Linear Transformations and Linear Combinations

Overview

The two axioms of linearity are stated for a single sum and a single scalar multiple. Applying them repeatedly extends both to any finite linear combination at once, giving the identity that a transformation may be moved inside a weighted sum of vectors, or extracted from one. The proof is a short induction and the statement is unsurprising, yet it is invoked more often than any other result about linear transformations, because it is the mechanism by which every structural theorem is proved.

The immediate corollary is the one that changes practice. If $B = \{\vec{u}_1, \ldots, \vec{u}_n\}$ is a basis of the domain $U$ , then every vector $\vec{w} \in U$ has a unique expansion $\vec{w} = \sum_i a_i \vec{u}_i$ . Pushing $T$ through that expansion gives $T(\vec{w}) = \sum_i a_i T(\vec{u}_i)$ , so the $n$ values $T(\vec{u}_1), \ldots, T(\vec{u}_n)$ determine every output the transformation will ever produce. A linear transformation on an $n$ -dimensional space carries exactly $n$ vectors' worth of information, no matter how complicated its internal description.

In engineering terms this is the licence for characterisation by testing. A structure is loaded with a small number of independent load cases and its response measured; a network is probed with a set of independent excitations; a plant is stepped through a set of independent input directions. Provided those cases span the input space and the system is linear, the response to every untested input follows by arithmetic. The number of tests required is the dimension of the input space, and no amount of extra testing adds information beyond that.

The one cost that the standard basis hides is the coordinate step. When the test cases are the standard unit vectors, the coefficients of an arbitrary input are simply its components and no work is needed. When the test cases are combined loads, mixed excitations or any other non-standard independent set — which is the usual experimental situation — predicting a new response first requires solving a linear system to express the new input in terms of the tested ones. That solve is where the arithmetic goes.

Definition

Linear Combination

Given vectors $\vec{u}_1, \ldots, \vec{u}_t$ in a vector space $U$ and scalars $\alpha_1, \ldots, \alpha_t$ , the vector $\alpha_1\vec{u}_1 + \alpha_2\vec{u}_2 + \cdots + \alpha_t\vec{u}_t$ is a linear combination of those vectors. The scalars are the coefficients or weights. Every operation available in a vector space produces linear combinations and nothing else, which is why a map compatible with linear combinations is compatible with everything the space can express.

Basis

A set $B = \{\vec{u}_1, \ldots, \vec{u}_n\}$ that is linearly independent and spans $U$ . Equivalently, every vector of $U$ is a linear combination of the elements of $B$ in exactly one way. Uniqueness of that expansion is what makes the coefficients well defined and hence makes the value of a linear transformation well defined by its action on $B$ .

Spanning alone is not enough for the coefficients to be unique, but it is enough to determine $T$ — different expansions of the same vector must give the same image, precisely because $T$ is linear.

Coordinates with Respect to a Basis

CRB

For a basis $B$ of $U$ and a vector $\vec{w} \in U$ , the unique scalars $a_1, \ldots, a_n$ with $\vec{w} = \sum_i a_i \vec{u}_i$ are the coordinates of $\vec{w}$ relative to $B$ . Collected into a column they form the coordinate vector. Computing them is the practical bottleneck of prediction: for column vectors it is the solve $P\vec{a} = \vec{w}$ , where $P$ has the basis vectors as its columns.

Concepts

Pushing a transformation through a combination

Let $T\colon U \to V$ be linear, let $\vec{u}_1, \ldots, \vec{u}_t$ be vectors of $U$ and $a_1, \ldots, a_t$ scalars. Additivity splits the transformation of the sum into a sum of transformations of the individual terms; homogeneity then extracts each scalar. The result is $T\left(\sum_i a_i \vec{u}_i\right) = \sum_i a_i T(\vec{u}_i)$ , valid for any finite $t$ by induction. Both directions of this identity get used constantly: pushing $T$ into a combination when the individual images are known, and pulling it out when a sum of images needs to be recognised as the image of a single vector.

