope.
Two outcomes recur. If the parameter directions miss the polytope, a projective change
makes the image of the parallel map lower dimensional; below the vertex threshold, this
image is a genuine weak summand, whose decomposition can be reconstructed in every
realization. If the parameter directions meet the polytope, the small number of remaining
vertices forces a facet satisfying Shephard’s criterion. With three parameters,
the same alternative is obtained by studying their three centers of concurrence.
The graph reduction at the end of the proof shows that these cases exhaust the
possibilities.
Throughout, and is full dimensional. A polytope is universally decomposable if every
realization of its combinatorial type is decomposable. A weak summand of is a polytope
for which for some and some polytope .
2. Parallel maps and combinatorial forcing
A map is parallel if
on every edge . It is a homothety if on all vertices. Kallay’s criterion says that is
decomposable exactly when it admits a parallel map that is not a homothety [3]. One way
to see the converse is to form the continuous function that is linear on each cone of the
normal fan and has gradient on the cone at . Adding a sufficiently large multiple of the
support function of makes it convex; after rescaling, it is the support function of a
summand.
Call a geometric subgraph rigid when all its parallel maps are homotheties. Triangles
are rigid, a rigid subgraph absorbs a vertex with two neighbors in it, and two rigid
subgraphs with two common vertices merge. We shall also use rigidity of a consisting of
actual polytope edges. If one of its four-cycles is nonplanar, that cycle is rigid and absorbs
the fifth vertex. If all four-cycles are planar, the five vertices lie in one plane, which would
put three edge rays at a vertex in a pointed two-dimensional cone, contradicting
extremality of the middle ray.
Lemma 2 (Dense rigidity and concurrence).A geometric graph on polytope
vertices, with minimum degree at least , is rigid. The same numerical condition forces
concurrence of a matching of edge lines if its cofaciality graph has minimum degree and every witnessing two-face contains exactly two matching edges.
Proof.The graph has a triangle, since its minimum degree exceeds . Close that
triangle under two-neighbor absorption, and suppose the resulting set has size . Every
outside vertex has at most one neighbor in it. The degree bound then gives
where the middle expression counts the set and the distinct outside neighbors
required by its vertices. But for . The case is a triangle.
For the line version, three pairwise coplanar matching lines are concurrent unless
all lie in one plane. In the latter case their witnessing faces would coincide and
contain all three matching edges, contrary to the hypothesis. Once two lines through a
common point have been obtained, any line cofacial with both must pass through that
point; otherwise the same planar contradiction occurs. The preceding two-neighbor
ordering therefore propagates concurrence to the entire family. □
In particular, a decomposable -polytope has at least vertices. A pyramid is rigid, since
its apex and every base edge form a triangle, and these triangles are joined along edges by
connectedness of the base graph.
A facet has Shephard’s property if every vertex of has exactly one neighbor outside .
Shephard’s criterion states that such a facet makes decomposable whenever at least two
vertices lie outside it [4]. It can be seen by moving its supporting plane slightly while
retaining all other supporting planes; the facet vertices move along their unique exterior
edges, giving a nontrivial parallel map. Thus every simple nonsimplex is universally
decomposable.
Lemma 3.If a facet has a rigid complement in a decomposable polytope, it has
Shephard’s property.
Proof.Subtract the homothety on the complement from a nontrivial parallel map.
If a facet vertex had two exterior neighbors, its value would be zero. The normal
component of the map is constant on the facet, so it would then be zero everywhere
on that facet. Every facet vertex has an exterior neighbor; its value is parallel to that
exterior edge and has zero normal component, hence vanishes. This would make the
whole map trivial. □
Parallel maps are projectively invariant. For
with positive denominator on , the transported map is
A homothety is sent to ; thus its center transforms as a projective point, including a
center at infinity.
We shall use Balinski’s theorem that the graph of a -polytope is -connected [1], and two
numerical facts. The excess
of a nonsimple -polytope is at least , and at equality the polytope has a facet with
Shephard’s property [4, Theorems 3.3 and 4.10]. A simplicial four-polytope on vertices has
at least edges [2].
3. Proper images and persistent summands
A low-dimensional image of a parallel map need not be convex in the right sense to define
a summand. The vertex bound removes this difficulty; the quotient fibers must
carry the entire image, and one of them is too small to support a nonhomothetic
map.
