Matrix Difference Equation¶

Propagate a vector-valued state at discrete indices through matrix-valued lag operators, using transition products, fixed points, and spectral structure to solve and diagnose the evolution.

Version: v2 · 2026-09-06 · History
Domain-specific #: 2240
Origin domain: discrete dynamical systems
Subdomain: linear systems

Core Idea¶

A matrix difference equation determines a vector-valued state at one discrete index from states at earlier indices through matrix-valued coefficients. Its basic constant first-order form is

\[ x_{k+1}=A x_k+b, \]

where \(x_k\in\mathbb R^n\) or \(\mathbb C^n\), \(A\in\mathbb F^{n\times n}\), and \(b\) is a constant forcing vector. A controlled form replaces \(b\) by \(Bu_k\); a time-varying form permits \(A_k,B_k\). Given the recurrence and enough initial data, forward substitution closes the trajectory. In the homogeneous constant case \(x_{k+1}=Ax_k\), the solution is \(x_k=A^k x_0\). Matrix-valued recurrences are a qualified extension only when vectorization or compatible left/right linear operators are declared. MIT's discrete-systems treatment develops the general time-varying counterpart as an ordered state-transition product rather than a single matrix power.^[1]

The recurrence joins four ideas that should not be collapsed: a discrete clock, a sufficient state, a linear transition operator, and an initial-value contract. The matrix does more than store numbers. Its powers or ordered products propagate the state; its invariant subspaces organize modes; and, in the constant finite-dimensional case, its eigenvalues govern asymptotic behavior. For \(x_{k+1}=Ax_k\), convergence to zero for every initial state is equivalent to \(\rho(A)<1\), where \(\rho\) is spectral radius.^[2]

Constant forcing introduces a fixed-point question. If \(I-A\) is invertible, \(x^*=(I-A)^{-1}b\) is the unique fixed point and deviations \(z_k=x_k-x^*\) satisfy \(z_{k+1}=Az_k\). The forced problem can therefore inherit the homogeneous stability verdict. If \(I-A\) is singular, fixed points may be absent or nonunique; the inverse formula is not licensed.

Higher-order equations have the form

\[ x_k=A_1x_{k-1}+\cdots+A_px_{k-p}+b_k. \]

For \(x_k,b_k\in\mathbb F^n\), \(A_i\in\mathbb F^{n\times n}\), and order \(p\), stack

\[ X_k=\begin{bmatrix}x_k\\x_{k-1}\\\vdots\\x_{k-p+1}\end{bmatrix}\in\mathbb F^{pn}. \]

Then the order-\(p\) recurrence becomes

\[ X_{k+1}= \underbrace{\begin{bmatrix} A_1&A_2&\cdots&A_{p-1}&A_p\\ I_n&0&\cdots&0&0\\ 0&I_n&\ddots&0&0\\ \vdots&\ddots&\ddots&\vdots&\vdots\\ 0&0&\cdots&I_n&0 \end{bmatrix}}_{C\in\mathbb F^{pn\times pn}}X_k+ \underbrace{\begin{bmatrix}b_{k+1}\\0\\\vdots\\0\end{bmatrix}}_{g_{k+1}\in\mathbb F^{pn}}. \]

For the second-order instance \(x_k=A_1x_{k-1}+A_2x_{k-2}+b_k\), the sufficient state is \(X_k=(x_k,x_{k-1})^{\mathsf T}\), the companion operator is \(C=\begin{bmatrix}A_1&A_2\\I_n&0\end{bmatrix}\), and the stacked forcing is \(g_{k+1}=(b_{k+1},0)^{\mathsf T}\). Mapped back: the two vector blocks are the lag memory, the first block row performs the recurrence, the identity block shifts the newest state into memory, and the stacked forcing affects only the new-state block. MIT demonstrates this construction even for the Fibonacci recurrence.^[3]

The abstraction is domain-specific. The portable skeleton is State and State Transition; Matrix supplies the algebraic substrate. Discrete indices, linear matrix coefficients, lag order, matrix powers/products, companion stacking, fixed-point reduction, and unit-disk spectral diagnostics remain constitutive technical commitments.

