Exponential decay is the decrease of a quantity at a rate proportional to its current value. In its standard continuous form, it is described by
where is the initial quantity, is elapsed time, and is the decay constant. The defining feature is not a constant amount lost per unit time, but a constant fraction lost over any fixed interval. A quantity that decreases by half every hour, for example, follows the same proportional rule whether its initial value is one thousand or one million.
The model connects elementary calculus with physical mechanisms, statistical distributions, and the dynamics of complex systems. It describes radioactive populations, first-order chemical reactions, electrical discharge, and certain forms of thermal relaxation. In probability, the same exponential form describes survival under a constant event rate. Its usefulness depends on specific assumptions: a proportional removal mechanism, a constant rate parameter, and an appropriate choice of the quantity that is decaying. Departures from those assumptions lead to mixtures of exponentials, changing-rate models, or fundamentally different decay laws.
Mathematical definition and characterization
The exponential function gives the standard decay law
with and . Here is Euler’s number, approximately . The exponent must be dimensionless, so has units reciprocal to those of time. If time is measured in seconds, is measured in inverse seconds. The same mathematics applies when the independent variable is distance, depth, or another nonnegative coordinate; the coefficient then has the corresponding inverse units.
The governing ordinary differential equation is
Its negative sign expresses loss, while the factor makes that loss proportional to the amount currently present. For positive , separating variables and integrating gives
and hence
Equivalently, the logarithmic rate of change is constant:
This formulation distinguishes proportional decay from a linear decrease, for which itself is constant.
A second characterization concerns successive intervals. For any ,
Thus the normalized remaining fraction satisfies
Among positive continuous functions, this multiplicative rule implies an exponential form. Continuity can be replaced by other mild regularity conditions, such as measurability or monotonicity. The rule expresses a time-homogeneous mechanism: an interval of length multiplies the quantity by the same factor regardless of when the interval begins.
For positive initial values, the curve is strictly decreasing and convex:
Its slope becomes progressively less negative as the remaining amount shrinks. It approaches zero as but never reaches zero at finite time. This is a property of the continuous model, not a claim that a finite collection of physical objects can contain fractions of its last object. More generally, can represent a signed deviation from equilibrium. A negative deviation then approaches zero from below, with its magnitude decaying exponentially even though the signed value increases.
Half-life, time constant, and measures of duration
Several equivalent parameters express the speed of exponential decay. The half-life is the time required for the remaining quantity to fall to half its current value. Substitution into the decay law gives
The same half-life applies after any elapsed time. After half-lives,
Five half-lives leave of the initial quantity, while ten leave approximately . These percentages are independent of the initial amount.
The time constant, usually denoted , is
At , the remaining fraction is , approximately ; about has been lost. Consequently,
The term e-folding time often denotes this same interval. It is especially convenient in differential equations because the normalized exponent becomes simply . A process with a time constant of five seconds has a half-life of approximately seconds, not five seconds.
For any prescribed remaining fraction , where , the required duration is
Reaching , , and of the initial value requires approximately , , and , respectively. Each additional tenfold reduction therefore requires the same extra duration. If a detector registers a quantity only while it exceeds a threshold , the ideal threshold-crossing time is
Thus practical disappearance depends on the detection threshold even though the mathematical curve never vanishes.
An integral provides another useful measure:
The area under the normalized curve is therefore the time constant. When that curve represents a survival probability, this area equals the mean lifetime. When it represents concentration, the area is an integrated exposure rather than a duration by itself. These interpretations share the same formula but differ dimensionally. The decay constant, time constant, half-life, and integrated area should consequently not be interchanged without specifying the quantity being measured and its units.
Discrete losses and the continuous-time limit
Many processes are observed or defined at separated times. If the same fraction , with , is removed during each interval, the remaining values obey the recurrence relation
Its solution is
This geometric sequence is the discrete counterpart of continuous exponential decay. A stock reduced by at each step retains of its initial amount after steps. The absolute losses become smaller at each step even though the proportional loss remains unchanged.
If each step lasts , the continuous exponential passing through the discrete values has rate
Conversely, a continuous decay constant implies a fractional loss
over each sampling interval. These relations are exact. The familiar approximation requires . For instance, a loss per hour corresponds to , not . Percentage changes compound multiplicatively; they cannot generally be added across intervals.
The continuous law also arises as a limiting process:
For sufficiently large , each factor represents a small proportional loss over a short interval. This construction connects exponential decay to repeated multiplication and explains why many microscopic removal mechanisms produce an exponential macroscopic law. It also shows why the differential equation is appropriate when the removal fraction in a very short interval is approximately proportional to that interval’s duration.
Discrete approximation introduces a separate issue in numerical analysis. Applying the forward Euler method to the differential equation gives
where is the numerical step size. To keep nonnegative initial values nonnegative requires . Absolute stability requires , but the range produces alternating signs that have no interpretation as a nonnegative decaying population. By contrast, the exact update remains positive for every . Distinguishing a genuine discrete removal law from a numerical approximation prevents errors in interpreting percentage losses and in simulating rapidly decaying systems.