A basis determines the transformation

If $B = \{\vec{u}_1, \ldots, \vec{u}_n\}$ is a basis of $U$ and $\vec{w} \in U$ , then $\vec{w}$ has a unique expansion $\vec{w} = \sum_i a_i \vec{u}_i$ . Applying the previous identity gives $T(\vec{w}) = \sum_i a_i T(\vec{u}_i)$ . Consequently no formula for $T$ is needed: the $n$ output vectors $T(\vec{u}_i)$ , together with linearity, are a complete description. This also gives the sharpest possible equality test for linear maps — two linear transformations with the same domain and codomain are equal if and only if they agree on a single basis.

Existence of the extension

The converse also holds and is worth stating explicitly. Given a basis $B$ of $U$ and an arbitrary list of $n$ vectors $\vec{v}_1, \ldots, \vec{v}_n$ in $V$ — chosen freely, with no independence or spanning requirement — there is exactly one linear transformation $T\colon U \to V$ with $T(\vec{u}_i) = \vec{v}_i$ . It is defined by $T\left(\sum_i a_i \vec{u}_i\right) = \sum_i a_i \vec{v}_i$ , which is well defined precisely because the coordinates $a_i$ are unique. So specifying a linear transformation is the same task as choosing $n$ output vectors.

Spanning sets work too, but redundantly

If the known set spans $U$ without being independent, the transformation is still determined: every vector is some combination of the set, and the identity applies. What is lost is uniqueness of the coefficients, so the same vector may be expanded in several ways. Linearity guarantees all such expansions give the same image, which is a useful consistency check on measured data: if two different combinations of tested load cases produce the same total load but predict different responses, the system is not linear or the measurements are inconsistent.

Where the arithmetic actually goes

Prediction has two stages: find the coordinates of the new input in the known basis, then form the weighted sum of the known outputs. When the basis is the standard unit vectors, stage one is free — the coordinates are the components — and the whole computation collapses into the matrix-vector product. For any other basis, stage one is a linear solve costing $O(n^3)$ once, or $O(n^2)$ per prediction if the basis matrix is factored and reused. For repeated prediction, factor once.

Relation to the matrix representation

For a transformation between column vector spaces, taking $B$ to be the standard basis makes $T(\vec{u}_i) = T(\vec{e}_i)$ , and collecting those images as columns produces the representing matrix. The identity of this page is the proof of that construction. Choosing a different basis produces a different matrix for the same transformation, related to the standard one by the change-of-basis matrix built from the basis vectors, which is the starting point of the general theory of matrix representations.

The limits of the argument

Two hypotheses do real work and neither can be dropped. The transformation must be linear: for an affine or saturating system the identity fails and predictions from combined tests are simply wrong. The known set must span the domain: if the tested inputs span only a subspace, responses to inputs outside it are not determined by any amount of data from within it. Recording the span of the executed test set, and comparing it against the operating envelope, is therefore part of test planning rather than an afterthought.

Procedure: predict a response from tested cases

Confirm linearity over the operating rangeSuperposition is valid only where the system is linear. Record the range — below yield, below saturation, within converter range — and check that both the tested cases and the target lie inside it.

Check the test set spans the input spaceAssemble the tested input vectors as the columns of

P

and confirm

\operatorname{rank}(P) = n

. A rank-deficient test set leaves some inputs unpredictable, and no post-processing can repair that.

Express the target input in the tested basisSolve

P\vec{a} = \vec{w}

for the coefficients

\vec{a}

. Factor

P

once if many predictions are needed; the same factorisation serves every target.

Form the weighted sum of measured responsesCompute

T(\vec{w}) = \sum_i a_i\, T(\vec{u}_i)

using the recorded output vectors. No further model evaluation is required.

Sanity-check by an independent routeReconstruct

\vec{w}

from the coefficients and confirm it matches the target, then verify the prediction against a formula or an additional measurement where one is available.

Equations

Transformation of a linear combination

EQ-LTLC-01

T (a_{1} {\vec{u}}_{1} + a_{2} {\vec{u}}_{2} + \dots + a_{t} {\vec{u}}_{t}) = a_{1} T ({\vec{u}}_{1}) + a_{2} T ({\vec{u}}_{2}) + \dots + a_{t} T ({\vec{u}}_{t})

Valid for any finite number of terms, obtained by induction from additivity and homogeneity. This is the superposition principle in its general algebraic form.

