Does having two possible outcomes mean each has probability one-half?

No. Their probabilities depend on the model and evidence.

Can a continuous uniform distribution assign zero probability to each exact point?

Yes; intervals can still have positive probability.

Is this the same as indifference between two products?

No. Equal preference concerns desirability; this principle concerns probability assignments.

AI Glossary / Graphs and symbolic AI / Advanced

Principle of indifference

Definition

The principle of indifference assigns equal probabilities to alternatives when the available information provides no reason to favor one over another. For a specified finite set of n mutually exclusive and exhaustive alternatives, this gives each probability 1/n. Also called the principle of insufficient reason, it requires care about which alternatives are being compared. Lack of information alone does not establish that a physical process produces equal frequencies.

What it means

Equal probabilities can express symmetry among the chosen possibilities. However, the choice of possibilities is itself part of the model. Dividing the same underlying outcomes into different groups and assigning equal probabilities to those groups can produce incompatible answers. The principle is therefore not a universal instruction to assign the same probability to every event that can be named.

How it works

Identify alternatives that cannot occur together and collectively cover the relevant possibilities. Specify the information supporting their equal treatment. With n finite alternatives, normalization gives probability 1/n to each, and an event containing k of them has probability k/n.
When evidence distinguishes the alternatives, reconsider the equal assignment. For continuous quantities, 'uniform' must refer to a specified variable over a specified range or to an appropriate measure. A distribution uniform in one variable generally does not remain uniform after a nonlinear transformation. Merely saying that every exact value has probability zero also fails to identify a unique continuous distribution. These limitations concern how a probability model is selected, not a failure of probability arithmetic once a consistent model has been specified.

Three practical examples

A test harness deliberately selects one of four configurations with equal probability. Its implemented randomization supports a uniform model over those configurations, while a production workload may follow different proportions.
An analyst uses equal prior probabilities for a small set of mutually exclusive hypotheses as a provisional assumption. The analyst documents that choice and examines whether later evidence or domain knowledge favors some hypotheses.
A simulation designer distinguishes random selection of an individual machine from random selection of a machine category. Categories containing different numbers of machines need not receive equal probabilities under uniform machine selection.

Think of it this way

Imagine placing one ticket for each option into a well-mixed draw. Equal chances follow from the chosen ticket arrangement and draw model. Combining several tickets under one label does not make that label as likely as a label containing only one ticket. The labels and the elementary possibilities are different levels of description.

Worked example

A fictional test system has four servers: A, B, C and D. Its selection procedure chooses each server with probability 1/4. Servers A, B and C are in the North group; D is in the South group.
The probability of selecting North is 1/4 + 1/4 + 1/4 = 3/4, while South has probability 1/4. These probabilities follow from the stated uniform selection among four individual servers. Treating the two group names as equally likely would instead give North and South probability 1/2 each, contradicting that selection model.
There is a different procedure that does produce equal group probabilities: first choose North or South with probability 1/2, then choose uniformly among servers within the selected group. Under this procedure, A, B and C each have probability (1/2)(1/3) = 1/6, while D has probability 1/2. The probabilities sum to 1, but they describe a different selection mechanism. Neither procedure is made equivalent to the other simply by calling both 'random.'
Suppose handling a test on A, B or C consumes 2 units of a resource, while handling it on D consumes 8. Uniform server selection gives expected consumption (3/4)(2) + (1/4)(8) = 3.5 units. Uniform group selection gives (1/2)(2) + (1/2)(8) = 5 units. The modeling choice therefore changes a practical forecast.
For a separate continuous illustration, let x lie between 0 and 1. If x is uniform, the probability that x is at most 0.5 is 0.5. Instead assume y = x squared is uniform between 0 and 1. The same event is y at most 0.25, so its probability is 0.25. This is a change of distributional assumption under a nonlinear transformation, not merely converting meters to centimeters.

Illustrative case study

In a fictional infrastructure forecast, one analyst assumes equal traffic among four servers while another assumes equal traffic between their two groups. Their resource estimates disagree at 3.5 versus 5 units per test. Both calculations are internally consistent; they answer questions about different traffic-generating procedures.
The team documents the actual routing rule and checks observed allocations against it. If routing later changes, it revises the probability model rather than retaining equal weights by habit. The investigation also records uncertainty about future workload separately from the mechanism used to allocate individual tests.

Common misconceptions and mistakes

Unknown probabilities are not automatically equal probabilities. Likewise, an event and its complement need not have probability one-half merely because there are two verbal descriptions. Uniformity over elementary outcomes does not imply uniformity over arbitrary groups of outcomes.
Another mistake is assuming that a uniform distribution remains uniform under every change of variable. Linear unit conversions preserve uniformity over the correspondingly transformed interval; the squared-variable example is nonlinear. Finally, equal prior probabilities do not prevent unequal posterior probabilities after evidence arrives. An initial symmetry assumption is not a commitment to ignore informative observations.

Business use

The principle can help articulate provisional assumptions in simulations and decision models, but those assumptions should be visible. Name the alternatives, explain why equal treatment is being considered and test the effect of plausible unequal weights. Distinguish an intentionally uniform experimental procedure from a forecast of naturally occurring demand. When results depend strongly on the chosen partition or parameterization, obtain better information or report the sensitivity instead of presenting one uniform model as an established fact.

Check your understanding

Under uniform server selection, what is North's probability?

Answer: Three-quarters.

Under uniform group selection, what is D's probability?

Answer: One-half.

Why do the continuous examples disagree?

Answer: Uniformity is imposed on different variables, creating different probability distributions.

Related terms

Uniform distribution
Prior probability
Sample space
Partition
Sensitivity analysis
symmetry
probability measure
Bayesian updating
Bertrand's paradox

Learn next

Move server C to the South group and recalculate both selection procedures. Explain why the probabilities then agree in this example and why equal group size is relevant.

Sources

Conor Mayo-Wilson, Classical Probability and the Principle of Indifference: https://faculty.washington.edu/conormw/Teaching/Files/Probability/Lectures/2_Prob_Lecture.pdf AIMA glossary, principle of indifference: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Principle of contrast Next: Principle of marginality