A queue in heavy traffic is not simply a busy queue. It is a sequence of systems whose offered load approaches available capacity under a declared scaling. The raw queue may grow while its rescaled workload approaches a non-degenerate stochastic process.
Diffusion approximations replace many small net-input changes by continuous Brownian fluctuations, with reflection preventing negative workload. In multiclass networks, state-space collapse can make a high-dimensional queue vector behave, after scaling, as though it lies near a lower-dimensional workload surface.
The Gold Standard is to state the asymptotic regime, prove or test the collapse relation, keep the lifting map visible, compare finite-system data with the approximation and refuse transfer when scheduling, routing, moments or bottlenecks fall outside the theorem.
A practical orientation
Use this guide when an exact queueing model is unwieldy but the system operates close to one or more bottlenecks. The worked example uses two customer classes, one pooled resource, 15-minute observations and a declared workload-lifting map.
Choose a reading route
Work and transfer the example
Chapters 7–9 calculate a two-class case and change load and discipline.
Chapter contents
- 1. Define heavy traffic as a sequence, not an adjective
- 2. Centre cumulative input before taking a diffusion scale
- 3. Use reflection to enforce the queue boundary
- 4. See why a queue vector can have fewer effective dimensions
- 5. Make the lifting map explicit
- 6. Distinguish multiplicative and diffusion-scale collapse
- 7. Worked example: two classes near one bottleneck
- 8. Build the one-dimensional diffusion approximation
- 9. Transfer check: change load, variability and discipline
- 10. Compare the approximation with the finite queue
- 11. State the limits and preserve the evidence
- 12. FAQs and the next reading route
1. Define heavy traffic as a sequence, not an adjective
Let system n have arrival and service parameters that may change with n. Heavy traffic requires a precise limiting statement, commonly that utilisation approaches one while time and space are accelerated and normalised so random fluctuations remain visible. Saying that a single observed system is at 96% utilisation does not by itself establish a theorem.
Record which resources become critical, how fast spare capacity shrinks and what is held fixed. A single-server sequence might satisfy one minus utilisation of order one over the square root of n. A network may instead have several bottlenecks and a workload cone with more than one effective dimension.
- Index the system sequence.
- Name every critical resource.
- State time and space scaling.
- Separate fixed primitives from n-dependent parameters.
- Do not infer a limit from one utilisation number.
2. Centre cumulative input before taking a diffusion scale
Cumulative arrivals and potential service contain a deterministic first-order trend and stochastic fluctuations. Fluid scaling exposes the trend. Diffusion scaling subtracts the nominal trend and magnifies the remaining fluctuations, often by a square-root factor under finite-variance assumptions.
The covariance of the limiting net-input process comes from arrival variability, service variability and routing dependence. Two queues with the same mean load can have different diffusion behaviour because their interarrival or service-time variance differs.
| Object | First-order question | Diffusion question |
|---|---|---|
| Arrivals | What is the mean rate? | How does centred input fluctuate? |
| Service | What capacity is available? | How variable are completions? |
| Routing | Where does mean flow go? | Which fluctuations are correlated? |
| Workload | Is net drift negative? | How does reflection shape delay? |
3. Use reflection to enforce the queue boundary
An unconstrained Brownian approximation can become negative, but workload cannot. The Skorokhod reflection map adds the smallest nondecreasing regulator needed to keep the process in the nonnegative region. In one dimension, the regulator grows only when the reflected workload is at zero.
In a network, the reflection directions and matrix matter. A mathematically valid semimartingale reflecting Brownian motion requires conditions on that geometry; an arbitrary covariance matrix plus a zero floor is not automatically the correct network limit.
| Term | Meaning | Visible effect |
|---|---|---|
| Net-input drift | Capacity margin after centring | Pulls workload down or up |
| Covariance | Joint fluctuation scale | Controls spread and correlation |
| Reflection | Boundary regulator | Prevents negative workload |
| Initial state | Scaled starting condition | Shapes transient behaviour |
4. See why a queue vector can have fewer effective dimensions
A multiclass system may record one queue for every class while only a few resource workloads constrain the near-critical motion. Under suitable service disciplines and stability conditions, fast directions relax and the scaled queue vector stays near a lower-dimensional manifold. This is state-space collapse.
