Find the mechanism before the formula
A monitor can receive every update and still act on stale information. Age of information measures how old the newest delivered observation is, so it exposes a queueing question that throughput and packet delay alone cannot answer.
We will draw the age sawtooth from generation and delivery times, calculate time-average and peak age, compare two service settings, and explain why generating updates faster can eventually make freshness worse.
This guide belongs to the Gold Standard of Queueing Theory, the Gold Standard of Education and the How X Works hub.
Choose a reading route
Start with the question closest to the decision you need to make. Every route returns to the same worked evidence and transfer checks.
| Route | Chapters | Use it when |
|---|---|---|
| 1 | 1–3 | You need the timestamp model and the age path. |
| 2 | 4–5 | You want average and peak-age calculations. |
| 3 | 6–8 | You need the queueing trade-off and changed-context check. |
| 4 | 9–10 | You must choose disciplines and respect model limits. |
| 5 | 11–12 | You want FAQs and a practical evidence checklist. |
Chapter contents
Open the group you need, then use the chapter links. Teaching sections remain open on the page.
Build the model
Test the mechanism
1. Freshness begins with two timestamps
Let update i be generated at time gi and delivered at time di. At time t, let u(t) be the generation time of the freshest update delivered by t. The age of information is Δ(t)=t−u(t).
Age therefore uses the source timestamp, not merely time since the last delivery. A packet generated long ago can arrive now and still leave a large age. Sequence numbers are useful for rejecting obsolete deliveries, but they do not replace a trustworthy clock or generation timestamp.
- Record generation and delivery time in one time base.
- Define what “freshest” means when packets arrive out of order.
- State whether discarded or failed updates enter the metric.
- Keep source sampling separate from network transmission.
2. Read the sawtooth before taking an average
Between deliveries, u(t) stays fixed while t advances, so age rises with slope one. When a newer update arrives, age drops to its system time di−gi. Repeating rises and drops produces the familiar sawtooth path.
The height just before a successful age drop is a peak age. A late packet that is older than the current information causes no drop if the monitor keeps only the freshest timestamp. This is why delivery order and packet-management rules must appear in the model.
A graph of Δ(t) is often more diagnostic than one mean: it shows long stale plateaus, irregular service, and deliveries that did not refresh the monitor.
3. Time-average age is an area calculation
Over an observation interval [0,T], the time-average age is (1/T)∫0TΔ(t)dt. For a sample path, divide the sawtooth into triangles and trapezoids, add their areas, and include the opening and closing partial pieces.
Peak age instead averages selected maxima immediately before informative deliveries. The two summaries answer different questions. Average age measures the stale exposure across time; peak age focuses on recurring worst points but can miss a long rare excursion if the sampling rule is not stated.
| Metric | Uses | Can miss |
|---|---|---|
| Mean delay | One packet’s delivery lag | Time without a fresh delivered state |
| Average age | Area under Δ(t) | Tail shape without another summary |
| Peak age | Pre-drop maxima | Non-peak exposure and rare censoring |
| Violation probability | Time Δ(t) exceeds a threshold | Magnitude above the threshold |
4. Worked example: four-second sampling and one-second service
Suppose updates are generated at times 0, 4 and 8 seconds. Each needs one second of service, so deliveries occur at 1, 5 and 9. Immediately after each delivery, age is 1 second. Just before the next delivery it has risen to 5 seconds.
From t=1 to t=5 the age rises linearly from 1 to 5. The area is 4(1+5)/2=12 second-squared, so the cycle-average age is 12/4=3 seconds. The peak age for the cycle is 5 seconds.
The result is not the one-second service delay. It includes the four-second interval during which the delivered observation gets older. Changing the observation window or initial condition requires adding the boundary pieces explicitly.
5. Peak age is tied to intergeneration time and delay
For consecutive informative updates delivered in order, the peak just before update i arrives can be written as di−gi−1. It contains the next update’s system time plus the gap between the two generation times.
This identity explains why a low packet delay does not by itself guarantee a low peak age: a source that samples rarely leaves a long generation gap. It also explains why a high sampling rate can fail, because queued updates increase system time and may arrive already stale.
Report whether peaks are measured before every delivery or only before deliveries that lower age. With obsolete packets, those definitions produce different data sets.
6. More updates can make information older
At a very low generation rate, the server is often idle and age grows while no new sample exists. Raising the rate initially improves freshness. Near capacity, however, waiting grows and delivered packets carry older source timestamps. The age curve can therefore have an interior minimum.
