How Mathematics Improves The World | Searching the Right Place First When Someone Is Missing
Someone is missing.
The search area is large.
Time is not.
A person may be near the last known point.
Or have followed a trail.
Or turned downhill.
Or moved toward water.
A phone ping may place them somewhere inside a broad sector.
A witness report may be reliable.
Or mistaken.
A helicopter can search open ground quickly.
Dense forest is slower.
A person can be inside a searched area and still not be seen.
Search and rescue therefore faces two different uncertainties:
- Where is the person likely to be?
- If we search there, how likely are we to detect them?
Mathematical search theory keeps those questions separate, then combines them.
That distinction can decide where scarce search effort goes first.
Quick Read
Search planning divides a possible search region into areas or probability cells. Each area receives a Probability of Containment or, in some land-search terminology, Probability of Area: the probability the missing person or search object is actually there. A second quantity, Probability of Detection, estimates the probability that the search method would detect the person if they were in that area.
The U.S. Coast Guard defines Probability of Success as the product of those two terms:
POS = POC × POD.
A region can have high probability of containing the person and still be a poor immediate search choice if detection is extremely difficult. Another may have lower containment probability but be quick and highly searchable. Search allocation therefore considers both location likelihood and search effectiveness.
After an unsuccessful search, the probability in that area should fall—but not necessarily to zero. If POD was 60%, failing to find the person still leaves a 40% chance of missing them even if they were there. Bayes’ theorem updates the map. New clues such as phone data, tracks or witness information can update it again. A 2026 paper in the Journal of Search and Rescue develops this logic explicitly, maintaining a spatial belief map, updating it after unsuccessful searches, and prioritising areas using POA × POD.
This is decision support, not an instruction manual replacing trained search leadership. Actual search-and-rescue operations depend on jurisdiction, incident command, terrain, weather, medical urgency, available teams and established professional protocols.
One-sentence answer: Mathematics improves the world by converting incomplete clues and imperfect search coverage into an updateable probability map, helping trained search planners place limited people, aircraft, dogs and time where they have the greatest justified chance of finding the missing person.
The Search Area Is Not Equally Likely Everywhere
A person was last seen at a trailhead.
Ten kilometres away in every direction is theoretically possible.
It is not equally plausible.
Trails channel movement.
Cliffs block it.
Roads attract or redirect it.
Water sources matter.
Age, equipment, experience and circumstances affect likely travel.
Search planning therefore constructs a probability distribution rather than colouring the whole map equally.
Probability of Containment: Where Do We Think the Person Is?
Let the search region be divided into areas A1, A2, …, An.
Assign probabilities:
P(A1) + P(A2) + … + P(An) = 1
if the set covers all current possibilities represented by the model.
The U.S. Coast Guard uses the term Probability of Containment, POC, for this concept.
Land SAR practice has also used Probability of Area, POA.
The number represents belief based on available evidence, not a physical property of the forest.
Probability Maps Combine Different Kinds of Evidence
Possible evidence includes:
- last known position;
- intended route;
- witness information;
- tracks or personal items;
- phone or beacon information when lawfully available;
- terrain;
- weather;
- lost-person behaviour data;
- movement constraints.
Some evidence is quantitative.
Some comes from trained judgement.
The search map is a structured belief state.
It should change when evidence changes.
Probability of Detection: Searching an Area Does Not Guarantee Finding the Person
A search team walks through an open field.
Detection may be high.
The same team searches thick scrub.
Detection falls.
A helicopter sees a bright object in open terrain.
It may miss a person beneath canopy.
A dog and a visual line search have different sensing mechanisms.
POD is conditional:
POD = P(detect | object is in searched area).
It measures search effectiveness, not location likelihood.
Probability of Success: Multiply Location and Detection
The U.S. Coast Guard states the simple relationship:
POS = POC × POD.
Area A:
POC = 60%.
POD = 20%.
Immediate POS = 12%.
Area B:
POC = 30%.
POD = 70%.
Immediate POS = 21%.
Area B has lower probability of containing the person and higher immediate chance of a successful search.
That does not mean B must always be searched first.
Travel time, urgency, clues and future search effects matter too.
But the multiplication exposes a trade-off intuition alone can miss.
An Unsuccessful Search Does Not Make Probability Zero
Area A has POC 40%.
A team searches with POD 50%.
No person is found.