Collapse is a claim about distance after scaling. It does not say that class queues are equal, that class identity is irrelevant or that every snapshot lies exactly on a line. The remaining dimensions can still matter for finite-system performance and protected outcomes.
- High-dimensional queue vector retained.
- Critical workload dimensions identified.
- Distance to the lifting surface measured.
- Fast and slow directions separated.
- Finite residuals not erased.
5. Make the lifting map explicit
Workload alone does not reconstruct class queues unless a lifting rule is supplied. The rule can arise from an optimisation, an invariant manifold or a theorem for the discipline. Write it as a map from workload to the predicted queue vector and keep its assumptions beside it.
If the service discipline changes, the lifting map may change even when aggregate workload does not. A priority rule, first-come-first-served rule and processor-sharing rule need not distribute the same workload across classes.
| Input to lifting map | Output | Assumption exposed |
|---|---|---|
| Resource workload | Class queue vector | Discipline and service rates |
| Workload cone point | Near-invariant state | Critical resources |
| Scaled time | Predicted path | Limit regime |
| Policy version | Allocation surface | Control rule |
6. Distinguish multiplicative and diffusion-scale collapse
Multiplicative collapse controls the queue-vector distance relative to the system magnitude. Diffusion-scale collapse is stronger for a diffusion theorem: after the same normalisation used for the limit, the residual must vanish in the required sense.
A small percentage residual can still be large on diffusion scale. Report both the absolute scaled residual and its ratio to total scaled workload. Do not switch definitions after seeing the data.
| Check | Question | Failure hidden by the other check |
|---|---|---|
| Relative residual | Is error small compared with the state? | Large diffusion-scale error |
| Scaled residual | Does error vanish under theorem scaling? | Large percentage error near zero workload |
| Path supremum | Is the whole interval controlled? | Brief excursions |
| Terminal residual | Is the final snapshot close? | Earlier breakdown |
7. Worked example: two classes near one bottleneck
Class A needs 6 service-minutes per job and class B needs 12. The declared invariant-manifold rule assigns two thirds of pooled workload to A and one third to B. At a workload of 720 service-minutes, the lifted queue vector is therefore 80 A jobs and 20 B jobs.
A 15-minute observation gives 77 A jobs and 22 B jobs. Their workload is 77 times 6 plus 22 times 12, or 726 service-minutes. Applying the lifting rule at 726 predicts 80.7 A and 20.2 B. The residual is minus 3.7 A and plus 1.8 B: close in workload allocation, but not zero.
Across 96 intervals, the root-mean-square scaled residual falls from 0.31 in the first warm-up block to 0.09 after the fast class-mix transient settles. That is evidence consistent with collapse for this dataset, not a proof of the asymptotic theorem.
| Quantity | Observed | Lifted at W = 726 |
|---|---|---|
| Class A jobs | 77 | 80.7 |
| Class B jobs | 22 | 20.2 |
| Total workload | 726 min | 726 min |
| Residual vector | — | (-3.7, +1.8) |
8. Build the one-dimensional diffusion approximation
Estimate net-input drift and variance from primitives or from a model-consistent experiment. Suppose centred 15-minute workload increments have mean minus 8 service-minutes and standard deviation 52. The approximation evolves as a Brownian motion with that drift and variance, reflected at zero.