Structural Signature¶

Sig role-phrases:

the vector-valued state sequence — \(x_k\), the jointly sufficient vector propagated from index to index; a matrix-valued extension must declare vectorization or compatible left/right operators
the discrete index and orientation — integer \(k\), with an explicit convention such as present-to-next \(x_k\mapsto x_{k+1}\)
the lag order and memory closure — the number of past states required, or the stacked state that converts those lags into first-order form
the transition or lag matrices — \(A\), \(A_k\), or \(A_1,\ldots,A_p\), the linear operators connecting earlier states to the next
the forcing or input channel — \(b_k\) or \(B_ku_k\), when evolution is not homogeneous
the initial-condition set — \(x_0\) for first order or \(p\) compatible starting states for order \(p\)
the propagation operator — \(A^k\) for constant first-order homogeneous evolution or an ordered product/state-transition matrix in the time-varying case
the equilibrium or reference trajectory — a fixed point, particular solution, or operating trajectory against which deviations can be studied
the spectral and conditioning diagnostics — eigenvalues, spectral radius, Jordan or invariant-subspace structure, norms, and transient amplification governing long-run and finite-horizon behavior

The roles are jointly diagnostic. A list of matrices is not a matrix difference equation without indexed propagation and initial data. A scalar recurrence can be embedded as a one-dimensional case, but the named family earns its identity from vector-state and operator structure. A nonlinear vector recurrence \(x_{k+1}=f(x_k)\) is a different family unless \(f\) is linear or affine, or a local linearization is explicitly being discussed.

For time-varying homogeneous evolution \(x_{k+1}=A_kx_k\), the transition from index \(j\) to \(k>j\) is

\[ \Phi(k,j)=A_{k-1}A_{k-2}\cdots A_j, \qquad x_k=\Phi(k,j)x_j. \]

Order matters because matrix multiplication need not commute. Replacing this ordered product by the power \(A_k^{k-j}\) is generally wrong. Driven responses add a sum of each past input propagated from its injection time to the present.^[1]

What It Is Not¶

Not any difference equation. Scalar nonlinear, stochastic, logical, or set-valued recurrences need not have a vector state with linear matrix coefficients.
Not a matrix differential equation. A differential equation evolves on a continuous index and uses derivatives and matrix exponentials; a difference equation uses discrete shifts and matrix powers or products.
Not a finite-difference numerical method by definition. A discretized differential equation can produce a matrix recurrence, but many matrix difference equations directly model sampled populations, inventories, signals, or control states rather than approximating a continuous equation.
Not generic state change. State and State Transition does not entail linearity, discrete time, a coefficient matrix, initial-value propagation, or a spectral stability test.
Not a matrix alone. The same matrix can represent a graph, bilinear form, data table, or static linear map. It becomes a transition matrix only inside the indexed recurrence contract.
Not necessarily an algorithm seeking a solution. Forward evaluation is iterative computation, but the recurrence may be the model of the phenomenon itself rather than an optimization or root-finding procedure.
Not necessarily returning to a prior value. An unstable trajectory may grow without revisiting any previous state. It still strictly instantiates Recurrence because each state is generated from indexed lags; literal reappearance is not required.
Not automatically stable when \(\rho(A)\le1\). Unit-modulus eigenvalues with nontrivial Jordan blocks can make \(A^k\) grow. Strict \(\rho(A)<1\) gives asymptotic decay in the finite-dimensional constant case.
Not protected from transient amplification by asymptotic stability. A nonnormal \(A\) can amplify some initial conditions for many steps even though every eigenvalue lies inside the unit circle.
Not the same as input-output stability. Decay of every unforced state, bounded-input bounded-output behavior, numerical-scheme stability, and robustness to coefficient uncertainty are related but distinct verdicts.

Scope of Application¶

Discrete-time control and signal processing. State-space models \(x_{k+1}=Ax_k+Bu_k\), \(y_k=Cx_k+Du_k\) propagate sampled plant or filter states. Transition matrices, controllability/observability calculations, pole placement, and state estimation all depend on this recurrence closure.^[1]

Population and ecological projection. Age- or stage-structured populations use a projection matrix to move counts between stages and add births. Matrix powers forecast the distribution; dominant eigenvalues indicate long-run growth or decay, subject to the model's stationarity assumptions.

Economics and operations. Multisector stocks, lagged prices, inventories, and linearized macroeconomic states can be represented by affine or forced matrix difference equations. Fixed points and eigenmodes distinguish convergence, oscillation, and divergence.

