Linear Algebra for Economists Explained: Matrices, Eigenvalues and the Steady State of a Labor Market

Understanding linear algebra for economists — simply put

Economic models rarely involve one equation. Prices in one market affect quantities in another, every industry buys from every other industry, and a regression explains one variable with several others. Linear algebra is the mathematics for handling many linked equations together. Its basic objects are vectors, lists of numbers such as the output of each industry, and matrices, tables of numbers that describe how one list of numbers turns into another.

Once a problem is written in matrix form, a small set of tools does most of the work. Matrix multiplication moves a system forward. The inverse of a matrix solves a system of equations. The determinant says whether a solution exists. Eigenvalues and eigenvectors describe where a repeated process ends up and how fast it gets there.

How does it actually help?

Take the labor market. Every quarter, some employed people lose their jobs, some unemployed people find work, and some people leave or join the labor force. Eurostat publishes these flows for the EU. Arrange them as a transition matrix P, where each row is a starting state and each column is the state a quarter later:

Multiplying the vector of shares by P gives next quarter's shares. Multiplying again gives the quarter after. The question economists care about is where this process settles: a vector $\pi$ that the matrix leaves unchanged, $\pi = \pi P$. In linear algebra terms, $\pi$ is an eigenvector of P for the eigenvalue 1. It is the labor market's steady state, and its unemployment share relative to the labor force is the unemployment rate the flows imply in the long run.

Why does this matter?

The same tools run through the whole of economics. The OLS regression estimator is a matrix formula. The Leontief input-output model, which national statistics offices use to trace how demand for one product ripples through the supply chain, is a matrix inverse. Comparative statics with several equations solves a linear system. Second-order conditions in constrained optimization are statements about the signs of determinants. Economists who are comfortable with matrices can read and build models that would be unmanageable equation by equation.

What makes the results reliable?

Matrix methods rest on conditions that are easy to state: a system of equations has a unique solution only if its matrix is invertible, and a transition matrix settles to a single steady state only if every state can eventually be reached from every other. When the data satisfy these conditions, the matrix answers are exact consequences of the inputs. The conditions, and what fails without each one, are set out after the worked example.

The takeaway

Linear algebra is the bookkeeping system for models with many moving parts. Writing a model in matrix form makes its solution, its stability and its long-run behavior visible in a few numbers. The rest of this guide builds a real transition matrix from Eurostat data and solves it, then sets out the general mechanics.

The basic tools, worked by hand

The EU example below uses three tools: solving a system of equations, multiplying a vector by a matrix, and finding eigenvalues. Each is easiest to learn first on a 2×2 case.

A system of equations. Suppose two linked markets clear when $2p_1 + p_2 = 5$ and $p_1 + 3p_2 = 10$, where $p_1$ and $p_2$ are the two prices (illustrative numbers). In matrix form this is $Ax = b$, where A has rows (2, 1) and (1, 3), x = ($p_1$, $p_2$) and b = (5, 10).

The determinant. For a 2×2 matrix with rows $(a_{11}, a_{12})$ and $(a_{21}, a_{22})$, $\det A = a_{11}a_{22} - a_{12}a_{21}$. Here det A = 2 × 3 − 1 × 1 = 5. It is not zero, so the system has exactly one solution.

The inverse. For a 2×2 matrix, swap the two diagonal entries, change the sign of the two off-diagonal entries, and divide by the determinant:

So $p_1 = 1$ and $p_2 = 3$. Substituting back confirms it: 2 + 3 = 5 and 1 + 9 = 10.

Eigenvalues. The eigenvalues $\mu$ of A solve $\det(A - \mu I) = 0$. For a 2×2 matrix this is the quadratic $\mu^2 - (a_{11} + a_{22})\mu + \det A = 0$. Here $\mu^2 - 5\mu + 5 = 0$, so $\mu = (5 \pm \sqrt{5})/2$, about 3.62 and 1.38.

