Drift AoII: Metrics and Dynamics

• Drift Age of Incorrect Information (AoII) is a metric that penalizes both the magnitude of mismatch and its persistence, providing a nuanced measure of data staleness.
• It is modeled via increment-reset processes and one-step drift recursions, which inform threshold-based transmission strategies across diverse communication models.
• Drift AoII analyses are extended through MDP, CMDP, POMDP, and SMDP formulations to optimize error management and resource utilization in systems such as video streaming and remote monitoring.

Age of Incorrect Information (AoII) is a freshness-and-mismatch metric for remote monitoring in which the penalty depends on both whether the monitor is wrong and how long the mismatch persists. In the general formulation, if $X_t$ denotes the true source, $\hat X_t$ its estimate at the receiver, $g(X_t,\hat X_t)\ge 0$ a mismatch function, $f(t-V_t)\ge 0$ a time-dissatisfaction function, and $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$ the last time the receiver was correct, then

$$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$

Equivalently, with

$S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$

one has $\Delta_{\rm AoII}(t)=f(S_t)$. “Drift AoII” (Editor’s term) denotes analyses that focus on the time-evolution of this process: one-step recursions, expected drift conditions, and cycle-based age accumulation under delayed, unreliable, pull-based, push-based, and multi-user communication models (Maatouk et al., 2020, Chen et al., 2023, Cosandal et al., 2024).

1. Formal definition and relation to adjacent metrics

AoII was introduced as an enabler for semantics-empowered communication, a communication paradigm centered around data’s role and its usefulness to the communication’s goal. Its defining feature is that it penalizes both the severity of a mismatch and the duration of that mismatch. In the formulation above, severity is carried by $g(X_t,\hat X_t)$, while duration is carried by the nondecreasing $f(\cdot)$ (Maatouk et al., 2020).

This differs sharply from the standard Age of Information (AoI), for which

$\hat X_t$0

where $\hat X_t$1 is the generation time of the last received packet. AoI grows whenever no fresh packet arrives, regardless of whether the receiver’s estimate is already correct. It also differs from error-based metrics such as

$\hat X_t$2

which vanish when $\hat X_t$3 but do not penalize the duration of a sustained error. AoII combines both aspects (Maatouk et al., 2020).

Several later papers specialize the general definition into increment-reset processes. With zero-one mismatch,

$\hat X_t$4

the instantaneous AoII is often written as

$\hat X_t$5

with

$\hat X_t$6

or, in continuous time,

$\hat X_t$7

These equivalent forms emphasize a common structural property: AoII ages only while the monitor is incorrect and is reset when source and estimate are synchronized again (Bountrogiannis et al., 2023, Cosandal et al., 2024).

2. Drift recursions and increment-reset dynamics

In discrete-time two-state tracking over a random-delay channel, with $\hat X_t$8 and $\hat X_t$9, the AoII process can be written as

$g(X_t,\hat X_t)\ge 0$0

and evolves according to

$g(X_t,\hat X_t)\ge 0$1

Its one-step expected drift is

$g(X_t,\hat X_t)\ge 0$2

The drift is therefore positive as long as $g(X_t,\hat X_t)\ge 0$3, which gives a direct explanation for threshold-based transmission rules (Chen et al., 2023).

For an $g(X_t,\hat X_t)\ge 0$4-ary symmetric Markov source with no transmission, the AoII process satisfies

$g(X_t,\hat X_t)\ge 0$5

so that

$g(X_t,\hat X_t)\ge 0$6

Under HARQ transmission, with retransmission count $g(X_t,\hat X_t)\ge 0$7 and success probability $g(X_t,\hat X_t)\ge 0$8, the drift above threshold becomes

$g(X_t,\hat X_t)\ge 0$9

The induced Markov chain is thus a birth-reset process whose reset probability is controlled by both source dynamics and HARQ reliability (Bountrogiannis et al., 2023).

In the power-constrained unreliable-channel model, the state is $f(t-V_t)\ge 0$0, where $f(t-V_t)\ge 0$1. The sample-by-sample update is

$f(t-V_t)\ge 0$2

and the one-step expected drift under action $f(t-V_t)\ge 0$3 is

$f(t-V_t)\ge 0$4

Here drift depends jointly on the current mismatch magnitude, the accumulated AoII, the channel success probability, and the Markovian source transitions (Chen et al., 2021).

A plausible implication is that “drift AoII” is not a single formula but a family of increment-reset laws. Across models, the common analytical task is to determine when the expected increase due to continued mismatch is outweighed by the expected reset induced by a sampling or transmission action.

