The Lag Operator and Difference Equations

Time-series models are full of statements like "today's value depends on yesterday's value, which depends on the day before, and so on." Writing that out in words gets unwieldy fast, especially once a model reaches several lags. The lag operator $L$ is just a piece of notation that fixes this: applied to a series $x_t$, it shifts you back one period, $Lx_t = x_{t-1}$. Apply it twice and you go back two periods, $L^2 x_t = x_{t-2}$, and so on for any number of lags.

That small trick turns recursive time-series equations into ordinary algebra. An AR(1) model, normally written $x_t = \phi x_{t-1} + \varepsilon_t$, becomes $(1 - \phi L)x_t = \varepsilon_t$ — a difference equation in lag-operator form. The expression $(1 - \phi L)$ can now be manipulated like a polynomial: factored, inverted, multiplied by other lag polynomials for more complex models. Whether a model is stable (its effects die out rather than explode) comes down to where the roots of that polynomial sit relative to 1, which is exactly the question that unit-root tests are built to answer.

Worked example. Take the AR(1) equation $(1 - 0.6L)x_t = \varepsilon_t$. Formally "inverting" $(1 - 0.6L)$ gives the infinite sum $1 + 0.6L + 0.6^2L^2 + 0.6^3L^3 + \dots$, so $x_t = \varepsilon_t + 0.6\varepsilon_{t-1} + 0.36\varepsilon_{t-2} + \dots$ — today's value is a fading, geometrically shrinking blend of all past shocks. Because $0.6 < 1$, the weights shrink to zero and the series is stable; if the coefficient were $1.2$ instead, the weights would grow without bound.