A steady state by hand. Take a two-state transition matrix with rows (0.9, 0.1) and (0.5, 0.5) (illustrative). The steady state $\pi = (\pi_1, \pi_2)$ solves $\pi = \pi P$. The first column gives $0.9\pi_1 + 0.5\pi_2 = \pi_1$, or $0.1\pi_1 = 0.5\pi_2$, and the adding-up condition gives $\pi_1 + \pi_2 = 1$. So $\pi_2 = 1/6 \approx 0.167$ and $\pi_1 = 5/6 \approx 0.833$. Check: 0.833 × 0.9 + 0.167 × 0.5 = 0.833 and 0.833 × 0.1 + 0.167 × 0.5 = 0.167, so πP = π. This π is the left eigenvector of P for the eigenvalue 1. The other eigenvalue is the trace minus 1, 0.9 + 0.5 − 1 = 0.4, so any deviation from π shrinks by 60% each period.

Multiplying a row vector by a matrix. Each entry of $xP$ is a weighted sum down one column of P. In the labor market below, next quarter's unemployed share is the employed share times $p_{EU}$, plus the unemployed share times $p_{UU}$, plus the inactive share times $p_{NU}$.

A real example: EU labor market flows

Eurostat's labor market flow statistics follow people aged 15 to 74 across consecutive quarters of the EU Labour Force Survey. Between the second and third quarters of 2025, 23.0% of people who were unemployed found a job, 52.4% remained unemployed and 24.5% left the labor force. Among the employed, 1.2% became unemployed and 2.2% left the labor force. Among people outside the labor force, 3.7% moved into employment and 3.4% into unemployment. The figures are seasonally adjusted.

The published percentages are rounded, so they do not always sum exactly to 100%: the three shares for the unemployed add up to 99.9%. In the matrix below the probability of staying in the same state is set as 100% minus the two published outflows, so that every row sums to exactly 100%.

Reading the result

The steady state $\pi$ solves $\pi = \pi P$ together with the condition that the three shares add up to one. That is a system of three linear equations in three unknowns. Write the column equations for employment and unemployment, $-0.034\pi_E + 0.230\pi_U + 0.037\pi_N = 0$ and $0.012\pi_E - 0.475\pi_U + 0.034\pi_N = 0$, add $\pi_E + \pi_U + \pi_N = 1$, and solve by substitution as in the two-state case. The solution is:

The implied unemployment rate is the unemployed share divided by the employed plus unemployed shares. Computed from the unrounded shares, it is 5.9%.

The steady state can be checked by hand with one row-times-column multiplication. Next quarter's unemployed share is 0.629 × 0.012 + 0.040 × 0.525 + 0.332 × 0.034 ≈ 0.0075 + 0.0210 + 0.0113 = 0.0398, which rounds back to 0.040: the flows into unemployment equal the flows out of it.

The flows imply a long-run unemployment rate of 5.9%. Eurostat's monthly estimates of the EU unemployment rate for the same quarter were 5.9% in July 2025 and 6.0% in August and September 2025. The two are close, but this is a consistency check, not an independent test: the flows and the unemployment rate come from the same Labour Force Survey, and when unemployment is roughly stable the observed rate has to sit near the rate its own flows imply. The steady state becomes informative when the two diverge, because the gap shows which way unemployment is heading.

The eigenvalues of P add a second piece of information. For a 3×3 matrix they solve a cubic equation, which is solved numerically in practice; here they are 1, 0.919 and 0.501. The eigenvalue of 1 belongs to the steady state. The other two govern how fast the labor market returns to it after a shock: any deviation decomposes into parts that shrink by a factor of 0.919 and 0.501 each quarter.

The shock raises the unemployment rate from 5.9% to 6.7%. About half of the excess disappears in the first quarter, as the 0.501 eigenvalue predicts, because unemployment is a state people leave quickly: on these flows nearly half of the unemployed leave unemployment each quarter, either for a job or for inactivity. Most of the remaining excess keeps shrinking at close to that pace, so about 90% of the rise is gone after four quarters. A small residual, under a tenth of the original rise, fades at the slower pace of the 0.919 eigenvalue, because that part of the adjustment runs through people leaving and re-entering the labor force.