Summation form

EQ-LTLC-02

T (\sum_{i}^{=} a_{i} {\vec{u}}_{i}) = \sum_{i}^{=} a_{i} T ({\vec{u}}_{i})

The compact statement. The scalars pass through $T$ untouched; only the vectors are transformed.

Unique expansion in a basis

EQ-LTLC-03

\vec{w} = a_{1} {\vec{u}}_{1} + a_{2} {\vec{u}}_{2} + \dots + a_{n} {\vec{u}}_{n}, B = {{\vec{u}}_{1}, {\vec{u}}_{2}, \dots, {\vec{u}}_{n}}

For a basis $B$ the coefficients exist and are unique, which is what makes the definition of a transformation by its values on $B$ unambiguous.

Transformation defined on a basis

EQ-LTLC-04

T (\vec{w}) = \sum_{i}^{=} a_{i} T ({\vec{u}}_{i})

Knowing the $n$ images $T(\vec{u}_i)$ and the coordinates $a_i$ of $\vec{w}$ determines $T(\vec{w})$ completely, with no formula for $T$ required.

The coordinate solve

EQ-LTLC-05

P \vec{a} = \vec{w}, P = [{\vec{u}}_{1} ∣ {\vec{u}}_{2} ∣ \dots ∣ {\vec{u}}_{n}]

For column vector spaces, finding the coordinates is a square non-singular linear system. $P$ is invertible exactly because $B$ is a basis.

Equality test for linear transformations

EQ-LTLC-06

T ({\vec{u}}_{i}) = S ({\vec{u}}_{i}) \forall {\vec{u}}_{i} \in B \Leftrightarrow T = S

Two linear transformations with the same domain and codomain are equal precisely when they agree on one basis. Checking $n$ vectors settles a statement about the whole space.

Prediction as a matrix product

EQ-LTLC-07

T (\vec{w}) = R P^{- 1} \vec{w}, R = [T ({\vec{u}}_{1}) ∣ \dots ∣ T ({\vec{u}}_{n})]

Collecting the two stages, the standard matrix representation of $T$ is $RP^{-1}$ : solve for coordinates, then weight the measured responses. When $B$ is the standard basis $P = I_n$ and the matrix is just $R$ .

Variable Definitions

Symbols used on this page
Symbol	Name	Meaning	Domain / type
$T$	Linear transformation	The map whose values are being predicted	linear map from U to V
$B$	Basis of the domain	Independent spanning set on which $T$ is known	set of n vectors in U
${\vec{u}}_{i}$	Basis vector	A tested input, the $i$ -th element of $B$	element of U
$T ({\vec{u}}_{i})$	Measured response	The recorded output for tested input $\vec{u}_i$	element of V
$a_{i}$	Coordinate	Coefficient of $\vec{u}_i$ in the expansion of the target vector	complex scalar
$\vec{w}$	Target input	The untested input whose response is to be predicted	element of U
$P$	Basis matrix	Columns are the tested input vectors; non-singular when $B$ is a basis	n x n non-singular matrix
$R$	Response matrix	Columns are the measured responses $T(\vec{u}_i)$	m x n matrix
$n$	Domain dimension	Size of the basis, hence the number of tests needed	positive integer

Worked Numerical Example

Problem statement

A bracket is instrumented with two strain gauges. Three combined-load test cases have been executed — each applies a mixture of the three actuator forces, because the rig cannot drive one actuator alone. Determine the gauge response to an untested load combination using superposition alone.