Kaul, Yates and Gruteser showed this distinction in an FCFS queue: the rate that minimises age need not maximise throughput or minimise delivery delay. The design variable is not simply “send as fast as possible”.
The optimum depends on service-time variability, discipline, packet management, source correlation and the chosen age penalty. A rate found for M/M/1 is not a universal setting for a sensor network.
7. Transfer check: service takes three seconds
Keep generation every four seconds but change deterministic service from one second to three. Deliveries occur at 3, 7 and 11 seconds, so age drops to 3 after each delivery and rises to 7 just before the next.
The cycle-average age becomes 4(3+7)/2 divided by 4, which is 5 seconds. Peak age becomes 7 seconds. The same geometric method transfers, while both the drop height and the peak change with service time.
If service exceeded the four-second generation interval, a FIFO backlog would accumulate. Then the simple repeating cycle would disappear and each delivery would need its own generation and completion timestamps.
8. Zero-wait is not always age-optimal
A zero-wait source generates a new update as soon as the previous one finishes service. This removes deliberate idle time and often looks obviously fresh. Yet Sun and colleagues showed that zero-wait can be suboptimal under general service distributions and age penalties.
Waiting can avoid sampling at an unfavourable moment or reduce the effect of highly variable and correlated service. That does not mean waiting is always better. It means the policy must be derived from the service process and information objective rather than from throughput intuition.
A changed-context test should compare at least zero-wait, periodic and threshold-based policies on the same timestamp trace, using the same observation window and boundary treatment.
9. Queue discipline changes which information survives
FCFS preserves order but may deliver a train of obsolete updates after congestion. Last-come-first-served with preemption can favour the newest sample, while replacement buffers keep only one waiting update. Each rule changes the age path and may waste different amounts of service.
Preemption is attractive when an interrupted update has little residual value and service can restart cheaply. It can be harmful when almost-complete work is expensive to discard. Multiple sources add fairness questions: one aggressive stream can keep its own age low while starving another.
- Name the buffer size and replacement rule.
- State whether service is preempted or only waiting packets are replaced.
- Measure each source, not just a pooled mean.
- Include completion cost, energy and lost work where relevant.
10. Limits and misconceptions
- Age is not the same as packet delay, inter-delivery time or clock time since receipt.
- A lower mean age does not guarantee fewer extreme stale intervals.
- Peak-age definitions must say which deliveries count.
- Faster generation can increase age when it creates waiting.
- Clock skew and timestamp errors can corrupt the metric.
- A policy optimal for one source and one server may be unfair or unstable in a network.
Age is a decision metric built on a model of information value. If an old update remains useful or different variables decay at different rates, use a weighted or nonlinear age penalty and explain the choice.
11. Frequently asked questions
Can age ever decrease without a delivery?
Not in the standard definition. Between informative deliveries the freshest generation time is fixed, so age increases at unit rate. A model that predicts another slope is measuring a different quantity.
Is peak age the maximum age ever seen?
Usually not. Peak age commonly means the value just before an informative delivery. A sample maximum over an observation window is a separate extreme-value statistic.
Should stale packets always be dropped?
Not automatically. Dropping can improve freshness, but packets may carry reliability, audit or cumulative information. The application must state whether newer data supersede older data.
12. A reproducible freshness checklist
| Stage | Evidence to retain |
|---|---|
| Timestamps | Generation, delivery and clock basis |
| Path | Age drops and obsolete arrivals |
| Metric | Average, peak and threshold definition |
| Queue | Service law, buffer and discipline |
| Policy | Sampling, waiting and replacement rule |
| Validation | Window boundaries, uncertainty and per-source results |
The gold standard is to reconstruct the sawtooth first, calculate the stated metric second, and only then compare policies. That order makes it difficult to mistake a short delivery delay for genuinely fresh information.
Reading routes · Previous chapter · Contents · Continue reading
Continue through the Gold Standard ecosystem
Return to the Gold Standard of Queueing Theory, the Gold Standard of Education parent, or the How X Works hub.
Read next in this subject
- Time Averages and Arrival Averages Under Nonhomogeneous Arrivals
- When Poisson Arrivals See Time Averages
- Priority Rules, Class Waits and Hidden Trade-Offs
- Mean-Field Limits, Propagation of Chaos and Large Service Systems