The Gold Standard of Water Treatment | Why Stage Averages Need the Same Surviving Population

Begin with a specified population of counted target units in one defined feed. The units might represent an abstract indicator in a paper exercise. Keep their definition and measurement basis unchanged through the calculation.

We will distinguish an original feed, the population remaining after Stage A and the population remaining after Stage B. A percentage whose denominator is the original feed answers a different question from a percentage whose denominator is the material entering Stage B.

That distinction matters even when the units are identical. Ten units per original ten thousand and ten units per intermediate hundred describe different remaining fractions. A stage label does not supply its denominator automatically.

For the main examples, assume no new target input, no bypass, no dilution or concentration change, and no change in the counted target’s definition between stages. These conditions let the count ratios form a linked account. Later chapters show why removing them requires additional balances rather than a new interpretation of the same percentages.

Contents · Next chapter

Let N₀ be the original target count, N₁ the count after A and N₂ the count after B. When N₀ and N₁ are positive, N₂/N₀ = (N₁/N₀)(N₂/N₁).

This is cancellation of a common intermediate count. It does not require an assumption that the stages act independently. The second factor already refers to the actual population reaching B in the defined sequence.

Call the first factor SA and the second SB|A. Then total remaining fraction is SA SB|A. The vertical bar reminds us that the second quantity is conditional on the intermediate population and the relevant operating context.

A separately measured B-alone fraction, SB, may describe the original feed rather than A’s survivors. Substituting SB for SB|A is an extra step requiring evidence or a model. The multiplication rule itself is not the problem; the problem is supplying it with a fraction that answers another question.

If N₁ is zero, the count account should be handled directly rather than dividing by zero to create an intermediate fraction.

Contents · Previous chapter · Next chapter

Divide an original count of 10,000 target units into 5,000 Class L and 5,000 Class H units. These labels simply distinguish two stipulated response classes. They are not biological categories or claims that actual treatment response can be fully classified in advance.

Suppose Stage A leaves 10% of Class L and 90% of Class H. It therefore leaves 500 L and 4,500 H units, total 5,000. Its overall remaining fraction on the original feed is 0.50.

The surviving population is now 10% L and 90% H. It is no longer the original equal mixture. A later stage tested on an equal mixture has not necessarily been tested on the composition that actually reaches it.

The class weights here are proportions of counted target units, not necessarily proportions of water flow or sample volume. A class contributing half the incoming water could contain more or less than half the target count. To substitute volume weights, the relevant target concentrations would need to make that conversion valid. Weighting begins with the population named in the question.

This selection effect is the engine of the example. Nothing has been added to the feed, and no new type is invented during treatment. Unequal retention has changed the relative weights of the types already present. The next calculation must preserve those changed weights.

Contents · Previous chapter · Next chapter

For one stipulated model, let Stage B also leave 10% of L and 90% of H, including when applied to the A survivors. That class-specific transfer is an explicit assumption of this teaching model, not something inferred from B’s average alone.

On the original equal feed, B alone would leave 500 + 4,500 = 5,000 units, so its standalone remaining fraction is also 0.50.

After A, however, B leaves 10% of the surviving 500 L units and 90% of the surviving 4,500 H units. The result is 50 + 4,050 = 4,100. Overall remaining fraction is 0.41.

Multiplying the two standalone averages gives 0.50 × 0.50 = 0.25, or 2,500 units. That understates the remaining count in this model by 1,600 units. The arithmetic error is not in multiplication. It lies in treating a second-stage average on an equal feed as though it described a feed now dominated by H.

Contents · Previous chapter · Next chapter

Stage B in the linked sequence leaves 4,100 of its 5,000 incoming units. Thus SB|A = 0.82. Multiplying SA SB|A = 0.50 × 0.82 gives the correct overall 0.41.

The same second-stage value follows from the changed class weights: 0.10 × 0.10 + 0.90 × 0.90 = 0.82. One tenth of the incoming survivors is L, and nine tenths is H.

This independent reconstruction is useful because it connects the conditional fraction to the physical population account. It is not an unexplained adjustment factor inserted to make the totals agree.

