Two numbers close this chapter out: how much a control device removes, and whether the measurement proving it is admissible at all.

Collection efficiency: $\eta = \dfrac{C_{in} - C_{out}}{C_{in}}$ — eta equals C-in minus C-out, over C-in. $C_{in}$ is the dust loading entering the device and $C_{out}$ the loading leaving, both in mg/m³ and both on the same basis; $\eta$ is a bare fraction, and the percent sign supplies the ×100. Its complement, $C_{out}/C_{in}$, is penetration — and penetration is the honest way to compare good devices. Going from 99 % to 99.9 % sounds like a tenth of a percent and is a tenfold cut in what leaves the stack. Rules of thumb: a single cyclone rarely beats 90 % on total mass, a fabric filter reaches 99.9 %, and both do considerably worse on the fine fraction that PM2.5 is measured in.

Isokinetic sampling: $Q_n = v_s A_n$ — Q-n equals v-s A-n. $v_s$ is the local stack velocity at the sampling point in m/s, $A_n$ the nozzle tip area, and $Q_n$ the flow the pump must draw so that gas enters the nozzle at exactly the speed it was already travelling. The unit bridge is worth memorising outright: 1 m/s through 1 cm² is 6 L/min.

Why it matters is pure inertia. Draw too slowly and gas spills around the nozzle while the heavy particles carry straight in — the sample is enriched and the loading is over-stated. Draw too fast and fine gas is pulled in from the sides without its share of the coarse particles — the sample is diluted and the loading is under-stated. EPA Method 5 accepts 90 to 110 % of isokinetic and rejects the run outside that band, whatever the scaffold cost to erect. Gases are unaffected: the bias is entirely a particle effect, so a run can be perfectly good for NOx and worthless for particulate at the same time.

The bias has a fixed sign for a given error direction, which is why it does not average out over a traverse — and why 'close enough' is not a verdict anyone is allowed to reach.