Invertibility Conditions For General MA Models

Why Invertibility Matters

A moving-average model describes an observed series as a finite combination of current and past shocks. Invertibility asks a related but different question: can those shocks be recovered stably from the current and past observations? The answer matters because the shocks summarize the new information entering the model. A stable recovery makes that sequence interpretable and supports a canonical representation of the observed dependence.

For a general MA(q), the test is based on the roots of its lag polynomial. The direction of the test is important: the roots must be outside the unit circle. A root on the circle fails the strict condition, just as a root inside it does.

Definition: Let $\{\varepsilon_t\}$ be white noise with mean zero and variance $\sigma^2$, where $0<\sigma^2<\infty$. A finite-order MA(q) model is $$ Y_t=\mu+\varepsilon_t+\theta_1\varepsilon_{t-1}+\cdots+\theta_q\varepsilon_{t-q} =\mu+\theta(B)\varepsilon_t, $$ where $B\varepsilon_t=\varepsilon_{t-1}$ and $\theta(z)=1+\theta_1z+\cdots+\theta_qz^q$. The representation is invertible if the centered observations can be used to express $\varepsilon_t$ as a convergent, stable linear combination of $Y_t-\mu,Y_{t-1}-\mu,\ldots$. In the standard root test, stability is ensured by an absolutely summable sequence of inverse-filter coefficients.

The polynomial convention here has constant term 1 and plus signs for the MA coefficients. Always form the polynomial from the model as written. If a model is written using minus signs, those signs carry through to its polynomial. The conditions below concern roots $z$ satisfying $\theta(z)=0$, not the reciprocals of those roots.

The Root Condition

Theorem: For an MA(q) model with a nonzero coefficient of $\varepsilon_{t-q}$, the model is invertible if and only if every root $r$ of $\theta(z)=0$, counted with multiplicity, satisfies $|r|>1$. A root with $|r|=1$ does not satisfy the condition. For a pure white-noise model, $q=0$ and $\theta(z)=1$, so there are no roots and the model is invertible.

The theorem applies to real and complex roots alike. For a real-coefficient polynomial, nonreal roots occur in conjugate pairs; the pair has the same modulus, so both roots are either inside, on, or outside the unit circle. If a root is repeated, its multiplicity does not change the root-location test: every occurrence must be outside.

Write the polynomial.
From the MA equation, form $\theta(z)=1+\theta_1z+\cdots+\theta_qz^q$, preserving each coefficient's sign.

Find all roots.
Solve $\theta(z)=0$, using factorization, the quadratic formula, or another appropriate method.

Compare each modulus with 1.
The model is invertible only when every root has modulus strictly greater than 1.

State the interpretation.
If the condition holds, the shocks have a stable representation in current and past centered observations. If it fails, this MA representation is not invertible in the standard causal sense.

Why the Root Test Works

Write the desired inverse filter as a power series: $$ \frac{1}{\theta(z)}=\sum_{j=0}^{\infty}\pi_j z^j. $$ If the series is stable, substituting the lag operator gives $$ \varepsilon_t=\sum_{j=0}^{\infty}\pi_j(Y_{t-j}-\mu). $$ To see when this is possible, factor the polynomial using its roots $r_1,\ldots,r_q$, with multiplicities included. Since its constant term is 1, $$ \theta(z)=\prod_{j=1}^{q}\left(1-\frac{z}{r_j}\right). $$

If all roots satisfy $|r_j|>1$, then each factor has a convergent geometric expansion for $|z|<|r_j|$: $$ \frac{1}{1-z/r_j}=\sum_{k=0}^{\infty}\left(\frac{z}{r_j}\right)^k. $$ The product therefore has a power series whose radius of convergence extends beyond 1. Its coefficients decay sufficiently fast to be absolutely summable. For repeated roots, multiplying the corresponding series can introduce factors that grow polynomially in $k$, but the geometric decay still makes the coefficients absolutely summable when $|r_j|>1$. Thus the inverse filter is stable.

Conversely, if a root lies inside or on the unit circle, $1/\theta(z)$ has a singularity at a point with modulus at most 1. Its power series about zero cannot converge on a disk extending beyond that point. It cannot supply the standard stable causal inverse. This explains the strict inequality in the test; a root exactly on the unit circle is not an acceptable borderline case.

Worked Example: A First-Order Model With an Inside Root

A hypothetical monthly claims index is modeled as $Y_t=\mu+\varepsilon_t+1.25\varepsilon_{t-1}$, where $\{\varepsilon_t\}$ is mean-zero white noise with finite variance $\sigma^2$. Determine whether this representation is invertible and explain what the root implies.

Method. Form the lag polynomial and locate its root. The common mistake to avoid is treating a coefficient with absolute value greater than 1 as evidence that the root is outside the unit circle; the root is obtained by solving the polynomial equation.

The polynomial is $\theta(z)=1+1.25z$. Setting it equal to zero gives $$ 1+1.25z=0,\qquad z=-\frac{1}{1.25}=-0.8. $$ Its modulus is $|-0.8|=0.8<1$. Therefore, this MA(1) representation is not invertible.

The same conclusion appears from the inverse equation. In lag notation, $Y_t-\mu=(1+1.25B)\varepsilon_t$, so formal inversion would require $1/(1+1.25B)$. Its geometric expansion about zero has ratio $-1.25B$, whose coefficients do not decay: $$ \frac{1}{1+1.25B}=\sum_{j=0}^{\infty}(-1.25)^jB^j. $$ The coefficient magnitudes grow, rather than being absolutely summable. This confirms that the shocks cannot be recovered through a stable current-and-past inverse filter.

