Iterated Nerve Complexes for k-Fold Sensor Coverage

1. Introduction

A sensor network is, for coverage purposes, a finite collection of regions $D_1,\dots ,D_n\subset {\mathbb{R}}^2$, each the detection footprint of one node, and the basic question is whether their union covers some domain of interest. When node positions are unknown or unreliable, a remarkable fact makes this question tractable from connectivity data alone. Ghrist and Muhammad, and de Silva and Ghrist in subsequent work, showed that the topology of the union of the $D_i$ can be tested from the nerve of the cover, a purely combinatorial object recording which subsets of nodes have overlapping footprints, with no metric information beyond an overlap test [1,2,3]. The underlying tool is a theorem of Borsuk: the union of a cover by convex (or, more generally, contractible-intersection) sets is homotopy equivalent to the nerve of the cover [5]. A nontrivial first homology class in the nerve signals a hole in the coverage that no individual overlap test would reveal (see [8] for standard background on simplicial homology).

Ordinary coverage, meaning every point seen by at least one sensor, is often not what a deployment actually needs. Reliability against node failure, handover in cellular networks, and multi-static sensing all require k-fold coverage: every point of the domain seen by at least k distinct sensors, so that the loss of any $k-1$ nodes still leaves the domain covered [10]. The existing homological treatment of this problem, due to Vergne, Decreusefond, and Martins, works by an iterative reduction: k-coverage is interpreted as k disjoint layers of ordinary 1-coverage, found one at a time by a simplicial reduction algorithm [11,12]. This is effective, but it uses an auxiliary partitioning problem at each layer, and the partition is not unique.

We give a different, direct construction. From the ordinary nerve complex N of the cover, the same object already needed for 1-coverage checking, we build a derived complex ${\Delta }_k\left(N\right)$, the k-shadow complex, whose vertices are the $(k-1)$-dimensional faces of N and whose higher faces are read off directly from N’s own combinatorics, with no new geometric computation and no search for disjoint layers. We prove that the region of ${\mathbb{R}}^2$ covered at least k times is homotopy equivalent to ${\Delta }_k\left(N\right)$ (Theorem 3), by a direct, two-line application of Borsuk’s nerve lemma to a secondary cover built from N’s own faces. For $k=1$, ${\Delta }_1\left(N\right)=N$, recovering the de Silva–Ghrist criterion exactly as a special case.

We check the construction computationally for $k=1$ and $k=2$ against an independently computed ground truth, and report a real implementation bug: a discretized intersection test that produced false negatives on thin, marginal overlaps, found through this cross-check and corrected with an exact closed-form test.1 We then extend the same check to four additional sensor fields with qualitatively different geometry, asking whether the original agreement on a single field reflects a property of the construction or an artifact of that one field.

The remainder of the paper is organized as follows. Section 2 fixes notation for the nerve complex and states the two classical theorems the construction rests on. Section 3 defines ${\Delta }_k\left(N\right)$. Section 4 proves the main equivalence. Section 5 draws the homological corollary used for checking coverage. Section 6 reports the computational experiment, beginning with a single field and its implementation history (Section 6.1, Section 6.2 and Section 6.3) before extending to four further fields (Section 6.4, Section 6.5, Section 6.6 and Section 6.7). Section 7 positions the construction against the existing k-coverage literature. Section 8 records the construction’s limitations honestly, including the scaling limit diagnosed in Section 6.7. Section 9 concludes.

2. Preliminaries

Throughout, $D_1,\dots ,D_n\subset {\mathbb{R}}^2$ are closed convex regions (in particular, closed disks of a common radius, though nothing in this section uses equal radii).

Definition 1 

(Nerve complex). The nerve N of the cover ${\left\{D_i\right\}}_{i=1}^n$ is the abstract simplicial complex on vertex set $\{1,\dots ,n\}$ whose faces are exactly the subsets $\sigma \subseteq \{1,\dots ,n\}$ with ${\bigcap }_{i\in \sigma }{D}_i\ne ⌀$.

Theorem 1 

(Nerve Lemma; Borsuk 1948 [5]). If every nonempty finite intersection ${\bigcap }_{i\in \sigma }{D}_i$ is contractible, then the union ${\bigcup }_i{D}_i$ is homotopy equivalent to the geometric realization of the nerve N.

For convex sets in ${\mathbb{R}}^2$, every nonempty finite intersection is itself convex, hence contractible, so the Nerve Lemma always applies to a cover by convex regions; this is the classical Čech theorem used throughout the sensor-coverage literature [2,9].2

The second classical fact we use is computational rather than topological. It lets the entire nerve be recovered from its 2-skeleton (vertices, edges, and triangles) alone, without testing higher-order intersections directly.

Theorem 2 

(Helly 1923 [6]). Let $X_1,\dots ,X_m$ be convex sets in ${\mathbb{R}}^2$. If every 3 of them have a common point, then all m of them have a common point.

3

Corollary 1 

(Two-skeleton determines the nerve). For a cover by convex sets in ${\mathbb{R}}^2$, a candidate vertex set σ with $\left|\sigma \right|\ge 3$ is a face of N if and only if every 3-element subset of σ is a face of N.

Proof. 

If $\sigma$ is a face, every subset of $\sigma$ is a face, in particular every 3-element subset. If every 3-element subset of $\sigma$ is a face, then every 3 of the convex sets ${\left\{D_i\right\}}_{i\in \sigma }$ share a common point, so by the Helly theorem all of them do, and $\sigma$ is a face. □

We use Corollary 1 in Section 6 only as a computational shortcut: it never changes which sets are faces, it only changes how cheaply that fact can be verified.

3. The k-Shadow Complex

Fix $k\ge 1$. Write $m\left(x\right):=\#\{i:x\in D_i\}$ for the coverage multiplicity at a point $x\in {\bigcup }_i{D}_i$, and $R_{\ge k}:=\{x:m\left(x\right)\ge k\}$.