Expressed as removals, A removes 50% of the original feed while B removes 18% of what reaches it. Those percentages cannot be added into 68% overall removal. Total removal is 1 − 0.41 = 59%. Stage percentages must remain attached to their distinct denominators whether they describe removal or remaining material.

Contents · Previous chapter · Next chapter

Keep Stage A unchanged but let B leave 90% of L and 10% of H. On the original equal feed, B alone still leaves 0.50 overall. The two standalone stage averages therefore look identical to those in the previous example.

After A, B now leaves 90% of 500 L units and 10% of 4,500 H units, giving 450 + 450 = 900. Overall remaining fraction is 0.09, and SB|A = 900/5,000 = 0.18.

The same pair of standalone 0.50 averages has therefore accompanied totals of 0.41 and 0.09 in two explicitly different class-response models. The missing information is how the stage responses align within the population.

This comparison does not assert that one real treatment is preferable. We have stipulated response numbers to expose an information gap. A real decision must establish appropriate performance for its targets and conditions rather than select a process because a fictional pattern appears complementary.

Contents · Previous chapter · Next chapter

The common Stage A leaves the same 500 L and 4,500 H units in each row. Only B’s stipulated class-specific remaining fractions change.

Use the table to reconstruct the conditional B fractions 0.82 and 0.18. Both use the same 5,000-unit intermediate denominator. Then multiply each by 0.50 and recover the final fractions.

The table establishes two mathematical possibilities under their stated assumptions. It does not estimate the frequency of either scenario, describe actual organisms or justify a specific performance credit. Keeping that scope beside the worked result allows the arithmetic to be useful without turning it into an unsupported water-quality claim.

Contents · Previous chapter · Next chapter

Let the original class proportions be wᵢ, summing to one. Let aᵢ be the fraction of class i remaining after A. Let bᵢ describe B’s remaining fraction for that class under the stipulated linked-stage model.

The first remaining fraction is SA = Σwᵢaᵢ. The standalone B average on the original class mixture is SB = Σwᵢbᵢ. If the class-specific transfer assumptions apply, the final fraction is SAB = Σwᵢaᵢbᵢ.

When SA is positive, the proportion of class i among A survivors is wᵢaᵢ/SA. Averaging B’s class fractions with those new weights gives SB|A = SAB/SA.

The formula follows the population transformation explicitly. It also exposes what is assumed: the classes must be defined consistently, their original weights must match the question, and the supplied bᵢ values must apply to the actual second-stage context. A sophisticated summation does not make those premises empirically true.

Contents · Previous chapter · Next chapter

Treat aᵢ and bᵢ as paired values over the original class weights. Their weighted covariance is Σwᵢaᵢbᵢ − (Σwᵢaᵢ)(Σwᵢbᵢ).

Thus SAB = SA SB + Covw(a,b). In the same-pattern example, the covariance is 0.41 − 0.25 = 0.16. In the complementary-pattern example it is 0.09 − 0.25 = −0.16.

Positive covariance here means that classes remaining more readily through A also remain more readily through B in the stipulated model. Negative covariance means that the class patterns tend to oppose one another. This is a descriptive relationship between the supplied class response values, not a causal diagnosis of a real barrier.

The decomposition is useful because it identifies exactly what the standalone-average product omits. It should not become a reason to assume covariance zero when it is inconvenient to measure. An unresolved relationship remains unresolved even when the two separate averages are reported with many decimal places.

A further bound follows from the weighted variances of the class responses: the magnitude of their covariance cannot exceed the square root of Varw(a)Varw(b). This is the usual finite weighted Cauchy–Schwarz inequality applied to the two centred response lists.

In the equal-mixture examples, both lists have mean 0.50 and variance 0.16. The covariance magnitude is therefore at most 0.16, so the final fraction lies within 0.25 ± 0.16, or 0.09 through 0.41, for response lists with those stated second moments. The same-pattern and opposite-pattern constructions attain the two ends.

