Jorge Quitério Search
Mathematics

The stationarity region of the AR(2) model

The stationarity region of a second-order autoregression is a triangle in the plane of its two coefficients. Below a parabola within it, the process has a pseudo-cycle; near the edges, shocks persist for a long time. An interactive map shows the behaviour for any pair of coefficients.

A first-order autoregression has one form of memory: geometric decay. With a second lag the range of behaviour is much wider. A second-order process can drift smoothly, alternate in sign, or oscillate with an irregular period, as in Yule's model of the sunspot cycle.

This range of behaviour can be represented on a single map. Each AR(2) model is a point (ϕ1,ϕ2)(\phi_1, \phi_2) in the plane, and a triangle with a parabola inside it divides the plane into kinds of behaviour. The map allows the behaviour of a fitted model to be read directly.

Everyday systems show each kind of behaviour. A car body after a run of potholes bounces in a damped oscillation that the next bump restarts before it has died out; measured at regular intervals, its height behaves approximately like an AR(2) in the oscillating region. A shop that over-orders after every shortage and under-orders after every surplus produces stock levels that alternate up and down from one order to the next. The temperature of a well-insulated building drifts slowly with no rhythm at all. The triangle places all three on one map.

The second-order model is where autoregression began: Yule (1927) fitted it to sunspot numbers to replace the idea of hidden periodic cycles with damped oscillations driven by random shocks. Walker (1931) extended the method, giving the Yule-Walker equations their name. Box and Jenkins (1970) made the stationarity triangle part of the standard identification toolkit, and it appears in every edition of their book and in textbooks such as Brockwell and Davis (1991). The modern view, through the roots of a characteristic polynomial, is the same one that explains the three regimes of the AR(1) estimator.

The model and its polynomial

The AR(2) model is

Xt=ϕ1Xt−1+ϕ2Xt−2+et. X_t = \phi_1 X_{t-1} + \phi_2 X_{t-2} + e_t .

Its behaviour is determined by the roots of the characteristic polynomial Φ(z)=1−ϕ1z−ϕ2z2\Phi(z) = 1 - \phi_1 z - \phi_2 z^2. The process is stationary when both roots lie outside the unit circle, or equivalently when the inverse roots λ1,λ2\lambda_1, \lambda_2, which solve λ2−ϕ1λ−ϕ2=0\lambda^2 - \phi_1\lambda - \phi_2 = 0, lie inside it.

The triangle

Working out when both inverse roots have modulus below 1 gives three linear inequalities:

ϕ1+ϕ2<1,ϕ2−ϕ1<1,∣ϕ2∣<1. \phi_1 + \phi_2 < 1, \qquad \phi_2 - \phi_1 < 1, \qquad |\phi_2| < 1 .

They cut out a triangle with corners at (−2,−1)(-2, -1), (2,−1)(2, -1) and (0,1)(0, 1). Inside, shocks fade and the process returns to its mean. On the right edge, ϕ1+ϕ2=1\phi_1 + \phi_2 = 1, one root is exactly 1: a unit root, the random-walk border. Outside, the process explodes.

The parabola

The inverse roots are real when the discriminant ϕ12+4ϕ2\phi_1^2 + 4\phi_2 is non-negative and complex otherwise. The parabola ϕ2=−ϕ12/4\phi_2 = -\phi_1^2/4 separates the two. Above it, the autocorrelation decays without changing rhythm, smoothly if the dominant root is positive and with alternating signs if it is negative. Below it, the roots are a complex pair re±iθr e^{\pm i\theta}, and the autocorrelation is a damped cosine,

ρh∝rhcos⁡(hθ+ψ),r=−ϕ2,cos⁡θ=ϕ12−ϕ2. \rho_h \propto r^h \cos(h\theta + \psi), \qquad r = \sqrt{-\phi_2}, \qquad \cos\theta = \frac{\phi_1}{2\sqrt{-\phi_2}} .

That is a pseudo-cycle with period 2π/θ2\pi/\theta and damping rr per step. It is not a true cycle: random shocks keep restarting it, so its length wanders from one swing to the next. The car body behaves this way, and so do business cycles: expansions and recessions recur, but not with a fixed period, which is consistent with a pseudo-cycle and not with a hidden periodic one.