Random lifetimes and constant hazard
Exponential decay has a probabilistic interpretation when individual objects disappear at unpredictable times. Let be a nonnegative random variable representing an object’s lifetime. Its survival function is
If each surviving object has the same constant instantaneous event rate , then
The resulting exponential distribution has probability density
Its mean is , and its variance is .
The relevant instantaneous rate is the hazard:
For a continuous lifetime distribution, wherever . Exponential survival is exactly the case . The condition concerns objects that have survived until , not the original population as a whole. The unconditional density declines because fewer objects remain available to experience the event, even though each survivor’s hazard is unchanged.
A distinctive consequence is memorylessness:
Conditional on survival to age , the distribution of additional lifetime is identical to the original lifetime distribution. Among nondegenerate continuous distributions on the nonnegative real line, this property characterizes the exponential distribution. It does not describe wear, aging, or accumulated damage unless those mechanisms happen to leave the hazard constant. Memorylessness concerns the specified event model and does not imply that every physical property of an object is independent of its history.
If initially present objects have mutually independent exponential lifetimes with the same rate, their number remaining at time has a binomial distribution:
Hence
The smooth decay curve is an expectation, whereas a realized population follows a downward staircase. Every finite collection with such lifetimes becomes extinct in finite time with probability one, despite its positive expected population at every finite time. Its mean extinction time is , where .
Exponential waiting times also occur in a homogeneous Poisson process, whose successive event intervals are independent and share a fixed rate. That construction differs from a finite decaying population: in the latter, the total event rate is when objects remain and therefore decreases after each loss. Both use exponential waiting times, but only the first has a constant total event rate.
Radioactive transformation and chemical reactions
Radioactive decay is a central physical application. For a population of identical unstable nuclei under fixed conditions, a constant transformation probability per unit time gives
The activity, or expected number of transformations per unit time, is
Activity is measured in becquerels, with one becquerel equal to one transformation per second. Activity is not the same as radiation energy output: different transformations can release different energies, and the energy absorbed elsewhere depends on emission and transport processes.
The radioactive transformation theory developed by Ernest Rutherford and Frederick Soddy in 1902–1903 established a framework in which radioactive substances change into other substances at characteristic rates. Carbon-14 provides a familiar application: its physical half-life is approximately 5,730 years. Radiocarbon dating uses the changing abundance of carbon-14 after carbon exchange with the environment ceases. Actual calendar-age determination also requires calibration because atmospheric carbon-14 abundance has varied. Conventional radiocarbon ages use a historical half-life convention of 5,568 years, a distinction between the reporting system and the modern physical half-life.
Sequential transformations generally produce more than one exponential. If species 1 transforms into species 2, which also decays, then
For and ,
The daughter population initially rises before declining because it is being produced as well as removed. When the two rate constants are equal, the limiting expression is . These equations are special cases of decay-chain equations analyzed by Harry Bateman in 1910.
In chemical kinetics, a first-order reaction obeys
where is reactant concentration and is a rate constant. Under constant conditions, . Its half-life is independent of initial concentration, unlike that of a simple second-order reaction. A mechanism involving two reactants can display pseudo-first-order behavior if one remains in such large excess that its concentration is effectively constant. The resulting effective rate constant incorporates that excess concentration. Exponential concentration loss therefore does not, by itself, identify an elementary molecular mechanism; different mechanisms can produce the same observed rate law over a restricted range.
Electrical, thermal, and spatial relaxation
A resistor–capacitor circuit provides a direct engineering realization. When an ideal capacitor with capacitance discharges through a resistance , charge conservation and the component laws give
Thus
and the time constant is . A resistance of and capacitance of give a one-second time constant. The charge follows the same exponential, but the stored energy,
decays as . A decay time therefore depends on whether the measured quantity is amplitude, charge, or energy.
Thermal relaxation often concerns a deviation from equilibrium rather than a quantity tending to absolute zero. Newton’s law of cooling assumes that heat transfer is proportional to the difference between an object’s temperature and a fixed ambient temperature :
Here is mass, is specific heat capacity, is an effective heat-transfer coefficient, and is exposed area. If these are constant,
The model requires an approximately uniform object temperature and an appropriate heat-transfer law. Large internal gradients, variable convection, or strong thermal radiation can invalidate a single-exponential description.
The same mathematical pattern describes damping envelopes. In a linear underdamped oscillator,
the displacement has the form
The displacement itself is not monotone: it oscillates through zero. Its envelope decays exponentially. For weak damping, energy averaged over oscillations decreases approximately as . At critical damping, polynomial factors multiply the exponential, so even a simple linear oscillator need not exhibit a pure exponential waveform.
Exponential attenuation can also occur in space. The Beer–Lambert law for a homogeneous absorbing medium gives
where is intensity, is path length, and is an attenuation coefficient. The half-value thickness is . This follows when each short layer removes a fixed fraction of the incident quantity under the model’s conditions. It must be distinguished from geometric spreading: an ideal point source in three dimensions produces inverse-square intensity even without absorption. When both effects operate, measured intensity may contain both a geometric factor and an exponential attenuation factor, rather than either law alone.
