Structural Freshness

• Structural Freshness is defined as a measure where information’s timeliness is determined by structural configurations—such as network topology, team collaboration graphs, and cache architectures—rather than solely by elapsed time.
• It is applied across various fields including team science, wireless systems, remote estimation, and memory caching, where structural factors like new collaboration ties or topology-dependent scheduling directly impact performance and reliability.
• Empirical findings indicate diverse optimal thresholds, with team collaboration freshness peaking around 29.72% for citation impact in small teams and wireless system designs showing up to 200-fold improvements in maintaining freshness under structural policies.

Searching arXiv for recent and foundational papers on “structural freshness” and related formulations. Structural freshness denotes a family of freshness formulations in which timeliness is determined by system structure rather than by elapsed time alone. Across the literature, the term does not have a single invariant meaning. In team science, it refers to new collaboration ties inside a team; in gossip and wireless systems, it refers to topology-dependent synchronization or age processes; in cache and secure-memory systems, it refers to write-driven maintenance or replay-resistant version structures; and in remote estimation, it refers to estimator architectures that determine whether information remains correct between queries. A plausible synthesis is that structural freshness characterizes how relational, topological, architectural, or policy structure shapes whether current information remains useful, deliverable, or correct (Liu et al., 2022, Bastopcu et al., 2021, Dong et al., 2024, Liyanaarachchi et al., 29 Jan 2026).

1. Conceptual scope and boundary conditions

Structural freshness is distinct from formulations in which freshness is purely temporal. The clearest counterexample appears in context-aware recommendation. "Freshness-Aware Thompson Sampling" defines freshness as “the strength of strangeness or the amount of forgotten experience,” operationalized through time elapsed since the last click and the number of previous clicks, using the memory-retention term

$Mr(d) = e^{-\frac{t(d)}{rsm(d)}}$

with $t(d)$ the time since the last click and $rsm(d)$ the number of times the document has been clicked (Bouneffouf, 2014). The same paper states explicitly that this freshness mechanism is temporal, grounded in recency of interaction and frequency of past clicks, not in document structure.

A second non-structural usage appears in session-based recommendation freshness. "Feedback-based Approach to Introduce Freshness in Recommendations" treats freshness as “diversity in the time domain,” measured by overlap between the current recommendation set and the set of products recommended in the past $k$ calls: $Freshness = \frac{|R \setminus A_k|}{t}$ where $R$ is the current recommendation set, $A_k$ the set of products recommended in the previous $k$ calls, and $t$ the number of recommendations served in the current request (Malladi et al., 2016). This is temporal diversity across sessions rather than a structural property of the underlying items or model.

By contrast, several later works use structural freshness to refer to network topology, collaboration structure, estimator structure, or freshness-maintenance architecture. This suggests that the term is best understood as domain-specific: it designates structural determinants of freshness, but the relevant structure may be social, algorithmic, graph-theoretic, architectural, or control-theoretic (Boekhout et al., 16 Jul 2025, Srivastava et al., 25 Apr 2025, Liyanaarachchi et al., 26 Jan 2026).

2. Scientific teams: freshness as collaboration structure

In team science, structural freshness is operationalized as the extent to which a team is composed of new collaboration ties rather than repeated prior ties. "Team formation and team performance: The balance between team freshness and repeat collaboration" treats a paper’s byline as a scientific team and classifies an author pair as a new collaboration relation when the two authors have not collaborated in the previous 20 years. Team freshness is then the fraction of missing links in the team’s prior collaboration network that are created by the focal paper. For a team of $n$ authors, the denominator is all possible author pairs $t(d)$0, and the paper illustrates the measure with an example in which 4 out of 10 possible author pairs are new, yielding freshness $t(d)$1 (Liu et al., 2022).