Two cautions apply. The matrix uses one quarter's flows, and real transition rates change over the business cycle; a model with time-varying rates would be needed to forecast. And the steady state is a property of these flow rates, not a forecast of where unemployment will be in a particular quarter.

The assumptions

The system is square and its matrix is invertible. A system of n linear equations in n unknowns, $Ax = b$, has a unique solution exactly when the determinant of A is non-zero. If the determinant is zero, at least one equation repeats information in the others, and the system has either no solution or infinitely many. In the steady-state calculation, $\pi = \pi P$ on its own is such a degenerate system, since the rows of $P - I$ are linearly dependent, which is why the adding-up condition has to replace one of its equations.

The transition matrix is constant. The Markov chain model assumes that the probability of moving between states depends only on the current state, not on how long someone has been in it or on the date. If transition rates change, for example because job-finding rates fall in a recession, the steady state moves, and powers of a single matrix no longer describe the path.

Every state can be reached from every other. A transition matrix in which all states communicate and the process does not cycle has a unique steady state, and the process converges to it from any starting point. If the states split into separate groups that cannot reach one another, there is more than one steady state and the long-run outcome depends on where the process started; if the process cycles between states, it need not settle at all.

Rows sum to one and entries are non-negative. These are the conditions for a valid transition matrix. They guarantee that 1 is an eigenvalue and that no eigenvalue is larger than 1 in absolute value; together with the previous condition, this makes the steady state the long-run limit rather than a point the process moves away from.

The formula

The derivations below go beyond what is needed for the example, but they are the version used in graduate courses and research.

Systems of equations. A system of n linear equations in n unknowns is written compactly as:

For one unknown at a time, Cramer's rule gives $x_i = \det(A_i)/\det(A)$, where $A_i$ is A with its i-th column replaced by b. It is the standard way to derive comparative statics results by hand. In the two-market example, $p_1 = \det\begin{pmatrix}5 & 1\\ 10 & 3\end{pmatrix}/5 = (15 - 10)/5 = 1$.

The steady state. With P the transition matrix and I the identity matrix, the steady state solves:

Because $P - I$ is singular, one of its columns is replaced by a column of ones and the matching zero on the right-hand side by 1, which imposes the adding-up condition and gives a system with a non-zero determinant and a unique solution. For the EU matrix this gives $\pi = (0.629,\ 0.040,\ 0.332)$.

Eigenvalues and eigenvectors. A scalar $\mu$ and a non-zero vector v satisfy:

Because the labor market shares form a row vector multiplied from the left, the relevant eigenvectors are left eigenvectors: row vectors v with $vP = \mu v$, which are the ordinary eigenvectors of the transpose $P^{\top}$. When P has distinct eigenvalues, any starting vector can be written as the steady state plus one component for each other eigenvalue:

Every term except $\pi$ shrinks geometrically, at a rate set by its eigenvalue. The second-largest eigenvalue in absolute value sets the slowest rate, so a deviation from the steady state halves roughly every $\ln(0.5)/\ln(\mu_2)$ periods. With $\mu_2 = 0.919$ that is about eight quarters for the slowest component of the EU matrix.

The Leontief input-output model. Let A be the matrix of technical coefficients, where $a_{ij}$ is the input from industry i needed per unit of output of industry j, and let d be the vector of final demand. Total output x must cover both intermediate use and final demand:

Each entry of the Leontief inverse gives the total output of one industry needed, directly and through every round of the supply chain, to deliver one unit of final demand for another industry's product. It is the basis of the output multipliers that statistics offices publish with their input-output tables. For example, with two industries and coefficients $a_{11} = 0.2$, $a_{12} = 0.3$, $a_{21} = 0.1$ and $a_{22} = 0.4$ (illustrative), $\det(I - A) = 0.8 \times 0.6 - 0.3 \times 0.1 = 0.45$, and final demand of 100 and 50 requires total output of about 166.7 and 111.1.

