Triad: Three-Way Structures in Multi-disciplinary Research

• Triad is a versatile three-part construct that appears in geometry as tangent circles and extends to dynamic systems, network motifs, and seismological sequences.
• In dynamic systems, triads denote minimal mode interactions in Fourier and wave analyses, offering insights into energy cascades and resonant phenomena.
• In applied fields, triads underpin methods in orientation estimation, trusted timing, and 3D MRI vision models, linking theoretical constructs to practical implementations.

Searching arXiv for papers on “Triad” and closely related uses across disciplines. Triad is a research term with multiple precise meanings that depend on disciplinary context. In the literature considered here, it denotes a triad of circles attached to an acute triangle; a minimal three-mode interaction in Fourier, wave, and resonant-basin dynamics; a three-node subgraph or motif in network analysis; the foreshock–main shock–aftershock organization of earthquake sequences; the Three-Axis Attitude Determination algorithm in orientation estimation; a trusted timestamp dispatcher built from cooperating enclaves; and, as a proper name, a 3D MRI vision foundation model and a family of generating-function constructions linking Macdonald-type and Baker–Akhiezer-type polynomials (Suppa et al., 2023, Benavides et al., 4 May 2026, Jia et al., 2016, Zotov et al., 2021, Marina et al., 2016, Fernandez et al., 2023, Wang et al., 19 Feb 2025, Mironov et al., 10 Mar 2025).

1. Euclidean and algebraic triads

In Euclidean geometry, Suppa and Rabinowitz study a triad of circles associated with an acute triangle $\triangle ABC$. For each side, one erects the semicircle with that side as diameter inside the triangle, and then defines three interior circles $\gamma_a,\gamma_b,\gamma_c$, each tangent to two sides of the triangle and externally tangent to the semicircle on the opposite side. If $r$ is the inradius and $\rho_a,\rho_b,\rho_c$ are the triad radii, then

$$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$

and the configuration satisfies

$$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$

where $\rho_i$ and $\rho_o$ are the radii of the inner and outer Apollonius circles of the triad. The paper also identifies the radical center of $\gamma_a,\gamma_b,\gamma_c$ with the Gergonne point $X_7$, proves concurrency at the Paasche point $\gamma_a,\gamma_b,\gamma_c$0, and places the centers of the inner and outer Apollonius circles on the line $\gamma_a,\gamma_b,\gamma_c$1 with ratio $\gamma_a,\gamma_b,\gamma_c$2 (Suppa et al., 2023).

In algebraic and integrable-systems usage, the "triad" of "Diamond of triads" denotes the embedding of symmetric polynomials and Baker–Akhiezer-type polynomials into a Noumi–Shiraishi-type power series. The basic triad is associated with the vector representation of the Ding–Iohara–Miki algebra; its symmetric reduction produces Macdonald-type polynomials, while the specialization $\gamma_a,\gamma_b,\gamma_c$3 yields Baker–Akhiezer-type polynomials. The paper then lifts the basic construction to two elliptic generalizations and to a bi-elliptic ELS triad, thereby organizing basic, elliptic, Shiraishi, and ELS constructions into a "diamond" connected to elliptic and bi-elliptic DIM algebras and to Seiberg–Witten theory with adjoint matter in various dimensions (Mironov et al., 10 Mar 2025).

2. Triads as units of nonlinear interaction

In two-dimensional turbulence, a triad is a triple of Fourier modes $\gamma_a,\gamma_b,\gamma_c$4 satisfying

$\gamma_a,\gamma_b,\gamma_c$5

Because the nonlinearity is quadratic, all nonlinear transfers of energy and enstrophy decompose into sums over such triads. The natural phase variable is the triad phase

$\gamma_a,\gamma_b,\gamma_c$6

and the triad transfer function is

$\gamma_a,\gamma_b,\gamma_c$7

The 2026 paper develops a stochastic closure in which neighboring triads are treated as weakly correlated noise, leading to a noisy Adler-type equation, a von Mises steady-state phase PDF, and a closed transfer rule whose sign is determined by the sign structure of $\gamma_a,\gamma_b,\gamma_c$8. For inertial-range spectra $\gamma_a,\gamma_b,\gamma_c$9, the paper shows that $r$0 for $r$1, which yields $r$2 and $r$3, i.e. inverse energy cascade and forward enstrophy cascade in 2D turbulence (Benavides et al., 4 May 2026).