Definition 2 

(k-shadow complex). Let N be the nerve of ${\left\{D_i\right\}}_{i=1}^n$ and let ${\mathcal{F}}_{k-1}\left(N\right)$ denote the set of $(k-1)$-dimensional faces of N (i.e. the k-element subsets $\sigma \subseteq \{1,\dots ,n\}$ with ${\bigcap }_{i\in \sigma }{D}_i\ne ⌀$). The k-shadow complex ${\Delta }_k\left(N\right)$ is the abstract simplicial complex with vertex set ${\mathcal{F}}_{k-1}\left(N\right)$, in which a finite collection $\{{\sigma }_1,\dots ,{\sigma }_m\}\subseteq {\mathcal{F}}_{k-1}\left(N\right)$ spans a face if and only if ${\sigma }_1\cup \cdots \cup {\sigma }_m$ is itself a face of N.

4

Remark 1 

($k=1$ is the identity). For $k=1$, ${\mathcal{F}}_0\left(N\right)$ is the vertex set $\{1,\dots ,n\}$ of N itself, and a finite set of vertices $\{v_1,\dots ,v_m\}$ spans a face of ${\Delta }_1\left(N\right)$ exactly when $\{v_1,\dots ,v_m\}$ is a face of N. Hence ${\Delta }_1\left(N\right)=N$. The construction below therefore generalizes, rather than replaces, the ordinary nerve.

4. Main Theorem

Theorem 3 

(k-fold coverage is the k-shadow complex). Let ${\left\{D_i\right\}}_{i=1}^n$ be closed convex regions in ${\mathbb{R}}^2$ with nerve N, and fix $k\ge 1$. Then

$$R_{\ge k}\simeq {\Delta }_k\left(N\right),$$

where ≃ denotes homotopy equivalence.

Proof. 

For $\sigma \in {\mathcal{F}}_{k-1}\left(N\right)$ write $R_{\sigma }:={\bigcap }_{i\in \sigma }{D}_i$, a nonempty convex (hence contractible) set by definition of ${\mathcal{F}}_{k-1}\left(N\right)$.

Step 1: ${\left\{R_{\sigma }\right\}}_{\sigma \in {\mathcal{F}}_{k-1}\left(N\right)}$ covers exactly $R_{\ge k}$. If $x\in R_{\sigma }$ for some $\sigma \in {\mathcal{F}}_{k-1}\left(N\right)$, then x lies in the k sets ${\left\{D_i\right\}}_{i\in \sigma }$, so $m\left(x\right)\ge k$ and $x\in R_{\ge k}$. Conversely, if $m\left(x\right)\ge k$, choose any k indices $i_1,\dots ,i_k$ with $x\in D_{i_j}$ for each j; then $\sigma =\{i_1,\dots ,i_k\}$ satisfies $x\in {\bigcap }_{i\in \sigma }{D}_i$, so $\sigma \in {\mathcal{F}}_{k-1}\left(N\right)$ and $x\in R_{\sigma }$. Hence ${\bigcup }_{\sigma }{R}_{\sigma }=R_{\ge k}$ exactly, with no points missed or added on either side.

Step 2: the cover $\left\{R_{\sigma }\right\}$ satisfies the Nerve Lemma hypotheses, and its nerve is ${\Delta }_k\left(N\right)$. For any finite $T\subseteq {\mathcal{F}}_{k-1}\left(N\right)$,

$$\bigcap _{\sigma \in T}{R}_{\sigma }=\bigcap _{\sigma \in T}\bigcap _{i\in \sigma }{D}_i=\bigcap _{i\in {\bigcup }_{\sigma \in T}\sigma }{D}_i,$$

which is convex (an intersection of convex sets), hence contractible whenever it is nonempty. It is nonempty exactly when ${\bigcup }_{\sigma \in T}\sigma$ is a face of N, which, by Definition 2, is exactly the condition for T to span a face of ${\Delta }_k\left(N\right)$. So every nonempty finite intersection of the $R_{\sigma }$ is contractible, and the simplicial complex recording which finite subsets of $\left\{R_{\sigma }\right\}$ have nonempty intersection is, by construction, ${\Delta }_k\left(N\right)$ itself.

Step 3: apply the Nerve Lemma. The cover ${\left\{R_{\sigma }\right\}}_{\sigma \in {\mathcal{F}}_{k-1}\left(N\right)}$ of $R_{\ge k}$ (Step 1) satisfies the contractible-intersection hypothesis of the Nerve Lemma (Step 2), so $R_{\ge k}={\bigcup }_{\sigma }{R}_{\sigma }$ is homotopy equivalent to the nerve of this cover, which is ${\Delta }_k\left(N\right)$. □

By Remark 1, taking $k=1$ in Theorem 3 recovers exactly the classical fact that ${\bigcup }_i{D}_i\simeq N$, the case already used by the de Silva–Ghrist coverage criterion. The theorem places no further hypothesis on k beyond convexity of the original regions; ${\Delta }_k\left(N\right)$ is well defined and the equivalence holds for every $k\le {max}_i\left\{\mathrm{size}\mathrm{of}\mathrm{a}\mathrm{maximal}\mathrm{clique}\mathrm{containing}i\right\}$, beyond which ${\mathcal{F}}_{k-1}\left(N\right)$ is empty and both sides are empty.

5. A Homological Test for k-Fold Coverage

Corollary 2 

(Betti-number test). $b_0\left({\Delta }_k\left(N\right)\right)$ equals the number of connected components of $R_{\ge k}$, and $b_1\left({\Delta }_k\left(N\right)\right)$ equals the number of bounded complementary regions enclosed by $R_{\ge k}$: the number of independent regions where the coverage multiplicity drops below k surrounded by area where it does not.

Proof. 