Can A now be eliminated?
No.
If the person was there, the search had a 50% chance of missing them.
The negative result is evidence against A.
It is not proof of absence.
This is why Bayesian updating is so important in search theory.
Bayes After a Failed Search
Suppose area A initially has probability p and search POD is q.
Probability of an unsuccessful search is:
P(no find) = 1 − pq.
The updated probability that the person remains in A given no find is:
P(A | no find) = p(1−q)/(1−pq).
Other areas gain relative probability because the searched area became less likely.
This exact logic appears in land-search probability-shifting guidance and modern Bayesian SAR research.
A Numerical Example
Initial P(A)=0.4.
POD=0.5.
No find.
P(A | no find)=0.4×0.5/(1−0.4×0.5)=0.2/0.8=0.25.
Area A falls from 40% to 25%.
It does not fall to zero.
A second thorough search may still be rational, especially if new evidence raises its probability or other areas are exhausted.
The Negative Result Has Strength
Search A with POD 10%.
No find.
Weak evidence against A.
Search A with POD 95%.
No find.
Strong evidence against A.
“We searched there” is incomplete information.
The meaningful statement is:
We searched there with approximately this probability of detection and did not find the person.
Search quality changes evidential value.
Coverage: Search Effort Per Area
Search theory relates detection to coverage.
Conceptually:
coverage = effective searched area / target area.
For parallel search tracks, coverage factor is often expressed using effective sweep width W and track spacing S:
C = W/S.
Closer spacing raises coverage and generally raises POD.
It also consumes more time.
Search effort is a resource-allocation problem.
Sweep Width: A Sensor Has an Effective Reach
Searchers do not detect everything inside a sharp geometric strip.
Detection probability falls with distance from the search track.
Effective sweep width compresses that detection curve into an equivalent width.
It differs by:
- terrain;
- target visibility;
- search modality;
- weather;
- lighting;
- searcher capability.
Research continues to validate how well sweep-width models estimate POD for ground teams and dogs.
Measurement uncertainty should remain attached to POD estimates.
Search Effort Has Diminishing Returns
First pass through an area can find obvious clues.
Second pass increases cumulative detection.
Eventually each additional hour produces smaller probability gain.
If independent searches each have POD q, cumulative POD after n identical independent passes is:
1 − (1−q)n.
Real searches are not perfectly independent.
The formula still reveals diminishing residual probability.
At some point, moving effort elsewhere can produce greater expected success.
Search Allocation: Spend the Next Hour Where It Buys the Most Success Probability
Sector A is highly likely and difficult.
Sector B is moderately likely and easy.
Sector C is unlikely but contains a dangerous cliff where survival time may be short.
Pure POS maximisation is one useful lens.
Real incident command can also consider urgency, rescue feasibility, clue value and safety of search personnel.
The objective function must represent the mission, not merely mathematical convenience.
Bayesian Updating With a New Clue
A witness reports seeing the missing person near Sector C.
How should probabilities change?
Ask how likely that report would be if the person were in each sector.
Bayes:
P(C | report) ∝ P(report | C)P(C).
A highly reliable clue can dominate prior beliefs.
An ambiguous clue should move them modestly.
Evidence has weight proportional to how differently it would behave under competing hypotheses.
Likelihood Ratios: How Informative Is This Clue?
Suppose a clue is five times more likely if the person is in Sector A than if they are elsewhere.
Likelihood ratio = 5.
Posterior odds equal prior odds multiplied by the likelihood ratio.
posterior odds = prior odds × likelihood ratio.
This form of Bayes is intuitive for integrating clues.
Strong evidence multiplies odds.
Weak evidence barely moves them.
Cell Phone Location: A Circle Is Not a Point
A network observation may place a device inside a broad region.
GNSS information may be more precise when available.
Different technologies have different error distributions.
Search planners should treat location data as uncertainty regions rather than precise dots unless the technology supports that precision.
The 2023 Bayesian GIS search-and-rescue case study demonstrates combining a cell-tower ping with other evidence to update a lost-person probability map.
A map marker is a probability surface wearing an icon.
Lost-Person Behaviour: Population Statistics Become Priors
Historical searches show patterns.
Different categories of missing persons tend to travel different distances and use terrain differently.
Such data can inform prior probability distributions.
But population patterns do not determine one individual.
A prior is a starting distribution.
