ALearning Material
An order-of-magnitude estimate is a calculation performed with inputs rounded hard enough for the arithmetic to be done mentally. Its result is good to about a factor of two and takes about thirty seconds, which often makes it more useful than an exact answer that arrives a week later. An estimate also gives an independent expectation against which a detailed answer can be tested. Where the previous lesson limited the precision a result may claim, an estimate gives up precision on purpose.
The method has three steps. Every input is rounded to one significant figure, preferably to a power of ten. The arithmetic is done in exponents, which the powers lesson turned into addition. The result is then compared with the value the design requires, and that comparison produces the decision.
A mobile robot needs either a 12 V battery of 5 Ah or one of 50 Ah. It drives two motors that each draw about 2 A while moving, the electronics take about 1 A, and it must run for four hours. Rounded hard, the charge required is:
The answer is tens of amp hours. The 5 Ah pack is too small by a factor of four, and no careful calculation will change that, while the 50 Ah pack is plausible with margin. The conclusion needed no datasheet, duty cycle or efficiency figure, and it is the decision that had to be made at this stage. A detailed calculation comes later and refines a choice that is already in the right decade.
Estimates also decide system architectures. A camera produces 1920 by 1080 pixels at 30 frames per second, with three bytes per pixel, and the question is whether it can stream uncompressed over a 100 Mbit/s link. rounds to pixels, since two million is close enough and easy to carry:
which is about bits per second after multiplying by eight. The link offers . The demand exceeds the link by more than a factor of ten, so the question is settled: the stream cannot be sent uncompressed, and the design needs compression, a faster link or a lower frame rate. The rounding was generous in places and tight in others, and it made no difference, because the gap was an order of magnitude. An estimate needs to be more accurate than the margin of the decision, and no more accurate than that.
The same rule sets the limit of the method. An estimate that lands close to the threshold has not decided anything. Had the link been 2 Gbit/s, a rough answer of 1.4 Gbit/s would lie within the error of the rounding, and the detailed questions (framing overhead, protocol efficiency, burst behaviour) would still be open. An estimate near the threshold is a signal to do the detailed calculation.
Two techniques improve an estimate without slowing it down. Bracketing computes a value that is certainly too small and one that is certainly too large, and checks whether the threshold lies inside or outside that range. A designer who cannot estimate a robot's power draw can usually say that it is more than 10 W and less than 1000 W, and for many decisions that is enough. Anchoring derives an unknown figure from a known one instead of recalling it: a person is about 2 m tall and weighs about 80 kg, a kettle draws about 2 kW, mains voltage is a couple of hundred volts, and walking pace is about 1 m/s. Anchors of this kind turn an unknown into a comparison.
The most frequent use of an estimate is a sanity check on a figure from someone else. A junior engineer reports that a conveyor motor will consume 45 kW. The conveyor moves 15 tonnes per hour up a 2 m rise. An energy check, with free rounding, gives:
The lifting power is under a hundred watts. Friction, drive losses and acceleration multiply it several times, so a few hundred watts is credible, and a kilowatt or two allows a generous margin. A figure of 45 kW is two orders of magnitude above the physics, so the calculation behind it contains a mistake, most likely a units error of the kind described in the previous lessons. The estimate does not locate the mistake. It establishes that a mistake exists, and that is what sends someone back to the working.
The value is used instead of 9.81 on purpose. It introduces a 2% error and removes a multiplication that would be hard to do mentally, and 2% is irrelevant against a discrepancy of 50 to 1. Knowing which approximations cost nothing is most of what makes an estimate fast.
Formulas & method.
- The procedure. Round each input to one significant figure or a power of ten, and do the arithmetic in exponents. Compare the result with what the decision requires.
- Useful roundings. , , one year s, one day s and .
- Anchors worth memorising. A person, 2 m and 80 kg. A kettle, 2 kW. Mains, a couple of hundred volts. Walking pace, 1 m/s. Water, 1000 kg per cubic metre, so 1 litre is 1 kg.
- Bracketing. Produce a lower bound you are sure of and an upper bound you are sure of. If the threshold lies outside the bracket, the decision is made. If it lies inside, the estimate has not settled it.
- Energy check for anything that lifts. . Use it on any conveyor, pump, hoist or lift to bound the power from below, since no arrangement can use less than this.
- When an estimate decides. Only when the gap to the threshold is larger than the rounding error, which for one-significant-figure inputs is roughly a factor of 2 to 3. A result within that band of the threshold has decided nothing.
- Reading a discrepancy. A factor of about 2 usually means neglected losses or efficiency. A factor of 10, 60, 1000 or 3600 usually means a units or prefix error. A factor of or 2 in a rotational problem usually means radians against revolutions.
Where this shows up.