Lemma 4.Suppose and a nonconstant parallel map has image of affine dimension
. After a translation, a change of sign, and a positive scaling, it is a weak summand
map.
Proof.Translate the image into an -space . The map is constant on every edge not
parallel to . Choose a functional vanishing on and exposing a vertex of the quotient .
A strictly increasing edge path from any vertex ends in that exposed fiber, preserving
along the way. Thus each quotient-vertex fiber carries the entire image. A parallel
map on a connected face takes values in a translate of its direction space, so each
such fiber has dimension exactly .
There are at least disjoint vertex fibers. If all their restrictions were
nonhomothetic, Lemma 2 would give
a contradiction. On one fiber the map is therefore with ; the same conclusion is
automatic when .
To recover convexity from this one fiber, define on the normal cone of each vertex
. Parallelism makes the pieces agree on their common walls. The gradients lie in , so
factors through . Choose exposing . For large , a vertex exposed by lies in and
maximizes there. Evaluating at this direction therefore gives
Choose the sign so that this factor is convex. A sufficiently small positive multiple
can then be subtracted from without losing convexity, giving the required weak
summand. □
Lemma 5 (Two homotheties).If and a nontrivial parallel map is covered by two
homotheties, is universally decomposable. In particular, the conclusion holds when has a segment weak summand.
Proof.The lines of edges joining the two classes concur at the center of their
difference. Neither class can be a singleton, since two-neighbor absorption would then
make the map homothetic. A center in would make the crossing edges a star or
a single edge, contradicting -connectivity. The center is therefore exterior and the
crossing edges form a matching of size at most .
Every two-face incident to a crossing edge contains exactly two crossing edges.
Distinct two-faces give distinct partners, so the cofaciality graph has minimum degree
. Lemma 2 forces a common projective center in every realization. It remains exterior
because two disjoint actual edges cannot meet inside the polytope. Choose two
distinct homotheties with this difference center; they define a nontrivial parallel map
on the two classes in the new realization.
For a segment summand, the summand map takes two values and is already
covered by two constant homotheties. □
We next reconstruct a simplex summand from its edge centers. The role of two disjoint
simplex faces is to make the compatibility among these centers combinatorial.
Lemma 6 (Simplex reconstruction).Suppose the vertices are colored by the
vertices of an -simplex, every color occurs at least twice, and every crossing family has a combinatorially forced projective center . If there are two disjoint simplex faces
with one vertex of each color, every realization admits a nontrivial parallel map that
is homothetic on each color.
Proof.Corresponding triangles in the two simplex faces force to be distinct and
collinear. In one simplex, with homogeneous vertices , the centers span a hyperplane
avoiding all the , and the triangle collinearities put every in it. If defines that
hyperplane, set . Then represents . The linear identification turns the into
compatible homothety parameters. Distinct parameters and at least two vertices in
each class ensure nontriviality. □
Proposition 7.Suppose and has a weak summand of dimension with . Then is universally decomposable. The same holds for a two-dimensional weak summand
when .
Proof.Rescale the summand and write , , . Quotient by the direction space of . Each
quotient-vertex fiber contains as a summand and therefore has at least vertices. A
vertex of receives the label of its vertex in .
Codimension at least three. Suppose . Since
must be a simplex. At most edges at a vertex change its label, one for each
incident simplex edge, so every color has at least vertices. For a fixed simplex edge,
its parallel copies form a matching, and
An edge belongs to at least two-faces. At most have a triangular image in , one
for each possible third label; every other two-face contains exactly two members of
the fixed family. Its cofaciality graph has minimum degree , so Lemma 2 forces its
center in every realization.
A quotient-vertex fiber is either a simplex or has at least vertices. If at most one
were a simplex, then
For this is impossible, and two simplex fibers give the conclusion by Lemma 6.
When , the only remaining equality consists of one triangular and quadrilateral
fibers over a quotient simplex, with no additional vertices. The facet opposite the
triangle has Shephard’s property; any edge changing quotient position preserves the
label, and the triangle has exactly one vertex of each label.