This is more information than the two standalone averages alone supply. The earlier fixed-mark bounds allow zero through 0.50 because they retain only the two marginal event fractions. Different inputs justify different bound widths. A narrower numerical interval should be credited to the additional response information, rather than presented as a universal bound for any pair of fifty-percent stages.

Contents · Previous chapter · Next chapter

If every class has the same B remaining fraction b, then SAB = bΣwᵢaᵢ = bSA. The standalone B fraction is also b, so the product SA SB is valid under that constant-response model.

Likewise, zero weighted covariance of the supplied class responses makes the product valid for this particular population and model. That is weaker than claiming that every aspect of the stages or target units is statistically independent.

A direct linked measurement of SB|A also avoids the substitution problem. It supplies the actual second factor rather than trying to infer it from a standalone average.

State which justification is being used. Constant response, a supported joint model and a direct intermediate-to-final measurement are different evidence routes. A report that says independence without identifying the variables, population and conditions may leave the central assumption obscure. The exact linked identity needs no such vague label when the intermediate counts are appropriately defined.

Contents · Previous chapter · Next chapter

For positive remaining fraction S, define the mathematical log reduction ℓ = −log₁₀S. The linked identity gives ℓAB = −log₁₀SA − log₁₀SB|A. Addition follows from the logarithm of a product. A zero log reduction means a remaining fraction of one in this definition, not zero remaining target units.

In the same-pattern example, the true total is −log₁₀0.41, approximately 0.387216. Adding the two standalone values −log₁₀0.50 gives approximately 0.602060, which describes the unsupported product 0.25 instead.

For the complementary model, the total is −log₁₀0.09, approximately 1.045757. The same standalone log sum misses in the other direction.

This does not invalidate the use of appropriately established serial log reductions. It shows why the numbers being added must apply to the linked target and process conditions. A change from fractions to logarithms cannot repair a population mismatch that was already present in the fractions.

Contents · Previous chapter · Next chapter

The US EPA’s Surface Water Treatment Rule fact sheet describes combining barrier reductions for a specified microorganism within its stated regulatory context. It also ties treatment credit to operating and design conditions.

Our abstract examples do not revise those rules or assign regulatory credit. They illustrate a narrower modelling question: whether separately supplied averages actually describe the linked populations to which someone wants to apply them.

An applicable validated stage result can already incorporate the relevant target and operating conditions. It should not be discarded because a classroom mixture example shows that unsupported transfers are possible. Conversely, an unqualified standalone percentage should not be promoted to a validated process credit merely because it resembles one numerically.

Keep the authority, target and evidence status visible. Source checked on 25 October 2026. The fictional count models here are educational constructions, not a current compliance assessment for any water supply.

Contents · Previous chapter · Next chapter

The existing guide on remaining-fraction bounds from residence time and time limits examines another aggregation problem. It specifies a first-order response along different time histories and derives bounds from a mean and endpoints.

The present problem instead asks how an earlier stage changes the target population presented to a later one. We have supplied class-specific remaining fractions rather than deriving them from contact-time histories.

Both require careful weighting, but the missing information differs. A residence-time distribution cannot automatically reveal how susceptibility to A aligns with susceptibility to B. A table of response classes cannot automatically establish a real reactor’s flow history.

Use the question to select the model. Avoid collecting more precise evidence about the wrong summary simply because it is easier to obtain. A meaningful next measurement should address the relationship that makes the desired combined result uncertain.

Contents · Previous chapter · Next chapter

Keep aL = 0.10, aH = 0.90 and the same B response pattern, but let the original feed be 20% L and 80% H. Stage A’s remaining fraction becomes 0.20 × 0.10 + 0.80 × 0.90 = 0.74.

The standalone B average is also 0.74 in the same-pattern model. The actual final fraction is 0.20 × 0.01 + 0.80 × 0.81 = 0.65. The standalone product is 0.5476, again different.

After A, the H share is 0.72/0.74 = 36/37, approximately 97.2973%. B’s conditional remaining fraction is 0.65/0.74, approximately 0.878378.

This changed exercise shows why a stage-average certificate from one challenge mixture may not describe another mixture under a model with heterogeneous response. Whether actual performance transfers requires the appropriate validation evidence. Our arithmetic identifies the possible dependence on composition; it does not establish how a named treatment responds to a real source change.