- Robotics. Battery sizing, payload against motor torque, and whether a planning algorithm can run inside a control period are all decided in the first minute by an estimate and only then refined.
- Automation. A valve's flow, a heater's duty and a tank's residence time are estimated before anything is specified, and a supplier's figure is sanity-checked before it becomes a purchase order.
- Electronics. Whether a regulator needs a heatsink is a power estimate, and whether a trace can carry a current is a comparison against an anchor, not a calculation.
- Digital design. Gate counts, memory footprints and timing budgets are all estimated in powers of two, and is the approximation that lets that be done in decimal.
Why it exists. Most engineering decisions choose between options that differ by a large factor, and a method that separates them only after a week of detailed work answers the question too late. An estimate decides those cases immediately. For the cases it cannot decide, it shows that they are close, which identifies the few questions that need the detailed treatment.
Mental model. An estimate asks which decade the answer lies in. The possible answers lie along a ruler marked in powers of ten, with the decision threshold somewhere on it, and an estimate places the answer on that ruler to within a notch or two. If the threshold is several notches away, the question is settled. If it is within a notch, the estimate shows that a more precise method is needed.
Common misunderstandings.
- "An estimate is a guess." A guess has no derivation. An estimate is a calculation with hard-rounded inputs, and every step can be shown and challenged, so an error in an estimate can be found, while an error in a guess cannot.
- "A more careful estimate is always better." Extra care on inputs that are already far more accurate than the decision's margin adds nothing, and it turns a thirty-second check into an afternoon's work, a cost large enough that the check may not be done at all.
- "If the estimate and the detailed answer disagree, the detailed one wins." The detailed calculation has more steps and therefore more places to go wrong, and a disagreement by a factor of 60 or 3600 usually points to an error in the detailed working, not in the estimate.
Connections. The exponent arithmetic here applies the powers lesson at speed, and the hard rounding reverses the precision lesson by discarding digits on purpose. Dimensional analysis from the units lesson complements estimation: units catch structural errors and estimates catch errors of magnitude, and neither catches the other's. Every later topic uses estimation before its own methods: a control engineer estimates a time constant before tuning, and a chip designer estimates a gate count before synthesis.
BImmediate Active Recall
QUERYWhat are the three steps of an order-of-magnitude estimate, and which one produces the decision?
REVEAL
Round every input to one significant figure or a power of ten, do the arithmetic in exponents, and compare the result with what the decision requires. The comparison produces the decision: an estimate with nothing to test it against has decided nothing.
QUERYWhen does an estimate actually settle a question?
REVEAL
Only when the gap to the threshold is larger than the rounding error, which for one-significant-figure inputs is roughly a factor of 2 to 3. A result near the threshold shows that the detailed calculation is needed, which is itself useful information.
QUERYEstimate whether 1920x1080 at 30 fps and 3 bytes per pixel fits a 100 Mbit/s link.
REVEAL
Round to pixels: B/s, which is about bit/s against the link's . More than ten times too much, so it cannot be done uncompressed and the rounding could not have changed the conclusion.
QUERYA conveyor moving 15 t/h up 2 m is quoted at 45 kW. Check it.
REVEAL
of lifting power. Losses might multiply that several times, so a few hundred watts to a kilowatt or two is credible. 45 kW is two orders of magnitude above the physics, so the quoted figure contains a mistake, not a pessimistic allowance.
QUERYName the two techniques that make an estimate better without making it slower.
REVEAL
Bracketing: produce a bound you are sure is too low and one you are sure is too high, and see whether the threshold falls inside or outside. Anchoring: derive an unknown from a figure you cannot forget, such as a person at 2 m and 80 kg, a kettle at 2 kW, or water at 1 kg per litre.
QUERYWhat does the size of a discrepancy tell you about its likely cause?
REVEAL
About 2 suggests neglected losses or efficiency. 10, 60, 1000 or 3600 suggests a units or prefix error. A factor of or 2 in a rotational problem suggests radians confused with revolutions. The size of the factor points to the kind of mistake before the working is reopened.
CConceptual Questions
Answer in your own words first, then compare with the model answer.
A detailed spreadsheet says a pump needs 7.2 kW and your thirty-second estimate says 700 W. Your colleague says the spreadsheet must be right because it accounts for far more effects. Evaluate that reasoning.