Codimension two. Now . The quotient is a triangle, since four vertex fibers would
require vertices. Also
so is indecomposable. Every vertex fiber is homothetic to or has at least vertices,
and at least one is homothetic. If these three fibers contain all vertices, the facet
opposite a homothetic fiber has Shephard’s property by the same label argument.
Suppose additional vertices occur and one fiber is nonhomothetic. Counting forces
two simplex fibers, a -vertex fiber, and one extra vertex . An edge through cannot
change its label, since that would require a second vertex at the same quotient point.
At most two vertices of its label occur in the simplex fibers and at most in the
remaining fiber. Hence is simple. The hyperplane through its neighbors cuts off only
, so deletion has a combinatorially determined type with a new simplex facet. It is
the preceding Shephard type. In every realization a nontrivial parallel map on the
deletion is homothetic on that simplex facet and extends by the same homothety to
the restored vertex. Thus restoration preserves universal decomposability.
It remains that all three vertex fibers are homothetic and extra vertices occur.
Their convex hull has fibers , where and are affine barycentric interpolants. An
extra vertex violates a side inequality
with exposing a facet of . The parallel exposed face of beyond this inequality
contains only extra vertices and retains that facet as a summand, so it needs at least
vertices. Thus
Equality forces to be a simplex and all extra vertices to form one -simplex ,
parallel to a facet of , over one nonvertex quotient point. Its labels are precisely those
of that facet.
The previous two-face count gives each edge family cofaciality degree at least
two. A family involving the missing label has exactly three members, so its lines
are forced concurrent. Every other family has four members, including one edge
of . A two-face through that edge with triangular image must lie in , because a
label-changing edge preserves the quotient position. There are at most such faces.
The extra edge therefore has cofaciality degree at least ; the other degrees are at
least two. This graph contains minus at most one edge, so two-neighbor propagation
forces concurrence. The simplex reconstruction lemma applies.
The three-dimensional endpoint. A planar summand of a three-polytope on at
most seven vertices must be a triangle, since its two endpoint fibers already use twice
its number of vertices. If one endpoint fiber is not homothetic to the triangle, it has
at least four vertices, leaving no others; the opposite facet is Shephard. Otherwise
both fibers are homothetic. Any additional vertex would violate a side inequality of
their hull, and the exposed face beyond it retains a triangle edge as a summand,
requiring two additional vertices. This is impossible. Either endpoint facet is therefore
Shephard. □
4. The geometry of homothety parameters
Fix a nontrivial parallel map and a cover of its vertices by homotheties
Enlarge each class to the full set and delete redundant classes. Every remaining class
then has a private vertex. Put
The use of homogeneous coordinates identifies with a subset of . A vertex is central if ;
a noncentral vertex belongs to exactly one class, since two homotheties agreeing at have
difference proportional to .
For an edge , with , and edge scalar , parallelism gives the identity
This puts a crossing edge and its homothety difference in the same projective line. It is
the common source of the dimension counts and the concurrence arguments
below.
Lemma 8 (The class bound).Every essential class has at least noncentral
vertices.
Proof.At a private vertex , an edge to has in (2); otherwise would also belong to
. Thus . The edge star at spans the ambient space, so
Central vertices contribute nothing modulo , and has dimension . □
Let be the number of central vertices, and write the number of noncentral vertices of
as . The class bound gives . With , a cover by essential classes therefore satisfies
These counts are defined before we know whether the central vertices form a
face.
Equality produces a simplex of quotient directions and, more significantly, faces in
.
Lemma 9 (Fiber faces).Suppose a class has exactly noncentral vertices . Their
classes form a basis modulo , they form a clique, and
is a -face for each . These faces meet in the face . The corresponding quotient
coordinates are nonnegative on , and every equation defines a facet.
Proof.Every neighbor of other than its classmates has quotient proportional to , by
(2). The star spans the quotient, so every classmate must be a neighbor. Normalize
the quotient basis by . For , this functional is nonnegative on all edge rays at and
hence on ; varying proves nonnegativity of every .
The sum exposes . At , only the classmate edges leave this face, so its dimension
is at least , and the defining projective subspace gives the reverse inequality. At a
different , only the edge to can leave , making that face a facet. Their common zero
set is . □
If is empty, the positive denominator sends to infinity. Formula (1) then makes all
homothety scales equal. After subtracting their common homothety, the image has
dimension at most the number of classes minus one. This is precisely the situation handled
by Section 3.
