Triplewise Contact Tracing

• Triplewise Contact Tracing is a higher-order approach that extends conventional pairwise tracing by using third-order interactions or triples to capture deeper transmission pathways.
• It encompasses multiple formulations—such as level-3 tracing in temporal graphs, motif-based epidemic models, and gathering-based methods with sideward tracing—to address diverse operational objectives.
• Some protocols integrate robust cryptographic and privacy-preserving workflows and dynamic contact queues to manage data scalability, epidemic thresholds, and compliance challenges.

Searching arXiv for the cited papers and topic to ground the article in the referenced literature. arxiv_search(query="(Andreoletti et al., 2020) OR (Mahapatra et al., 2020) OR (Meijere et al., 4 Mar 2026) OR (Mancastroppa et al., 2021) OR (Loh, 2020)", max_results=10, sort_by="relevance") arxiv_search(query="triplewise contact tracing", max_results=10, sort_by="relevance") Triplewise Contact Tracing denotes a set of higher-order contact-tracing formulations that extend conventional direct-contact tracing by incorporating third-order structure. In the arXiv literature, the term has been used for at least five distinct but related ideas: third-hop tracing in temporal contact graphs, tracing conditioned on a three-node motif such as $[TIT]$, gathering-level tracing that adds a “sideward” pathway beyond forward and backward tracing, a flipped-perspective alerting scheme that reports distance-$3$ proximity to infection in a recent interaction graph, and a privacy-preserving workflow spanning multiple Mobile Operators, a Governmental Authority, and end-users. A plausible implication is that “triplewise” functions less as the name of a single protocol than as an umbrella for contact-tracing schemes in which epidemiological inference depends on triples, three-hop neighborhoods, or three-party coordination rather than pairwise exposure alone (Mahapatra et al., 2020, Meijere et al., 4 Mar 2026, Mancastroppa et al., 2021, Loh, 2020, Andreoletti et al., 2020).

1. Scope of the term

The principal usages of Triplewise Contact Tracing differ in mechanism, data model, and operational objective. Some formulations are graph-theoretic and define triplewise tracing as level-$3$ neighborhood discovery; some are epidemic-dynamical and require two tracing-triggering neighbors for an infected node to be traced; some are operational and exploit whole-gathering tracing to identify infections lateral to the index case; some are notification-oriented and expose the graph distance to infection; and some are privacy-preserving systems in which three classes of actors jointly compute exposure without revealing raw mobility or infection status.

Research line	Triplewise meaning	Core object
Multilevel digital tracing	Level-$3$ contact tracing	Three-hop, time-respecting paths
Network epidemic model	Tracing on a $[TIT]$ triple	Higher-order motif
Gathering tracing	Forward + backward + sideward	Traced simplex
Flipped perspective	$k=3$ relationship-distance alert	Recent interaction graph
Multi-operator workflow	MO–GA–user protocol	Privacy-preserving score computation

The shared conceptual shift is from pairwise exposure notification to higher-order inference. In one line of work, this means augmenting a contact graph with temporal edge labels and tracing paths of length $3$; in another, it means tracing only when two treated neighbors simultaneously surround an infected node; in another, it means tracing all participants in a gathering so that infections unrelated by a direct transmission edge to the index can still be found. This suggests that the common denominator is not a single epidemiological semantics, but the use of third-order relational structure as the operative unit of tracing (Mahapatra et al., 2020, Meijere et al., 4 Mar 2026, Mancastroppa et al., 2021, Loh, 2020, Andreoletti et al., 2020).

2. Temporal graph formulations and level-3 tracing

In the multilevel digital contact-tracing framework, the underlying object is a dynamic contact graph

$G(t)=(V,E(t),\{Q_e(t)\}),$

where $V$ is the set of smartphone-enabled individuals, $E(t)$ is the undirected edge set of close contacts active in the last $3$0 time slots, and each edge carries a binary circular contact queue $3$1. The queue stores whether a qualifying contact occurred in each recent slot within the incubation window. If $3$2 is the incubation period in days and $3$3 is the slot duration in minutes, then

$3$4

For the COVID-19 parametrization used in the paper, $3$5 days and $3$6 minutes, giving $3$7.

