Physics · Article

Orders of magnitude and Fermi estimation

Build a rough estimate from simpler quantities, check its units and scale, and show how assumed ranges and model choices limit the answer.

Previous article: Dimensions, scaling, and similarity

Back to Physics

After reading, you can

Write positive quantities in scientific notation and state the scale convention being used.
Break an estimation question into simpler inputs and an explicit relationship.
Check counting factors, units and arithmetic before trusting an estimate.
Calculate conditional bounds for a product of positive inputs.
Distinguish assumed model bounds from measured results and statistical confidence.
Identify which input range or omitted contribution deserves the next check.

Start with a quantity and a question

A Fermi estimate uses a reasoned model and rough inputs to estimate a quantity. It often breaks a large question into smaller ones whose answers can be estimated more readily. State what you want to estimate: total liquid volume, for example, or the active time needed for a sequence of tasks.

Specify the situation and what is included. An estimate of active work time may leave out setup and pauses; elapsed completion time may include them. Assumptions or further information are needed when the question is underspecified.

Every numerical case in this lesson is invented. Counts, representative inputs and ranges are stipulated for practice, rather than taken from observations. In an investigation, record where each input came from and how you could check it. A calculation from assumed inputs is not itself a measurement.

Use powers of ten with an explicit convention

Scientific notation separates a number into a coefficient and a power of ten. For a positive numerical value, write c × 10^n, where 1 ≤ c < 10 and n is an integer. For example, 7200 s = 7.2 × 10^3 s. The exponent 3 means multiply by 10 × 10 × 10 = 1000. Also, 10^0 = 1 and 10^-3 = 1/1000 = 0.001.

Here we report a rough scale by its decade band: from 10^n inclusive up to 10^(n+1) excluded, in the stated unit. Thus 7200 s is in the thousands-of-seconds band, represented by 10³ s. This is our declared reporting convention. Some textbooks instead round the coefficient to 1 or 10 by a specified rule. State the convention when a boundary matters.

A band is not an exact value or a guarantee of closeness. Values near a boundary can have different band labels while being close together. To compare two quantities of the same kind, first use the same unit and divide their values. A ratio of 10 is a tenfold difference; 100 is two successive tenfold differences. Changing the unit can change the numerical exponent without changing the quantity.

Our decade-band convention, illustrated with positive durations in seconds
Power	Numerical value of the power	Band in seconds
10^-3	0.001	From 0.001 s inclusive to 0.01 s excluded
10^0	1	From 1 s inclusive to 10 s excluded
10^3	1000	From 1000 s inclusive to 10 000 s excluded
10^6	1 000 000	From 1 000 000 s inclusive to 10 000 000 s excluded

Build the relationship before choosing numbers

Write a relationship connecting the required quantity to simpler inputs. Repeated contributions can be combined using a count and a representative contribution per item. For one-at-a-time tasks, with no overlap, a useful model is active time = number of tasks × average active time per task.

The average in this relationship is the total active time divided by the number of tasks. The individual tasks need not take identical times. Before they have been timed, a chosen average is an assumption. If tasks run in parallel, their summed active times need not equal elapsed completion time. Setup, waiting and repeated tasks need their own treatment.

A count may itself be decomposed: total pieces = trays × average pieces per tray. Define each factor so the same piece is counted once. Choose representative inputs, calculate, and then examine how other allowed inputs would change the result. More calculator digits cannot check whether the chosen factors describe the intended situation.

Carry the units through the calculation

Multiply and divide units along with numerical values; matching units cancel. The result must have the required dimension. Passing this check does not establish that the model is physically correct.

For the time model, a count multiplied by seconds per task leaves seconds. Labels such as trays, pieces and tasks help track the counting even though counts have dimension one. For a volume model, vessels × average litres per vessel leaves litres. Multiplying a vessel count by litres without defining the per-vessel contribution can conceal a mistaken interpretation.

Convert units consistently before comparing results. Here 1 min = 60 s and 1 L = 1000 mL. A calculation expressed in millilitres must be converted before comparison with a limit expressed in litres. A factor-of-1000 mistake changes the numerical scale substantially; a unit label helps reveal it.

Show what the assumed ranges allow

Input bounds can be used to bound a model result. For a product of positive inputs, increasing any one factor while holding the others fixed increases the result. Multiply the lower endpoints for a lower bound and the upper endpoints for an upper bound.

These bounds are conditional: they hold if every input lies within its stated range and the relationship applies. If inputs are related, the resulting interval still contains all allowed product values, but its endpoints might not be jointly attainable. Additional constraints can make the interval narrower. No statistical independence is needed for this bound calculation.