REVEAL MODEL ANSWER
The reasoning treats complexity as evidence of correctness, which reverses the real relationship. The spreadsheet has more steps, more inputs and more unit conversions than the estimate, so it has more places for a mistake to enter and no independent check on any of them. The estimate has few steps, each of which can be inspected in a moment. A disagreement by a factor of ten is therefore a reason to suspect the spreadsheet first. Ten is a typical factor of a prefix or conversion error, not of a physical effect. Real losses rarely multiply a figure by ten: pump and motor efficiencies together might account for a factor of two or three, and even a generous allowance for friction and fittings rarely reaches an order of magnitude. The next step is to reconcile the two results instead of choosing one. The spreadsheet's hydraulic power is the intermediate value that should match the estimate most closely, since it depends only on flow and head and on no efficiency assumption. If the spreadsheet's hydraulic power differs from the estimate by a factor of ten, the error lies upstream of every loss calculation and is almost certainly a unit.
Why is an estimate that lands close to the decision threshold still a useful result, given that it has not decided anything?
REVEAL MODEL ANSWER
A marginal estimate shows where the effort should go, which is scarce information at the start of a project. Most candidate designs are either clearly feasible or clearly impossible, and the estimate settles those in seconds, leaving a small number of marginal questions. Knowing that a question is marginal leads to action: the detailed calculation is worth an afternoon, the design has no margin so each assumption now matters, and a supplier's tolerance or a temperature coefficient could move the answer across the threshold. A design response is also available before any further calculation: change the problem so that it is no longer marginal, by choosing a faster link, a larger motor or a lower frame rate. A marginal estimate treated as a failure of the method leads to a design that rests on the detailed number alone, which is the situation in which a single wrong assumption decides the outcome with nothing to catch it.
Dimensional analysis and estimation are both quick checks. Explain what each catches that the other cannot, and why using both is not redundant.
REVEAL MODEL ANSWER
The two checks test different properties. Dimensional analysis tests the structure of a calculation: whether the quantities have been combined in a way that could describe the physics at all. It rejects a formula that divides where it should multiply, because the result carries the wrong units, and it does this without knowing any of the numbers. It cannot see any dimensionless quantity, so a factor of two, a missing efficiency, or an answer that is correct in form and a thousand times too large all pass it cleanly, because a prefix error leaves the units intact. Estimation tests magnitude: whether the number is in the right decade. It catches the factor of a thousand that the units check passed, and it catches a forgotten efficiency, but it cannot show that a height was multiplied where it should have been divided if the resulting magnitude happens to look plausible. The two failure modes are therefore close to complementary: structural errors that preserve magnitude, and magnitude errors that preserve structure. Running both costs a couple of minutes and covers most of the ways a calculation goes wrong before it reaches anyone who will act on it.
DPractice Problems
P1 (easy). Estimate, showing your rounding: (a) how many seconds there are in a year; (b) how much a cubic metre of water weighs; (c) how long a 2 kW kettle takes to raise 1 litre of water by 80 degrees, given that water needs about 4000 J per kg per degree.
P2 (medium). A warehouse robot carries 20 kg, moves at 1.5 m/s, and must run an 8 hour shift. (a) Estimate the mechanical power needed to overcome rolling resistance, taking the coefficient as roughly 0.02 and the robot's own mass as 60 kg. (b) Estimate the battery capacity needed at 24 V, assuming the drive is about 50% efficient overall and the electronics add roughly 20 W. (c) State which of your assumptions the answer is most sensitive to.
P3 (harder). A supplier claims their servo drive dissipates 12 W while holding a 5 N m load stationary on a 50:1 gearbox. (a) Estimate the torque the motor must hold, and explain why a stationary load still costs power. (b) Estimate the motor current if its torque constant is roughly 0.05 N m per amp, and the resistive loss if its winding resistance is about 1 ohm. (c) Compare with the claim and state what you would ask the supplier.
Solutionsclick to reveal
P1. (a) A year is about 365 days, a day about 24 hours, an hour 3600 s. Round: , so about . The exact figure is , so the estimate is within 5%. (b) A litre of water is 1 kg and a cubic metre is 1000 litres, so , which is exact rather than estimated because the metric system was designed that way. (c) Energy needed is J. At 2000 W that is , call it three minutes, which matches everyday experience and is the anchor being checked.
P2. (a) Total mass is kg, weight about N, rolling resistance about N. Power is force times speed: . (b) Mechanical 24 W at 50% efficiency is about 48 W drawn, plus 20 W of electronics, so roughly . Over 8 hours that is Wh, and at 24 V, , so specify something in the 25 to 30 Ah range for margin. (c) The answer is most sensitive to the duty cycle, which has been assumed to be 100%. A robot that actually drives for a third of the shift needs a third of this. After that, the efficiency figure and the rolling coefficient each move the answer by a factor of up to two, while the payload mass barely matters since it is a quarter of the total.