Specific evidence should update it.
Statistics supports judgement precisely by admitting exceptions.
Terrain as a Probability Model
Steep cliffs reduce movement in one direction.
Trails increase it along another.
Rivers can attract or block.
Dense vegetation changes detection probability separately from movement probability.
Geographic information systems encode slope, land cover, roads and hydrology.
One map layer estimates where the person may move.
Another estimates how hard they are to detect there.
Do not confuse the two.
Movement Models: The Person May Not Stay Where the Map Put Them
A static probability map assumes the subject remains fixed.
A mobile person changes location while teams search.
Transition probabilities can move probability mass among neighbouring cells through time.
pt+1 = T pt
where T is a movement transition matrix in a simplified discrete model.
Paths, slope and behaviour modify T.
Search planning becomes a hidden-state tracking problem.
Particles: Simulate Thousands of Possible People
Another approach represents location uncertainty with many simulated particles.
Each particle is one plausible subject path.
Move particles according to behavioural and environmental models.
Weight them according to new evidence.
Remove or reduce particles in thoroughly searched regions.
Particle filtering is used widely for nonlinear tracking problems.
Search theory and navigation share mathematical machinery.
Search Paths: Where Should the Aircraft Fly?
Once a high-priority area is selected, a search pattern must cover it efficiently.
Parallel sweeps.
Expanding squares.
Sector searches.
Actual pattern choice is governed by professional SAR doctrine, terrain and platform.
The mathematical core is coverage:
place the sensor path so high-probability area receives effective search effort without wasting excessive distance on already low-value regions.
Travel Time Is Part of Search Cost
Sector A has highest POS but requires two hours to reach.
Sector B is slightly lower and ten minutes away.
If survival probability decreases quickly with time, B can become the rational first action.
Search optimisation may therefore maximise expected probability of timely rescue, not merely probability of eventual detection.
Time changes the objective.
Survival Curves: Finding Later Is Not the Same Outcome
Exposure, injury, dehydration and weather can reduce survival probability over time.
A sector with slightly lower immediate location probability but much faster access can have greater expected rescue value.
The objective can be conceptualised as:
expected rescue benefit = location probability × detection probability × survival value at detection time.
Actual incident management uses richer operational judgement.
The Mathematics reveals why urgency belongs inside allocation.
Search Team Safety Is a Hard Constraint
A high-probability cliff sector cannot simply be searched recklessly.
Weather.
Avalanche.
Night.
Water.
Fire.
Incident commanders balance rescue urgency against risk to responders.
A mathematical optimiser should treat unsafe options as infeasible or heavily constrained, not merely expensive.
The rescuer is not a disposable search resource.
Multi-Team Allocation: Different Sensors Belong in Different Places
A drone may search open terrain quickly.
A dog team may perform better in another environment.
Technical rope teams access steep terrain.
Ground searchers inspect dense areas slowly.
Each team has a sector-specific POD and cost.
Assignment becomes a matrix:
POD(team i, sector j).
The best search plan matches sensing capability to terrain rather than treating every team as interchangeable.
Multiple Clues Can Be Correlated
Two witnesses repeat the same story.
Independent evidence?
Maybe they spoke to each other.
Two phone-derived locations may rely on the same underlying network observation.
Multiplying likelihood ratios as if evidence were independent can overstate certainty.
Bayesian models need dependency awareness.
Two copies of one clue are not two clues.
False Clues: Evidence Can Move Probability the Wrong Way
A footprint belongs to another hiker.
A witness misremembers clothing.
A phone is not with the missing person.
Evidence models should include false-positive probability.
A clue’s value depends on specificity.
How often would we see this clue if the person were not here?
Bayes forces that counter-question.
Search Logs: Memory Must Be External
Large searches involve many teams and operational periods.
Where did Team 6 actually walk?
At what spacing?
What visibility?
Which clues were found?
GPS tracks and search records preserve coverage history.
Without logs, probability updating relies on imperfect human memory.
A search map should remember what has actually been searched.
Re-Searching Can Be Rational
An area was searched yesterday.
Why search again?
Because:
- POD may have been low;
- the subject may have moved into it later;
- new evidence may raise its POC;
- a new sensor may have higher POD;
- lighting or weather may improve.
The 2026 Bayesian SAR framework emphasises exactly this point: when POD is below 1, an unsuccessful search does not erase posterior probability.