Keep the unrounded endpoint calculations when checking a limit. Rounding a lower endpoint upward or an upper endpoint downward can remove allowed values. For a rough summary, label the approximation and retain the conditional bounds alongside it.

Say what the interval means

An interval calculated from stipulated limits does not assign probabilities to values inside it. It therefore does not justify a claim of 95% statistical confidence. In metrology, a measurement coverage interval carries a stated probability and is distinguished from a statistical confidence interval. This lesson's model bounds provide neither of those statistical constructions.

A representative estimate is the result of the selected representative inputs. It need not be the interval's midpoint, an average of the possible answers or the most likely answer. We have not supplied information that assigns those meanings.

Likewise, exact arithmetic within an invented model does not show the precision of real measurements. If an assumed range misses the actual inputs, or the model omits an important contribution, the real result may fall outside the calculated interval. Report the quantity, assumptions, selected inputs and bounds together.

Check the assumptions that matter to the decision

Ask which factors need better information and which neglected contributions could matter. One simple check is to vary one input through its allowed range while keeping the others fixed. Compare the resulting answer ratios. The largest effect in this comparison is a candidate for the next input check, not a universal rule for every model.

Also test the structure of the calculation: could a contribution have been omitted, counted twice or assigned to the wrong time interval? A narrow input range cannot repair the wrong relationship. If a stipulated upper bound exceeds an allowed limit, the estimate alone cannot guarantee that the limit will be met.

Before using the result, revisit units, arithmetic and reasoning. Compare it with relevant independently checked information when available. In a real investigation, checking a count or timing representative tasks could improve the inputs; the present invented values are practice data, not the results of such checks.

From assumed inputs to a conditional volume estimate

1. Define the quantity
Estimate the total liquid placed in the once-filled vessels. Keep the stated exclusions attached to that question.
2. Connect the factors and units
Total fill volume = vessel count × average fill volume per vessel. The counting labels cancel, leaving litres.
3. Choose a representative estimate
The selected inputs give 30 × 0.20 L = 6 L. This result has not been established as the most likely volume.
4. Retain the conditional bounds
The positive endpoint products are 25 × 0.15 L = 3.75 L and 35 × 0.25 L = 8.75 L. The interval holds if the input limits and fill model apply.
5. Choose the next check
Check the count, average fill and what the quantity includes. The interval is a model bound; no confidence level has been calculated.

Original invented plan: 25–35 vessels and average fill 0.15–0.25 L per vessel, all endpoints inclusive. Each vessel is filled once. Representative inputs are 30 vessels and 0.20 L per vessel. Only liquid placed in the vessels is counted; spill and liquid retained elsewhere are excluded. These are stipulated inputs, with no measurements or probabilities supplied.

Worked examples

Build an active-time estimate from three factors

An invented plan has 12 trays, with an assumed average of 25 pieces per tray. Each piece receives one inspection. Assume an average active inspection time of 18 s per piece, with inspections performed one at a time. Estimate total active inspection time, excluding setup, pauses and rechecks.

Decompose the count: 12 trays × 25 pieces/tray = 300 pieces.
Multiply by the assumed time contribution: 300 pieces × 18 s/piece = 5400 s.
Convert with 60 s per minute: 5400 s / (60 s/min) = 90 min.
Write the result as 5.4 × 10^3 s. A rough summary is about 5 × 10^3 s, or several thousand seconds. The unrounded model result is 5400 s.

The selected inputs give 5400 s, or 90 min, of active inspection time. Under our convention this is in the thousands-of-seconds band. The rounded summary reflects the intended rough estimate; neither the calculation nor its digits establish actual inspection times. Completion time would also need any excluded contributions.

Bound a total volume from two positive inputs

For an invented filling plan, assume 25–35 vessels inclusive, each filled once, with average volume per vessel between 0.15 L and 0.25 L inclusive. Use 30 vessels and 0.20 L per vessel for a representative estimate. The required quantity is liquid placed in the vessels, excluding any spill or liquid retained elsewhere. What estimate and conditional bounds follow?

Use total fill volume = vessel count × average fill volume per vessel. The counting labels cancel, leaving litres.
The representative inputs give 30 × 0.20 L = 6 L.
For positive factors, the lower bound is 25 × 0.15 L = 3.75 L, and the upper bound is 35 × 0.25 L = 8.75 L.
Using 1 L = 1000 mL gives the same bounds as 3750–8750 mL. The numerical scale changes with the unit; the volumes do not.

The representative estimate is 6 L, with conditional model bounds of 3.75–8.75 L inclusive. These exact endpoint products follow from the stipulated ranges. They are not measured volumes, a 95% confidence interval or a statement that values near 6 L are more likely. The bounds cover only the defined fill volume.