The contact queue admits two equivalent updates. In shift-and-inject form,

$3$8

with wrap mask $3$9 and slot bit $3$0. In circular-buffer form, one updates index $3$1 and writes $3$2. The contact bit is determined by whether two individuals have remained within the disease-specific threshold $3$3 for at least $3$4 minutes during the slot. If a queue becomes $3$5, the corresponding edge can be removed, keeping the graph sparse and dynamic.

Triplewise tracing in this framework is level-$3$6 tracing. For a seed set $3$7 of infected individuals, the level-$3$8 neighborhood is

$3$9

Thus $3$0 are direct contacts, $3$1 are second-level contacts, and $3$2 are triplewise contacts. Eligibility is temporal rather than merely topological. A path $3$3 is time-respecting if the edge slots satisfy $3$4. The framework enforces this through a bit-level TraceOperator on aligned CCQs: for two consecutive edges, the operator returns true if either $3$5, or $3$6 but the oldest-one index in the first queue is at least the latest-one index in the second.

This representation supports breadth-first expansion to depth $3$7 while preserving temporal order. It also supports infection-pathway construction and optional risk scoring through binary, summed, or recency-decayed edge risks. The storage analysis in the paper gives $3$8 bits overall, with adjacency plus $3$9 bits for CCQs; in the case study with $[TIT]$0, $[TIT]$1 days, $[TIT]$2 minutes, and average close contacts per incubation window $[TIT]$3, storage is approximately $[TIT]$4 GB. The formalism is therefore not merely conceptual: it is intended as a tractable temporal data structure for enumerating level-$[TIT]$5, level-$[TIT]$6, and level-$[TIT]$7 contact lists together with time-respecting transmission pathways (Mahapatra et al., 2020).

3. Higher-order epidemic models and analytical thresholds

In the network-based mean-field model, triplewise contact tracing is not a graph-distance notion but a motif-conditioned transition. Individuals occupy states $[TIT]$8, $[TIT]$9, $k=3$0, and $k=3$1, where $k=3$2 denotes treated individuals who are non-infectious and actively trigger tracing among their neighbors. Pairwise tracing acts on an $k=3$3 edge: an infected individual moves to $k=3$4 at rate $k=3$5 conditional on direct contact with one treated neighbor. Triplewise tracing acts on a $k=3$6 triple: an infected individual moves to $k=3$7 at rate $k=3$8 conditional on simultaneous connection to two tracing-triggering neighbors.

After standard moment closure,

$k=3$9

and rescaling by the natural timescale $3$0, the closed system includes

$3$1

Here $3$2 is the time-scaled infection rate, $3$3 is the time-scaled termination rate of tracing activity, $3$4 is the compliance fraction entering $3$5, and $3$6 is the average degree.

The threshold problem is reduced through fast variables,

$3$7

which stabilize rapidly near the start of the outbreak. The epidemic threshold is located by solving the quasi-steady algebraic system and imposing zero prevalence growth,

$3$8

For pure pairwise tracing $3$9, the critical tracing level is

$G(t)=(V,E(t),\{Q_e(t)\}),$0

where

$G(t)=(V,E(t),\{Q_e(t)\}),$1

If $G(t)=(V,E(t),\{Q_e(t)\}),$2, then $G(t)=(V,E(t),\{Q_e(t)\}),$3. For pure triplewise tracing $G(t)=(V,E(t),\{Q_e(t)\}),$4, the critical level is

$G(t)=(V,E(t),\{Q_e(t)\}),$5

where

$G(t)=(V,E(t),\{Q_e(t)\}),$6

If $G(t)=(V,E(t),\{Q_e(t)\}),$7, then $G(t)=(V,E(t),\{Q_e(t)\}),$8.

Several consequences follow directly from these formulas. Triplewise tracing is more stringent than pairwise tracing, since the required tracing rate is larger and the minimum compliance bound is higher in many regimes. The model also departs from analyses that assume tracing is arbitrarily fast. Outside the fast-tracing regime, the exact pairwise threshold can be orders of magnitude larger than the Eames–Keeling limit, and triplewise tracing requires an adjusted fast-tracing scaling, with $G(t)=(V,E(t),\{Q_e(t)\}),$9 alongside $V$0, to prevent $V$1 motifs from becoming too rare. When pairwise and triplewise tracing operate jointly, the control boundary is obtained by numerically solving the quasi-steady equations for $V$2 and evaluating the zero-growth condition. The paper reports a slight reinforcement beyond a naive linear combination of the pure mechanisms, arising from motif correlations (Meijere et al., 4 Mar 2026).