Homotopy equivalence preserves all homology groups, in particular $H_0$ and $H_1$, whose ranks count connected components and independent 1-cycles respectively; for a planar region these are exactly the stated topological features. □

For $k=1$, Corollary 2 is exactly the de Silva–Ghrist criterion: $b_1\left(N\right)=0$ shows that the network has no coverage gaps. For $k\ge 2$, the same statement, computed on ${\Delta }_k\left(N\right)$ instead of N, checks k-fold redundancy, meaning fault tolerance against the simultaneous failure of up to $k-1$ sensors, using no information beyond what was already computed to check ordinary coverage.

Figure 1 gives a hand-checkable instance of Theorem 3 at $k=2$: four disks in a necklace, overlapping pairwise but never triple-wise, give a 2-fold-covered region consisting of four disjoint lens-shaped patches, and ${\Delta }_2\left(N\right)$ – four isolated points, since no triple overlap exists to connect them – has exactly $b_0=4$, $b_1=0$, in exact agreement.

6. Computational Check

Figure 1 is small enough to check by hand. To test the construction on a case too large for manual checking, we built a synthetic network of 91 sensors of equal radius $R=1.55$: a double ring of 46 boundary sensors (radii $5.9$ and $7.4$ from a common centre, dense enough that the boundary band is solidly multiply covered) plus a $7\times 7$ interior grid at spacing $1.4$ with small random jitter, from which a cluster of 4 mutually adjacent sensors near the centre was deliberately removed to create a region of reduced coverage. We refer to this as Field A throughout the paper.5 Section 6.1, Section 6.2 and Section 6.3 establish the corrected pipeline and its agreement with ground truth on Field A. Section 6.4, Section 6.5, Section 6.6 and Section 6.7 extend the test to four further, independently generated fields.

6.1. An Implementation Bug, Found and Corrected

Our first implementation tested whether a candidate set of disks shares a point by sampling a local grid and checking whether every disk in the set covers some sampled point. This is the natural first approach, and it is wrong in exactly the cases this experiment is designed to probe: thin, marginal overlaps near the threshold distance $|p_i-p_j|=2R$ can be missed by a grid of any fixed finite resolution, and because higher faces of N were built from already-found lower faces (Corollary 1), a single missed marginal edge silently removed an entire branch of faces downstream.

The bug was caught, not assumed absent, by computing ${\Delta }_2\left(N\right)$ two ways: once from the cached, discretized intersection data, and once by re-testing every candidate union directly. The two disagreed in 240 places.6 We replaced the discretized test with an exact one available for equal-radius disks: k disks of common radius R share a point if and only if the minimum enclosing circle of their centres has radius at most R, and for a point set of size at most 3 the minimum enclosing circle is determined in closed form by at most 3 of the points (a pair, as a diameter, or a triple, as a circumcircle), so the test is exact and requires no discretization at any set size relevant here. After this correction, the two independent computations of ${\Delta }_2\left(N\right)$ agreed exactly, with zero disagreements on direct re-test.

6.2. Results on Field A

With the exact intersection test, the nerve N of Field A has 91 vertices, 493 edges, and 922 triangles, extending (via Corollary 1) to faces of dimension up to 6.

For $k=1$, ${\Delta }_1\left(N\right)=N$ gives $b_0=1$, $b_1=1$: one connected network, one coverage gap, located at the centre of the removed cluster. Ground truth was computed independently by densely sampling the plane, marking points with $m\left(x\right)\ge 1$, and counting connected components of the foreground and of the background components not touching the sampling window’s border. It gives $b_0=1$, $b_1=1$ at the same location: an exact match.

For $k=2$, ${\Delta }_2\left(N\right)$ has 493 vertices, 5124 edges, and 23848 triangles (computed in full, not approximated), giving $b_0=1$, $b_1=8$: one connected 2-fold-covered region with 8 enclosed regions of reduced redundancy. One is the large deliberately engineered gap (area $\approx 3.75$ in the units of the construction), and the other seven are smaller and unintended, arising where the double-ring fence falls just short of full double coverage at certain angular positions. Ground truth, computed the same way as for $k=1$ but with $m\left(x\right)\ge 2$, gives $b_0=1$, $b_1=8$ at the same eight locations: again an exact match.

Figure 2. Field A, $k=2$. Green marks the region covered at least twice; red marks the region covered exactly once. Blue dots are the 91 sensor positions: 46 on the double-ring boundary fence and 45 on the jittered interior grid, after the deliberate removal of four mutually adjacent grid sensors near the centre. The eight orange circles mark the eight enclosed background components that ${\Delta }_2\left(N\right)$ and the independent raster ground truth both identify, in exact agreement on both count and location: one large component at the centre, with area approximately $3.75$ in the units of the construction, corresponding directly to the four removed sensors, and seven small components scattered around the perimeter, each arising where the double-ring fence falls just short of full double coverage at a particular angular position rather than from any deliberate design choice. The seven small holes were not anticipated when the field was constructed; their presence and exact location were discovered only once the computation was run, which is part of why the agreement between ${\Delta }_2\left(N\right)$ and the independent ground truth on all eight locations, not only the one engineered hole, is informative rather than a foregone conclusion.

Figure 2. Field A, $k=2$. Green marks the region covered at least twice; red marks the region covered exactly once. Blue dots are the 91 sensor positions: 46 on the double-ring boundary fence and 45 on the jittered interior grid, after the deliberate removal of four mutually adjacent grid sensors near the centre. The eight orange circles mark the eight enclosed background components that ${\Delta }_2\left(N\right)$ and the independent raster ground truth both identify, in exact agreement on both count and location: one large component at the centre, with area approximately $3.75$ in the units of the construction, corresponding directly to the four removed sensors, and seven small components scattered around the perimeter, each arising where the double-ring fence falls just short of full double coverage at a particular angular position rather than from any deliberate design choice. The seven small holes were not anticipated when the field was constructed; their presence and exact location were discovered only once the computation was run, which is part of why the agreement between ${\Delta }_2\left(N\right)$ and the independent ground truth on all eight locations, not only the one engineered hole, is informative rather than a foregone conclusion.