Numerical time stepping. Spatial discretization of a linear evolution equation can yield \(u^{n+1}=F u^n+g^n\). Here the recurrence is also a numerical scheme, and spectral/norm behavior of \(F\) helps diagnose stability. This application must keep physical-model stability separate from discretization stability.

Markov and stochastic linear systems. Probability vectors can evolve by a transition matrix, and state estimates or covariances can follow linear or Riccati-like recurrences. A stochastic process, however, may require noise laws and expectation/covariance semantics beyond the deterministic matrix equation itself.

Higher-order linear recurrences. Autoregressive, vibration, filter, and sampled-system equations with multiple lags are handled by companion stacking. UBC's linear-algebra treatment explicitly distinguishes first-order, second-order, and inhomogeneous matrix difference equations.^[4]

The family is bounded to discrete indexed linear or affine propagation. Nonlinear maps, switching logic, stochastic kernels, differential inclusions, and operator recurrences in infinite-dimensional spaces require additional nodes or explicit qualification.

Clarity¶

Matrix difference equations turn the vague statement “several quantities influence one another over time” into an executable contract. The state vector names what must be remembered, the matrices name which components feed which next components, the forcing names what enters from outside, and the initial conditions fix the trajectory. A disagreement can then be located: wrong state, wrong lag, wrong coefficient, wrong input, or wrong starting point.

They also clarify three questions that prose often confuses. Existence of a trajectory is normally settled by recurrence plus initial data. Existence of a fixed point requires solving \((I-A)x^*=b\). Convergence to that fixed point requires a stability condition on the homogeneous deviation dynamics. A fixed point can exist and still be unstable; a system can be stable around zero while persistent forcing prevents the state from converging to zero.

The stacked-state construction clarifies memory. A second-order rule is not “non-Markovian” after its state is correctly defined: the pair \((x_k,x_{k-1})\) is sufficient for the next step. The modeler must decide whether the named state contains enough lag history, rather than attributing missing information to mysterious dynamics.

Manages Complexity¶

Coupled scalar equations expand quickly: \(n\) variables across \(p\) lags create \(np\) histories and many cross-effects. Matrix notation compresses them into one recurrence without hiding their linear structure. Powers and transition products replace repeated substitution; eigenspaces or Schur forms separate modes; companion stacking reduces many orders to one analysis interface.

The fixed-point transformation \(z_k=x_k-x^*\) removes constant forcing and reuses the homogeneous toolkit. Instead of separately tracking each affine trajectory, an analyst solves one equilibrium equation and studies deviations. The same reduction supports comparative statics: change \(b\), recompute the fixed point, and retain \(A\)'s convergence structure when the transition mechanism is unchanged.

Spectral structure compresses long-horizon reasoning. Rather than simulate every possible initial state forever, the constant finite-dimensional system can be classified by eigenvalues and invariant subspaces. Yet the abstraction also records when this shortcut is unsafe: time variation requires products; nonnormality can create transients; defective unit-circle modes can grow; and coefficient uncertainty can move eigenvalues across the boundary.

Failure is routable by role. A dimension mismatch implicates state or coefficient definition. Different predicted trajectories under the same inputs implicate initial conditions or index convention. Persistent offset implicates forcing or fixed-point specification. Long transient growth with stable eigenvalues implicates nonnormality. Divergence after stacking implicates companion-spectrum or lag estimates. This routing makes revision cheaper than treating the model as one opaque formula.

Abstract Reasoning¶

Initial-state prediction. For a homogeneous linear recurrence, doubling \(x_0\) doubles every \(x_k\). If two trajectories share the same input, their difference follows the homogeneous equation, so stability predicts whether initial-condition disagreement is forgotten.

Fixed-point prediction. If \(x^*=(I-A)^{-1}b\) exists and \(\rho(A)<1\), every trajectory converges to \(x^*\). Changing \(b\) moves the equilibrium but does not change the homogeneous convergence rates when \(A\) is fixed.

Mode prediction. An initial state aligned with an eigenvector \(v\) evolves as \(\lambda^k v\). Magnitude below one decays, above one grows, negative real \(\lambda\) alternates sign, and complex-conjugate modes rotate/oscillate while their modulus sets the envelope.

Lag-state prediction. If the next value depends on two prior values, storing only the latest value cannot determine the future. Stacking both lags restores sufficiency and makes the eigenvalues of the companion matrix the characteristic roots.