This measure is explicitly structural: it does not count merely whether a team contains newcomers, but whether the dyadic collaboration graph among team members is new. High freshness means more unfamiliar ties and more new relational structure; low freshness means more embeddedness in prior ties and more repeat collaboration. The paper emphasizes that both sides have advantages and liabilities. Freshness can bring new knowledge, broader search space, new perspectives, and conflict that may stimulate creativity, whereas excessive freshness can bring coordination costs, lack of trust, adaptation costs, and communication problems. Repeat collaboration can produce efficiency, trust, reciprocity, and smoother knowledge transfer, but too much repeat collaboration can cause myopic learning, homogeneity, reduced search breadth, and less creative abrasion (Liu et al., 2022).

Using more than 43 million papers from 1950–2018, the study reports that team freshness increased over time, with freshness growing from 1970 to 2000, remaining roughly stable in the 2000s, and increasing again after 2010. By 2018, average freshness reached about 0.443. The proportion of papers with freshness $t(d)$2 declined sharply, while the proportion with freshness $t(d)$3 remained below about 20%. The paper interprets this as evidence that science is not moving toward pure novelty in collaboration structure, but toward a balance of fresh and repeated ties (Liu et al., 2022).

The association with performance is non-monotone. Using $t(d)$4 and $t(d)$5 as dependent variables and an ordinary least squares specification with freshness and freshness squared, the paper finds an inverted-U-shaped relationship between freshness and citations for both short-term and long-term impact. For $t(d)$6, the turning point is reported as 29.72% freshness. The same study reports that the optimal freshness varies by discipline, from 8.83% in mathematics to 45.91% in history, and that the team-size moderation is substantial: teams with fewer than 10 authors exhibit the inverted-U-shaped relationship, while teams with more than 9 authors show a positive linear relationship between freshness and citations (Liu et al., 2022).

A later network-driven study extends the structural argument from dyadic novelty inside a single team to overlap among persistent teams. "Freshness, Persistence and Success of Scientific Teams" identifies persistent teams as temporal maximal cliques in a persistent collaboration network, where an author pair is deemed persistently collaborating when it co-authors at least 3 publications within a 5-year period. Structural freshness is then associated with freshness impulses: overlapping collaborations that arise during or after the focal team’s formation and inject new collaborative ties into the team’s persistent structure. The paper distinguishes persistence impulses, freshness impulses, and synchronous impulses according to the timing of overlap relative to the focal team (Boekhout et al., 16 Jul 2025).

The study identifies 10,232,084 scientific teams, reports an average team duration of 4.37 years, and states that teams are responsible for over 70% of global scientific output. It further reports that the probability that a publication becomes top 1% cited decreases continuously after the second year of a team’s lifespan, that many successful teams obtain their first highly cited paper in year 1, and that open teams with impulses are more likely than closed teams to produce highly cited work. Persistence impulses are associated with earlier success, whereas freshness impulses are associated with later success, especially for top 1% outcomes (Boekhout et al., 16 Jul 2025). A plausible implication is that structural freshness in team science is not simply a static team-composition variable, but a dynamic property of how teams remain open to recombination while retaining some continuity.

3. Communication networks: freshness as topology, age structure, and synchronization state

In communication systems, structural freshness is often expressed through age-based or correctness-based state variables whose evolution depends on topology. "Relative Age of Information: Maintaining Freshness while Considering the Most Recently Generated Information" introduces Relative AoI for systems in which each newly arrived message obsoletes all previous messages. Classical AoI is

$t(d)$7

while Relative AoI is defined as

$t(d)$8

The identity

$t(d)$9

shows that AoI combines two effects: the lag between the latest arrival and the latest delivered arrival, and the age of the latest arrival itself. Relative AoI isolates the first effect (Kesidis et al., 2018).

For the GI/GI/1/1-PO queue, the paper derives the average Relative AoI

$rsm(d)$0

the busy-cycle length

$rsm(d)$1

and the relation

$rsm(d)$2

The paper states that Relative AoI is a structural freshness metric because it captures the time during which a newer-but-not-yet-delivered message is present in the system. In D/M/1/1-PO and M/M/1/1-PO, it reports $rsm(d)$3 (Kesidis et al., 2018).