Record the executed test cases and their responses
Each test case is a vector of three actuator forces; each response is a pair of gauge readings in microstrain. The transformation from forces to strains is linear over the elastic range, which is the hypothesis everything below depends on.
$T ([\begin{matrix} 1 \\ 1 \\ 0 \end{matrix}]) = [\begin{matrix} 3 \\ - 1 \end{matrix}], T ([\begin{matrix} 0 \\ 1 \\ 1 \end{matrix}]) = [\begin{matrix} 2 \\ 4 \end{matrix}], T ([\begin{matrix} 1 \\ 0 \\ 1 \end{matrix}]) = [\begin{matrix} - 5 \\ 2 \end{matrix}]$
Confirm the test set is a basis
Assemble the three load vectors as the columns of $P$ and check that it is non-singular. Its determinant is $2 \ne 0$ , so the three executed cases are linearly independent and span $\mathbb{C}^3$ . Every possible load combination is therefore predictable from these three tests.
$P = [\begin{matrix} 1 & 0 & 1 \\ 1 & 1 & 0 \\ 0 & 1 & 1 \end{matrix}], \det (P) = 2 \neq 0$
State the target load case
The design condition to be assessed applies actuator forces $(5, 1, 2)$ , a combination that was never tested directly. No new measurement will be taken.
Solve for the coordinates
Row-reduce the augmented system $[\, P \mid \vec{w} \,]$ . The reduced form is the identity with the coefficients in the final column, so the target load is $2$ times case one, minus case two, plus $3$ times case three.
$[\begin{matrix} 1 & 0 & 1 & 5 \\ 1 & 1 & 0 & 1 \\ 0 & 1 & 1 & 2 \end{matrix}] ⟶ [\begin{matrix} 1 & 0 & 0 & 2 \\ 0 & 1 & 0 & - 1 \\ 0 & 0 & 1 & 3 \end{matrix}], \vec{a} = [\begin{matrix} 2 \\ - 1 \\ 3 \end{matrix}]$
Verify the decomposition before using it
Reconstruct the target from the coefficients as a check on the solve: $2(1,1,0)^{t} - (0,1,1)^{t} + 3(1,0,1)^{t} = (2 + 3,\; 2 - 1,\; -1 + 3)^{t} = (5, 1, 2)^{t}$ . The decomposition is correct.
Superpose the measured responses
Apply the transformation to the expansion. Linearity moves $T$ inside the combination, and each term becomes a scaled copy of a response that has already been measured.
$T (\vec{w}) = 2 [\begin{matrix} 3 \\ - 1 \end{matrix}] - 1 [\begin{matrix} 2 \\ 4 \end{matrix}] + 3 [\begin{matrix} - 5 \\ 2 \end{matrix}] = [\begin{matrix} 6 - 2 - 15 \\ - 2 - 4 + 6 \end{matrix}] = [\begin{matrix} - 11 \\ 0 \end{matrix}]$
Recover the standard-basis matrix for repeated use
If many further predictions are needed, convert once to the standard representation $A = RP^{-1}$ , whose columns are the responses to unit forces at each actuator. These were never measured directly; they are inferred from the combined tests.
$A = R P^{- 1} = [\begin{matrix} - 2 & 5 & - 3 \\ - \frac{3}{2} & \frac{1}{2} & \frac{7}{2} \end{matrix}]$
Cross-check the prediction
Apply the recovered matrix to the target load directly: gauge one gives $-2(5) + 5(1) - 3(2) = -11$ , and gauge two gives $-\tfrac{3}{2}(5) + \tfrac{1}{2}(1) + \tfrac{7}{2}(2) = -7.5 + 0.5 + 7 = 0$ . Both agree with the superposition result, confirming the coordinate solve and the weighted sum independently.
$Cross-check the prediction$

Result

The untested design load produces $-11$ microstrain at gauge one and, notably, exactly zero at gauge two: the contributions of the three actuators cancel there. Three combined-load tests have characterised the bracket completely, so every further load case is answered by arithmetic rather than by rig time. The zero at gauge two is a warning rather than a comfort — that gauge is blind to this particular load direction, and a monitoring scheme relying on it would miss the condition entirely.

Applications & Industry Use

Structural engineering

Design load combinations from unit load cases

Analysis codes prescribe combinations such as dead plus live plus a factored wind case. Each elementary case is solved once and the combinations formed as weighted sums of the stored results. The saving is large, since the number of prescribed combinations typically exceeds the number of elementary cases by an order of magnitude.

Electrical engineering

Superposition analysis of multi-source networks

A linear network with several independent sources is analysed one source at a time, with the others suppressed, and the branch results summed. The elementary source excitations form a basis of the source space, so the method is exactly the theorem of this page applied to a circuit.

Signal processing

Basis decomposition and filtering

Decomposing a signal into a basis — sinusoids, wavelets, or the columns of a learned dictionary — and passing each element through a linear filter, then recombining, gives the same result as filtering the original signal. This is why a frequency response characterises a linear filter completely and why fast transform methods are legitimate.