Contents · Previous chapter · Next chapter

Use original class weights 0.40 and 0.60, with A fractions 0.20 and 0.80 and B fractions 0.70 and 0.30. Then SA = 0.56, SB = 0.46 and SAB = 0.40 × 0.20 × 0.70 + 0.60 × 0.80 × 0.30 = 0.20.

In this fixed-class, transferable-response model, products aᵢbᵢ commute, so reversing the order leaves the final fraction 0.20. The intermediate and conditional fractions differ: B after A leaves 0.20/0.56 = 5/14, whereas A after B leaves 0.20/0.46 = 10/23.

That commutation is a property of the stipulated model. Real stages can alter chemistry, particles or later response conditions, invalidating the fixed bᵢ or aᵢ transfer assumptions. One should not infer operational interchangeability from our algebra.

The useful transfer habit is to name which result comes from arithmetic and which comes from a physical assumption. Equal final products in a paper table do not establish equal practical processes.

Contents · Previous chapter · Next chapter

Another abstract representation assigns each original unit two binary marks: whether it would remain through A and whether it would remain through B, under a fixed-response, noninteracting thought model. Both marks must refer to the same original population.

Let the marginal probabilities of those marks be a and b. The probability j of both marks is bounded by max(0,a + b − 1) ≤ j ≤ min(a,b). These follow from nonnegative probabilities of both, A only, B only and neither.

With a = b = 0.50, j can range from zero to 0.50. The standalone product 0.25 is one possible value, not a uniquely determined result from the marginals alone.

This model is deliberately limited. In a real treatment sequence, the first stage can change the second-stage response, so fixed hypothetical marks may not be an appropriate representation. The bounds apply to the stated common-population event model, not automatically to arbitrary separate tests or interacting processes.

Contents · Previous chapter · Next chapter

For a = b = 0.50, one extreme assigns half the original units both remaining marks and half neither mark. Then j = 0.50: the two marks coincide exactly.

The other extreme assigns half A-only and half B-only. Then j = zero: no original unit has both marks. Each marginal remains one half, just as in the first construction.

These explicit finite allocations prove that the interval endpoints are attainable within the fixed-mark model. Intermediate mixtures of the two allocations produce intermediate j values. The bounds are therefore more than loose numerical cautions.

Do not translate zero marked survivors in a fictional table into a claim of complete microbial safety. The table is a mathematical allocation over supplied units. Measurement limits, unobserved units, other targets and real process validity remain outside that finite construction. A sharp abstract bound is useful only when its scope remains attached.

The linked account extends to a third stage without changing the principle. With actual intermediate counts, total remaining fraction is SA multiplied by SB|A and then SC|AB. The third fraction refers to the population remaining after both earlier stages.

For a three-stage version of the main same-pattern model, let C leave ninety percent of L and ten percent of H. The final fraction is 0.50(0.10 × 0.10 × 0.90) + 0.50(0.90 × 0.90 × 0.10) = 0.045.

The standalone original-feed average of C is still 0.50, but its linked fraction after A and B is 0.045/0.41 = 9/82, approximately 0.109756. Multiplying the three standalone averages would give 0.125. Multiplying 0.50 × 0.82 × (9/82) instead gives the correct 0.045.

The third stage has not introduced a new arithmetic rule. It has made the population-selection history longer. Preserve that history or justify an adequate state representation; do not reset the challenge mixture to its original weights after each stage merely because every stage report has a familiar percentage format.

Contents · Previous chapter · Next chapter

If water volume stays unchanged, count ratios and concentration ratios can coincide. When volumes change, the correspondence needs a separate balance.

For example, a hypothetical intermediate stream contains 100 target units in one litre. Adding an equal volume of target-free water halves concentration to fifty units per litre while leaving the target count at one hundred. That is dilution, not removal in the count account.

If N₂/N₁ is the intended remaining-count fraction, substituting a concentration ratio without accounting for volume would report a false count reduction. Conversely, a concentration question should retain the actual volume and mixing assumptions rather than silently switch to total inventory.