Check. The polynomial has one root, at $-0.8$, and that root is inside the unit circle. Both the root test and the inverse-series behavior give the same decision.

Worked Example: A Factored MA(2) With Two Outside Roots

A hypothetical quarterly payment series has centered MA polynomial $\theta(z)=(1-\tfrac12z)(1-\tfrac13z)$. Determine invertibility and find the first three coefficients of the inverse filter. Assume the innovations are mean-zero white noise with finite variance.

Method. Use the factors to identify the roots, then expand the reciprocal to verify stable recovery. The trap is confusing the factor values $1/2$ and $1/3$ with the roots themselves.

Multiplying the factors gives $$ \theta(z)=1-\left(\frac12+\frac13\right)z+\frac16z^2 =1-\frac56z+\frac16z^2. $$ The roots come from $1-\tfrac12z=0$ and $1-\tfrac13z=0$, so they are $z=2$ and $z=3$. Their moduli are 2 and 3, both strictly greater than 1. The model is invertible.

For the inverse, use partial fractions: $$ \frac{1}{(1-\frac12B)(1-\frac13B)} =\frac{3}{1-\frac12B}-\frac{2}{1-\frac13B}. $$ Expanding each factor as a geometric series gives $$ \frac{1}{\theta(B)} =\sum_{k=0}^{\infty}\left(3\left(\frac12\right)^k-2\left(\frac13\right)^k\right)B^k. $$ Thus $\pi_0=3-2=1$, $\pi_1=3/2-2/3=5/6$, and $\pi_2=3/4-2/9=19/36$. The coefficients are combinations of two decaying geometric sequences, so their absolute values have a finite sum. The inverse is stable.

Check. Multiplying $1-\tfrac56B+\tfrac16B^2$ by the inverse series gives constant coefficient 1 and zero coefficients at positive powers of $B$, as required for an inverse.

Worked Example: Complex Roots on the Unit Circle

For a hypothetical process, suppose $Y_t=\mu+\varepsilon_t-0.5\varepsilon_{t-1}+\varepsilon_{t-2}$, with mean-zero white-noise innovations of finite variance. Decide invertibility by calculating both roots and their moduli. The main trap is an incorrect division in the quadratic formula, which can change a boundary case into an apparent outside-root case.

Method. The polynomial is $\theta(z)=1-0.5z+z^2$. Set it equal to zero and apply the quadratic formula with $a=1$, $b=-0.5$, and $c=1$: $$ z=\frac{0.5\pm\sqrt{(-0.5)^2-4(1)(1)}}{2} =\frac{0.5\pm\sqrt{-3.75}}{2} =\frac{0.5\pm i\sqrt{15}/2}{2} =\frac14\pm i\frac{\sqrt{15}}4. $$ For either root, the squared modulus is $$ |z|^2=\left(\frac14\right)^2+\left(\frac{\sqrt{15}}4\right)^2 =\frac{1}{16}+\frac{15}{16}=1. $$ Hence both roots have modulus 1. They lie on the unit circle, so the model is not invertible.

Check. The roots are conjugates, as expected for a real-coefficient polynomial. Their moduli equal 1 exactly, not 2, and the theorem requires a strict modulus greater than 1.

What Invertibility Does—and Does Not—Say

Invertibility concerns whether the innovations can be recovered stably from present and past observations. It is not a requirement for the finite MA process to be defined. With finite-variance white noise, the finite sum defining an MA(q) has finite variance and is weakly stationary for any finite coefficients, whether or not its polynomial roots satisfy the invertibility condition.

A noninvertible model can still describe a stationary process, but its shocks are not represented by a stable causal inverse of the observations under this test. Invertibility therefore selects a useful, stable representation; it does not say that every noninvertible parameter choice is meaningless. The distinction is especially useful when interpreting an estimated MA model: the sign and size of its coefficients alone do not settle invertibility, because the condition is about the locations of all polynomial roots.

The root test also does not require the innovations to be normally distributed. The assumptions used here are mean zero, constant finite variance, and zero autocovariance at nonzero lags, as in the standard white-noise setup. Normality would be an additional distributional assumption, not part of this invertibility condition.

Common exam trap: Do not test whether the polynomial's coefficients are inside or outside the unit circle. Solve $\theta(z)=0$, calculate the modulus of each root, and require $|r|>1$ for every root. A root on the unit circle fails, even if all other roots are outside.

Check Your Understanding

For each model, form or use the stated lag polynomial, locate its roots, and apply the strict root condition. Answers identify the root locations and the invertibility decision.

An MA(1) model has $\theta(z)=1-0.4z$. Find its root and determine invertibility. (Answer: The root is $2.5$, so the model is invertible.)
An MA(1) model has $\theta(z)=1+1.1z$. Find its root and determine invertibility. (Answer: The root is $-10/11$, with modulus less than 1, so the model is not invertible.)
An MA(2) model has $\theta(z)=1-\tfrac34z+\tfrac18z^2$. Factor the polynomial, find both roots, and determine invertibility. (Answer: $\theta(z)=(1-\tfrac12z)(1-\tfrac14z)$; the roots are 2 and 4, so the model is invertible.)
An MA(2) model has $\theta(z)=1+z+z^2$. Find the complex roots' moduli and determine invertibility. (Answer: The roots are $-\tfrac12\pm i\tfrac{\sqrt3}{2}$, each with modulus 1, so the model is not invertible.)
An MA(3) model has $\theta(z)=(1-\tfrac15z)^2(1+\tfrac12z)$. Identify the roots with multiplicity and decide invertibility. (Answer: The roots are 5 with multiplicity 2 and $-2$ with multiplicity 1; all have modulus greater than 1, so the model is invertible.)

What You'll Learn