We attempted $k=3$ on Field A; ${\Delta }_3\left(N\right)$ has 922 vertices (the triangles of N), and enumerating its higher faces by brute-force pairwise and triple-wise testing did not complete in practical time with the unoptimized implementation used here.7 We report this honestly in Section 8 rather than omit it. The direct geometric sampler, which does not require building ${\Delta }_3\left(N\right)$ at all, gives $b_0=5$, $b_1=60$ for $k=3$ on Field A; this value is reported for completeness but is not checked against ${\Delta }_3\left(N\right)$ for the reason just stated.

6.3. Why a Single Field Is Not Enough

The agreement reported in Section 6.2 is exact, but it is a single data point: one synthetic network, one realisation of randomness (fixed seed 3), one choice of radius, and one hand-engineered fence geometry designed specifically to produce a clean, well-separated gap structure. A construction that is correct in general should continue to agree with independent ground truth on fields it was not designed around. The remainder of this section tests exactly that, on four additional fields generated independently of Field A and of each other.

6.4. Multi-Topology Validation

We generated four further sensor fields, summarised in Table 1, chosen to vary the qualitative geometry of the deployment rather than merely re-running Field A with new noise.8

• Field B: the identical ring-and-grid construction as Field A, but with an independent random seed (seed 7 in place of seed 3), governing only the interior grid jitter and the choice of four removed sensors. This tests sensitivity to randomness alone, holding the qualitative topology fixed.
• Field C: 95 sensors placed by an unstructured uniform random distribution over a square domain, with no engineered fence and no minimum inter-sensor spacing enforced. This tests the construction under a deployment geometry with no designed structure at all.
• Field D: a jittered hexagonal lattice of 79 sensors with a deliberate rectangular dead zone carved out by removing all sensors whose centre falls inside a fixed rectangle. This tests a regular, non-circular topology with an engineered gap of a different shape than Field A’s circular fence.
• Field E: 95 sensors split between a dense Gaussian cluster (55 sensors, standard deviation $0.9$) and a sparse uniform halo (40 sensors). This tests strong local density heterogeneity within a single field.

For each field, we ran the identical corrected pipeline used for Field A in Section 6.2: build the nerve N with the exact minimum-enclosing-circle intersection test, compute ${\Delta }_1\left(N\right)=N$ and ${\Delta }_2\left(N\right)$, and compare against an independently computed raster ground truth. Table 2 reports the result for each field that completed.

Fields B and D reproduce the exact agreement of Field A on two genuinely different topologies: a different random realisation of the same engineered geometry, and a structurally different regular lattice with a differently shaped engineered gap. This is evidence that the exact agreement on Field A was not an artifact specific to its particular geometry.

Figure 3. Nerve size (vertices, edges, triangles, log scale) for the four fields whose nerve construction completed within the time available. Fields A and B, the two ring-and-grid fields, have nearly identical nerve size despite the different random seed: 493 versus 496 edges and 922 versus 928 triangles, a difference small enough to be attributable entirely to which specific sensors happened to fall close to the overlap threshold under each seed’s jitter. Field C’s nerve is visibly denser at every level: 619 edges and 1677 triangles against Field A’s 493 and 922, despite having only four more sensors in total (95 against 91). This is the structural root of both the near-tangent-contact issue documented in Section 6.5 and the larger ${\Delta }_2\left(N\right)$ computation reported in Table 2: a denser nerve has more edges, and the cost of building ${\Delta }_2\left(N\right)$ grows with the square of the edge count. Field D’s regular hexagonal geometry produces the sparsest nerve of the four relative to its vertex count, 197 edges and only 118 triangles on 79 vertices, consistent with a lattice in which each sensor has a small, fixed number of close neighbours rather than the more variable neighbour count of a jittered or random field.

Figure 3. Nerve size (vertices, edges, triangles, log scale) for the four fields whose nerve construction completed within the time available. Fields A and B, the two ring-and-grid fields, have nearly identical nerve size despite the different random seed: 493 versus 496 edges and 922 versus 928 triangles, a difference small enough to be attributable entirely to which specific sensors happened to fall close to the overlap threshold under each seed’s jitter. Field C’s nerve is visibly denser at every level: 619 edges and 1677 triangles against Field A’s 493 and 922, despite having only four more sensors in total (95 against 91). This is the structural root of both the near-tangent-contact issue documented in Section 6.5 and the larger ${\Delta }_2\left(N\right)$ computation reported in Table 2: a denser nerve has more edges, and the cost of building ${\Delta }_2\left(N\right)$ grows with the square of the edge count. Field D’s regular hexagonal geometry produces the sparsest nerve of the four relative to its vertex count, 197 edges and only 118 triangles on 79 vertices, consistent with a lattice in which each sensor has a small, fixed number of close neighbours rather than the more variable neighbour count of a jittered or random field.

Field C disagrees with ground truth at both $k=1$ and $k=2$, and we treat this disagreement as a result in its own right rather than a failure to be corrected away, for the reasons given in Section 6.5. Field D, by contrast, gives the largest $b_1$ of any field tested (40, against Field A and B’s 8), and Figure 5 shows why: the regular hexagonal lattice leaves a small, periodic single-coverage gap between every group of six neighbouring sensors, in addition to the one large engineered rectangular dead zone at the centre. ${\Delta }_2\left(N\right)$ and the independent ground truth again agree exactly on all 40.