“Already searched” is not a binary state.
Search Saturation: Eventually Another Area Becomes More Valuable
Search the same easy field repeatedly.
Its posterior probability falls.
Cumulative POD rises.
At some point, an initially lower-probability forest becomes the better allocation of the next hour.
Bayesian updating naturally redirects effort as evidence accumulates.
Search strategy should move with belief.
Information Search and Physical Search Compete for Resources
Send ten more people into the field?
Or spend thirty minutes verifying a phone record that could eliminate half the search area?
Investigation can be more valuable than immediate physical coverage when a clue has high information value.
The Coast Guard land-SAR guidance emphasises that investigation should continue throughout the search because new information can reshape the search area.
Searching for information is part of searching for the person.
Value of Information: Which Question Should We Answer Next?
A database check can confirm whether the missing person carried a vehicle.
A witness interview can distinguish two trails.
Which investigation is worth doing first?
Decision theory calculates expected value of information:
expected improvement in decisions after observing the information − cost of obtaining it.
The best next action may be a question, not a search line.
POMDP Thinking: Search When State Is Hidden and Actions Change Future Knowledge
The person’s true location is hidden.
The team maintains a belief distribution.
An action searches one area.
The result changes the belief.
Then another action is chosen.
This is the structure of a partially observable Markov decision process.
The 2026 Journal of Search and Rescue framework points toward this longer-horizon formulation.
The optimal action can be valuable partly because of what it teaches, not only what it immediately finds.
Monte Carlo Search Planning: Simulate Thousands of Possible Outcomes
Sample a plausible initial location.
Simulate movement.
Simulate weather.
Simulate search detection.
Repeat thousands of times.
Compare candidate search allocations by expected discovery time or probability of success.
Simulation is useful when analytical equations become too complicated.
It should remain anchored to realistic behaviour and detection models.
Map Resolution: Smaller Cells Are Not Automatically Better
Divide a search region into one-hectare cells.
Very detailed.
Do we have enough evidence to distinguish probability at that resolution?
Maybe not.
False precision appears when grid resolution exceeds data resolution.
Search regions should reflect meaningful changes in probability and terrain, not simply the smallest GIS pixel available.
Calibration: Does a 30% Search Area Really Contain the Subject 30% of the Time?
Probability maps can be tested across historical incidents.
Among sectors assigned 30% probability, how often did the subject actually lie there?
If only 5%, planners are overconfident.
If 60%, underconfident.
Calibration turns expert judgement into a measurable forecasting system.
Good judgement should improve through feedback.
Human Bias: Search Planners Are Not Immune to Psychology
People anchor on the first theory.
Overweight vivid clues.
Under-update after failed searches.
Or clear an area mentally because “we already looked there”.
Research on SAR probability judgement has documented biases in how planners update probabilities.
A formal Bayesian ledger does not remove judgement.
It gives judgement a structure that can be checked.
Consensus Mapping: Several Minds, One Distribution
Search planners may independently assign probabilities, then discuss differences.
One knows local terrain.
One knows behavioural statistics.
One knows the investigation.
Structured consensus can reduce domination by the loudest voice.
The final map remains subjective in part.
Subjective probability can still be rigorous when assumptions are documented and updated coherently.
Search Theory Has Famous Historical Examples
Bayesian search methods have been associated with difficult searches for lost submarines, aircraft and other objects.
The general method is reusable:
- construct plausible scenarios;
- assign location probabilities;
- model detection;
- allocate search effort;
- update after results.
Whether the object is a vessel at sea or a missing hiker, uncertainty has the same mathematical grammar.
A Classroom Thought Experiment: Search the Hidden Coin
Hide one coin in one of four opaque boxes.
Prior probabilities:
A 40%, B 30%, C 20%, D 10%.
Searching A has only 50% detection because it contains shredded paper.
Searching B has 90% detection.
Immediate POS:
A = 20%.
B = 27%.
Search B first.
If not found, update probabilities.
Students experience the difference between “most likely location” and “best place to search next”.
A Second Thought Experiment: Why “Not Found” Is Not “Not There”
Hide a small bead in tall grass inside a tray.
Give a student ten seconds to search.
They fail.
Was the bead absent?
No.
Repeat many times and estimate POD.
Now an unsuccessful search has a measurable strength.
The experiment teaches evidence under imperfect detection.
