Periodic Point Sets under Isometry

• Periodic Point Sets are defined as finite motifs repeated over a lattice in Euclidean space, with isometry invariance ensuring identical properties under rigid motions.
• Continuous invariants such as density functions, AMD, and PDD provide robust metrics that enable both rapid search and precise discrimination of periodic structures.
• These isometry-based descriptors are pivotal in crystallography and materials science, overcoming instability in unit-cell classifications and aiding large-scale database analysis.

Searching arXiv for the cited papers to ground the article in current records. Periodic point sets are subsets of Euclidean space obtained by repeating a finite motif by all translations of a lattice, and they model solid crystalline materials whose structures are determined in a rigid form. Their natural equivalence is rigid motion or, more generally, isometry preserving inter-point distances, so the central problem is to study periodic sets through descriptors, metrics, and classification procedures that do not depend on arbitrary choices of unit cell or coordinates (Anosova et al., 2022). Recent work in Periodic Geometry treats isometry classes of periodic point sets as a continuous space, motivated by the fact that atomic vibrations and measurement noise make discrete crystallographic classifications unstable under perturbations (Anosova et al., 2021).

1. Crystallographic model and isometric equivalence

A periodic point set in $\mathbb{R}^n$ is described by a basis $v_1,\dots,v_n$, the associated lattice

$$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$

the unit cell

$$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$

and a finite motif $M=\{p_1,\dots,p_m\}\subset U$. The periodic point set is the Minkowski sum

$S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$

that is, a finite motif repeated at all lattice translations (Anosova et al., 2022). Equivalent formulations appear across the literature, including the representation of periodic sets as finite unions of lattice cosets and as periodic Delone sets invariant under a full-rank lattice (Coulangeon et al., 2018).

The relevant equivalence notion is isometry of Euclidean space. In $\mathbb{R}^n$,

$$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$

so translations, rotations, and reflections are all allowed (Anosova et al., 2022). Two periodic point sets $S,S'\subset\mathbb{R}^n$ are considered the same up to rigid motion if there exists an isometry $\phi$ such that $v_1,\dots,v_n$0 (Anosova et al., 2022). This is stronger than equivalence under a particular unit-cell description, because the same periodic set can be represented by infinitely many bases and motifs, and even primitive-cell descriptions can change discontinuously under perturbations (Anosova et al., 2021).

This setting suggests a general program: an isometry invariant $v_1,\dots,v_n$1 should satisfy $v_1,\dots,v_n$2 whenever $v_1,\dots,v_n$3 is obtained from $v_1,\dots,v_n$4 by an isometry, should vary continuously under perturbations, and ideally should support reconstruction or exact comparison of periodic sets. A recurring theme in the cited work is that past invariants based on symmetry groups, reduced cells, or fixed local cut-offs are discontinuous or representation-dependent, whereas the newer invariants are designed to live naturally on isometry classes (Anosova et al., 2021).

2. Continuous invariants in Periodic Geometry

A central family of invariants is given by density functions. For a periodic set $v_1,\dots,v_n$5 with unit cell $v_1,\dots,v_n$6, for each integer $v_1,\dots,v_n$7 and radius $v_1,\dots,v_n$8, one considers all closed balls of radius $v_1,\dots,v_n$9 centered at points of $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$0, lets $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$1 be the region covered by exactly $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$2 balls, and defines

$$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$3

The infinite sequence $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$4 is the density fingerprint (Anosova et al., 2022). Because isometries preserve distances, balls, and coverage multiplicities, $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$5 is an isometry invariant (Anosova et al., 2022). The monograph on Periodic Geometry presents density functions as one of the key new continuous coordinates on the space of isometry classes of periodic point sets (Anosova et al., 2021).

A second family is based on ordered nearest-neighbor distances. For a periodic set $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$6 with motif $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$7, if $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$8 is the distance from $$\Lambda = \Big\{\sum_{i=1}^n c_i v_i : c_i\in\mathbb{Z}\Big\},$$9 to its $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$0-th nearest neighbor in the infinite periodic set, then

$$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$1

defines the Average Minimum Distance of order $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$2 (Widdowson et al., 2020). These invariants are isometry invariant and satisfy the Lipschitz bound

$$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$3

under bottleneck perturbations (Widdowson et al., 2020). The same paper shows that AMD can distinguish some periodic sets that have identical density functions and that AMD can be computed in near linear time in the input size (Widdowson et al., 2020).