A sharper structural formulation appears in gossip networks. "Gossiping with Binary Freshness Metric" defines binary freshness at node $rsm(d)$4 by

$rsm(d)$5

and for a set $rsm(d)$6,

$rsm(d)$7

The steady-state average freshness is

$rsm(d)$8

Using a stochastic hybrid systems recursion, the paper shows that the topology of the network strongly determines scaling laws. In a disconnected network, the average freshness of a single node is

$rsm(d)$9

so freshness decreases as $k$0. In a ring network, the scaling remains $k$1, with asymptotic constant $k$2. In a fully connected network, the scaling depends on $k$3: it is $k$4 when $k$5, $k$6 when $k$7, and $k$8 when $k$9 (Bastopcu et al., 2021). The structural message is explicit: more connectivity yields better synchronization with the source.

"Information Freshness in Dynamic Gossip Networks" extends the topological argument to switching graphs using version age,

$Freshness = \frac{|R \setminus A_k|}{t}$0

with long-term average

$Freshness = \frac{|R \setminus A_k|}{t}$1

When the topology switches between two graph sequences $Freshness = \frac{|R \setminus A_k|}{t}$2 and $Freshness = \frac{|R \setminus A_k|}{t}$3 according to a two-state CTMC with holding-time scale $Freshness = \frac{|R \setminus A_k|}{t}$4, and the two static topologies have average version age scales $Freshness = \frac{|R \setminus A_k|}{t}$5 and $Freshness = \frac{|R \setminus A_k|}{t}$6 with $Freshness = \frac{|R \setminus A_k|}{t}$7, the paper proves that if $Freshness = \frac{|R \setminus A_k|}{t}$8, then the varying-topology network has average version age $Freshness = \frac{|R \setminus A_k|}{t}$9. It further motivates the notion of a typical set of nodes because a vanishingly small number of atypical nodes can disproportionately worsen the network-wide average (Srivastava et al., 25 Apr 2025). This moves structural freshness from static topology to the interaction of topology and topology-switching timescale.

Wireless-system work often uses AoI as the operative freshness state but still frames freshness structurally. "WiFresh: Age-of-Information from Theory to Implementation" defines

$R$0

and the expected network AoI as

$R$1

The central claim is that freshness depends on system structure: FCFS queues and random access cause stale information under congestion, whereas LCFS queues, polling multiple access, and Max-Weight scheduling preserve freshness. The scheduling index is

$R$2

and the selected source is

$R$3

The paper reports that WiFresh can improve information freshness by two orders of magnitude compared with standard WiFi and gives several quantified comparisons, including improvement factors of at least 200 over equivalent standard WiFi for high packet rates in a ten-source network (Kadota et al., 2020).

Other wireless work makes the structural dependence more explicit through policy taxonomies or geometric interference models. "Optimizing Information Freshness in Wireless Networks: A Stochastic Geometry Approach" analyzes peak AoI in a PPP network with local-information scheduling, and derives a fixed-point characterization of the locally adaptive access probability $R$4 based on a stopping set $R$5 (Yang et al., 2020). "Fresh Multiple Access: A Unified Framework Based on Large Models and Mean-Field Approximations" embeds AoII and peak AoII into a tandem reservation/transmission queue and uses large sparse Markov chains and mean-field approximations to analyze freshness-oriented multiple access (Hui et al., 2023). "A Survey of Freshness-Aware Wireless Networking with Reinforcement Learning" organizes freshness measures into native, function-based, and application-oriented families and classifies RL policies into update-control RL, medium-access RL, risk-sensitive RL, and multi-agent RL (Alibotaiken et al., 24 Dec 2025). These works collectively suggest that, in wireless systems, structural freshness is inseparable from queue discipline, topology, access control, interference geometry, and the policy layer at which freshness is optimized.