Wave literature uses the same term for three-wave couplings. For deep-water surface gravity waves with dispersion $r$4, exact resonant triads do not exist, but quadratic terms still generate non-resonant triad interactions that are dynamically significant over finite times. The 2024 deep-water study finds that triad contributions can follow the trend of quartet resonances with comparable magnitude for most wavenumbers, while dominating the rapid filling of low-wavenumber, low-energy portions of the spectrum and producing both bound and free modes at the same wavenumber (Zhang et al., 2024). By contrast, for gravity waves confined to a finite-depth cylinder, exact resonant triads can exist. For any three correlated modes with Laplacian wavenumbers $r$5, there exists a positive finite critical depth $r$6 such that $r$7 if and only if

$r$8

and that critical depth is unique. The resulting amplitude equations are canonical three-wave equations, unforced triad evolution is always periodic, and externally forced triads may be periodic, quasi-periodic, or chaotic (Durey et al., 2023).

3. Triads in graph and network analysis

In graph theory, triad usually denotes a subgraph induced by three vertices, but the exact object depends on graph type. For large directed graphs, a triad has $r$9 possible labeled edge states, which are condensed into a $\rho_a,\rho_b,\rho_c$0-element triad census by considering isomorphic cases. Scalable triadic analysis therefore centers on counting how many times each directed three-node configuration occurs. The 2012 systems paper implements an $\rho_a,\rho_b,\rho_c$1 Batagelj–Mrvar census algorithm on shared-memory architectures, using compact adjacency structures and parallel local census arrays, and evaluates it on the Cray XMT, HP Superdome, and AMD multi-core NUMA systems for graphs ranging from millions to more than $\rho_a,\rho_b,\rho_c$2 vertices (Jr. et al., 2012).

Motif analysis refines this by asking which triad patterns are statistically overrepresented. In directed networks there are $\rho_a,\rho_b,\rho_c$3 connected non-isomorphic triad patterns globally, but once one distinguishes the focal node there are $\rho_a,\rho_b,\rho_c$4 connected node-specific triad patterns. NoSPaM defines, for each node $\rho_a,\rho_b,\rho_c$5, a node-specific Z-score

$\rho_a,\rho_b,\rho_c$6

relative to a degree-preserving null ensemble. Empirically, the paper finds that motifs are distributed highly heterogeneously: even the feed-forward loop, often treated as a global motif, is concentrated around a relatively small subset of nodes rather than being uniformly distributed across the network (Winkler et al., 2014).

A different network-science usage appears in metadata-group detection for undirected, unweighted graphs. There, a triad motif is specifically a three-vertex path with exactly two edges,

$\rho_a,\rho_b,\rho_c$7

and a substructure is called triad-rich if every pair of vertices participates in at least one triadic interaction. The paper proves that this is equivalent to being an induced connected subgraph of diameter $\rho_a,\rho_b,\rho_c$8, i.e. a $\rho_a,\rho_b,\rho_c$9-club, and develops DIVANC, an edge-division algorithm based on edge Niche Centrality,

$$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$0

together with a $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$1-hop overlapping strategy for overlapping groups (Jia et al., 2016).

4. Tectonic earthquake triads

In seismology, triad is a formal classification of local earthquake sequences into foreshocks, main shock, and aftershocks. The main shock magnitude is denoted $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$2; the maximum foreshock and aftershock magnitudes are $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$3 and $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$4; and the corresponding event counts in a prescribed window are $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$5 and $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$6. The spatial window is a circular epicentral zone of radius $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$7 defined by

$$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$8

with $$\rho_a=r\Bigl(1-\tan\frac{A}{2}\Bigr),\quad \rho_b=r\Bigl(1-\tan\frac{B}{2}\Bigr),\quad \rho_c=r\Bigl(1-\tan\frac{C}{2}\Bigr),$$9 in km, and the temporal window is 24 h before to 24 h after the main shock. A classical triad satisfies $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$0, $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$1, and $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$2. A mirror triad reverses the count inequality to $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$3, often with no aftershocks at all, while a symmetric triad satisfies $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$4. The paper also isolates "Grande terremoto solitario" with $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$5 as a degenerate case of the triadic framework (Zotov et al., 2021).

Using the USGS/NEIC global catalog for 1973–2019, with main shocks $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$6 and depth $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$7 km, the study identifies $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$8 main shocks and reports that, among triads excluding GTS, classical triads comprise about $$\rho_i^2=\rho_a^2+\rho_b^2+\rho_c^2,\qquad 3\rho_o=\rho_i+4r,$$9, mirror triads about $\rho_i$0, and symmetric triads about $\rho_i$1. Mirror triads exhibit an Omori-like pattern on the foreshock side and a foreshock analogue of Bath’s law through $\rho_i$2. Their proposed mechanical interpretation uses the threshold law

$\rho_i$3

with heterogeneity in $\rho_i$4 allowing smaller faults to fail before the largest fault, thereby shifting clustering from the aftershock side to the foreshock side (Zotov et al., 2021).