3. MDP, CMDP, POMDP, and SMDP formulations

The core single-link AoII control problem is commonly cast as an average-cost Markov decision process. In the foundational rate-constrained discrete-time model, one defines

$f(t-V_t)\ge 0$5

uses the per-stage Lagrangian cost

$f(t-V_t)\ge 0$6

and solves the Bellman equation

$f(t-V_t)\ge 0$7

The differential cost-to-go $f(t-V_t)\ge 0$8 is nondecreasing in $f(t-V_t)\ge 0$9, and for fixed $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$0 an optimal policy is a deterministic threshold policy,

$V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$1

For the constrained problem, the optimal policy is a randomized threshold policy that randomizes between two adjacent thresholds so that the average update rate equals the constraint exactly. The associated parameter search uses binary search on the threshold and bisection on the Lagrange multiplier (Maatouk et al., 2020).

Under an average power constraint for an unreliable channel, the constrained problem is formulated as a CMDP with state space

$V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$2

instant cost $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$3, and Bellman equation

$V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$4

The optimal policy is a deterministic threshold policy parameterized by a non-increasing vector $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$5, and the constrained optimum is achieved by randomizing between two deterministic threshold policies at visits to the regeneration state $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$6 (Chen et al., 2021).

For the HARQ model, the CMDP state is $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$7, where $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$8 is the retransmission count. If $V_t=\max\{u\le t: g(X_u,\hat X_u)=0\}$9, then the optimal action is always $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$0, so the optimal policy is to never transmit. If $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$1, then for each $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$2 there is a unique threshold $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$3, and the constrained optimum is a randomized mixture of two discrete threshold-based policies that randomize on at most one state (Bountrogiannis et al., 2023).

In random-delay systems, the control problem is again an MDP. Under an easy-to-verify condition, the optimal decision is to initiate a transmission whenever the channel is idle and AoII is not zero, equivalently a threshold policy with $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$4. When the condition is not invoked directly, relative value iteration on a truncated state-space provides a computable approximation sequence to the true optimal policy (Chen et al., 2022, Chen et al., 2023).

A more general extension allows an arbitrary nonnegative function $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$5 of the current AoII conditioned on the current estimate $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$6, leading to a push-based SMDP over synchronization cycles. In that setting the out-of-sync interval is modeled by a dual-regime absorbing Markov chain (DR-AMC), the absorption time is a dual-regime phase-type (DR-PH) distribution, and standard SMDP policy-iteration is used to optimize multi-threshold policies $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$7 (Cosandal et al., 14 Apr 2025).

4. Delay, continuous-time, and coding-based drift models

Random transmission delay changes the AoII analysis because decisions are made when the channel is idle, but the receiver state changes only after service completion. For a two-state Markov source with threshold $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$8, an embedded Markov chain on the AoII at idle epochs has transition probabilities $$\Delta_{\rm AoII}(t)=f(t-V_t)\times g(X_t,\hat X_t).$$9. Although the AoII state space is infinite, a finite-dimensional reduction is possible: $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$0 together with the tail sum

$S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$1

satisfy a closed system of $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$2 linear equations, supplemented by the normalization

$S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$3

Closed-form expressions are available for $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$4 and $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$5, and the calculations show that there are better strategies than the transmitter constantly transmitting new updates (Chen et al., 2022).

The related optimization problem with bounded delay and discard-after-threshold service yields the same central lesson. Threshold policies can be analyzed exactly by a Markov chain on $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$6, where $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$7 is AoII, $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$8 is time in service, and $S_t=(t-V_t)\,g(X_t,\hat X_t)\in\mathbb N,$9 is the channel state. Relative value iteration on truncated state-spaces then provides computable policies, while policy improvement shows that the threshold-$\Delta_{\rm AoII}(t)=f(S_t)$0 rule is optimal under a simple condition that is verified numerically over a wide range of $\Delta_{\rm AoII}(t)=f(S_t)$1 (Chen et al., 2023).

For continuous-time Markov chains, push-based and pull-based sampling are analyzed through synchronization points. If $\Delta_{\rm AoII}(t)=f(S_t)$2 denotes synchronization at state $\Delta_{\rm AoII}(t)=f(S_t)$3, the sequence of synchronization points forms an embedded DTMC with transition matrix $\Delta_{\rm AoII}(t)=f(S_t)$4 and stationary vector $\Delta_{\rm AoII}(t)=f(S_t)$5. The out-of-sync interval started at $\Delta_{\rm AoII}(t)=f(S_t)$6 is modeled by an absorbing CTMC with generator

$\Delta_{\rm AoII}(t)=f(S_t)$7

so that the inter-synchronization time $\Delta_{\rm AoII}(t)=f(S_t)$8 is phase-type. The key cycle quantities are

$\Delta_{\rm AoII}(t)=f(S_t)$9

and the long-run average AoII is

$g(X_t,\hat X_t)$0

In push-based operation, thresholds $g(X_t,\hat X_t)$1 produce an explicit rate–AoII trade-off through KKT conditions (Cosandal et al., 2024).