Figure 4. Field B, $k=2$, generated and visualised with the same pipeline and the same visual convention as Figure 2. The gap structure near the centre differs visibly from Field A’s (compare the white wedge-shaped region at centre here against the near-circular gap in Figure 2), reflecting the independent random jitter of seed 7 against seed 3, while the topological invariants are identical: $b_0=1$, $b_1=8$ for both ${\Delta }_2\left(N\right)$ and the independent ground truth, with all eight enclosed components again in exact correspondence.

Figure 4. Field B, $k=2$, generated and visualised with the same pipeline and the same visual convention as Figure 2. The gap structure near the centre differs visibly from Field A’s (compare the white wedge-shaped region at centre here against the near-circular gap in Figure 2), reflecting the independent random jitter of seed 7 against seed 3, while the topological invariants are identical: $b_0=1$, $b_1=8$ for both ${\Delta }_2\left(N\right)$ and the independent ground truth, with all eight enclosed components again in exact correspondence.

Figure 5. Field D, $k=2$. The jittered hexagonal lattice produces 39 small, regularly spaced single-coverage gaps, one inside almost every hexagonal cell of the lattice and visible as the ring of red surrounding each interior sensor, plus one large rectangular dead zone at the centre, where the two sensors of the original $9\times 9$ candidate lattice whose centres fell inside the design rectangle $[5.0,7.5]\times [4.5,6.5]$ were removed, leaving 79 sensors in total. The dead zone is more than an order of magnitude larger in area than any single periodic gap, but contributes only one component to $b_1$, the same as each of the 39 small gaps: Betti number counts components, not their size, which is part of why ${\Delta }_2\left(N\right)$, a purely combinatorial object with no notion of area built in, can recover the count exactly without ever computing an area. All 40 enclosed background components are circled in orange; ${\Delta }_2\left(N\right)$ and the independent raster ground truth agree exactly on the count and the location of every one, the largest $b_1$ of any field tested.

Figure 5. Field D, $k=2$. The jittered hexagonal lattice produces 39 small, regularly spaced single-coverage gaps, one inside almost every hexagonal cell of the lattice and visible as the ring of red surrounding each interior sensor, plus one large rectangular dead zone at the centre, where the two sensors of the original $9\times 9$ candidate lattice whose centres fell inside the design rectangle $[5.0,7.5]\times [4.5,6.5]$ were removed, leaving 79 sensors in total. The dead zone is more than an order of magnitude larger in area than any single periodic gap, but contributes only one component to $b_1$, the same as each of the 39 small gaps: Betti number counts components, not their size, which is part of why ${\Delta }_2\left(N\right)$, a purely combinatorial object with no notion of area built in, can recover the count exactly without ever computing an area. All 40 enclosed background components are circled in orange; ${\Delta }_2\left(N\right)$ and the independent raster ground truth agree exactly on the count and the location of every one, the largest $b_1$ of any field tested.

6.5. Field C: A Disagreement Traced to Near-Tangent Contacts

Unlike Fields A, B, and D, Field C has no engineered minimum spacing between sensors: positions are drawn independently and uniformly over a square domain. With 95 sensors of radius $1.4$ in a $12\times 12$ square, the expected total disk area exceeds the domain area by a factor of approximately 4, so the field is, on average, heavily overcovered; what makes it a useful test is not its coverage density but the absence of any constraint keeping sensor pairs away from the threshold separation $|p_i-p_j|=2R$ at which two disks transition from overlapping to disjoint.

We searched all $\left({\displaystyle \genfrac{}{}{0pt}{}{95}{2}}\right)=4465$ sensor pairs in Field C for separations within $0.05$ length units of exact tangency ($2R=2.8$) and found 20 such pairs, the closest at a gap of $0.0035$ units below threshold. This is expected: with no engineered spacing, a field of this size will by chance place some pairs arbitrarily close to the threshold, and Field C does.

We verified directly that the exact intersection test classifies every one of these 20 near-tangent pairs correctly as overlapping (the correct classification, since each pair’s separation is strictly less than $2R$): each pair appears in the nerve’s edge set, as it must by the exactness of the minimum-enclosing-circle criterion regardless of how close to tangency the pair lies.

The raster ground-truth sampler, in contrast, does not handle these near-tangent pairs reliably. We inspected the ten background components the $k=1$ raster sampler reports as enclosed holes (against ${\Delta }_1\left(N\right)$’s correct value of 3) and found that every one of them is a degenerate sliver of area at most $0.056$ square units, several occupying only 1–2 raster cells, located precisely at the near-tangent contact points identified above. Table 3 records the eight smallest. A resolution-convergence check (Section 6.6) confirms these raster counts are themselves unstable under grid refinement, which is direct independent evidence that the raster sampler, not the exact nerve construction, is the unreliable instrument at these near-degenerate contacts.

This is, in effect, a second and independent instance of the same phenomenon that produced the bug described in Section 6.1: a discretized procedure failing precisely at thin, marginal overlaps, while the exact geometric test does not. In Section 6.1, the unreliable discretized procedure was the nerve construction itself, and the fix was to replace it with the exact test described there. Here, the unreliable discretized procedure is the independent ground-truth sampler used to check the (already exact) nerve construction. The lesson is the same in both cases: any fixed-resolution raster test, whether used to build the complex or merely to check it, is vulnerable at near-degenerate contacts, while the closed-form minimum-enclosing-circle test used throughout the corrected pipeline is not. We regard the $k=1,2$ disagreement on Field C as evidence in the construction’s favour, not against it, once the cause is traced to its source.

Figure 6. Left: Field C’s coverage region $R_{\ge 1}$, with the orange box marking the location of one cluster of near-tangent sensor pairs. Right: a zoomed view of the boxed region, with the full disk boundary of every nearby sensor drawn explicitly. The thin white sliver between two nearly tangent disk boundaries is the source of one of the degenerate raster components listed in Table 3: it is a genuine gap in coverage (the disks really do not quite meet there), but one whose width is far below the raster grid spacing used to sample it, so its detection by the raster sampler depends sensitively on exactly where the grid lines happen to fall relative to the contact point. The exact nerve construction has no such dependence: it tests the relevant disks directly via the closed-form minimum-enclosing-circle criterion, independent of any grid.