Primary Mathematics: Search Begins With Fractions on a Map
Primary students can understand:
- fractions;
- percentages;
- maps;
- area;
- multiplication;
- updating after evidence.
40% chance here.
50% chance of seeing the person if there.
20% chance of immediate success.
A rescue problem begins with elementary probability.
Secondary Mathematics: Bayes Enters the Search Map
Secondary students add:
- conditional probability;
- Bayes’ theorem;
- exponential detection models;
- matrices;
- optimisation;
- expected value.
A failed search becomes conditional probability.
Movement becomes a transition matrix.
Team assignment becomes optimisation.
Search theory becomes applied decision science.
Advanced Mathematics: Search as Bayesian Sequential Decision-Making
Modern search theory draws on:
- Bayesian inference;
- probability maps;
- operations research;
- optimal control;
- Markov models;
- particle filters;
- Monte Carlo simulation;
- decision theory;
- POMDPs.
The hidden state is location.
Search actions generate imperfect observations.
Belief updates.
The next action changes.
This is the mathematics of searching while learning.
Why This Improves the World
1. It prevents search effort being spread evenly when probability is not even
Probability maps focus scarce resources where evidence supports them most strongly.
2. It separates location uncertainty from detection difficulty
A difficult high-probability sector and an easy moderate-probability sector can be compared explicitly.
3. It learns from unsuccessful searches
Bayesian updating reduces—but does not automatically erase—the probability of areas already searched.
4. It makes different search assets comparable
Sector-specific POD and time cost help planners match teams, aircraft and sensors to terrain.
5. It makes new clues operational immediately
Evidence can update the probability map instead of remaining only a narrative note.
6. It makes uncertainty discussable
Incident teams can challenge probabilities, assumptions and detection estimates explicitly instead of arguing from intuition alone.
What Mathematics Does Not Do
A probability map does not know where the person truly is.
It does not make unreliable clues reliable.
It does not make POD estimates exact.
It does not replace trained incident commanders or jurisdictional SAR procedures.
It does not justify unsafe search assignments.
It does not turn population behaviour statistics into certainty about one individual.
And a sector with the highest numerical score is not automatically the correct operational choice if the model omitted urgency, safety, access or critical evidence.
Frequently Asked Questions
What is Probability of Area or Probability of Containment?
It is the estimated probability that the missing person or search object is inside a particular search area, given current evidence and assumptions.
What is Probability of Detection?
POD is the conditional probability that a search would detect the person if the person were actually in the searched area. Terrain, sensor, visibility and search effort affect it.
What is Probability of Success?
In basic search theory, POS is the probability the target is in the area multiplied by the probability of detecting it there: POS = POC × POD.
If an area was searched and nothing was found, should it be eliminated?
Not automatically. Unless POD was effectively 100%, the target could have been present and missed. Bayesian updating reduces the area’s probability according to search effectiveness.
Why would search teams revisit an area?
Because the previous POD may have been low, the subject may have moved, new evidence may raise the area’s probability, or a different search method may detect the person more effectively.
Sources and Further Reading
- U.S. Coast Guard, Search and Rescue Addendum, defining search effectiveness and the relationship POS = POC × POD.
- U.S. Coast Guard, The Theory of Search, explaining the scientific basis of search planning.
- Rossmo, Velarde and Mahood, Optimizing Wilderness Search and Rescue: A Bayesian GIS Analysis, 2023.
- Dutta, Sherlock Bayes: The Curious Case of the Vanishing Posterior, Journal of Search and Rescue, 2026, on Bayesian POA updating, POD and sequential sector prioritisation.
Continue Through eduKateSG
Continue with How Mathematics Works. Compare this article with Counting Wild Animals Without Seeing Every Animal: both separate presence from detection probability. It also connects to Finding a Planet We Cannot See, where indirect evidence and Bayesian reasoning again recover something hidden from incomplete observations.
Final Thought: Search Is the Mathematics of Not Knowing Yet
The person is somewhere.
The map does not know where.
So probability is spread across it.
A team searches one sector.
Nothing.
The map changes.
A clue arrives.
The map changes again.
A helicopter can see one terrain well.
A ground team can search another better.
Resources move.
Belief narrows.
Until probability becomes a person found at one real coordinate.
Mathematics improves the world here by giving disciplined form to uncertainty while there is still time to act.