Time-variation diagnostic. If coefficients change with \(k\), eigenvalues of each \(A_k\) considered separately do not generally determine the product's stability. Inspect ordered products, induced norms, Lyapunov conditions, or the particular switching structure.

Transient diagnostic. If simulation grows before decaying despite \(\rho(A)<1\), do not conclude that the eigenvalue computation is wrong. Inspect nonnormality, eigenvector conditioning, pseudospectral sensitivity, and the horizon at which asymptotic behavior becomes visible.

Intervention prediction. Altering \(b\) changes the operating point; altering \(A\) changes both equilibrium and modes; state feedback \(u_k=Kx_k\) replaces \(A\) by \(A+BK\). The location of the intervention therefore predicts which outputs can change.^[2]

Model-boundary diagnostic. If coefficients depend on \(x_k\), the global equation is nonlinear. A Jacobian matrix may yield a local matrix difference equation near an operating point, but its stability verdict is local unless stronger nonlinear results are supplied.

Knowledge Transfer¶

The full mechanism transfers literally across sampled control systems, stage-structured populations, multisector inventories, linear signal models, network dynamics, and discretized linear evolution: identify a sufficient vector state, write the lag operators, declare forcing and initial data, propagate, and inspect modes. Names and units change; the matrix recurrence and its diagnostics do not.

The strongest transfer is methodological. Companion stacking turns lag into state. Particular-plus-homogeneous decomposition separates forcing from internal dynamics. Spectral decomposition separates modes. Fixed-point subtraction turns affine evolution into homogeneous deviation dynamics. These operations can move intact from a population projection to a sampled controller.

Transfer must stop at the substrate boundary. A workflow with stages is not a matrix difference equation unless its state is numeric, transitions are linear or affine, and the same coefficient semantics apply. A nonlinear epidemic map or a stochastic hidden-state model may be locally approximated by a matrix recurrence, but the approximation does not inherit global validity. The parent State and State Transition carries the broader insight after matrix linearity is removed.

Examples¶

Canonical¶

Let

\[ A=\begin{bmatrix}0.5&0.2\\0&0.8\end{bmatrix}, \qquad x_0=\begin{bmatrix}10\\5\end{bmatrix}, \qquad x_{k+1}=Ax_k. \]

Then

\[ x_1=\begin{bmatrix}6\\4\end{bmatrix}, \qquad x_2=\begin{bmatrix}3.8\\3.2\end{bmatrix}. \]

The eigenvalues are \(0.5\) and \(0.8\), so \(\rho(A)=0.8<1\). Indeed, the second component is \(5(0.8)^k\), and the first is

\[ \frac{10}{3}(0.8)^k+\frac{20}{3}(0.5)^k, \]

which gives 10, 6, and 3.8 at \(k=0,1,2\) and tends to zero.

Mapped back: \(x_k\) is the state sequence; \(k\to k+1\) is the discrete orientation; order one is the memory closure; \(A\) is the transition matrix; zero input is the forcing channel; \(x_0\) is the initial condition; \(A^k\) is the propagation operator; zero is the equilibrium; and eigenvalues \(0.5,0.8\) are the spectral diagnostics. The off-diagonal \(0.2\) couples the second component into the first without changing the stable verdict.

Applied / In Practice¶

Suppose a two-stage inventory or population model uses

\[ x_{k+1}=\begin{bmatrix}0.7&0.1\\0.2&0.6\end{bmatrix}x_k+ \begin{bmatrix}10\\5\end{bmatrix}, \qquad x_0=\begin{bmatrix}20\\10\end{bmatrix}. \]

The first two updates are \(x_1=(25,15)^{\mathsf T}\) and \(x_2=(29,19)^{\mathsf T}\). Since

\[ (I-A)^{-1}=\begin{bmatrix}4&1\\2&3\end{bmatrix}, \]

the unique fixed point is \(x^*=(45,35)^{\mathsf T}\). The transition eigenvalues are \(0.8\) and \(0.5\), so deviations \(x_k-x^*\) decay and the trajectory converges to that forced equilibrium.

Mapped back: the two stocks or stages form the state; each reporting period is the discrete index; the 2-by-2 matrix is the transition operator; \((10,5)^{\mathsf T}\) is the forcing; the reported opening stocks are the initial condition; repeated powers propagate the trajectory; \((45,35)^{\mathsf T}\) is the reference equilibrium; and the unit-disk eigenvalues are the convergence diagnostic. A policy changing replenishment changes the forcing and equilibrium; a policy changing retention or cross-stage transfer changes \(A\) and therefore the modes as well.