A third family is the Pointwise Distance Distribution. For a periodic set $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$4, one forms for each motif point the ordered distances to its $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$5-nearest neighbors in the full periodic set, lexicographically sorts the resulting rows, and collapses identical rows with weights. The resulting matrix $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$6 is an isometry invariant, is Lipschitz continuous under bottleneck perturbations, and is complete for all periodic sets in general position according to the 2021 development cited in the database description (Widdowson et al., 2021). A later machine-learning adaptation states that PDD is a continuous and generically complete isometry invariant for periodic point sets and that it distinguished all more than 660 thousand periodic crystals in the Cambridge Structural Database as purely periodic sets of points without atomic types (Balasingham et al., 2024).

These invariant families are complementary. Density functions emphasize higher-order overlap geometry of balls, AMD emphasizes ordered nearest-neighbor scales, and PDD encodes weighted local distance profiles. This suggests a layered picture of periodic-point-set classification: coarse and fast descriptors for search, richer fingerprints for discrimination, and complete invariants for exact classification.

3. One-dimensional periodic sequences and explicit density formulas

In dimension $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$7, a periodic point set is a periodic sequence

$$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$8

with $$U(v_1,\dots,v_n)=\Big\{\sum_{i=1}^n c_i v_i : c_i\in[0,1)\Big\},$$9, period $M=\{p_1,\dots,p_m\}\subset U$0, and gaps

$M=\{p_1,\dots,p_m\}\subset U$1

(Anosova et al., 2022). In this setting, balls are intervals, every $M=\{p_1,\dots,p_m\}\subset U$2 is piecewise linear in $M=\{p_1,\dots,p_m\}\subset U$3, and the paper “Density functions of periodic sequences” gives a complete explicit description of $M=\{p_1,\dots,p_m\}\subset U$4 for all $M=\{p_1,\dots,p_m\}\subset U$5 for any 1D periodic sequence (Anosova et al., 2022).

For the uncovered-region density $M=\{p_1,\dots,p_m\}\subset U$6, let $M=\{p_1,\dots,p_m\}\subset U$7 be the sorted gaps. Then $M=\{p_1,\dots,p_m\}\subset U$8 is piecewise linear with corner points $M=\{p_1,\dots,p_m\}\subset U$9 and

$S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$0

with last corner $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$1 (Anosova et al., 2022). A striking consequence is that $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$2 is entirely determined by the multiset of gap lengths $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$3, not by their order along the line (Anosova et al., 2022). This immediately produces many non-isometric sequences with the same $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$4.

For $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$5, each $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$6 decomposes as a sum of trapezoid functions $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$7, one for each contiguous block of $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$8 points. If $S=M+\Lambda=\{u+v\mid u\in M,\ v\in\Lambda\},$9 and $\mathbb{R}^n$0, then

$\mathbb{R}^n$1

where each $\mathbb{R}^n$2 is determined by the triple $\mathbb{R}^n$3, with first and last entries symmetric (Anosova et al., 2022). This gives a purely combinatorial description of the 1D geometry in terms of the cyclic gap sequence.

The same paper proves refined symmetry relations: $\mathbb{R}^n$4 and

$\mathbb{R}^n$5

(Anosova et al., 2022). It also derives integrated densities

$\mathbb{R}^n$6

with closed forms

$\mathbb{R}^n$7

(Anosova et al., 2022).

The 1D theory is not only descriptive but also classificatory. The paper shows that the full density fingerprint is not complete in dimension $\mathbb{R}^n$8: the homometric pair

$\mathbb{R}^n$9

satisfies $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$0 for all $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$1, hence $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$2, despite being non-isometric (Anosova et al., 2022). On the positive side, if all gaps $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$3 are distinct, then the first density function $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$4 uniquely determines the sequence $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$5 up to isometry, so $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$6 is a complete isometry invariant for all generic 1D periodic sequences (Anosova et al., 2022).

A subsequent extension allows different initial radii $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$7, motivated by atomic radii and by continuous events occupying disjoint intervals in time series. The generalized density functions remain isometry invariants, admit explicit 1D formulas, and are strictly stronger than the zero-radius case: they distinguish periodic sequences that have identical densities when all radii vanish (Anosova et al., 2023). This suggests that “decorated” periodic point sets can carry stronger continuous fingerprints than bare point sets.

4. Metrics and complete classification frameworks

A major development is the passage from invariants to metrics on isometry classes. One line of work introduces exactly computable, continuous metrics on isometry classes of finite and 1-periodic sequences (Kurlin, 2022). In 1D, a periodic sequence with successive distances $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$8 is classified up to isometry by the lexicographically smallest cyclic-or-reversed list $$\mathrm{Isom}(\mathbb{R}^n)=\{x\mapsto Ax+b: A\in O(n),\ b\in\mathbb{R}^n\},$$9, and the elastic metric

$S,S'\subset\mathbb{R}^n$0

is a genuine metric on isometry classes (Kurlin, 2022). The key novelty in the periodic case is continuity under perturbations that change the minimum period, and the paper proves

$S,S'\subset\mathbb{R}^n$1

(Kurlin, 2022). For high-dimensional 1-periodic sequences, the analogous invariant is the time-value invariant $S,S'\subset\mathbb{R}^n$2, and the corresponding elastic metric is computable in $S,S'\subset\mathbb{R}^n$3 time (Kurlin, 2022).