For the other case, suppose the parameters are affinely independent and a
minimum class exists, so Lemma 9 applies. Now , and a central vertex belongs
to at most two classes. Let be the set of classes containing a central vertex .
Central-to-noncentral edges can only join vertices in the same class, while crossing edges
between noncentral vertices preserve their quotient direction. Each of the faces
supplies at least neighbors of outside . Counting these distinct neighbors gives
This inequality measures the cost of a central vertex not having enough neighbors in
the central face.
Lemma 10 (A common central class).Suppose , every central vertex belongs to
one class , and
Then has Shephard’s property.
Proof.Among the other classes, choose one with . Its vertices are simple; there are
at most neighbors within the class and at most one towards each other parameter.
At one such vertex take the facet omitting its edge towards . This facet contains the
whole minimum class and all difference centers between the other classes. Propagating
along their crossing matchings puts at least vertices of every class , , into the facet.
An omitted vertex of such a class has at most neighbors in its own class, at most
in the other noncentral classes, and at most one towards . Its degree is therefore at
most
which is impossible. All classes other than lie in the facet, and each of their
vertices has its unique neighbor outside. Facet connectivity then excludes any vertex
inside the facet. This is Shephard’s property. □
5. Three homotheties
The class bound alone does not settle a cover by three homotheties. Here their difference
directions form one projective line, and two-faces control the three possible crossing
families.
Proposition 11.If , , and a nontrivial parallel map is covered by at most three
homotheties, then is universally decomposable.
Proof.Pass to essential full classes and assign each shared vertex to one of its classes,
obtaining a disjoint coloring. Two colors are covered by Lemma 5. If a crossing family
is absent, say , subtract and split the resulting map into its restrictions supported on
colors 0 and 1. Both are parallel and at least one is nontrivial, giving a two-class map.
If the three parameters are collinear, replace them by two values on their common
line; every crossing edge still respects the same center, so this also reduces to two
classes.
We may therefore assume that the parameters are affinely independent, that all
three crossing families are nonempty, and that their distinct centers lie on the line
. Lemma 8 gives at least noncentral vertices of each color. If misses , a positive
projectivity sends it to infinity and makes the image at most two dimensional.
Lemma 4, Lemma 5, and Proposition 7 apply. Henceforth assume .
Forcing an exterior center. An exterior center makes its crossing family a
matching. For an edge off , at most one incident two-face has plane containing
. Every other incident two-face uses just this family; another family would put
two distinct centers in its plane. Such a face contains exactly two members of the
matching. Distinct faces give distinct partner edges, so the off-line cofaciality graph
has minimum degree .
Once contains vertices, an off-line family with edges uses only noncentral vertices,
whence
For , this gives , and Lemma 2 forces its center in every realization. For , and
minimum degree one gives two cofacial lines, so they too have a forced intersection.
The new center remains exterior, since two disjoint actual edges cannot meet in . We
use this argument only after showing that contains a vertex.
A center is a vertex. There cannot be two vertex centers; their owner pairs share
a color, and deleting those two vertices disconnects that color’s noncentral vertices
from the others, contrary to Balinski’s theorem. Suppose the center of is a vertex
, assigned to color 0. If it has no color-0 neighbor, assign it to color 1 instead; the
family disappears and we are done. Otherwise has neighbors of both colors 0 and
1.
It has no color-2 neighbor. Such a neighbor would lie on , by (2), and would be
unique. Delete its vertex from the vertex figure at . The remaining graph is connected
and contains both colors 0 and 1, so has an edge joining them. Its corresponding
two-face contains one edge. Parity around its boundary then forces an edge of each
other crossing family, so its plane contains and hence the edge . But the two edges
of this face at were chosen to avoid , a contradiction.
The other centers must be exterior. Indeed, a center in the interior of an edge has
that edge as its sole crossing edge. Parity in two incident two-faces puts on its line.
The vertex would then be an endpoint of that edge, giving the forbidden color-2
neighbor. Thus both nonstar families lie off , and the preceding matching argument
forces their centers.