Figure 6. Left: Field C’s coverage region $R_{\ge 1}$, with the orange box marking the location of one cluster of near-tangent sensor pairs. Right: a zoomed view of the boxed region, with the full disk boundary of every nearby sensor drawn explicitly. The thin white sliver between two nearly tangent disk boundaries is the source of one of the degenerate raster components listed in Table 3: it is a genuine gap in coverage (the disks really do not quite meet there), but one whose width is far below the raster grid spacing used to sample it, so its detection by the raster sampler depends sensitively on exactly where the grid lines happen to fall relative to the contact point. The exact nerve construction has no such dependence: it tests the relevant disks directly via the closed-form minimum-enclosing-circle criterion, independent of any grid.

6.6. Resolution Stability of the Ground-Truth Sampler

The discrepancy on Field C motivated a check not used elsewhere in this paper: running the independent ground-truth sampler at three different grid resolutions (half, baseline, and double the spacing established in Section 6.2, i.e. $18/350$, $18/700$, and $18/1400$ length units per sample) and recording whether the resulting Betti numbers are stable. Table 4 reports the result for $k=1,2,3$ on every field.

Two findings follow. First, even on Field A, the field used throughout this paper, the $k=2$ and $k=3$ ground-truth Betti numbers are not fully stable under grid refinement, though the value at the resolution used in Section 6.2 happens to coincide with ${\Delta }_2\left(N\right)$ exactly at $k=2$. This was not previously checked or reported, and we report it here for completeness: the agreement at $k=2$ on Field A, while exact at the resolution used, sits near a resolution boundary rather than in a region of comfortable stability. Second, Field C is unstable at every k and every resolution tested, consistent with the near-tangent-contact diagnosis of Section 6.5: a field with contacts arbitrarily close to the threshold distance will not have a raster-stable ground truth at any practical fixed resolution, because the relevant length scale (the gap below threshold, as small as $0.0035$ units in Field C) is smaller than the grid spacing used. Field D, by contrast, is fully stable at every k, consistent with its regular lattice geometry keeping all contacts well clear of the threshold.

Figure 7. Raster ground-truth $b_1$ at half, baseline, and double the grid resolution established in Section 6.2, for Fields A, B, and C and $k=1,2,3$ (Field D is omitted from this comparison because its $b_1$ values, 1, 40, and 0, are stable at every resolution and on a different numerical scale that would compress the more informative A/B/C comparison; see Table 4 for Field D’s values). At $k=1$, Fields A and B are flat at $b_1=1$ across all three resolutions while Field C swings between 5 and 10. At $k=2$ and $k=3$, all three fields show some resolution dependence, but Field C’s range is consistently the widest relative to its baseline value, consistent with the near-tangent-contact diagnosis of Section 6.5.

Figure 7. Raster ground-truth $b_1$ at half, baseline, and double the grid resolution established in Section 6.2, for Fields A, B, and C and $k=1,2,3$ (Field D is omitted from this comparison because its $b_1$ values, 1, 40, and 0, are stable at every resolution and on a different numerical scale that would compress the more informative A/B/C comparison; see Table 4 for Field D’s values). At $k=1$, Fields A and B are flat at $b_1=1$ across all three resolutions while Field C swings between 5 and 10. At $k=2$ and $k=3$, all three fields show some resolution dependence, but Field C’s range is consistently the widest relative to its baseline value, consistent with the near-tangent-contact diagnosis of Section 6.5.

6.7. Field E and the True Size of ${\Delta }_2\left(N\right)$

Field E’s dense Gaussian cluster produces 1810 nerve edges and a maximal clique of size 41: 41 sensors whose pairwise disks all mutually overlap. An initial implementation enumerated, for each maximal clique, every subset up to size 8 and Helly-verified each one (Corollary 1); for a clique of size 41 this is $\left({\displaystyle \genfrac{}{}{0pt}{}{41}{8}}\right)\approx 9.55\times {10}^7$ candidate subsets from that clique alone, and the construction did not complete.

This is fixable, and we fixed it. $b_0$ and $b_1$ of ${\Delta }_2\left(N\right)$ depend only on its 1- and 2-skeleton (Corollary 2), and face membership of any candidate union is determined by the 2-skeleton of N alone, via the Helly shortcut of Corollary 1: there is no need to enumerate N’s higher faces, or its maximal cliques, at all. Replacing the clique-based enumeration with an on-demand, memoized oracle that tests each candidate union directly against N’s cached edges and triangles removes the clique-size obstruction entirely: N’s 2-skeleton for Field E now builds in about a second, independent of its largest clique, and the same oracle reproduces every number reported for Field A in Section 6.2 exactly, while running in under half a second.

Field E’s ${\Delta }_2\left(N\right)$ still does not complete, and the reason is different from, and more fundamental than, the clique-size obstruction that motivated the fix. We counted its meta-edges exactly: $836991$ out of $1637145$ candidate pairs of N-edges, meaning slightly over half of every possible pair of N-edges happens to union into a face of N. ${\Delta }_2\left(N\right)$ therefore has roughly $837,000$ vertices for Field E, against 493 for Field A. This is a real difference in the size of the object itself, not an artifact of how it is computed.9 A candidate search for meta-triangles over a graph with $837,000$ vertices has on the order of ${10}^{11}$ candidate triples in the worst case; even restricting attention only to pairs of meta-edges that already share a meta-vertex, as the implementation does, the candidate count exceeded $2.5\times {10}^8$ and was still climbing when we stopped the computation, well before the rank of the resulting boundary matrix had saturated.