Linear systems, modes, and exponential stability
A single decay equation extends naturally to systems of interacting quantities. For a constant matrix , the linear system
has solution
where the matrix exponential is defined by a convergent power series. The scalar law corresponds to the one-dimensional matrix . In higher dimensions, individual components usually combine several time scales. Concentrations in connected compartments, for example, can reflect both transfer between compartments and removal from the whole system.
If has a basis of eigenvectors, the solution can be expanded in modes:
Real negative eigenvalues give nonoscillatory decaying modes. Complex eigenvalues with negative real parts produce oscillations with exponentially decreasing envelopes. At long times, the mode with the least negative real part typically dominates, provided the initial state excites it and the measured quantity detects it. This explains why a complicated system can look approximately single-exponential late in an experiment while showing substantial curvature at earlier times.
The complete stability condition is more precise than requiring each observed component to decrease. For a finite-dimensional constant-coefficient system, the origin is exponentially stable exactly when every eigenvalue of has strictly negative real part. There then exist constants and such that
for all . Some components, and even a chosen norm, can initially increase before the eventual decay. This transient amplification is possible when eigenvectors are not orthogonal. If the matrix is not diagonalizable, terms such as can occur; they still satisfy an exponential bound with any sufficiently smaller decay rate.
Spatially distributed systems provide an infinite-dimensional counterpart. Solutions of the heat equation in a bounded region can often be expanded in spatial modes whose amplitudes decay exponentially. For a rod of length , with zero temperature deviation at its endpoints and thermal diffusivity , the th sine mode has amplitude proportional to
Fine-scale spatial variations disappear faster than broad ones. The overall temperature at a point is a sum of these modes, becoming approximately single-exponential only when the slowest observable mode dominates. With insulated rather than fixed-temperature boundaries, a constant mode persists, and it is the deviation from the conserved mean temperature that decays.
Estimating decay rates from observations
An exact single exponential becomes a straight line after taking the natural logarithm:
A plot of logarithmic quantity against ordinary time is therefore a basic diagnostic. Two exact positive observations determine
With more observations, linear regression on logarithms can estimate the slope. This procedure is appropriate only when the transformed measurement errors match the regression model; a visually straight semilogarithmic plot does not establish that its error assumptions hold.
The distinction between additive and multiplicative error matters. For
with additive errors of approximately constant variance, fitting the untransformed exponential by nonlinear least squares is generally more appropriate than fitting logarithms. If
then taking logarithms produces an additive-error model and may be suitable. Logging additive-noise data changes error variances and disproportionately affects small observations; nonpositive measurements cannot be logged at all. Error models therefore influence both estimated rates and their uncertainty.
For independently observed complete lifetimes assumed exponential, maximum likelihood estimation yields
If some lifetimes are right-censored, the observation records only that survival exceeded a specified time. Under independent, noninformative censoring, the estimator becomes
where is the number of observed events and is each object’s observed exposure time. The denominator is total time at risk, including time contributed by censored objects. Treating censored durations as completed lifetimes would distort the estimate.
Background signals and finite counting intervals require additional care. Measurements may follow
where is a nondecaying background. Failing to include makes late-time observations appear to decay too slowly. If events are counted over intervals, the expected count is an integral rather than an instantaneous rate. For an event rate , the expected count between and is
before adding background and detector effects. Appropriate confidence intervals depend on the observation model; counting variability, censoring, calibration uncertainty, and correlations between fitted parameters cannot be inferred from the fitted curve alone.
Changing rates, mixtures, and limits of the model
A time-dependent proportional loss rate gives
Although an exponential appears in this expression, the result is not generally a single exponential in time. Its instantaneous logarithmic slope changes with , and its half-life can depend on the starting time. In survival models, the integral of the hazard is called the cumulative hazard. Pure exponential survival is the special case in which that cumulative hazard is linear in time.
Heterogeneity can create changing apparent rates even when every component obeys a constant-rate law. A positive mixture has the form
Its effective rate is
Fast-decaying components disappear first, leaving progressively more weight on slow components. In fact,
where the weights are proportional to . This flattening needs no change in any individual component’s rate. Resolving several nearby rates from noisy data is nevertheless an ill-conditioned inverse problem: quite different mixtures can produce nearly indistinguishable observations over a finite interval.
Other functional forms represent different mechanisms. A power-law tail such as decays more slowly than every with fixed . A stretched exponential,
has a declining logarithmic rate,
Such forms occur in descriptions of heterogeneous relaxation, but a fitted stretched exponential does not uniquely identify a microscopic mechanism. Distinguishing these alternatives requires an adequate time range and an explicit treatment of measurement error and background.
Nonlinear loss also changes the law. For a second-order process,
the solution is
whose half-life depends on . Conversely, a nonlinear system can approach an equilibrium exponentially if its linearization has suitable negative rates. Writing near an equilibrium of gives
When , sufficiently small deviations have an approximately exponential relaxation. If that derivative vanishes, slower or otherwise nonexponential behavior can arise. An observed exponential tail can therefore reveal local relaxation near equilibrium without implying that the full process follows a proportional-loss law at all amplitudes.