P3. (a) Holding torque at the motor is , ignoring gearbox friction, which in reality helps here by holding some of the load. A stationary load still costs power because the motor must produce torque, and in a conventional motor torque requires current, and current through a winding dissipates as heat. No mechanical work is done, since the speed is zero, so every watt drawn is heat. (b) Current is . Resistive loss is . (c) The estimate says about 4 W from the winding, plus drive electronics losses that might add a watt or two, so something in the region of 5 to 7 W. The claim of 12 W is roughly twice the estimate, which is the signature of neglected losses, not of a units error: gearbox friction, drive switching losses, a winding hotter than nominal and therefore more resistive, and possibly a higher torque constant tolerance. The question for the supplier is what the 12 W is measured at, specifically the winding temperature and whether the figure includes the drive's own consumption, since a factor of two here is explainable and worth confirming instead of disputing.
EFeynman Exercise
Explain to someone who thinks rough calculations are unprofessional: what an order-of-magnitude estimate is, why throwing away accuracy on purpose is the point rather than a compromise, and how you would use one to check a number somebody else gave you. Write your explanation, then compare.
REVEAL MODEL ANSWER
An estimate is a real calculation done with deliberately crude numbers. Everything is rounded to one digit, or to the nearest power of ten, and the arithmetic is done mentally. It is not a guess, because every step is written down and can be challenged. Accuracy is given up on purpose because most decisions are not close. In a choice between a 5 amp hour battery and a 50 amp hour battery, a rough sum that says 20 makes the decision in half a minute. Working out 19.4 instead of 20 would not change which battery is bought, so the extra work adds nothing. If the gap between the answer and the threshold is larger than the roughness introduced by the rounding, the question is decided, and if it is not, a proper calculation is needed. Both outcomes are useful. Estimation is most useful for checking someone else's number. Suppose a report says a conveyor motor needs 45 kilowatts to lift 15 tonnes an hour by two metres. The lifting power is fifteen thousand kilograms, times ten for gravity, times two metres, divided by the 3600 seconds in an hour, which is about 80 watts. After friction and drive losses, a few hundred watts would be expected, or a couple of kilowatts with a generous allowance. Forty-five kilowatts is hundreds of times the physics, so the report contains a mistake, and that is known before any line of it is read. The mistake still has to be found, but knowing that one exists is the hard part of reviewing someone else's work.
FError Analysis Framework
The mistake: accepting a detailed calculation without an independent check of its magnitude, and treating the disagreement between an estimate and a spreadsheet as automatically settled in the spreadsheet's favour.
Why it happens: a detailed result looks correct, with many digits, named intermediate quantities and a tidy layout, while an estimate looks like a scribble, so complexity reads as diligence. Challenging a colleague's spreadsheet with arithmetic done on the back of a drawing also has a social cost, which is why the estimate should be done before the detailed work exists.
How to recognize it: a result that disagrees with a rough check by a factor of 10, 60, 1000 or 3600; a specification whose power, mass or time is implausible against a familiar anchor; a design where nobody can say what the answer ought to be, only what the tool computed; a number that survived review because everyone checked the arithmetic and nobody checked the size.
How to correct it: estimate before calculating, so that the expectation exists before the detailed number can influence it. When the two disagree, reconcile them instead of choosing one: find an intermediate quantity that both methods compute and compare it, since the error lies upstream of the first point where they differ. Read the size of the discrepancy as a clue to the kind of error. Record the estimate next to the detailed result, so that the next reader of the work inherits the check.
GMini Challenge
A proposal states that a fleet of 40 autonomous floor-cleaning robots will save a supermarket 80 000 kWh of electricity a year compared with the mains-powered scrubbers they replace. Each robot carries a 48 V, 40 Ah battery and is charged once per shift, three shifts a day, every day. Estimate whether the claimed saving is even possible, and say what the estimate does and does not establish.
REVEAL MODEL ANSWER
A saving cannot exceed the energy the old machines used, and the new fleet's own consumption bounds the comparison, so the estimate starts with the new fleet. One battery holds , call it 2 kWh. Three charges a day give about 6 kWh per robot per day, and 40 robots give 240 kWh a day. Over a year that is , so roughly 88 000 kWh a year consumed by the new fleet, before charger losses, which would push it above 100 000. The claimed saving of 80 000 kWh is therefore almost the same size as the new fleet's own consumption. For the claim to hold, the old scrubbers must have used about 170 000 kWh a year, nearly double what the robots use for the same work. That is possible, since an operator-driven machine may be less efficient, run longer or clean areas that do not need it, but it is a strong claim that the proposal has not supported. The estimate establishes the scale of what is being asserted and the assumption on which the claim depends, which is the old machines' consumption, not anything about the robots. It does not establish whether the claim is true, because there is no figure for the old equipment and only the new fleet has been bounded. The question to ask is therefore about the previous machines: what they drew, for how many hours, and whether that figure was measured or assumed. An estimate of this kind moves a review from the impressive part of a proposal to the part that has not been examined.
Quiz Check
Answer without looking back.