Proof of the Stationarity Condition for an AR(1) Process (|φ| < 1)

Q: Why is it called "weak" stationarity, and what is "strict" stationarity?

Weak (or covariance) stationarity requires constant mean, finite variance, and autocovariance that depends only on the lag, not time. Strict stationarity requires the entire joint probability distribution of the process to be time-invariant. Weak stationarity is often sufficient for practical applications and easier to prove.

Q: What happens to the AR(1) process if $ \phi = 1 $?

If $ \phi = 1 $, the process becomes a random walk ($ Y_t = c + Y_{t-1} + \epsilon_t $). In this case, past shocks have a permanent impact, the variance grows indefinitely with time, and the process is non-stationary. Differencing ($ \Delta Y_t = Y_t - Y_{t-1} $) can make it stationary.

Q: Can an AR(1) process with $ |\phi| > 1 $ ever be stationary?

No. If $ |\phi| > 1 $, the impact of past shocks ($ \phi^j \epsilon_{t-j} $) grows exponentially, leading to an explosive process where the variance is infinite. Such a process is clearly non-stationary.

Q: How crucial are the assumptions about the error term $ \epsilon_t $ (white noise) for this proof?

Extremely crucial. The assumptions $ E[\epsilon_t] = 0 $, $ Var(\epsilon_t) = \sigma_{\epsilon}^2 < \infty $, and $ Cov(\epsilon_t, \epsilon_s) = 0 $ for $ t \neq s $ are fundamental. Without $ E[\epsilon_t] = 0 $, the mean would be more complex. Without finite variance, the variance of $ Y_t $ wouldn't be finite. Most importantly, uncorrelated errors simplify $ Var(\sum X_i) = \sum Var(X_i) $, enabling the derivation of $ Var(Y_t) $.

Unravel the stationarity condition for AR(1) processes. Rigorous proof, cinematic intuition, and essential insights for time series analysis.

The Formal Theorem

Consider an Autoregressive process of order 1 (AR(1)) defined by

Y_t = c + \phi Y_{t-1} + \epsilon_t

, where

Y_t

is the observed value at time

t

\phi

is the autoregressive coefficient,

c

is a constant, and

\epsilon_t

is a white noise error term such that

E[\epsilon_t] = 0

Var(\epsilon_t) = \sigma_{\epsilon}^2 < \infty

, and

Cov(\epsilon_t, \epsilon_s) = 0

for

t \neq s

.\nIf the condition

|\phi| < 1

holds, then the AR(1) process

Y_t

is weakly stationary, meaning its mean, variance, and autocovariance function are constant over time. Specifically, the mean

E[Y_t]

and variance

Var(Y_t)

are given by:

\begin{aligned} E[Y_t] &= \frac{c}{1 - \phi} \\ Var(Y_t) &= \frac{\sigma_{\epsilon}^2}{1 - \phi^2} \end{aligned}

Analytical Intuition.

Imagine a grand, echoing hall where every new sound,

\epsilon_t

, is like a fresh whisper at time

t

. This whisper doesn't vanish instantly; it leaves an echo,

\phi Y_{t-1}

, that contributes to the sound at the next moment. The coefficient

\phi

acts as the 'fading factor' for this echo. If

|\phi|

is strictly less than 1, each subsequent echo of that original whisper becomes weaker, diminishing exponentially into the past. The hall thus reaches a steady, predictable hum – a stable, statistically constant background noise – because the influence of distant whispers eventually becomes imperceptible. However, if

|\phi|

were 1, every whisper would echo indefinitely at full strength, leading to an ever-growing cacophony. And if

|\phi| > 1

, each echo would amplify its predecessor, turning a whisper into a deafening, explosive roar. The stationarity condition,

|\phi| < 1

, precisely ensures this graceful diminishing return, allowing the system to settle into a rhythmic, statistically constant state where the impact of past events elegantly fades over time, giving rise to predictable long-run properties like stable mean and variance.

CAUTION

Institutional Warning.

Students often struggle to grasp that $|\phi| < 1$ is not merely a condition for finite variance, but fundamentally ensures that the influence of past values and initial conditions decays over time, preventing explosive behavior. They sometimes confuse this stability condition with the concept of a stable estimation of $\phi$ itself.

Institutional Deep Dive.

The notion of stationarity in time series analysis is fundamentally about predictability and statistical equilibrium. For an Autoregressive process of order 1,

Y_t = c + \phi Y_{t-1} + \epsilon_t

, the condition

|\phi| < 1

is the mathematical linchpin that guarantees this equilibrium. Let's dissect this through core logic, geometric mechanics, and institutional pitfalls.\n\n**Core Logic**:\nTo understand why

|\phi| < 1

is crucial, we employ iterative substitution. We can express

Y_t

not just in terms of its immediate past,

Y_{t-1}

, but as a sum of all past innovations,

\epsilon_t

.\nStarting with

Y_t = c + \phi Y_{t-1} + \epsilon_t

:\nSubstitute

Y_{t-1} = c + \phi Y_{t-2} + \epsilon_{t-1}

into the equation for

Y_t

:\n

Y_t = c + \phi(c + \phi Y_{t-2} + \epsilon_{t-1}) + \epsilon_t

Y_t = c(1 + \phi) + \phi^2 Y_{t-2} + \phi \epsilon_{t-1} + \epsilon_t

\nRepeating this substitution backward in time

k

times:\n

Y_t = c(1 + \phi + \phi^2 + \dots + \phi^{k-1}) + \phi^k Y_{t-k} + \sum_{j=0}^{k-1} \phi^j \epsilon_{t-j}