A broader classification framework is built around the isoset. The monograph “Introduction to Periodic Geometry and Topology” describes the isoset as a complete invariant built from local cluster isometry classes at a stable radius, together with weights, and then equips the space of isometry classes with a continuous metric via Earth Mover’s Distance on these weighted cluster classes (Anosova et al., 2021). This program is refined in the bridge-length paper, where the bridge length $S,S'\subset\mathbb{R}^n$4 is the minimum number $S,S'\subset\mathbb{R}^n$5 such that any two points $S,S'\subset\mathbb{R}^n$6 can be connected by a finite sequence with consecutive gaps at most $S,S'\subset\mathbb{R}^n$7, equivalently the smallest $S,S'\subset\mathbb{R}^n$8 such that the graph $S,S'\subset\mathbb{R}^n$9 is connected (McManus et al., 2024). The paper proposes a practical algorithm to compute $\phi$0 from a lattice basis and motif using labelled quotient graphs, cycle sums, and Smith Normal Form; if the last added edge has length $\phi$1, then $\phi$2 (McManus et al., 2024). The resulting complexity is

$\phi$3

where $\phi$4, $\phi$5 is the aspect ratio of the cell, and $\phi$6 is the time complexity of computing Smith Normal Form (McManus et al., 2024). Since the size and complexity of the isoset strongly depend on the bridge length, this algorithm is described as a key ingredient in a continuous isometry classification of periodic point sets (McManus et al., 2024).

PDD provides another route to complete or generically complete classification. The 2021 database-scale study defines $\phi$7 by collecting sorted nearest-neighbor distance rows for motif points and collapsing identical rows with weights; it proves invariance under isometry, Lipschitz continuity under bottleneck perturbations, and completeness of PDD for all periodic sets in general position (Widdowson et al., 2021). A later study adapts PDD for property prediction and states that PDD is a continuous and generically complete isometry invariant for periodic point sets, with practical success on large databases (Balasingham et al., 2024). More recent work introduces higher-order PDD variants $\phi$8, based on averages over $\phi$9-tuples, specifically to distinguish all known counter-examples to the completeness of past descriptors (Widdowson et al., 18 Sep 2025). The same paper proves exact completeness in 1D through the pointwise shift distribution $v_1,\dots,v_n$00, which is a complete invariant under rigid motion for all periodic sequences in $v_1,\dots,v_n$01 and computable in time $v_1,\dots,v_n$02 (Widdowson et al., 18 Sep 2025).

These developments collectively indicate two complementary directions. One direction emphasizes exact classification by complete invariants such as isosets, bridge-length-based constructions, and PDD in general position. The other emphasizes continuous metrics such as elastic metrics, Earth Mover’s Distance on PDD-type fingerprints, and bottleneck-style metrics on periodic-point-set spaces.

5. Dynamical, geometric, and coarse-geometric perspectives

Periodic point sets also appear naturally in topological dynamics. In the framework of uniformly discrete and relatively dense sets, periodic sets are Delone sets, and the hull $v_1,\dots,v_n$03 is the orbit closure of a point set under translations or under the Euclidean group $v_1,\dots,v_n$04 in the vague topology (Frettlöh et al., 2012). Almost repetitivity characterizes minimality of the associated dynamical system, and almost linear repetitivity implies unique ergodicity (Frettlöh et al., 2012). Periodic sets sit at the most regular end of this hierarchy: they are linearly repetitive in a very strong sense, hence their hulls are minimal and uniquely ergodic under both translations and Euclidean motions (Frettlöh et al., 2012). This gives a dynamical interpretation of periodicity under isometry: periodic point sets are not only rigid geometric objects but also highly regular points in natural topological dynamical systems.

Another perspective concerns the large-scale geometry of spaces of periodic point sets themselves. The paper “On the metric spaces of lattices and periodic point sets” considers the spaces $v_1,\dots,v_n$05 and $v_1,\dots,v_n$06 equipped with bottleneck distance $v_1,\dots,v_n$07 and Euclidean bottleneck distance $v_1,\dots,v_n$08, where

$v_1,\dots,v_n$09

(Garber et al., 2023). For periodic point sets $v_1,\dots,v_n$10, the following are equivalent: $v_1,\dots,v_n$11, $v_1,\dots,v_n$12, and $v_1,\dots,v_n$13 (Garber et al., 2023). Thus density is a complete invariant for finite bottleneck distance in the periodic setting. The same paper studies embeddability into Hilbert space and shows that certain spaces of periodic point sets of fixed density with only a packing bound or only a covering bound do not coarsely embed into any uniformly convex Banach space, in particular not into Hilbert space (Garber et al., 2023). By contrast, subclasses with bounded motif size and covering radius, or bounded unit-cell diameter, are bounded in $v_1,\dots,v_n$14 and hence trivially coarsely embeddable (Garber et al., 2023). A plausible implication is that the global geometry of the moduli space of periodic point sets is far more complicated than the local continuity of individual invariants might suggest.