4. Caches and memory systems: freshness as maintenance architecture and replay resistance

In caching, structural freshness appears when freshness guarantees are determined by the architecture used to maintain coherence with changing backend state. "Caching under Content Freshness Constraints" studies a request process in which each request for object $R$6 carries a freshness specification $R$7, interpreted as the maximum acceptable age of the cached content. A request is a hit only if the object is present and its age is at most $R$8. The steady-state hit rate for object $R$9 is $A_k$0, and the overall hit probability is

$A_k$1

The paper derives the universal upper bound

$A_k$2

with total upper bound

$A_k$3

It then analyzes LP, LRU, M-LP, M-LRU, and LEH, with LEH prioritizing contents by expected future hits

$A_k$4

for a cached item that has already resided in cache for $A_k$5 slots (Poojary et al., 2017). Here freshness is structural because stale contents are treated as cache redundancies and the replacement policy itself must reason about freshness slack.

A more explicit architecture-level formulation appears in "Revisiting Cache Freshness for Emerging Real-Time Applications". The paper argues that TTLs are inadequate for real-time freshness because, in a lazy/cache-aside design, writes bypass the cache and freshness maintained solely by time-based expiry becomes prohibitively costly as the bound $A_k$6 shrinks. It defines a freshness cost $A_k$7 and a staleness cost $A_k$8, with

$A_k$9

For TTL-expiry,

$k$0

For TTL-polling,

$k$1

The paper then proposes write-driven updates and invalidations. For updates,

$k$2

and for invalidation,

$k$3

Its adaptive choice between update and invalidate is governed by the threshold

$k$4

which simplifies as $k$5 to

$k$6

The paper explicitly states that a structurally fresh cache should coordinate with the backend on writes, send updates or invalidates only for keys that actually changed, and choose between those actions based on read/write mix and system bottlenecks (Mao et al., 2024).

In trusted memory, freshness becomes a replay-resistance property of the metadata structure. "Toleo: Scaling Freshness to Tera-scale Memory using CXL and PIM" defines freshness as the guarantee that an adversary cannot replay an old value in response to a memory read request. Conventional schemes maintain a version number for each cache block and protect those versions with a Merkle tree, but the paper argues that the resulting metadata access overhead becomes prohibitive with protected memory size. Toleo replaces the Merkle-tree root with trusted smart memory connected via CXL IDE, stores versions in that trusted device, and removes the need for an authenticated tree over version metadata (Dong et al., 2024).

The paper further separates freshness and nonce roles via a 64-bit full version split into 37 high-order bits of UV and a 27-bit stealth version, stores only the 27-bit stealth version in trusted memory, and uses page-granularity Trip compression. The three page representations are flat, uneven, and full, with data-to-version ratios of 1:341, 1:60, and 1:18, respectively; 92% of pages use flat format; and the average data-to-version ratio across workloads is 240:1. The paper states that one 168 GB Toleo smart memory device can provide freshness to a 28 TB CXL-expanded main memory pool with negligible performance overhead (Dong et al., 2024). In this setting, structural freshness means that replay protection scales because the structure of version storage and trust anchoring has changed, not because the system is polling more aggressively in time.

5. Remote estimation and control: freshness as estimator structure and state-space geometry

Remote-estimation work uses correctness-based freshness, but shifts the locus of optimization from update timing alone to estimator structure. "Multi-Stage Structured Estimators for Information Freshness" and "Beyond Martingale Estimators: Structured Estimators for Maximizing Information Freshness in Query-Based Update Systems" define binary freshness by

$k$7

with mean binary freshness

$k$8

The baseline martingale estimator is

$k$9

where $t$0 is the latest update time. Both papers argue that martingale estimators are far from optimal in pull-based systems, because they simply hold the most recently received sample even when another state becomes more probable (Liyanaarachchi et al., 26 Jan 2026, Liyanaarachchi et al., 29 Jan 2026).