\nFor the process to be truly stationary, its statistical properties must be independent of the starting point or initial condition

Y_{t-k}

. As

k

approaches infinity (i.e., looking infinitely far back in time), the term

\phi^k Y_{t-k}

must vanish for the influence of the initial condition to fade. This occurs precisely when

|\phi| < 1

, because

\lim_{k \to \infty} \phi^k = 0

.\nIf this condition holds, and assuming

Y_t

has been running for an infinitely long time (or started sufficiently far in the past), the sum

1 + \phi + \phi^2 + \dots

converges to

\frac{1}{1 - \phi}

(a geometric series sum). Similarly, the sum of past errors

\sum_{j=0}^{\infty} \phi^j \epsilon_{t-j}

also converges.\nThus, in the limit,

Y_t

can be expressed as an infinite moving average:\n

Y_t = \frac{c}{1 - \phi} + \sum_{j=0}^{\infty} \phi^j \epsilon_{t-j}

\nFrom this representation, we can readily derive the mean and variance.\n

E[Y_t] = E\left[\frac{c}{1 - \phi} + \sum_{j=0}^{\infty} \phi^j \epsilon_{t-j}\right]

= \frac{c}{1 - \phi} + \sum_{j=0}^{\infty} \phi^j E[\epsilon_{t-j}]

\nSince

E[\epsilon_t] = 0

for all

t

, then

E[Y_t] = \frac{c}{1 - \phi}

. This is a constant, independent of

t

.\nFor the variance:\n

Var(Y_t) = Var\left[\frac{c}{1 - \phi} + \sum_{j=0}^{\infty} \phi^j \epsilon_{t-j}\right]

\nThe constant term

\frac{c}{1 - \phi}

does not contribute to the variance. Also, since

\epsilon_t

are uncorrelated, the variance of their sum is the sum of their variances:\n

Var(Y_t) = \sum_{j=0}^{\infty} (\phi^j)^2 Var(\epsilon_{t-j})

= \sum_{j=0}^{\infty} \phi^{2j} \sigma_{\epsilon}^2

= \sigma_{\epsilon}^2 \sum_{j=0}^{\infty} (\phi^2)^j

\nThis is another geometric series with ratio

\phi^2

. For this series to converge, we require

|\phi^2| < 1

, which simplifies to

|\phi| < 1

. If this holds, the sum converges to

\frac{1}{1 - \phi^2}

.\nTherefore,

Var(Y_t) = \frac{\sigma_{\epsilon}^2}{1 - \phi^2}

. This is also a constant, independent of

t

.\nThe autocovariance function also proves to be time-invariant under

|\phi| < 1

.\n\n**Geometric Mechanics**:\nThe condition

|\phi| < 1

acts as a dampening mechanism. Imagine a physical system, like a pendulum. If you give it a push (an

\epsilon_t

shock), it swings. But friction (analogous to

1-\phi

) gradually reduces the amplitude until it settles back to its equilibrium. Each swing is smaller than the last, much like

\phi^j

gets smaller as

j

increases.\nIf

\phi = 0

Y_t = c + \epsilon_t

, a purely random process (white noise with a constant mean). The effect of any

\epsilon_t

is immediate and vanishes completely by

t+1

.\nIf

0 < \phi < 1

, a positive shock

\epsilon_t

causes

Y_t

to increase. This increase is carried over to

Y_{t+1}

\phi Y_t

, but with a reduced magnitude. This effect continues, but its influence decays exponentially. The series "remembers" past shocks but with decreasing clarity.\nIf

\phi = 1

, the process becomes a random walk (

Y_t = c + Y_{t-1} + \epsilon_t

). A shock

\epsilon_t

has a permanent impact; it shifts the level of the series indefinitely. The variance grows with time, meaning it's non-stationary.\nIf

\phi > 1

\phi < -1

, the process is explosive. A shock

\epsilon_t

is amplified in subsequent periods (

\phi^j

grows in magnitude), leading to an ever-increasing (or oscillating and increasing) deviation from the mean, making the variance infinite.\nThe crucial point is that

|\phi| < 1

ensures that the process has finite memory; the distant past ceases to have an appreciable influence on the present.\n\n**Institutional Pitfalls**:\nStudents often conflate weak stationarity with strict stationarity. While strict stationarity implies weak stationarity (if moments exist), the reverse is not always true. However, for an AR(1) process with i.i.d. Gaussian white noise

\epsilon_t

, weak stationarity *does* imply strict stationarity, because linear combinations of Gaussian variables are also Gaussian, and Gaussian processes are fully characterized by their first two moments.\nAnother common pitfall is neglecting the assumptions about