Decide which assumed input deserves a check

A separate invented serial-inspection model has 180–220 tasks and average active time 8–32 s per task, all endpoints inclusive. Use 200 tasks and 20 s per task as representative inputs. Compare the two input ranges' effects on active time. Does this model guarantee active time of at most 5000 s?

The representative result is 200 × 20 s = 4000 s.
Hold average time at 20 s and vary the count: 180 × 20 s = 3600 s through 220 × 20 s = 4400 s. The endpoint ratio is 4400/3600 = 11/9, about 1.22.
Hold the count at 200 and vary average time: 200 × 8 s = 1600 s through 200 × 32 s = 6400 s. The endpoint ratio is 6400/1600 = 4.
Allow both ranges: the conditional bounds are 180 × 8 s = 1440 s and 220 × 32 s = 7040 s.
The assumed per-task time range has the larger multiplicative effect in this comparison. Check that input and the serial-time model before treating the representative result as a completion guarantee.

The average-time range changes the answer by a factor of 4 in its individual check; the count range changes it by about 1.22. The full conditional interval is 1440–7040 s inclusive. Since it includes values above 5000 s, the stipulated model does not guarantee the allowed active-time limit. No probability of exceeding the limit has been calculated.

Try it yourself

Write the stipulated duration 0.0063 s in scientific notation and identify its decade band under this lesson's convention. If 1 ms = 0.001 s, does writing it as 6.3 ms change the duration?
Show the answer and reasoning
0.0063 s = 6.3 × 10^-3 s, in the 10^-3 s band. It is also 6.3 ms, with no change of duration.
The coefficient 6.3 lies from 1 inclusive to 10 excluded. The band runs from 0.001 s inclusive to 0.01 s excluded. Dividing by 0.001 s per millisecond changes the numerical unit representation, not the quantity.
Two stipulated positive lengths are 9 × 10^4 m and 9 × 10^2 m. How many times larger is the first? How many successive tenfold differences is that?
Show the answer and reasoning
It is 100 times larger, equal to two successive tenfold differences.
In the same unit, the ratio is 90 000/900 = 100 = 10 × 10. The equal coefficients cancel. For unequal coefficients, comparing the powers alone would omit part of the ratio.
In an invented model, 5 racks each hold 16 objects. Each object receives one 30 s task, performed one at a time with no setup, gaps or repeated tasks. What total active time follows, in seconds and minutes?
Show the answer and reasoning
2400 s, or 40 min.
The count is 5 × 16 = 80 objects. Then 80 × 30 s = 2400 s, and 2400 s / (60 s/min) = 40 min. Counting each object once and specifying serial tasks makes the sum correspond to the defined time. These are supplied model inputs.
An invented plan fills 10–12 vessels once, with average volume 0.40–0.50 L per vessel; all endpoints are included. The product model applies to fill volume. What are its bounds, and does a 5 L supply necessarily cover that fill volume?
Show the answer and reasoning
The bounds are 4–6 L inclusive. A 5 L supply is not guaranteed to cover every allowed fill volume.
The positive endpoint products are 10 × 0.40 L = 4 L and 12 × 0.50 L = 6 L. The allowed upper result exceeds 5 L. This conclusion concerns the stipulated fill model; no probability or allowance for spill has been supplied.
A stipulated active-time model gives 1000 s. A separate stipulated setup step takes 600 s and happens before the active work, with no overlap or other delays. What is the revised elapsed time? Would more digits in the original active-time calculation fix the omission?
Show the answer and reasoning
1600 s. More digits would not add the missing setup step.
For these non-overlapping contributions, add 600 s + 1000 s = 1600 s. The correction changes the relationship being used to describe elapsed time. Numerical precision in an incomplete model cannot supply the omitted contribution.
Conditional model intervals for two volumes are 4–6 L and 8–10 L, inclusive. Does the first have to be smaller if both models and input bounds hold? Would that ordering still follow if the second interval were instead 5–9 L?
Show the answer and reasoning
Yes for 4–6 L versus 8–10 L. No guaranteed ordering follows for 4–6 L versus 5–9 L.
In the first comparison, even the largest first volume, 6 L, is below the smallest second volume, 8 L. The revised intervals overlap: permitted choices include a first volume of 6 L and a second of 5 L, or a first of 4 L and a second of 9 L. No probability has been assigned to those choices.

What to carry forward

Define the quantity and what is included before choosing estimation inputs.
State the unit and the convention when reporting an order-of-magnitude scale.
Decompose the question into explicit factors, then check counting, units and arithmetic.
Positive product endpoints give conditional bounds when the input limits and model apply.
Assumed bounds and representative estimates do not establish measurements or statistical confidence.
Check influential input ranges and missing contributions against the decision the estimate must support.

References

Sources for the concepts and teaching approaches in this article.