The linked identity can be written for consistently defined concentrations too, but the interpretation changes when dilution or concentration is part of the process. State the target quantity before using the cancellation. Identical units printed on two laboratory rows do not by themselves establish that the underlying samples represent a comparable process inventory.

Contents · Previous chapter · Next chapter

Suppose the class model predicts 900 units after both stages, but an additional source contributes 100 units after the second stage. The final delivered count in that stipulated account is 1,000.

The ratio of final count to original 10,000 is now 0.10. It should not be attributed solely to the two-stage retention product 0.09. A new source term has entered the balance.

Likewise, a bypass stream needs its own path and weight. It is not repaired by reducing a stage’s removal percentage until the total appears to agree. The network structure and input locations determine how the counts combine.

This is a useful diagnostic exercise: draw each stream, identify where new target units enter and calculate each path on a compatible basis. A mismatch between observed and predicted final count can have several causes. A single fitted overall fraction does not identify whether the issue lies in stage response, bypass, new input or measurement interpretation.

Contents · Previous chapter · Next chapter

In a paper study, one can compare B on the original mixture with B on the stipulated A survivors. The two tests answer different questions because their input class weights differ.

A meaningful evidence plan identifies whether the intended use concerns standalone response, sequential response or a validated range of relevant inputs. It preserves input characterisation, operating conditions, target definition and the measurement method.

An actual treatment validation requires appropriate professional procedures. This article supplies no challenge-organism protocol, dose recommendation or acceptance criterion. The transferable planning question is which observation would distinguish the competing explanations for a combined result.

For the main model, learning only the two standalone 0.50 averages cannot distinguish 0.41 from 0.09. A linked intermediate-to-final count or justified class-response information can. The next useful evidence therefore concerns the surviving population or the joint response, rather than another copy of a standalone average that both models already share.

Contents · Previous chapter · Next chapter

Suppose a fictional final measurement reports below a stated detection limit rather than an exact count. That is an inequality about the measurement method’s result, not proof that the true final target amount is zero.

A log reduction calculated from a bounded final value is correspondingly bounded. Inventing an exact zero would create an infinite mathematical log reduction while discarding the method’s actual information.

The same caution applies to an intermediate measurement. If N₁ is not known exactly, the second-stage fraction N₂/N₁ may carry additional uncertainty or only admit bounds. Exact cancellation of symbols does not remove uncertainty in the observed quantities those symbols represent.

A careful report distinguishes model zeros, measured non-detections and inferred bounds. Those are different evidential states. The numerical examples here deliberately use exact stipulated counts so the population-selection mechanism can be understood before method limits and sampling uncertainty are added.

Contents · Previous chapter · Next chapter

Try original class weights 0.25 and 0.75. Stage A leaves fractions 0.20 and 0.60, while B leaves 0.80 and 0.20 under the same transferable class-response assumption. Find each standalone fraction, the final fraction and B’s conditional fraction after A.

SA = 0.25 × 0.20 + 0.75 × 0.60 = 0.50. SB = 0.25 × 0.80 + 0.75 × 0.20 = 0.35. The linked final fraction is 0.25 × 0.20 × 0.80 + 0.75 × 0.60 × 0.20 = 0.13.

Thus SB|A = 0.13/0.50 = 0.26. The standalone product is 0.175, which exceeds the linked result. The weighted covariance is −0.045.

Explain the direction before checking the decimal: the class that more readily survives A has the lower B remaining fraction. Reweighting towards that class makes B’s actual remaining fraction smaller than its original-feed average. This is the same method with different weights and responses, not a memorised answer from the fifty-fifty example.

Contents · Previous chapter · Next chapter

Begin with 20,000 fictional units: 5,000 in the first class and 15,000 in the second. A leaves 1,000 and 9,000, total 10,000. The survivor mixture is now 10% first class and 90% second class.

B then leaves 800 and 1,800, total 2,600. The final ratio is 2,600/20,000 = 0.13 and the second-stage conditional ratio is 2,600/10,000 = 0.26, matching the weighted formulas.