Yule's sunspot model, ϕ1=1.34\phi_1 = 1.34 and ϕ2=−0.66\phi_2 = -0.66, sits well below the parabola: r≈0.81r \approx 0.81 and a period of about 10.5 years.

Persistence across the region

The modulus of the largest inverse root measures memory. Near 0, shocks vanish almost immediately. Near 1, they linger for a long time. Colouring the triangle by that modulus shows a pale centre and dark edges: long persistence occurs near the corners and sides of the triangle.

Clicking in the coloured plane selects (ϕ1,ϕ2)(\phi_1, \phi_2); the lower charts show the autocorrelation function and a simulated path.

Suggested settings:

  1. Yule's sunspots. The preset lands below the parabola. The autocorrelation is a damped wave and the path swings with an irregular rhythm of about 10 steps; the readout gives 10.5.
  2. Smooth, no cycle. (0.5,0.3)(0.5, 0.3) is above the parabola: real roots, a slowly decaying autocorrelation, a path that drifts without oscillating.
  3. Alternating sign. (−0.9,0)(-0.9, 0) has a negative real root: the autocorrelation alternates in sign and the path alternates direction at almost every step.
  4. Near the edge. (1.6,−0.95)(1.6, -0.95) has modulus close to 1: a long, slowly decaying oscillation, nearly a sine wave. Just outside the triangle, the readout says "no" and the path is replaced by a warning, because the process would explode.
  5. Crossing the parabola. At ϕ1=1\phi_1 = 1, select a point just below the dashed line (for example ϕ2=−0.4\phi_2 = -0.4) and then just above it (ϕ2=−0.1\phi_2 = -0.1). Below, a long pseudo-cycle of about 10 steps; above, smooth decay with no rhythm at all. Near the line, the cycles are so long and so damped that the two behaviours are hard to tell apart.

Derivation of the triangle. The inverse roots solve g(λ)=λ2−ϕ1λ−ϕ2=0g(\lambda) = \lambda^2 - \phi_1 \lambda - \phi_2 = 0, and their product is −ϕ2-\phi_2. If they are complex, they have equal modulus −ϕ2\sqrt{-\phi_2}, which is below 1 exactly when ϕ2>−1\phi_2 > -1. If they are real, both lie in (−1,1)(-1, 1) exactly when the parabola gg is positive at ±1\pm 1 and its vertex ϕ1/2\phi_1/2 lies between them: g(1)=1−ϕ1−ϕ2>0g(1) = 1 - \phi_1 - \phi_2 > 0, g(−1)=1+ϕ1−ϕ2>0g(-1) = 1 + \phi_1 - \phi_2 > 0 and ∣ϕ1∣<2|\phi_1| < 2. Together with ∣ϕ2∣<1|\phi_2| < 1 (the product of two numbers in (−1,1)(-1, 1)), these give the three inequalities, and the complex case lies inside the same triangle because there ϕ2<−ϕ12/4≤0\phi_2 < -\phi_1^2/4 \le 0, so ∣ϕ1∣<2−ϕ2<2|\phi_1| < 2\sqrt{-\phi_2} < 2 and both g(±1)>0g(\pm 1) > 0.

Reading a fitted AR(2) model through this map is quicker than reading its coefficients. The relevant questions are whether it lies inside the triangle and how close to the edge, whether it lies above or below the parabola, and, if below, what its period and damping are. The answers describe the type of persistence in the data.

References

  1. G. U. Yule, "On a method of investigating periodicities in disturbed series, with special reference to Wolfer's sunspot numbers", Philosophical Transactions of the Royal Society A, 226, 267-298, 1927.
  2. G. Walker, "On periodicity in series of related terms", Proceedings of the Royal Society A, 131(818), 518-532, 1931.
  3. G. E. P. Box and G. M. Jenkins, Time Series Analysis: Forecasting and Control, Holden-Day, 1970.
  4. P. J. Brockwell and R. A. Davis, Time Series: Theory and Methods, 2nd ed., Springer, 1991.