In the lower-star filtration (LSF), simplices are assigned filtration values according to the maximum-vertex rule, with each simplex entering the filtration when its highest-valued vertex appears. A limitation of this construction is that some higher-dimensional simplices may enter simultaneously with their boundary faces, causing certain topological features to be filled immediately upon formation and thus not appear in the resulting persistence diagram. To address this limitation, we propose three alternative filtration methods: the ε-lower-star filtration (ELSF), the Lp filtration (LpF), and the softmax filtration (SMF). {ELSF introduces a dimension-dependent delay to the maximum-vertex rule,} allowing higher-dimensional simplices to enter the filtration after their boundary faces, whereas {LpF and SMF use aggregation functions that depend on all vertex values rather} than solely on their maximum. By modifying how filtration values are assigned, these methods allow simplices to enter the filtration more progressively, so that topological features such as loops and voids may persist over longer filtration intervals rather than being filled immediately. Consequently, the resulting persistence diagrams may include persistence features that are not observed under LSF, potentially providing useful information for distinguishing data populations. Experimental results on simulated and real-world datasets demonstrate that the proposed filtration methods are computationally efficient and can provide greater sensitivity to structural differences in the data than LSF.
Figures & tables
Figure 1: Geometric representations of simplices. From left to right: a 0 -simplex (point), a 1 -simplex (edge), a 2 -simplex (filled triangle), and a 3 -simplex (tetrahedron).
Figure 2: A toy graph G with five vertices and seven edges, together with three filled triangles (v2,v5,v3) , (v1,v2,v4) , and (v2,v4,v3) , forming a two-dimensional simplicial complex with five 0 -simplices, seven 1 -simplices, and three 2 -simplices.
Figure 3: Geometric evolution of the lower-star filtration across increasing filtration thresholds. Simplices are added according to the maximum function value among their vertices.
Figure 4: Geometric evolution of the ε -lower-star filtration across increasing filtration thresholds.
Figure 5: Geometric evolution of the L1 -filtration across increasing filtration thresholds. Simplices are added according to the sum of the function values of all vertices within each simplex.
Figure 6: Geometric evolution of the softmax filtration across increasing filtration thresholds. Simplices are added using a softmax aggregation of the vertex function values.
Figure 7: Persistence diagrams corresponding to the lower-star filtration (LSF), ε -lower-star filtration (ELSF), Lp -filtration (LpF), and softmax filtration (SMF).
2N
LSF
ELSF (ε=0.1)
LpF (p=1)
LpF (p=2)
LpF (p=10)
SMF (β=1)
SMF (β=2)
100
46.25
45.89
24.81
24.77
25.06
25.09
26.46
200
88.24
87.40
48.47
39.56
39.39
40.02
41.64
300
108.35
106.84
58.35
58.35
59.98
61.14
63.47
400
139.65
137.19
78.23
78.52
80.96
81.54
85.13
500
181.02
178.33
98.64
98.98
101.96
103.64
109.66
600
247.23
210.43
121.34
122.44
126.15
128.96
135.77
Table 1: Median runtime (seconds) for persistence diagram computation across varying dataset sizes (total number of graphs 2N ). The fastest method per row is indicated in blue, and the slowest in red.
Figure 8: Empirical power of the modified energy-based permutation test in homological dimension H1 under simulated Dirichlet RDPG graph distributions for varying separation parameter r . Results are shown for LSF, ELSF, LpF, and SMF using degree and betweenness centrality as filter functions, together with their concatenated vector representations based on GPL, PL, and BF.
Figure 9: Empirical power of the modified energy-based permutation test for different filtration parameter choices using degree centrality as the filter function. Results are shown for BF and PL summaries across varying separation parameter r . The columns correspond to the proposed filtration methods, and the black curve represents LSF.
Figure 10: Empirical power of the modified energy-based permutation test for different filtration parameter choices using betweenness centrality as the filter function. Results are shown for BF and PL summaries across varying values for separation parameter r . The columns correspond to the proposed filtration methods, and the black curve represents LSF.
Dataset
LSF
ELSF
LpF
SMF
B0
D0
C0
B0
D0
C0
B0
D0
C0
B0
D0
C0
GPL- H0
DHFR
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
PROTEINS
✓
✓
✓
✓
✓
✓
✗
✓
✓
✓
✓
✓
IMDB-BINARY
✓
✓
✓
✗
✓
✓
✓
✓
✓
✓
✓
✓
REDDIT-BINARY
✗
✗
✗
✗
✓
✓
✗
✗
✗
✓
✓
✓
Table 2: Summary of hypothesis testing results based on the energy statistic for the real-world graph datasets in homological dimension H0 . Results are reported for the three topological summary functions (GPL, PL, and BF) across the four filtration methods (LSF, ELSF, LpF, and SMF). Filtrations are induced using betweenness and degree centrality as filter functions. The columns B0 and D0 denote the H0 topological summaries obtained using betweenness and degree centrality, respectively, while C0 denotes their concatenated representation. Test outcomes are denoted as follows: ✓ indicates rejection of the null hypothesis ( p<0.05 ), * indicates borderline evidence ( 0.05≤p<0.10 ), and ✗ indicates failure to reject the null hypothesis ( p≥0.10 ).