4. Gatherings, simplices, and sideward tracing

A distinct line of work defines triplewise tracing through traced gatherings rather than pairwise edges or $V$3 motifs. The network substrate is a simplicial activity-driven temporal network in which a gathering of size $V$4 is an $V$5-simplex, represented as a fully connected clique of $V$6 nodes created for one time step and then removed. Gathering sizes are drawn from a distribution $V$7, and the average number of contacts per individual per unit time is fixed by

$V$8

The epidemic process uses compartments $V$9, $E(t)$0, $E(t)$1, $E(t)$2, and $E(t)$3, with added states $E(t)$4 and $E(t)$5 for traced and quarantined asymptomatic individuals. Infection occurs on any active link with probability $E(t)$6 per contact; symptomatic-path infections follow $E(t)$7 with probability $E(t)$8, asymptomatic-path infections follow $E(t)$9 with probability $3$00; $3$01 occurs at rate $3$02; asymptomatics recover at rate $3$03; symptomatic individuals are immediately isolated and cease participating in gatherings.

Contact tracing is triggered when a presymptomatic individual becomes symptomatic. For each gathering attended in the previous $3$04 days, the entire simplex is traced with probability $3$05. This generates three mechanisms. Forward tracing identifies asymptomatic individuals infected by the index case. Backward tracing identifies the asymptomatic source that infected the index case. Sideward tracing identifies asymptomatic infections in the same gathering that neither infected nor were infected by the index case. Sideward tracing is therefore absent for $3$06 and appears only for gatherings with at least three participants.

In the linearized dynamics, the tracing contributions are

$3$07

$3$08

$3$09

The paper evaluates efficacy by the increase in the epidemic threshold $3$10 induced by tracing. Because backward and sideward terms depend nonlinearly on $3$11, a simple $3$12 factorization is not used; instead, the principal analytical observable is the shift in $3$13.

A central result is that tracing efficiency becomes non-monotonic in gathering size $3$14. Backward tracing is relatively more effective for small $3$15, whereas sideward tracing scales strongly with larger gatherings through both $3$16 and the term $3$17. The combination yields a peak at intermediate sizes: around $3$18 for single-size or sharply exponential $3$19, and around $3$20 for broader heavy-tailed $3$21. The empirical analysis on a University of Parma campus uses Wi-Fi Access Point data from 713 APs over two policy phases. Under symptomatic isolation only, the shift in gathering-size distribution implied

$3$22

and

$3$23

Under partial opening, tracing targeted at large gatherings was sufficient to match closure-level suppression with $3$24, corresponding to tracing all gatherings with $3$25 and about $3$26 of gatherings with $3$27. To remain subcritical against a variant with $3$28, $3$29 was sufficient, corresponding to tracing $3$30 and about $3$31 of gatherings with $3$32. Mechanism-wise, backward tracing contributes the largest share across strategies, but sideward tracing becomes major, and in very broad tails dominant, when large gatherings are present (Mancastroppa et al., 2021).

5. Flipped-perspective tracing and distance-3 alerts

In the flipped-perspective framework, triplewise tracing is an alerting problem on a recent interaction graph rather than a quarantine instruction for direct contacts. The operative variable is relationship distance: $3$33 is a direct recent physical contact of a newly reported positive case, $3$34 is a contact of a contact, and $3$35 is a contact of a contact of a contact. A user is not told who the positive case is; instead, the system displays how many relationships away infection has recently appeared.

The graph is built over a rolling 14-day window from pseudonymous proximity detections and co-Wi-Fi presence. An undirected edge exists if two users were within about 10 meters for at least 15 minutes via BLE or ultrasound, or were concurrently on the same Wi-Fi Access Point for at least 3 hours. On receipt of a token from a positive case or from a confirmed contact, the server constructs a filtered 14-day graph snapshot and runs a breadth-first search to depth $3$36. Distances are pinned at report time and displayed for a fixed fade-out period. The system emits only per-user distance counts rather than identities or paths.

The framework analyzes ring sizes on random graphs. If $3$37 is the degree of a random node, with mean $3$38 and excess-degree factor

$3$39

then

$3$40

and the closed $3$41-ball excluding ego is

$3$42

when $3$43. Under random adoption $3$44, the giant-component threshold is approximately

$3$45

while a simple deployment heuristic is $3$46, or, in dense schools and universities, $3$47. The paper argues that this can place practical critical mass below $3$48 in dense settings.