Figure 8. Left: Field E in full, 95 sensors split between a dense Gaussian cluster of 55 sensors at centre (standard deviation $0.9$) and a sparse uniform halo of 40 sensors spread over the wider domain. Red dots mark the 41 sensors forming the largest maximal clique: every pair among these 41 has overlapping disks, meaning 41 is the size of the largest fully mutually overlapping group anywhere in the field, drawn from the 55-sensor dense cluster. Right: the same 41-sensor clique shown alone, with every disk boundary drawn. This clique no longer determines the computational cost of building N, since the on-demand oracle described in this section tests face membership directly against the 2-skeleton rather than by enumerating subsets of cliques; what it illustrates instead is the geometric density responsible for ${\Delta }_2\left(N\right)$’s large vertex count, since the 820 mutually overlapping edges internal to this clique alone are a substantial fraction of Field E’s 1810 total N-edges. No comparable clique appears in Fields A, B, or D, whose largest cliques have 7, 8, and 3 sensors respectively; Field C’s largest clique, at 10 sensors, is closer in kind to these three than to Field E’s 41, which is part of why Field C’s ${\Delta }_2\left(N\right)$, despite its own near-tangent-contact difficulties (Section 6.5), completed in seconds while Field E’s did not.

Figure 8. Left: Field E in full, 95 sensors split between a dense Gaussian cluster of 55 sensors at centre (standard deviation $0.9$) and a sparse uniform halo of 40 sensors spread over the wider domain. Red dots mark the 41 sensors forming the largest maximal clique: every pair among these 41 has overlapping disks, meaning 41 is the size of the largest fully mutually overlapping group anywhere in the field, drawn from the 55-sensor dense cluster. Right: the same 41-sensor clique shown alone, with every disk boundary drawn. This clique no longer determines the computational cost of building N, since the on-demand oracle described in this section tests face membership directly against the 2-skeleton rather than by enumerating subsets of cliques; what it illustrates instead is the geometric density responsible for ${\Delta }_2\left(N\right)$’s large vertex count, since the 820 mutually overlapping edges internal to this clique alone are a substantial fraction of Field E’s 1810 total N-edges. No comparable clique appears in Fields A, B, or D, whose largest cliques have 7, 8, and 3 sensors respectively; Field C’s largest clique, at 10 sensors, is closer in kind to these three than to Field E’s 41, which is part of why Field C’s ${\Delta }_2\left(N\right)$, despite its own near-tangent-contact difficulties (Section 6.5), completed in seconds while Field E’s did not.

The mathematics places no restriction on clique size or nerve density: Theorem 3 holds for ${\Delta }_2\left(N\right)$ on Field E exactly as it does on every other field, and the true value of $(b_0,b_1)$ for Field E’s ${\Delta }_2\left(N\right)$ exists and is well defined, independent of whether any particular implementation can compute it in practical time. What we have established precisely is the size of the object that would need to be searched: $837,000$ vertices, over half of all possible N-edge pairs realised as meta-edges, and a meta-triangle search space on the order of hundreds of millions of candidates even after restricting to adjacent meta-edge pairs. Densely clustered deployments are not an unusual design choice in practice, since co-located redundant sensors are a common reliability strategy, and they are exactly the case in which ${\Delta }_2\left(N\right)$ itself, not merely its computation, becomes large. Section 8.1 discusses what completing this case would require.

7. Related Work

The homological approach to coverage checking originates with Ghrist and Muhammad and with de Silva and Ghrist [1,2,3], who showed that connectivity data alone, with no node localization, checks ordinary coverage via $H_1$ of the nerve (or, in their distributed setting, of a Rips complex bounding it from each side), provided the domain boundary is appropriately fenced [4]. Subsequent work made the computation of the relevant complex distributed and incremental [14,15,16], extended it to spherical domains [17], used persistence to localize holes more precisely [18,19], and addressed simplicial reduction for tractable computation on large networks [12,13]. Cavanna et al. generalize the coverage criterion itself to weighted and non-uniform sampling regimes [20], and more recent work continues to use related simplicial constructions for fault recovery in deployed networks [21]; the broader use of higher-order combinatorial structures across network science is surveyed in [22].

The one existing homological treatment of k-fold coverage specifically, due to Vergne, Decreusefond, and Martins, works by an iterative mechanism: k-coverage is checked by finding k pairwise disjoint subsets of sensors, each of which individually satisfies the ordinary 1-coverage criterion [11]. This requires solving a combinatorial partition problem at each of the k layers, and a given network may admit no such partition into exactly disjoint covering layers even when it is genuinely k-covered by overlapping (not disjoint) subsets. The construction in this paper takes a different route: ${\Delta }_k\left(N\right)$ is read off directly from the faces of N already computed for ordinary coverage, with no partition search of any kind, and Theorem 3 shows the resulting complex captures the k-fold coverage topology exactly, not merely sufficiently.

The general idea that two simplicial complexes built from a single combinatorial relation can be homotopy equivalent has classical roots in Dowker’s duality theorem for binary relations [7]. The k-shadow complex is not an instance of a Dowker complex in the strict sense, since it is obtained by applying the Nerve Lemma a second time to a derived cover, rather than from a single relation between two vertex sets. The underlying spirit, that the same topological information can be read off a relation in more than one combinatorially distinct way, is nonetheless the same one Dowker identified. To the best of our literature search, the direct ${\Delta }_k\left(N\right)$ construction and Theorem 3 have not previously appeared in this form; this is a search-limited statement, not a claim of exhaustive priority.

8. Limitations

The construction inherits the standing assumption of the entire nerve-theoretic approach to coverage: sensing regions must be convex (or at least have contractible finite intersections), which holds for disks of any radii but not for arbitrary, e.g. non-convex or sector-shaped, sensing footprints without further modification.