Energy minimization provides a further geometric viewpoint. A periodic point set $v_1,\dots,v_n$15 is $v_1,\dots,v_n$16-periodic if it is a finite union of lattice cosets, and energies in the Gaussian core model depend only on pairwise distances, hence are invariant under isometries (Coulangeon et al., 2018). The paper “Local Energy Optimality of Periodic Sets” characterizes periodic point sets that are $v_1,\dots,v_n$17-critical for all $v_1,\dots,v_n$18 in terms of balanced shells and weighted spherical $v_1,\dots,v_n$19-designs (Coulangeon et al., 2018). For the 2-periodic family $v_1,\dots,v_n$20, shell symmetry and spherical $v_1,\dots,v_n$21-design properties allow explicit Hessian analysis, leading to local $v_1,\dots,v_n$22-optimality in odd dimensions $v_1,\dots,v_n$23 for sufficiently large $v_1,\dots,v_n$24 (Coulangeon et al., 2018). This suggests that isometry classification and local energy landscape analysis are tightly linked through shell geometry and symmetry.

6. Applications, limitations, and open directions

The applications in crystallography and materials science are immediate. Continuous, isometry-invariant descriptors are intended to compare crystal structures across experiments, simulations, and noise, to detect duplicates and near-duplicates, and to organize large databases by geometry rather than by unstable cell descriptions (Anosova et al., 2022). PDD-based representations have already been used to compare nearly 1.5 million crystals from the world’s four largest databases within 2 hours on a modest desktop computer (Widdowson et al., 2021). The same line of work argues that the PDD will not allow anyone to claim a “new” material as a noisy disguise of a known crystal (Widdowson et al., 2021). Machine-learning work using PDD reports accuracy on par with state-of-the-art methods while being several times faster in both training and prediction time (Balasingham et al., 2024).

The bridge-length algorithm was tested on a large crystal dataset and is explicitly described as required for an efficient continuous classification of all periodic crystals and as a key step toward inverse design of materials from new invariant values (McManus et al., 2024). Higher-order PDD invariants were designed to distinguish all known counter-examples to the completeness of past descriptors and to confirm thousands of near-duplicates in the world’s largest databases of inorganic crystals within hours on a desktop computer (Widdowson et al., 18 Sep 2025). These claims indicate a shift from purely theoretical classification to operational data curation.

Several limitations remain explicit in the literature. Density fingerprints are complete in general position in $v_1,\dots,v_n$25, but incompleteness appears already in dimension $v_1,\dots,v_n$26 (Anosova et al., 2022). PDD is complete for all periodic sets in general position, yet higher-order constructions were needed to resolve known counterexamples to the completeness of first-order descriptors (Widdowson et al., 2021). The problem is open for $v_1,\dots,v_n$27-periodic sequences with $v_1,\dots,v_n$28 in the metric framework of exactly computable elastic metrics (Kurlin, 2022). Extension of explicit 1D density formulas to dimensions $v_1,\dots,v_n$29 is also open, because ball intersections become much more complex and the piecewise linear structure disappears (Anosova et al., 2022).

The open directions stated across the cited papers are coherent. One line asks for stronger or simpler complete invariants and for parameterizations of the space of isometry classes that support reconstruction and inverse design (Anosova et al., 2021). Another asks for higher-dimensional generalizations of explicit density formulas, for extremal configurations minimizing $v_1,\dots,v_n$30 or the maximum of $v_1,\dots,v_n$31, and for understanding when $v_1,\dots,v_n$32 has a unique local maximum (Anosova et al., 2022). A third line concerns broader geometric settings: quasi-periodic structures, Delone sets beyond strict periodicity, substitution tilings with dense tile orientations, and point sets with almost periodic modulations (Frettlöh et al., 2012).

Periodic point sets under isometry therefore form a domain in which discrete geometry, crystallography, topology, metric geometry, and data science now interact through a common language: lattices and motifs, rigid-motion equivalence, continuous invariants, and metrics on spaces of structures. The cited works show that this language is already strong enough to support explicit formulas in one dimension, exact algorithms for certain complete invariants, practical large-scale database analysis, and a developing global geometry of the moduli space itself (Anosova et al., 2021).