Least squares. With an $n \times k$ matrix X of regressors and a vector y of n observations, the OLS estimator is:

It exists only if $X^{\top}X$ is invertible, which fails when one regressor is an exact linear combination of others. That condition is perfect multicollinearity.

Definiteness. A symmetric matrix H is negative definite if $z^{\top} H z < 0$ for every non-zero vector z. A function whose Hessian matrix of second derivatives is negative definite is strictly concave, so a point where its gradient is zero is a maximum.

Checking definiteness through the signs of leading principal minors (the determinants of the top-left 1×1, 2×2, … blocks) is how second-order conditions are verified in optimization. The matrix A in the two-market example is symmetric with leading principal minors 2 and 5, both positive, so it is positive definite. The test reads: a symmetric matrix is positive definite if all its leading principal minors are positive, and negative definite if they alternate in sign starting with a negative one (first < 0, second > 0, third < 0, …). The matrix with rows (−2, 1) and (1, −3) has minors −2 and 5, so it is negative definite.

When this isn't the right tool

Non-linear relationships. Linear algebra handles linear systems exactly. Many economic models are non-linear, and matrix methods apply to them only after linearizing around an equilibrium, which is accurate for small deviations and can mislead for large ones.

Changing transition rates. A single transition matrix assumes constant flow rates. Labor market flows move with the business cycle, so a credible forecast of unemployment uses time-varying rates, estimated with time series methods.

Heterogeneity hidden in averages. An EU-wide matrix averages over countries, ages and skill levels with very different flow rates. Two populations with the same average matrix can have different long-run outcomes once they are modeled separately.

Numerical precision. Inverting a nearly singular matrix magnifies small errors in the data. In applied work, systems are solved with numerically stable methods rather than by computing an explicit inverse.

Practice problems

Problem 1. Solve the system $3x + y = 9$ and $x + 2y = 8$ with Cramer's rule, and check the answer.

Solution. The coefficient matrix has rows (3, 1) and (1, 2), so its determinant is 3 × 2 − 1 × 1 = 5. Replacing the first column with (9, 8) gives the determinant 9 × 2 − 1 × 8 = 10, so x = 10/5 = 2. Replacing the second column gives 3 × 8 − 9 × 1 = 15, so y = 15/5 = 3. Check: 6 + 3 = 9 and 2 + 6 = 8.

Problem 2. In a two-state labor market each quarter, 2% of the employed lose their jobs and 30% of the unemployed find one. Write the transition matrix, find the steady-state unemployment rate, and find the second eigenvalue.

Solution. The matrix has rows (0.98, 0.02) and (0.30, 0.70). In the steady state the flows balance: 0.02 × employed = 0.30 × unemployed, so the unemployment rate is 0.02/(0.02 + 0.30) = 6.25%. The eigenvalues of a 2×2 transition matrix are 1 and its trace minus 1, here 0.98 + 0.70 − 1 = 0.68, so a deviation from the steady state shrinks by 32% each quarter. Check: with 6.25% of the labor force unemployed, inflows 0.02 × 0.9375 = 0.01875 equal outflows 0.30 × 0.0625 = 0.01875.

Further reading

• Optimization and Lagrange Multipliers Explained — internal
• Comparative Statics Explained — internal
• OLS Regression Explained — internal
• Time Series Analysis Explained — internal
• Business Cycles Explained — internal
• Eurostat, Labour market flows in Q3 2025 — external reference
• Eurostat, Euro area unemployment at 6.3% (September 2025 data) — external reference

The transition rates are from Eurostat's labor market flow statistics for the second to third quarter of 2025 and the July to September 2025 EU unemployment rates from Eurostat's euro indicators release of 30 October 2025. The steady state, eigenvalues and shock path are EconoLens calculations from those rates.