The vertex-figure graph is two-connected, so there are at least two edges between
its color-0 and color-1 parts. The corresponding two-face planes each contain and an
edge of both nonstar families. In every realization their intersection therefore makes
and the two reconstructed centers collinear. Three distinct collinear centers give
compatible homothety differences. If the two exterior centers coincide, merge colors
0 and 1. Either way a nontrivial parallel map is reconstructed.
No center is a vertex. If a center lies in the interior of an edge , that family consists
only of . Every two-face through it uses the other two families by parity, and two
such planes show that is the line of . The other centers are exterior, since the same
actual edge cannot belong to two families.
The same configuration is forced if all centers are exterior. The graph of crossing
edges is a forest. To prove this, a cycle has one constant nonzero quotient direction
modulo , so lies in a plane containing . As a cycle of actual polytope edges, it bounds
the section of by that plane. Yet every supporting edge line meets outside , which
is incompatible with meeting that section. Thus there is no cycle.
Each crossing component spans at most one quotient direction, so there are at
least components. Any family containing an off-line edge has at least edges, by its
cofaciality degree. If all three did, the forest would have at least vertices. Hence
some family consists of the unique edge on .
In both cases, contains the endpoints of , and the matching argument forces the
other two centers in every realization. Each two-face through contains an edge of
both other families, so two such planes put their centers on the line of . Choose the
two homothety differences with these centers. Their third difference is also centered
on , and therefore respects the sole remaining crossing edge. Coincident centers give
a two-color map. All constructed maps are nontrivial because each color has at least
vertices and the homotheties do not all agree. □
6. The uniform parameter count
With four or more classes, the vertex budget is strong enough to force full affine
independence of the parameters. Equation (3) then leaves only two possible sizes for the
central face.
Proposition 12.Suppose and a nontrivial parallel map has an essential cover by homotheties, where . Then either a positive projectivity gives an image of dimension
at most , or has Shephard’s property.
The case . Now , and (3) forces every central vertex to be shared and to have degree in
. Thus is a clique and is a homothety . If were not a given parameter, the ownership sets
at different vertices would be disjoint, and at most one vertex could be shared; two lines
joining disjoint pairs of affinely independent parameters cannot intersect at .
Since , this is impossible. Therefore is a common central class and Lemma 10
applies.
The case . Here and . Lemma 2 makes rigid for , and the case is also rigid. For , any
triangle absorbs the remaining vertex, leaving only a central four-cycle as an
exception.
First suppose . If is a given parameter, use Lemma 10. Otherwise the ownership sets
are disjoint and at most one vertex is shared. A singly owned vertex has degree by (3);
a missing edge would therefore have two shared endpoints. It follows that is
complete.
If no vertex is shared, one class owns no central vertex. Its vertices supply the
fiber faces, each with one noncentral vertex of every color. In each fiber face the
vertex is universal, and the other vertices have minimum degree , hence form a
connected graph. The cone over a spanning tree is rigid. The noncentral vertices of
that fiber form the complement of the facet , so Lemma 3 makes that facet
Shephard.
If a central vertex is shared by classes , all other central vertices are singly owned. The
minimum - and -fiber systems coincide, since their matchings exhaust any singly owned
class. Their two vertices in each fiber share the central vertex and the other row vertices.
This contains a rigid , forcing , contrary to essentiality.
It remains that and is a four-cycle. Every central vertex is shared by (3). Along the
connected graph of , identity (2) puts its owner parameters in one translate of , of
dimension three. Thus at most four classes occur. Their owner pairs are distinct and
triangle-free, since a triangle of pairs would give three collinear polytope vertices. They
form the cycle , leaving one class unowned. Disjoint owner pairs cannot be adjacent in ,
because the edge homothety would meet two skew parameter lines; hence this is also the
physical cyclic order.
All noncentral vertices are simple. At an vertex take the facet omitting its edge
towards class 1. Its homogeneous span is
where . This facet contains exactly
Vertices in and 3 have their unique exterior neighbor in class 1; those in 2
and have it at ; those in 4 and have it at . The facet is therefore Shephard.
□
7. The two boundary covers
The range leaves two small configurations that can occur in dimensions four and five.
Both have the same seven-vertex residue. Let consist of two triangles meeting at a vertex
, and a disjoint edge .