VLSF coding introduces a different source of drift control: the feedback sequence itself. In that model the MDP state is $g(X_t,\hat X_t)$2, where $g(X_t,\hat X_t)$3 is current AoII, $g(X_t,\hat X_t)$4 is total symbols sent in the current transmission, and $g(X_t,\hat X_t)$5 is symbols sent since the last feedback. The long-run average-cost equation is

$g(X_t,\hat X_t)$6

and the feedback intervals $g(X_t,\hat X_t)$7 are read off from the optimal stationary policy. The paper shows that a lower average delay does not necessarily correspond to a lower average AoII, and periodic feedback can outperform the delay-optimal policy when the feedback schedule better matches source dynamics and obsolete-sample probabilities (Bountrogiannis et al., 2024).

5. Belief-state drift under partial observation and distributed access

When the monitor does not observe the source directly, AoII control becomes a partially observed problem. In multi-user real-time tracking of Markov remote sources, the scheduler maintains for each user a belief

$g(X_t,\hat X_t)$8

which evolves according to

$g(X_t,\hat X_t)$9

The resulting scheduling problem is formulated as a POMDP, relaxed via a subsidy for passivity, and decomposed into single-user average-cost MDPs. The single-arm problem has a threshold policy, is indexable, and admits closed-form Whittle indices for MAoII, yielding a scheduler of complexity $f(\cdot)$0 per slot (Kriouile et al., 2021).

A pull-based remote estimation system with non-zero packet transmission times requires a richer sufficient statistic: the joint age-state belief

$f(\cdot)$1

The monitor then uses the MAP estimator

$f(\cdot)$2

and solves a belief-MDP with instantaneous Lagrangian cost

$f(\cdot)$3

Two belief-dependent policies are developed: a threshold policy based on the instantaneous expected AoII,

$f(\cdot)$4

and a deep $f(\cdot)$5-learning policy using a feedforward network with two hidden layers of 60 units each and target-network updates every $f(\cdot)$6 steps (Cosandal et al., 2024).

In distributed medium access, the DELTA protocol uses local AoII and beliefs about other nodes’ AoII inferred from public acknowledgments. For node $f(\cdot)$7,

$f(\cdot)$8

and, during an anomalous state,

$f(\cdot)$9

If node $\hat X_t$00 transmits with probability $\hat X_t$01, then its slot-success probability is

$\hat X_t$02

and the expected one-step drift is

$\hat X_t$03

The belief-threshold test

$\hat X_t$04

determines whether a node transmits in the BT phase, thereby shaping the drift through distributed belief processing. A sufficient negative-drift condition is $\hat X_t$05 whenever $\hat X_t$06 (Chiariotti et al., 2024).

6. Comparative properties, applications, and recurring design lessons

A central theoretical comparison shows that, for the adopted source model and linear dissatisfaction $\hat X_t$07, the AoII-optimal policy is also error-optimal, whereas the converse is not necessarily true. In particular,

$\hat X_t$08

This formalizes the distinction between minimizing the probability of being wrong and minimizing the time-weighted persistence of wrongness (Maatouk et al., 2020).

The application studies are correspondingly heterogeneous. In video streaming over fading channels, with channel state $\hat X_t$09 and an $\hat X_t$10 matching the standard H.264 distortion-propagation model, the AoII-optimal policy yields up to $\hat X_t$11–$\hat X_t$12 lower average distortion than AoI-optimal or error-optimal policies. In machine overheating, with Weibull penalty

$\hat X_t$13

the AoII-optimal policy reduces long-run breakdown risk by $\hat X_t$14–$\hat X_t$15 versus AoI/error policies. In fire-monitoring dispatch, with

$\hat X_t$16

the AoII-optimal policy cuts average damage by $\hat X_t$17–$\hat X_t$18 (Maatouk et al., 2020).

Several misconceptions are explicitly contradicted by later work. Constantly transmitting new updates is not generally optimal when delay is nontrivial; better threshold strategies exist (Chen et al., 2022). Lower average delay does not necessarily correspond to lower average AoII, because feedback times and obsolete-sample probabilities alter reset opportunities in ways that average delay alone does not capture (Bountrogiannis et al., 2024). Communication can even be inadequate in reducing AoII: for the $\hat X_t$19-ary symmetric HARQ model, if $\hat X_t$20, the optimal policy is to never transmit (Bountrogiannis et al., 2023).

Belief-aware control also yields measurable gains. In the pull-based belief-MDP, MAP estimation reduces average AoII by up to $\hat X_t$21 at modest sampling rates relative to martingale estimation, while belief-dependent threshold or DRL sampling outperforms uniform or random baselines by another $\hat X_t$22. In DELTA, distributed belief processing is significantly more efficient and robust than classical random access and outperforms state-of-the-art scheduled schemes by at least $\hat X_t$23, even with imperfect feedback (Cosandal et al., 2024, Chiariotti et al., 2024).

Taken together, these results suggest a consistent interpretation of drift-centered AoII analysis. The decisive object is not freshness alone, but the stochastic geometry of desynchronization intervals: how quickly AoII grows while incorrect, how likely control is to trigger a reset, and how channel delay, coding, sampling constraints, retransmissions, and partial observability reshape that increment-reset balance.