\epsilon_t

. The proof relies heavily on

E[\epsilon_t] = 0

and

Var(\epsilon_t) = \sigma_{\epsilon}^2 < \infty

, and crucially,

Cov(\epsilon_t, \epsilon_s) = 0

for

t \neq s

. If the error terms themselves are autocorrelated or have time-varying variance, the derived stationarity conditions and properties may no longer hold. For instance, if

\epsilon_t

itself were an AR(1) process, the overall process

Y_t

would be more complex and its stationarity would depend on the properties of

\epsilon_t

. Finally, the requirement

\phi \neq 1

for the mean to be defined is critical, and

\phi \neq \pm 1

for the variance. The

|\phi| < 1

condition elegantly covers both these requirements while simultaneously ensuring the convergence of the infinite sums.

Academic Inquiries.

Why is it called "weak" stationarity, and what is "strict" stationarity?

Weak (or covariance) stationarity requires constant mean, finite variance, and autocovariance that depends only on the lag, not time. Strict stationarity requires the entire joint probability distribution of the process to be time-invariant. Weak stationarity is often sufficient for practical applications and easier to prove.

What happens to the AR(1) process if $\phi = 1$ ?

If $\phi = 1$ , the process becomes a random walk ( $Y_t = c + Y_{t-1} + \epsilon_t$ ). In this case, past shocks have a permanent impact, the variance grows indefinitely with time, and the process is non-stationary. Differencing ( $\Delta Y_t = Y_t - Y_{t-1}$ ) can make it stationary.

Can an AR(1) process with $|\phi| > 1$ ever be stationary?

No. If $|\phi| > 1$ , the impact of past shocks ( $\phi^j \epsilon_{t-j}$ ) grows exponentially, leading to an explosive process where the variance is infinite. Such a process is clearly non-stationary.

How crucial are the assumptions about the error term $\epsilon_t$ (white noise) for this proof?

Extremely crucial. The assumptions $E[\epsilon_t] = 0$ , $Var(\epsilon_t) = \sigma_{\epsilon}^2 < \infty$ , and $Cov(\epsilon_t, \epsilon_s) = 0$ for $t \neq s$ are fundamental. Without $E[\epsilon_t] = 0$ , the mean would be more complex. Without finite variance, the variance of $Y_t$ wouldn't be finite. Most importantly, uncorrelated errors simplify $Var(\sum X_i) = \sum Var(X_i)$ , enabling the derivation of $Var(Y_t)$ .

Standardized References.

Definitive Institutional SourceBox, G.E.P., Jenkins, G.M., Reinsel, G.C., and Ljung, G.M. (2015). Time Series Analysis: Forecasting and Control. 5th ed. Wiley.

Intermediate

Proof that Autocovariance Depends Only on Lag for Weakly Stationary Processes

Explore the rigorous proof that autocovariance for weakly stationary processes depends only on lag, understanding its deep implications for time series analysis.

Foundational

Derivation of the Autocorrelation Function (ACF) for a White Noise Process

Derive the Autocorrelation Function (ACF) for white noise. Understand its theoretical properties and intuitive meaning in Time Series Analysis.

Intermediate

Proof of the Invertibility Condition for an MA(1) Process (|θ| < 1)

Exploring the cinematic intuition of Proof of the Invertibility Condition for an MA(1) Process (|θ| < 1).

Intermediate

Derivation of the Mean for a Stationary AR(1) Process

Exploring the cinematic intuition of Derivation of the Mean for a Stationary AR(1) Process.

Institutional Citation

Reference this proof in your academic research or publications.

NICEFA Visual Mathematics. (2026). Proof of the Stationarity Condition for an AR(1) Process (|φ| < 1): Visual Proof & Intuition. Retrieved from https://www.nicefa.org/library/time-series-analysis/proof-of-the-stationarity-condition-for-an-ar-1--process--------1-

Dominate the Logic.

"Abstract theory is just a movement we haven't seen yet."

Master the Proof Early Access

The Formal Theorem

Analytical Intuition.

Institutional Warning.

Institutional Deep Dive.

Academic Inquiries.

Why is it called "weak" stationarity, and what is "strict" stationarity?

What happens to the AR(1) process if ϕ=1 \phi = 1 ϕ=1?

Can an AR(1) process with ∣ϕ∣>1 |\phi| > 1 ∣ϕ∣>1 ever be stationary?

How crucial are the assumptions about the error term ϵt \epsilon_t ϵt​ (white noise) for this proof?

Standardized References.

Related Proofs Cluster.

Proof that Autocovariance Depends Only on Lag for Weakly Stationary Processes

Derivation of the Autocorrelation Function (ACF) for a White Noise Process

Proof of the Invertibility Condition for an MA(1) Process (|θ| < 1)

Derivation of the Mean for a Stationary AR(1) Process

Institutional Citation

Dominate the Logic.

What happens to the AR(1) process if $\phi = 1$ ?

Can an AR(1) process with $|\phi| > 1$ ever be stationary?

How crucial are the assumptions about the error term $\epsilon_t$ (white noise) for this proof?