Figure 11: Representative chest X-ray images from the TBX11K dataset. From left to right: healthy (normal), sick but non-TB, and active TB cases.
Figure 12: Illustration of the image-to-simplicial-complex construction using a 5×5 image grid. The grayscale image patch is represented as a pixel grid, where larger intensity values appear on the left and smaller values appear on the right. Each pixel is treated as a vertex in a lattice, neighboring pixels are connected, and each square cell is subdivided into two triangles. This construction produces a triangulated simplicial complex suitable for filtrations induced by pixel intensities and subsequent persistent homology analysis.
Figure 13: Estimated power of the modified energy-based permutation test for chest X-ray image data as sample size increases using images with resolution 256×256 pixels at significance level α=0.05 . Results are shown for four filtration methods (LSF, ELSF, LpF, and SMF) and three topological summaries (GPL, PL, and BF) across homological dimensions H0 (top row) and H1 (bottom row).
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Parameter
H0
H1
dB
Sup Norm
Bound
dB
Sup Norm
Bound
ELSF
ε=0
1.828
2.791
2.791
2.727
2.791
2.791
ε=0.5
1.828
2.791
2.791
2.727
2.791
2.791
ε=1
1.828
2.791
2.791
2.727
2.791
2.791
ε=2
1.828
2.791
2.791
2.727
2.791
2.791
ε=5
2.281
2.791
2.791
2.727
2.791
2.791
Appendix
Table A1: Numerical stability results using degree centrality as the filter function.
Method
Parameter
H0
H1
dB
Sup Norm
Bound
dB
Sup Norm
Bound
ELSF
ε=0
2.526
4.279
4.279
3.041
4.279
4.279
ε=0.5
2.526
4.279
4.279
3.041
4.279
4.279
ε=1
2.526
4.279
4.279
3.041
4.279
4.279
ε=2
2.526
4.279
4.279
3.041
4.279
4.279
ε=5
2.526
4.279
4.279
3.041
4.279
4.279
Appendix
Table A2: Numerical stability results using betweenness centrality as the filter function.
Figure B1: Empirical power of the modified energy-based permutation test in homological dimension H0 under simulated Dirichlet RDPG graph distributions for varying separation parameter r . Results are shown for LSF, ELSF, LpF, and SMF using degree and betweenness centrality as filter functions, together with their concatenated vector representations based on GPL, PL, and BF.
Figure B2: Empirical power of the modified energy-based permutation test in homological dimension H2 under simulated Dirichlet RDPG graph distributions for varying separation parameter r . Results are shown for LSF, ELSF, LpF, and SMF using degree and betweenness centrality as filter functions, together with their concatenated vector representations based on GPL, PL, and BF.
Dataset
No. of Graphs
No. of Classes
Avg. Nodes
Avg. Edges
DHFR
756
2
42.43
44.54
PROTEINS
1113
2
39.31
72.82
IMDB-BINARY
1000
2
19.77
96.53
REDDIT-BINARY
2000
2
429.60
497.80
DD
1178
2
284.32
715.66
COX2
467
2
41.22
43.45
Appendix
Table C3: Summary of the real-world graph datasets used in Section 5.3 , including the number of graphs, number of classes, and average numbers of nodes and edges.
Dataset
LSF
ELSF
LpF
SMF
B1
D1
C1
B1
D1
C1
B1
D1
C1
B1
D1
C1
GPL- H1
DHFR
✓
✗
✓
✓
✗
✓
✓
✓
✓
✓
✗
✓
PROTEINS
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
IMDB-BINARY
✗
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
REDDIT-BINARY
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
Appendix
Table D4: Summary of hypothesis testing results based on the energy statistic for the real-world graph datasets in homological dimension H1 . Results are reported for the three topological summary functions (GPL, PL, and BF) across the four filtration methods (LSF, ELSF, LpF, and SMF). Filtrations are induced using betweenness and degree centrality as filter functions. The columns B1 and D1 denote the H1 topological summaries obtained using betweenness and degree centrality, respectively, while C1 denotes their concatenated representation. Test outcomes are denoted as follows: ✓ indicates rejection of the null hypothesis ( p<0.05 ), * indicates borderline evidence ( 0.05≤p<0.10 ), and ✗ indicates failure to reject the null hypothesis ( p≥0.10 ).
Figure E3: Estimated power of the energy-based permutation test for chest X-ray image data as sample size increases using images with resolution 64×64 pixels at significance level α=0.05 . Results are shown for four filtration methods (LSF, ELSF, LpF, and SMF) and three topological summaries (GPL, PL, and BF) across homological dimensions H0 (top row) and H1 (bottom row).
Figure E4: Estimated power of the energy-based permutation test for chest X-ray image data as sample size increases using images with resolution 128×128 pixels at significance level α=0.05 . Results are shown for four filtration methods (LSF, ELSF, LpF, and SMF) and three topological summaries (GPL, PL, and BF) across homological dimensions H0 (top row) and H1 (bottom row).