A seven-vertex graph of minimum degree three with no rigid set of four vertices has
such a cover. Its degrees are three or four, and odd order forces a degree-four vertex .
Equality in the neighborhood count used below makes its neighbor graph . Its
two nonneighbors are adjacent and each meets one endpoint of each , giving
.
Lemma 13.Suppose a nontrivial parallel map is covered, in dimension four, by a
rigid four-set and , or, in dimension five, by two disjoint rigid four-sets and . Then
is universally decomposable.
Proof.Fewer essential parameters are covered by Propositions 11 and 12, so assume
that all four or five parameters are essential and distinct. The shared vertex is
central. If , the class bound would require at least noncentral vertices in addition to
, exceeding . Hence and .
The noncentral parts of have exactly two vertices, so all excess lies in , or in .
Lemma 9 applies and the central face contains .
For , the budget is . At , inequality (3) gives , so . If , the homothety on that
edge is a common central class; otherwise its other endpoint would be singly owned
and require excess at least two, whereas . The common class is or , and Lemma 10
applies. If , the other two central vertices belong to ; the homothety on the central
triangle would be , giving three owners at , which is impossible.
For , we have and , so . Suppose . The central triangle has a homothety . If is
not a given parameter, its three ownership sets are disjoint. The vertex uses , and
the other two central vertices belong respectively to . Each original four-set has only
one central vertex, so . At most one of those two central vertices can also belong to ,
leaving a singly owned vertex. Inequality (3) requires excess at least two in its class,
whence , contrary to . Thus is a common central class, and Lemma 10 applies.
Suppose . If were homothetic, one of would own two of its vertices and hence its
homothety, again giving three owners at . Therefore is nonrigid. A face of dimension
at least three with at most five vertices is rigid by Lemma 2, so is a quadrilateral
or pentagon.
Both and own central vertices; each original four-set has at least two noncentral
vertices, and neither can account for all centers other than . Together with , these
are four owner colors. The three-dimensional space allows no fifth, leaving unowned.
Since and , equation (3) makes every central vertex shared. Their distinct owner
pairs form a triangle-free graph on four colors, forcing and the four-cycle from the
proof of Proposition 12.
The sole excess vertex belongs to or ; call this color 1 and use as the unowned
class. The hyperplane (5) again gives a Shephard facet, containing the other three
classes and the two central vertices . Its two ordinary color-1 vertices lie outside. If
the possible third color-1 vertex were inside, it would have no neighbor in the facet;
its classmates and incident centers are outside, and every other crossing edge has
center outside the hyperplane. Facet connectivity excludes it. This completes both
boundary cases. □
8. Rigid covers and completion of the proof
We finally show that the graph always supplies one of the covers already treated. Choose a
largest rigid set in each remaining induced graph. Each removal lowers the minimum
degree by at most one, because an outside vertex with two neighbors in the removed set
would have been absorbed into it.
Lemma 14.A graph of minimum degree three on at most seven polytope vertices
is covered by three rigid sets; on at most six vertices, two suffice.
Proof.There is a rigid set with at least three vertices. Otherwise the graph would
be triangle-free and -free. At a vertex of degree , its independent neighbors require
at least outgoing edges, whereas each other vertex can meet at most two of them.
Thus . For this forces a three-regular graph on seven vertices, impossible by parity.
Enlarge a rigid set maximally. Its complement has minimum degree two and
at most four vertices. A three-vertex complement is a triangle, and a four-vertex
complement has a perfect matching and is covered by two rigid edges. This proves
the claims. □
The counting estimate used in the remaining argument is elementary but effective. If all
rigid sets have size at most three, a degree- vertex in a graph of minimum degree satisfies
Indeed, its neighbor graph has components of size at most two, while every other vertex
meets at most two of those neighbors, by rigidity of . Thus . If rigid sets have size at most
four, the same argument, with neighbor components of size at most three, gives
8.1. The four-regular tail. A graph of minimum degree four on at most ten remaining
vertices is covered by three rigid sets, except for a configuration that cannot occur here. If
it has a rigid four-set, remove it and apply Lemma 14. Otherwise (6) makes it
four-regular and forces at least nine vertices. On nine vertices, equality gives a triangle;
removing it leaves six vertices of minimum degree three and again gives a three-set
cover.