The main constructive object is the $t$1-MAP estimator, a piecewise-constant multi-stage approximation to MAP. If the latest sample is state $t$2, the age $t$3 is partitioned by

$t$4

and the estimator uses stage value $t$5 when $t$6. One formulation is

$t$7

where $t$8. The stage label is chosen by

$t$9

with $n$0 (Liyanaarachchi et al., 26 Jan 2026).

For time-reversible CTMCs, both papers state that the MAP estimator is piecewise constant with finitely many transition points. This yields the principal structural result: the $n$1-MAP estimator can exactly represent MAP when the thresholds are chosen as the actual MAP transition points (Liyanaarachchi et al., 29 Jan 2026). The associated freshness decomposition is

$n$2

and the paper formulates constrained optimization of state-dependent sampling rates as

$n$3

for a single source, and weighted rate-allocation problems for multiple heterogeneous CTMCs (Liyanaarachchi et al., 26 Jan 2026, Liyanaarachchi et al., 29 Jan 2026). The structural element is therefore the estimator’s staged internal evolution between queries.

A control-theoretic counterpart appears in "Status Updating via Integrated Sensing and Communication: Freshness Optimisation". The paper uses a two-dimensional AoI state

$n$4

where $n$5 is the AoI at the source and $n$6 is the AoI at the base station. The stage cost is

$n$7

and the objective is the discounted infinite-horizon problem

$n$8

The paper proves that the optimal stationary deterministic policy has threshold form

$n$9

for a nondecreasing integer-valued switching curve $t(d)$00 (Soleymani et al., 30 Jan 2026). In this setting, structural freshness is the geometry of the optimal decision boundary in the two-dimensional AoI state space.

6. Interpretation, misconceptions, and recurrent themes

A recurring misconception is that freshness is always reducible to recency. The cited literature shows otherwise. Relative AoI depends on whether a fresher message has arrived but not yet been delivered (Kesidis et al., 2018). Binary freshness depends on whether a node has the latest source version, not on how long ago a sample was generated (Bastopcu et al., 2021). Team freshness depends on missing links in prior collaboration networks, not on when team members last published (Liu et al., 2022). Toleo’s freshness guarantee is replay resistance, enforced by version metadata and trust placement rather than by time since last access (Dong et al., 2024). Structured estimators maximize the fraction of time the estimate is exactly correct, which can improve even when query timing is unchanged (Liyanaarachchi et al., 29 Jan 2026).

A second misconception is that structural freshness is uniformly beneficial. Team-science evidence contradicts that view. The relation between team freshness and citations is inverted-U-shaped in small teams, and only in large teams does the relationship become positive linear (Liu et al., 2022). The persistent-team study similarly reports that the gains from impulses eventually flatten or reverse because of coordination costs (Boekhout et al., 16 Jul 2025). Cache work makes the same point in a different register: aggressive TTL reduction can technically improve freshness bounds while making overhead prohibitive, which is why write-reactive policies are proposed instead (Mao et al., 2024).

A third recurrent theme is that structure and dynamics interact. In dynamic gossip networks, switching rate relative to intrinsic freshness scales determines whether the network behaves like the better or worse topology (Srivastava et al., 25 Apr 2025). In ISAC, the decision to sense or communicate depends jointly on source AoI and base-station AoI (Soleymani et al., 30 Jan 2026). In wireless networks, local topology, queue state, and access policy jointly determine peak AoI (Yang et al., 2020). This suggests that structural freshness is rarely a purely static property; it is usually a property of structured dynamics.

Taken together, the literature supports a broad but disciplined characterization. Structural freshness is freshness mediated by configuration: the graph of prior collaborations, the topology of a gossip network, the policy architecture of a wireless system, the organization of cache maintenance, the trust structure of memory metadata, or the staged form of an estimator. Where freshness is purely recency- or memory-based, as in the recommender-system papers, the relevant work explicitly falls outside that structural usage (Bouneffouf, 2014, Malladi et al., 2016).