Can zero be a fixed point?

Yes, whenever T(0) = 0, but many mappings have nonzero fixed points.

Does repeated application always find one?

No. A map can have a fixed point while its iteration diverges or cycles from some starting values.

Is a small residual the same as a small error?

Not generally. An error bound needs additional properties of the map; the worked affine example has an explicit relation.

AI Glossary / Mathematics and optimization / Intermediate

Fixed point

Definition

A fixed point of a mapping is an input that the mapping leaves unchanged. If the mapping is T, a value x* is a fixed point when T(x*) = x*. The value can be a number, vector or more complex object. Fixed points help describe equilibria and stopping states in iterative algorithms, but being fixed does not by itself mean that a state is desirable, unique or reachable from a chosen starting point.

What it means

The word fixed refers to one particular transformation. A value can be fixed under one rule and changed by another. For example, multiplying by one leaves every number fixed, whereas adding one leaves no real number fixed. In machine learning and optimization, an update rule may have fixed points corresponding to solutions or stationary states. Establishing that correspondence requires an argument about the rule. The mere absence of change is not proof that a trained model is accurate or that a numerical procedure has solved the intended problem.

How it works

To find fixed points, solve T(x) = x within the allowed domain. Equivalently, solve the root equation T(x) − x = 0. An iterative procedure may try x_(k+1) = T(x_k), but the existence of a solution does not ensure that this procedure reaches it. The map's behavior near the solution determines whether nearby iterates move toward it, move away or cycle. Existence, uniqueness and convergence are separate questions. A numerical candidate should also be checked through its residual, the difference between applying the map and leaving the candidate unchanged.

Three practical examples

An iterative calibration rule can have an equilibrium setting at which another update produces no change.
A differentiable optimization method can express stationary points as fixed points of an appropriate gradient-step map.
A repeated evaluation procedure can track the norm of T(x) − x to measure how nearly its current candidate satisfies the fixed-point equation.

Think of it this way

Think of a thermostat setting that, under a specified adjustment rule, receives no further adjustment. It is fixed relative to that rule. The analogy is limited because physical thermostats have time delays, external disturbances and changing conditions. A mathematical fixed point is an exact relation for a defined mapping, and a stable numerical state need not represent comfortable or efficient operation.

Worked example

Let T(x) = 0.5x + 3 on the real numbers. Solving x = 0.5x + 3 gives x* = 6. Starting the iteration at x_0 = 0 produces 3, 4.5, 5.25 and 5.625. Each error from 6 is half the previous error. At x = 5.625, the residual is T(x) − x = 5.8125 − 5.625 = 0.1875, while the distance to the fixed point is 0.375. They are related but not equal. Now define S(x) = 2x − 6. It has the same fixed point 6, but starting at 5 produces 4, 2, −2 and −10, moving away. The equation's solution is unchanged while the iteration's behavior differs. Finally, T(x) = x² has fixed points 0 and 1, since x² − x = x(x − 1). Thus a map can have multiple fixed points. The translation T(x) = x + 1 has none on the real line. These examples show why a fixed-point claim must name the mapping and domain rather than merely report a number that stopped changing in software.

Illustrative case study

In a fictional analytics review, Oak Calibration uses repeated updates to estimate a sensor offset. The implementation stops when consecutive values differ by less than a chosen tolerance. A reviewer also evaluates the update residual at the returned value and checks the assumptions linking that value to the calibration objective. They discover that one alternative update has the correct fixed point but moves away from it near the initial guess. The case illustrates verification of a method, not a reported improvement in sensor accuracy.

Common misconceptions and mistakes

A fixed point is not the same as a constant function: a nonconstant function can leave particular inputs unchanged. It is not necessarily a root of T itself; it is a root of T(x) − x. A fixed point need not be attractive or unique. In optimization, a zero gradient can correspond to a maximum or saddle point when the objective is nonconvex. Finite-precision stagnation can also make successive stored values identical even when the mathematical equation is not satisfied to the required accuracy. Always distinguish exact statements from numerical tolerances.

Business use

When a vendor or technical team describes an iterative method, ask what its fixed points represent in the business model. Document the domain, update rule and termination criterion. Compare candidates against independent constraints or known small cases. Monitor residuals and resource limits, and report failure to meet the intended criterion instead of silently treating a maximum iteration count as success. If several fixed points exist, investigate how initialization affects which one is obtained. A mathematically consistent equilibrium still needs evaluation against the actual decision or prediction objective.

Check your understanding

Does the fixed point 6 make the two worked affine iterations equally reliable from arbitrary starting values?

Answer: No. The first halves the error, while the second doubles it. Sharing a fixed point does not mean sharing convergence behavior.

Related terms

Fixed-point equation
Fixed-point iteration
Residual
contraction mapping
equilibrium
stationary point

Learn next

draw y = T(x) and y = x for the two affine examples and locate their intersection. Then calculate several iterates on either side of 6. Explain why an intersection establishes a fixed point but does not establish that the iteration approaches it.

Sources

Sources: Alexander Jung, Federated Learning: From Theory to Practice, Section 4.8, https://arxiv.org/pdf/2505.19183 ; Parikh and Boyd, Proximal Algorithms, Section 2.3, https://stanford.edu/~boyd/papers/pdf/prox_algs.pdf ; SciPy, fixed_point reference, https://docs.scipy.org/doc/scipy/reference/generated/scipy.optimize.fixed_point.html . The affine comparisons, residual arithmetic and fictional calibration review are original.

Previous: Fixed model Anova Next: Fixed-point characterization