5. TRIAD in orientation estimation and Triad in trusted timing

In aerospace and robotics, TRIAD stands for Three-Axis Attitude Determination. It is a deterministic two-vector attitude algorithm that constructs orthonormal bases from two known, non-collinear reference vectors $\rho_i$5 and their measured body-frame counterparts $\rho_i$6: $\rho_i$7

$\rho_i$8

and then outputs the direction cosine matrix

$\rho_i$9

In the 2016 UAV paper, TRIAD is used as the observation model of a UKF-based AHRS whose state is $\rho_o$0, with gravity and magnetic field as references and accelerometer and magnetometer as measurements. The reported result is good real-time performance with low computational cost in a microcontroller, validated in simulation and fixed-wing UAV field experiments (Marina et al., 2016).

A 2025 extension uses the sub-optimal asymmetry of TRIAD to mitigate magnetometer contamination of pitch and roll in a Manifold EKF. Instead of feeding the raw normalized magnetometer $\rho_o$1 into the filter, the method computes

$\rho_o$2

and replaces the magnetometer channel by $\rho_o$3, with the corresponding inertial reference $\rho_o$4. In the reported robotic-arm experiments, at $\rho_o$5 roll the equal-noise Manifold EKF2 has about $\rho_o$6 error, the retuned $\rho_o$7 version about $\rho_o$8, and the TRIAD-aided Manifold EKF2 about $\rho_o$9 (Sadananda et al., 27 Sep 2025).

Triad is also the name of a trusted timestamp system for untrusted cloud environments. It runs mutually supportive enclave-based clock servers inside Intel SGX and SCONE, maintains cached timestamps from CPU counters while enclaves remain continuously active, marks time as tainted after asynchronous enclave exits, and uses peer enclaves or external time servers to untaint and continue a monotonic trusted timeline. The system formalizes bounded-error and monotonicity requirements as

$\gamma_a,\gamma_b,\gamma_c$0

with recovery-time error modeled as $\gamma_a,\gamma_b,\gamma_c$1. In the reported evaluation, mean peer RTT is $\gamma_a,\gamma_b,\gamma_c$2 with maximum $\gamma_a,\gamma_b,\gamma_c$3, mean external Roughtime RTT is $\gamma_a,\gamma_b,\gamma_c$4, and over a $\gamma_a,\gamma_b,\gamma_c$5-second run local timestamps dominate overwhelmingly, with peer and external retrievals being rare fallback events (Fernandez et al., 2023).

6. Triad as a proper name in medical imaging

In machine learning for medical imaging, Triad is a 3D MRI vision foundation model rather than a generic three-element object. It is trained on TriadMR-131K, a pre-training corpus of $\gamma_a,\gamma_b,\gamma_c$6 3D MRI volumes from brain, breast, and prostate, and combines an autoencoder-based visual backbone with organ-independent imaging descriptions extracted from DICOM metadata. During pre-training, those descriptions are encoded by a frozen GTR-T5-Large text encoder, and the visual representation is constrained by a log-ratio loss in addition to an $\gamma_a,\gamma_b,\gamma_c$7 reconstruction loss. The released family includes nnUNet and 3D Swin Transformer encoders and is evaluated on $\gamma_a,\gamma_b,\gamma_c$8 downstream datasets spanning segmentation, classification, and registration in within-domain and out-of-domain settings (Wang et al., 19 Feb 2025).

The abstract reports that nnUNet-Triad improves segmentation performance by $\gamma_a,\gamma_b,\gamma_c$9 compared to nnUNet-Scratch across $X_7$0 datasets, Swin-B-Triad improves classification by $X_7$1 over Swin-B-Scratch across five datasets, and SwinUNETR-Triad improves registration by $X_7$2 over SwinUNETR-Scratch across two datasets (Wang et al., 19 Feb 2025). In this usage, "Triad" functions as the proper name of a domain-specific model, but the naming still preserves the broader pattern found elsewhere in the literature: a compact structure that mediates between multiple representations, modalities, or interacting components.

Across these literatures, triad consistently names a three-part construction, but the object itself varies sharply: three tangent circles, three coupled modes, three nodes, three phases of a seismic sequence, two reference vectors plus a derived orthonormal basis, three mutually supporting enclave clocks, or a model title attached to a multi-component learning system. The term therefore belongs less to a single discipline than to a recurring research grammar of structured three-way relation.