On the original feed, B alone would leave 4,000 + 3,000 = 7,000, or 0.35. The three denominators are now visible through actual count rows rather than abstract labels.

Ask the learner to reconstruct one missing intermediate count and then explain why 7,000 cannot be inserted as the linked second-stage output. This checks whether the population relationship is understood. Repeating a multiplication without naming its input population is insufficient evidence of transfer.

Contents · Previous chapter · Next chapter

Our class examples concern target-response heterogeneity within a stipulated operating sequence. Another problem concerns whether two physical barriers fail together because they share power, controls or another dependency.

Those are different random objects. A target’s chance of remaining through functioning stages should not be conflated with a facility’s chance of losing the stages’ operating functions. Both can matter in a system assessment, but they require their own models and evidence.

A statement such as independent barriers can therefore be ambiguous. Does it refer to response mechanisms, validation evidence, equipment failure events, monitoring or management responsibilities? Specify the relationship before interpreting the phrase mathematically.

The Water Treatment Engineering guide provides the wider system route. This article supplies a local population-accounting lesson within that broader context. It does not turn one class-response calculation into a complete assessment of operational reliability.

Contents · Previous chapter · Next chapter

Start with the original target definition, population or concentration basis, and the counts or weights entering each stage. Identify which response fractions are measured on the actual intermediate stream and which are transferred from another challenge.

Write the linked identity before inserting numbers. If using a class model, retain the original weights, class-specific fractions and the assumption that those fractions apply in the sequence. Reconstruct at least one result through explicit counts.

Add source terms, bypass paths and volume changes where they belong. Keep detection limits and other bounded results visibly bounded. If logarithms are reported, retain the underlying fractions so the population connection remains inspectable.

A useful conclusion names both the result and its condition: the stipulated two-class model leaves thirteen percent because the second-stage fraction on the selected intermediate population is twenty-six percent. That is more informative than reporting two standalone averages and an unsupported combined removal.

Contents · Previous chapter · Next chapter

For repair, use the 10,000-unit count table and ask what reaches Stage B. If the learner uses the original equal mixture again, return to the 500 and 4,500 survivors before introducing covariance or logarithms.

For stabilisation, reverse B’s class pattern while retaining its original-feed average. The learner should predict that the combined result can change despite the unchanged standalone percentage. Then calculate the conditional fraction from the intermediate denominator.

For extension, change the original weights, add a source or remove exact intermediate measurement. Ask which conclusion remains identified and what new evidence would narrow the possibilities.

The learning objective is a complete population account. A learner who can explain why the second denominator changed can transfer the method to reliability, screening and other sequential-selection problems while checking whether those new settings justify the same fixed-response assumptions.

Contents · Previous chapter · Next chapter

Must serial remaining fractions always be independent?

No. The product of the first fraction and the actual conditional second fraction is a linked identity. An independence assumption becomes relevant when replacing that conditional fraction with a standalone one.

Are log reductions allowed to add?

For appropriate linked positive fractions, logarithms add exactly. The relevant question is whether the supplied stage values apply to the target, population and conditions being combined.

Does a worse second-stage average prove the stage degraded?

Not necessarily. In our example, unchanged class response combines with a changed incoming class mixture. Actual degradation is another hypothesis requiring evidence.

Can these teaching numbers establish water safety?

No. They describe stipulated abstract counts and response assumptions. Real safety and treatment credit require the appropriate targets, validated performance, operational evidence and responsible authority framework.

Contents · Previous chapter · Next chapter

Follow the process and its evidence

Return to Water Treatment, then use Water Treatment Engineering for the broader system. The membrane-integrity guide and UV-validation guide develop specific evidence questions rather than interchangeable average percentages.

Keep aggregation questions distinct

The linked residence-time article derives a different mathematical bound. Use it for a supplied time-response model; use this guide when an earlier stage changes the population that a later average must describe.

Transfer the learning

Continue through Gold Standard Education and How X Works. In any sequence, identify what remains, what enters next and which denominator makes the next performance claim meaningful.

Contents · Previous chapter · Return to Water Treatment