Simulate or solve the reflected model for workload quantiles, then lift each workload to class queues. Keep parameter uncertainty separate from diffusion randomness. A narrow simulated band based on uncertain drift is falsely precise.
| Input | Worked value | Role |
|---|---|---|
| Mean net increment | -8 service-min | Negative drift |
| Increment standard deviation | 52 service-min | Diffusion scale |
| Boundary | 0 | Reflection |
| Initial workload | 726 service-min | Transient start |
9. Transfer check: change load, variability and discipline
First reduce utilisation from the near-critical regime. The diffusion approximation should become less necessary, and finite-system discreteness may dominate. Next keep the mean load but double arrival variability; workload spread should increase even though utilisation is unchanged.
Finally change the service discipline while retaining the same pooled capacity. The aggregate workload approximation may remain useful, but the old lifting map must not be reused without justification. A successful transfer preserves only the pieces whose assumptions survive.
| Changed context | May remain transferable | Must be rechecked |
|---|---|---|
| Lower utilisation | Primitive workload identity | Heavy-traffic accuracy |
| Higher arrival variance | Mean drift | Covariance and tails |
| New discipline | Resource workload | Lifting map and collapse |
| Second bottleneck | Local primitives | Workload dimension and reflection |
10. Compare the approximation with the finite queue
Use held-out paths and compare workload distributions, delay quantiles, boundary mass, autocorrelation and collapse residuals. A good mean with poor upper quantiles is not an adequate service approximation. Check several traffic intensities to see whether error shrinks in the direction predicted by the limit.
Align initial conditions and observation windows. A stationary diffusion compared with a warming finite queue confounds approximation error with transient mismatch. Preserve random seeds and primitive samples when simulation is part of the check.
- Held-out paths.
- Matched initial conditions.
- Mean, quantiles and boundary mass.
- Autocorrelation and residual paths.
- Several distances from critical load.
11. State the limits and preserve the evidence
Finite-variance Brownian approximations can fail under heavy-tailed primitives, long-range dependence, unstable routing or disciplines outside the theorem. Multiple bottlenecks may require a multidimensional reflected process. Rare events far beyond the diffusion scale can need large-deviation or direct simulation methods.
Preserve primitive distributions, dependence assumptions, scaling sequence, critical-resource calculation, covariance construction, reflection data, lifting map, residual diagnostics, finite-system comparisons and exclusions. A formula without its regime is not reusable evidence.
- No universal heavy-traffic threshold.
- No automatic tail accuracy.
- No discipline-free lifting map.
- No collapse claim from one snapshot.
- No transfer without bottleneck review.
12. FAQs and the next reading route
Heavy-traffic limits compress a complicated queue only after the scaling, resource geometry and policy assumptions are made explicit. State-space collapse is the bridge from workload back to class queues; it is not permission to discard class information.
Read the Queueing Theory root, the Lindley-recursion guide and the quantile-forecast guide alongside this article. The Gold Standard parent and How X Works hub connect the same practice across subjects: expose the model, change context and state what the approximation cannot carry.
Does 95% utilisation mean heavy traffic?
It may motivate the approximation, but heavy traffic is a limiting regime with declared scaling, not a universal percentage threshold.
Does state-space collapse make class queues identical?
No. It says the scaled vector stays near a lower-dimensional surface determined by a lifting map.
Why can the diffusion match workload but miss class delays?
The workload limit and the class-level lifting rule rely on different assumptions; discipline or transient composition can break the second while leaving the first useful.
Continue through the verified ecosystem
Sources and scope note
- Peterson — A Heavy Traffic Limit Theorem for Networks of Queues with Multiple Customer Types — used for regulated Brownian workload limits and class-level results through state-space collapse.
- Williams — Diffusion Approximations for Open Multiclass Queueing Networks — used for sufficient conditions, reflected Brownian motion and the role of state-space collapse.
- Bramson — State Space Collapse with Application to Heavy Traffic Limits — used for multiplicative state-space collapse and discipline-dependent conditions.
The two-class queue, lifting rule and numerical values are original teaching examples. The cited papers support the mathematical concepts; they do not validate the example as a real service design.