The system also uses positive-test tokens and confirmed-contact tokens to amplify signal generation. If a positive has $3$49 traced contacts and each submits a token independently with probability $3$50, then the chance that at least one token is entered is

$3$51

which exceeds $3$52 when $3$53 and $3$54. In early deployment at Georgia Tech, around $3$55 voluntary adoption still yielded users seeing on the order of $3$56–$3$57 connected users across distance rings in their charts. The third ring expands coverage substantially: in heterogeneous networks,

$3$58

so when $3$59, the marginal increase from adding $3$60 is approximately $3$61 the sum of rings $3$62 and $3$63. The trade-off is reduced per-alert specificity, since signal-to-noise proxies decay with $3$64, but the intended benefit is earlier and broader self-protective behavior change rather than post-exposure quarantine alone (Loh, 2020).

6. Privacy-preserving multi-operator workflows

A further usage of Triplewise Contact Tracing maps a multi-operator COVID-19 tracing protocol into a three-sided privacy architecture involving multiple Mobile Operators (MOs), a Governmental Authority (GA), and end-users. Each user belongs to exactly one MO in the core protocol. MOs hold passive geolocation data and inferred contact flags; the GA holds official infection statuses $3$65 and the Paillier private key; end-users receive encrypted exposure scores and a masking token used to recover only their own score. The trust model is honest-but-curious: parties follow the protocol but may attempt to infer additional information.

The core spatial predicate is

$3$66

implemented in practice by securely comparing squared distances from estimated locations $3$67 without square roots. The time-resolved exposure score is

$3$68

and the optional place-based score for a user-selected area $3$69 is

$3$70

Conceptually, this is a private intersection between a user’s contact set and the positive set, but operationally it is realized through Shamir Secret Sharing and homomorphic summation rather than explicit PSI.

The protocol combines three cryptographic primitives. Shamir Secret Sharing with a $3$71 threshold splits each sensitive value into two shares and supports secure multiplication, equality, and comparison on shares. Paillier encryption provides additive homomorphism, with public encryption for all parties and GA-only decryption. An SSS-based Oblivious Transfer is used for selective identity revelation in the thresholded, GA-triggered mode. The message flow has four phases. In setup, GA generates Paillier keys; MOs agree on the sampling interval $3$72, the contact threshold $3$73, and a spatial partition into non-overlapping squares of side $3$74; GA sets an optional alert threshold $3$75. In Phase A, pairs of MOs securely compute cross-operator contacts by exchanging fresh position shares and evaluating SecDist and Comp, then reconstruct only the contact bit $3$76. In Phase B, GA sends encrypted infection statuses to MOs; MOs homomorphically accumulate encrypted scores for their users, build tables indexed by subscriber, and cannot decrypt any score. In Phase C, a user requests $3$77, multiplies it by a random $3$78, sends the masked ciphertext to GA for decryption, and removes the mask locally after GA returns $3$79. In Phase D, optional GA-triggered identification proceeds only for users with score at least $3$80: GA decrypts indexed scores without identities and then uses OT to recover identity shares for selected rows.

The privacy guarantees are correspondingly layered. MOs never learn plaintext infection status or decrypted scores. GA never learns mobility, contacts, locations, or user scores in the user-triggered path, and in the GA-triggered path learns identities only for users whose score is at least $3$81. In the overhead-reduced thresholded mode, GA sends an index range around a candidate row, and MO privacy is quantified by

$3$82

For example, $3$83 yields privacy of approximately $3$84.

The simulations use population $3$85 million, $3$86 MOs, $3$87 s, a 1-hour period, a region of $3$88 km$3$89, and $3$90 positives. The distance threshold is $3$91 m. In the contact-tracing phase, per-contact overhead is approximately $3$92 bits with $3$93 bits, average per-MO overhead stays below approximately $3$94 MB per area, and the maximum per-area overhead rises to approximately $3$95 MB at $3$96 m for $3$97. Average timing at $3$98 m is approximately $3$99 s and maximum timing approximately $3$00 s, permitting $3$01 s sampling for $3$02 m; at $3$03 m, maximum time rises to approximately $3$04 s. In the score-computation phase, GA$3$05MOs traffic is $3$06 MB per hour and MO-to-MO traffic is $3$07 MB per hour. User-triggered communication is approximately $3$08 KB per user. GA-triggered identity resolution under total privacy has overhead