The computational cost of ${\Delta }_k\left(N\right)$ grows quickly with k: its vertex set is ${\mathcal{F}}_{k-1}\left(N\right)$, already 493 for $k=2$ on the 91-node Field A, and the brute-force pairwise and triple-wise testing used here to build its higher faces did not complete in practical time at $k=3$ on the same network. The proof of Theorem 3 imposes no such restriction, since the equivalence holds for every k, but a practical implementation for larger k would need either a more efficient face-enumeration strategy or a restriction to computing only $b_0$ and $b_1$ directly (for which the full higher-dimensional face structure of ${\Delta }_k\left(N\right)$ is not strictly required), neither of which we have implemented here.

8.1. What the Fix Resolved, and What It Revealed

The multi-topology validation in Section 6.4 exposed a scaling limit not visible from Field A alone: dense local clustering, rather than overall network size, governs the cost of building N. The original obstruction was structural and specific: an implementation that enumerates subsets of each maximal clique up to a fixed dimension cap has cost growing combinatorially in the size of the largest clique present, independent of the number of sensors. We replaced this with an on-demand, memoized oracle that tests face membership of a candidate set directly against N’s 2-skeleton, using the Helly shortcut of Corollary 1 without ever enumerating N’s higher faces or its maximal cliques in advance.10 We verified the oracle reproduces every number reported for Field A in Section 6.2 exactly, including the full 5124 meta-edges and 23848 meta-triangles of ${\Delta }_2\left(N\right)$, while running in well under a second; the same oracle resolved Field E’s nerve construction in about a second, with no dependence on its 41-sensor maximal clique at all.

This isolated the real difficulty precisely. With the clique obstruction removed, ${\Delta }_2\left(N\right)$ for Field E still does not complete, because ${\Delta }_2\left(N\right)$ itself is roughly 1700 times larger than for Field A: $836991$ meta-vertices against Field A’s 493, with the meta-triangle search space scaling accordingly. This is not an implementation limitation in the sense the original diagnosis described; it is a statement about the size of the mathematical object ${\Delta }_2\left(N\right)$ assigns to a densely clustered field, which no algorithmic improvement to the search strategy changes, since the object being searched is itself large.

We looked for a route around this and found one that works in part. A meta-triangle of ${\Delta }_2\left(N\right)$ is three N-edges whose vertex union is a face of N, and this union falls into one of four patterns depending on how many of the three pairs of edges share a sensor: all three pairs share a vertex, exactly two do, exactly one does, or none do. The last case, three pairwise disjoint N-edges together spanning exactly six distinct sensors, can be found efficiently: rather than searching ${\Delta }_2\left(N\right)$’s own meta-graph, it suffices to enumerate the size-six faces of N directly, a search bounded by N’s own clique structure and independent of ${\Delta }_2\left(N\right)$’s size, and then check which triples of N-edges inside each such face form a perfect matching covering it. We verified this against Field A by brute force and found an exact match: 810 disjoint-edge meta-triangles, against 810 found this way. On Field A, however, this case accounts for only 810 of the $23848$ total meta-triangles, a little over three per cent; the remaining patterns, where at least one pair of edges shares a vertex, make up the bulk of the count and resisted every shortcut we tried, including one early attempt that appeared to work, in the sense of finding far fewer meta-triangles than the true total yet still recovering the correct rank on Field A, until we checked it directly against the brute-force count and found it undercounted by more than half. That attempt is not reported as a method here, because a technique that gives the right answer on one field by chance is not a technique.

A further reduction would need either a method that computes $b_0$ and $b_1$ without ever forming the boundary matrix of ${\Delta }_2\left(N\right)$ explicitly, for instance by working with the original disks directly rather than through the nerve, since Theorem 3 only requires homotopy equivalence to hold and not that this particular route to computing it be the cheapest available, or a complete case-by-case enumeration strategy for the three vertex-sharing patterns left unresolved here. Neither is attempted in this paper. We measured the achievable search rate for the full brute-force meta-triangle scan directly on Field E: approximately three hundred candidates per second once true set-intersection costs are accounted for, against an estimated total candidate count in the range of several hundred million to several billion, putting full completion on the order of days on the hardware available for this work, not hours.

The computational check in this paper covers five synthetic networks and $k\in \{1,2\}$, with $k=3$ ground truth (but not ${\Delta }_3\left(N\right)$) reported additionally on Field A. Validation against irregular sensing radii and real, non-synthetic deployment data remains future work.

9. Concluding Remarks

We have given a direct construction, the k-shadow complex ${\Delta }_k\left(N\right)$, that reduces k-fold coverage checking to a purely combinatorial operation on the nerve complex already required for ordinary coverage checking, with no new geometric data and no search for disjoint covering layers. The construction is exact, not heuristic: Theorem 3 proves a genuine homotopy equivalence, not merely a sufficient condition, via two applications of a classical 1948 theorem to an explicitly constructed secondary cover. The case $k=1$ recovers the established de Silva–Ghrist criterion exactly.

The computational check covers five independently generated sensor fields rather than one. Two reproduce exact agreement between ${\Delta }_k\left(N\right)$ and direct geometric sampling on genuinely different topologies. A third exposed a disagreement traced precisely to near-tangent sensor contacts and to instability in the raster ground-truth sampler at those contacts, not to any failure of the exact nerve construction itself, which we verified classifies every near-tangent pair correctly. A fourth exposed a real computational obstruction in an initial implementation, traced to a single maximal clique of size 41; correcting that obstruction did not complete the field, but instead revealed that ${\Delta }_2\left(N\right)$ itself has over $800,000$ vertices for this field, an intrinsic property of the object rather than of how it was computed. Taken together, these results go beyond a single confirming example: the construction’s exactness held up under deliberate testing against topologies it was not designed around, and where a check failed or did not complete, the failure was traced to its precise source rather than left unexplained. This is the same standard applied to the discretized intersection test corrected in Section 6.1.