Suppose the tail has ten vertices. Every vertex lies in one or two triangles. Some vertex
lies in exactly one, since twenty triangle incidences cannot be divisible by three. Equality
in (6) then makes each of its five nonneighbors adjacent to two vertices of its star. All
four-cycles are planar, since a nonplanar one would be rigid. Thus the entire tail lies in the
four-flat spanned by this star.
At this stage every tail vertex is simple in the original polytope; it has four tail
neighbors and at most one neighbor in each of the removed sets. At the chosen vertex, the
four tail edges therefore determine a four-face in that flat. All tail vertices lie in this face,
and their four tail edges exhaust their possible edges within it. Face connectivity
excludes any earlier vertex. We have obtained a simple four-polytope with ten
vertices.
No such polytope exists. Its simplicial dual has ten facets, so if it has vertices,
Dehn–Sommerville gives . Barnette’s bound gives , hence , contrary to . This excludes the
ten-vertex tail.
8.2. Dimensions at least six. For a nonsimple -polytope with and , a largest
rigid set has at least five vertices. Otherwise (7) at a nonsimple vertex gives
which is impossible for . For it forces degree seven; the neighbor partition then
requires at least outgoing edges, while the other vertices have capacity only
22.
After removing this first set, the residual graph has minimum degree and at most
vertices. For , inequality (6) forces a rigid set of size at least four. At , failure would force
a five-regular graph on fourteen vertices, each neighborhood consisting of and
one isolated vertex. Every vertex would lie in two triangles, giving twenty-eight
triangle incidences, again impossible. Continue removing at least four vertices until
.
At most ten vertices now remain, and the preceding tail argument covers them
by three rigid sets. In total we have at most sets. A nontrivial parallel map
is homothetic on each, so Propositions 11 and 12 apply. In the proper-image
alternative, its dimension is at most , and Section 3 gives universal decomposability.
In the other alternative the Shephard facet does so; the pyramid exception is
excluded by the nontrivial map. Simple types were already settled by Shephard’s
criterion.
8.3. Dimensions four and five. For , inequality (6) gives a rigid set of size at least
four in every nonsimple polytope on at most eleven vertices. If , Lemma 14 covers its
complement by two sets. If , the same holds unless the complement has seven vertices and
no rigid four-set. That is precisely the configuration from Section 7. Thus either
Proposition 11 or Lemma 13 applies.
For , the same inequality gives a rigid set of size at least four. If , its complement has
at most ten vertices and minimum degree four; the tail argument covers it by three sets.
The excluded four-regular tail is globally simple after this single removal, so the same
four-face contradiction applies.
Suppose . If all residual rigid sets have size at most three, (6) makes the residual graph
four-regular, so all its vertices are simple in . All excess lies in , and
The excess theorem gives Shephard’s property; a pyramid is again impossible.
Otherwise remove a largest residual rigid set of size at least four. If , the remaining at
most six vertices need two further sets. If , the remaining at most seven vertices need two
sets unless they form . We obtain either four classes, handled by Proposition 12, or the
second boundary cover of Lemma 13.
Proof of Theorem 1. In dimension three, Lemma 14 supplies at most three rigid
pieces, and Proposition 11 applies. The preceding reductions settle every larger
dimension. Thus below vertices, any decomposable realization forces decomposability
of the whole combinatorial type. Wang and Yost’s construction [5] gives a
conditionally decomposable type with exactly vertices in every dimension . □
References
[1]
M. L. Balinski, On the graph structure of convex polyhedra in -space, Pacific J. Math. 11 (1961),
431–434.
[2]
D. Barnette, A proof of the lower bound conjecture for convex polytopes, Pacific J. Math. 46
(1973), 349–354.
[3]
M. Kallay, Indecomposable polytopes, Israel J. Math. 41 (1982), 235–243.
[4]
G. Pineda-Villavicencio, J. Ugon and D. Yost, The excess degree of a polytope, SIAM J. Discrete
Math. 32 (2018), 2011–2046.
[5]
J. Wang and D. Yost, Lower bound results for conditionally decomposable polytopes, Adv. Geom.
(2026), DOI 10.1515/advgeom-2026-0006.