Structural Tensions¶

T1: State compression versus hidden memory. A compact vector simplifies propagation, but omitted lags make identical named states produce different futures. Diagnostic: do two histories with the same \(x_k\) always imply the same next state, or must lagged values be stacked?

T2: Spectral simplicity versus transient reality. Eigenvalues classify asymptotic constant-system behavior, while nonnormal matrices can amplify finite-horizon disturbances. Diagnostic: compare \(\rho(A)\) with norms of \(A^k\) and eigenvector conditioning over the decision horizon.

T3: Stationarity versus changing regimes. A constant \(A\) enables powers and clean modes; real systems may switch seasonally or structurally. Diagnostic: are residual changes random noise around one operator, or evidence that \(A_k\) must vary?

T4: Fixed-point clarity versus singular structure. The inverse formula is compact, but \(I-A\) can be singular. Diagnostic: test rank and compatibility before announcing a unique equilibrium.

T5: Model stability versus input behavior. Stable unforced dynamics forget initial conditions, yet persistent or explosive inputs can dominate outputs. Diagnostic: separate zero-input response from the forced convolution/state-transition sum.

T6: Linear tractability versus nonlinear fidelity. Linear recurrences are analyzable, but coefficients may depend on state. Diagnostic: does one \(A\) predict across the relevant region, or is it only a local Jacobian near an operating point?

T7: Exact propagation versus uncertain coefficients. Matrix powers give exact consequences of the declared model, while small coefficient errors can move sensitive modes. Diagnostic: perturb \(A\), inspect eigenvalue/pseudospectral sensitivity, and report estimation uncertainty.

T8: Autonomy versus reduction. The node is built from State and State Transition plus Matrix, yet it owns a recognized equation family with lag order, companion stacking, forcing, initial-value closure, matrix-power solutions, and spectral verdicts. Diagnostic: if those obligations disappear, route to the parents; if they remain jointly load-bearing, preserve Matrix Difference Equation.

Structural–Framed Character¶

Matrix Difference Equation is structural. Its state, index, matrices, inputs, and spectral claims are formal relations rather than evaluations. No trajectory is intrinsically good or bad; stability becomes desirable only after an application supplies a goal.

Its vocabulary travels intact across multiple technical domains. Engineers, ecologists, economists, and numerical analysts use the same state vector, transition matrix, initial condition, matrix power, eigenvalue, and fixed-point roles. The mathematical identity is recognized rather than imported as a social norm.

Its institutional origin is technical—linear algebra, difference equations, dynamic systems, and control—but the definition is not bound to one institution or professional practice. A matrix recurrence remains the same object outside a laboratory or agency.

Human choices frame model construction: which quantities enter the state, which sampling interval is used, whether coefficients are treated as constant, and what stability notion matters. Those choices affect validity without making the equation itself normative.

Its character: a formally structural, domain-bounded linear-dynamics object whose equations are substrate-neutral among numeric state systems but whose matrix, discrete-index, and spectral apparatus prevents promotion to a general prime.

Structural Core vs. Domain Accent¶

What is skeletal. A present state plus inputs determines a next state; an initial condition launches a trajectory; repeated transition composes local rules into long-horizon behavior. This is prime:state_and_state_transition.

What remains technical. The state is vector-valued, the transition is linear or affine, the index is discrete, lags are finite and stackable, propagation uses matrix products, and stability is read from operator structure. A matrix-valued extension must declare vectorization or compatible left/right linear operators. Those commitments require domain_specific:matrix and applied linear-systems vocabulary.

Why the candidate is not a prime. Substitute a legal case, workflow, chemical phase, or qualitative state machine for the numeric vector and the coefficient-matrix, matrix-power, eigenvalue, and unit-circle statements cease to be literal. Only the state-transition skeleton survives.