$3$09

with an example of approximately $3$10 MB for $3$11 and $3$12; using an $3$13-range reduces this to approximately $3$14 MB at approximately $3$15 privacy. Under total privacy and $3$16, obtaining identities for users with score at least $3$17 takes approximately $3$18 minutes; at approximately $3$19 privacy, approximately $3$20 s. The operational claim is therefore not absolute efficiency, but feasible early detection with explicit privacy–overhead trade-offs (Andreoletti et al., 2020).

7. Comparative issues, misconceptions, and open problems

A recurrent misconception is that Triplewise Contact Tracing names a single standardized protocol. The literature does not support that interpretation. It instead contains several non-equivalent constructions: a temporal depth-$3$21 neighborhood algorithm, a $3$22 motif-based epidemic control mechanism, whole-gathering tracing with sideward detection, a distance-$3$23 alerting system on pseudonymous interaction graphs, and a cryptographic MO–GA–user workflow. These constructions share a move beyond direct-contact tracing, but they do not share a single state space, objective function, or privacy architecture.

The comparative advantages are likewise heterogeneous. In relation to Bluetooth-based smartphone tracing such as GAEN or DP-3T, the multi-operator network-side workflow emphasizes passive positioning, no app installation, potential place-level exposure scoring, and no device energy burden, while noting that GAEN or DP-3T requires widespread app installation, approximately $3$24 of the population for effectiveness, and that Bluetooth proximity is noisy because RSSI is susceptible to multipath and device heterogeneity. The dynamic-graph and gathering-based approaches instead emphasize improved path reconstruction or the ability to exploit higher-order interactions that pairwise tracing cannot represent. The flipped-perspective system emphasizes pre-exposure warning and behavior change rather than post-exposure quarantine. The analytical pairwise-mean-field model emphasizes rigorous epidemic thresholds under partial compliance and beyond the fast-tracing regime (Andreoletti et al., 2020, Mahapatra et al., 2020, Meijere et al., 4 Mar 2026, Mancastroppa et al., 2021, Loh, 2020).

The limitations are correspondingly specific. The privacy-preserving workflow depends on ultra-dense 5G positioning assumptions, careful tuning of $3$25, $3$26, and $3$27, and an honest-but-curious threat model; it includes add-ons for specific malicious behaviors, but broader defenses such as composable security, verifiable MPC, and auditability are identified as potential improvements. The multilevel CCQ framework faces memory growth with large $3$28 and a fidelity–resource trade-off in the slot size $3$29. The $3$30 threshold model uses a static network, homogeneous rates, and moment closure, so clustering, multilayer structure, and richer temporal dynamics remain outside the formalism. The gathering model assumes perfect instantaneous tracing, perfect isolation, and no iterative tracing, and evaluates efficacy through threshold shifts rather than prevalence above threshold. The flipped-perspective system requires centralized distance computation and careful rate-limiting to mitigate re-identification and alert fatigue, especially at $3$31. More generally, higher-order tracing can improve recall but may increase false positives if eligibility is too permissive or if compliance, sensing quality, and timing are weakly controlled (Andreoletti et al., 2020, Mahapatra et al., 2020, Meijere et al., 4 Mar 2026, Mancastroppa et al., 2021, Loh, 2020).

Future directions identified across the literature are similarly diverse but convergent in spirit. They include optimizing space partitioning and inter-operator scheduling, extending privacy-preserving protocols to roaming or multi-MO subscriptions, formalizing leakage bounds beyond the metric $3$32, hardening against collusion, exploring PSI or PSI-CA as alternatives or complements to SSS and Paillier, compressing or replacing full CCQs when $3$33 is large, distributing processing for nationwide deployments, and analyzing higher-order tracing under more realistic network structure and temporal heterogeneity. A plausible synthesis is that Triplewise Contact Tracing has become a general research direction concerned with exploiting third-order relational information while controlling the computational, epidemiological, and privacy costs that such richer inference inevitably introduces (Andreoletti et al., 2020, Mahapatra et al., 2020, Meijere et al., 4 Mar 2026, Mancastroppa et al., 2021, Loh, 2020).