Why it is not a mere composite. State and Matrix placed side by side do not entail a difference equation. They do not specify discrete orientation, lag order, initial-value closure, forcing, companion stacking, particular/homogeneous decomposition, or spectral stability. The developed closure supplies reusable reasoning beyond its ingredients.

prime:state_and_state_transition — proposed strict subsumption parent. A matrix difference equation is a state-transition system specialized to discrete linear or affine propagation. For higher order, the stacked lag vector supplies the sufficient current state.
domain_specific:matrix — proposed strict presupposition. Transition coefficients, ordered products, powers, invariant subspaces, and eigenvalues require the matrix apparatus; Matrix is not a taxonomic genus of equations.
prime:recurrence — strict subsumption parent. Every matrix difference equation generates each indexed vector state from one or more earlier indexed states. Literal return or reappearance is not required by the live parent identity.
prime:iteration — related solution/execution pattern. Simulating the recurrence repeats an update, but the equation may model evolution rather than seek a computational fixed answer.
prime:equilibrium — related diagnostic. Fixed points can organize affine systems but are absent, nonunique, or unstable in legitimate instances.

Relationships to Other Abstractions¶

Current abstraction Matrix Difference Equation Domain-specific

Parents (3) — more general patterns this builds on

Matrix Difference Equation is a kind of Recurrence Prime

prime:recurrence — strict subsumption parent. Every matrix difference equation determines each indexed vector state from one or more earlier indexed states; literal return of a value is not required.
Matrix Difference Equation is a kind of State and State Transition Prime

prime:state_and_state_transition — proposed strict subsumption parent. A matrix difference equation is a state-transition system specialized to discrete linear or affine propagation.
Matrix Difference Equation presupposes Matrix Domain-specific

prime:state_and_state_transition — proposed strict subsumption parent. A matrix difference equation is a state-transition system specialized to discrete linear or affine propagation.

Neighborhood in Abstraction Space¶

Matrix Difference Equation sits in a sparse region of the domain-specific corpus (75^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Tell it from Matrix: ask whether a discrete indexed state is being propagated from initial data. A static array or linear map alone is Matrix.
Tell it from State and State Transition: ask whether vector linearity, coefficient matrices, lag order, and spectral analysis are constitutive. If not, use the prime.
Tell it from Recurrence: Matrix Difference Equation strictly specializes Recurrence by requiring vector-valued discrete linear or affine propagation; it need not revisit a prior value.
Tell it from Iteration: ask whether repeated updates describe the modeled system or an algorithm trying to solve another problem.
Tell it from Matrix Differential Equation: ask whether the independent variable is shifted discretely or differentiated continuously; powers and exponentials are not interchangeable.
Tell it from Finite-Difference Method: ask whether the recurrence is the model itself or was constructed to approximate derivatives on a grid.
Tell it from Markov Chain: ask whether the state is a probability distribution and the matrix is stochastic. Those constraints define a narrower probabilistic subtype.
Tell it from Vector Autoregression: ask whether observations are a stochastic time-series model with innovations and estimation assumptions. VAR is a statistical subtype, not the whole deterministic equation family.
Tell it from nonlinear discrete dynamics: ask whether one fixed matrix or affine operator governs the relevant region. A Jacobian recurrence is only a local approximation when coefficients depend on state.
Tell asymptotic from BIBO or numerical stability: ask which inputs, outputs, norms, discretization, and horizon the verdict quantifies.

References¶

[1] M. A. Dahleh, M. Dahleh, and G. Verghese, Lectures on Dynamic Systems and Control, Chapter 10, MIT OpenCourseWare. https://ocw.mit.edu/courses/6-241j-dynamic-systems-and-control-spring-2011/996025f6db0d90b00f11c44fc49b85f9_MIT6_241JS11_textbook.pdf. Verified 2026-08-26. registry ↩a ↩b ↩c

[2] Pablo A. Parrilo and Rekha R. Thomas, Semidefinite Optimization and Convex Algebraic Geometry, §2.2.1, SIAM book manuscript, 2012. https://sites.math.washington.edu/~thomas/frg/frgbook/SIAMBookFinalvNov12-2012.pdf. Verified 2026-08-26. registry ↩a ↩b

[3] MIT 6.310 Dynamic System Modeling and Control Design, “Second Order Homogeneous Linear Difference Equations,” 2025. https://introcontrol.mit.edu/fall25/prelabs/prelab2/sohldes. Verified 2026-08-26. registry ↩

[4] Thomas J. Baird and collaborators, Interactive Linear Algebra, “Discrete Dynamical Systems,” University of British Columbia. https://personal.math.ubc.ca/~tbjw/ila/dds.html. Verified 2026-08-26. registry ↩