Experimental mechanics

Characterisation from combined-load rig tests

Test rigs frequently cannot isolate a single actuator, so measured cases are combinations. Provided the executed cases are independent, the unit responses are recoverable by solving for coordinates, and the full influence matrix is reconstructed without ever running a single-actuator test.

Optics & imaging

Point spread functions and image formation

A linear shift-invariant imaging system is characterised by its response to a point source. An arbitrary scene is a superposition of point sources, so the image is the corresponding superposition of point spread functions — the integral form of the identity on this page, and the basis of deconvolution.

Thermal engineering

Superposition of heat sources

For a linear conduction problem, the temperature field due to several sources is the sum of the fields due to each source alone. Precomputing the field for each elementary source lets a designer evaluate any placement or power distribution instantly, which is standard practice in electronics thermal layout.

Design Considerations

Plan the test set to span, not merely to be realistic

The value of a test programme is set by the span of its executed cases, not by how representative each case looks. A set of realistic but nearly parallel load cases leaves most of the input space undetermined and yields an ill-conditioned coordinate solve. Choose cases to be well separated, and confirm the rank of the assembled test matrix before the rig is released.

Report the condition number of the test matrix

Prediction error scales with $\kappa(P)$ : nearly dependent test cases amplify measurement noise in the recovered coefficients. A test set that is technically independent but poorly conditioned gives predictions that are formally valid and practically worthless. Quote $\kappa(P)$ alongside the results.

Factor once when predictions are repeated

Each prediction requires solving $P\vec{a} = \vec{w}$ . Computing an LU factorisation of $P$ once reduces every subsequent prediction to $O(n^2)$ , and forming $A = RP^{-1}$ once reduces it to a single matrix-vector product. Explicitly inverting $P$ is the least accurate of the three routes and should be reserved for the case where the standard-basis matrix itself is the deliverable.

Extra tests buy noise reduction, not new information

Once the test set spans the domain, further cases add no new determinacy. They are still worth running, but for a different reason: a least-squares fit over redundant cases averages measurement error and provides a residual that detects non-linearity. Budget them as validation rather than as characterisation.

Validate the linearity assumption at the extremes

Superposition is exact only where the model is linear. Reserve at least one test at or near the design limit to be compared against the superposed prediction from the low-level cases. A systematic discrepancy that grows with amplitude is the signature of the non-linearity that invalidates the whole method.

Record the span, not just the results

A prediction for an input outside the span of the executed cases is an extrapolation with no justification from the data. Store the test matrix with the results so that any later query can be checked for membership of the tested subspace before an answer is issued.

Standards & Codes

Notation, interchange and numerical standards that govern how this material is written down, stored and computed in production systems.

Applicable standards, conventions and reference implementations
Reference	Title	Relevance to this topic
`ISO 80000-2`	Quantities and units — Part 2: Mathematics	Fixes the summation and indexing conventions used for linear combinations, including the placement of limits on $\sum$ and the italic convention for indices.
`EN 1990 (Eurocode 0)`	Basis of structural design	Prescribes the load combination rules that are, mathematically, weighted sums of elementary load case responses; their validity rests on the linearity assumed here.
`IEC 60050-131`	International Electrotechnical Vocabulary — Circuit theory	Defines the superposition theorem for electrical networks and the conditions under which sources may be treated one at a time.
`ISO 5725-2`	Accuracy of measurement methods and results — Basic methods	Provides the framework for quantifying repeatability of the measured responses, which propagates directly into the uncertainty of any superposed prediction.
`ISO/IEC 40314`	Mathematical Markup Language (MathML) Version 3.0	Encodes the summation and basis notation on this page semantically so that indices and limits are exposed correctly to assistive technology.

Material Selection

For a mathematical topic, "material" is the numeric representation: the scalar field, storage format and precision the computation is built from.

Representation and precision selection
Representation	Select when	Trade-off
Standard unit vector test basis	Each input channel can be excited independently at unit magnitude.	Coordinates are free and the response matrix is the standard representation directly, but single-channel excitation is often physically impossible or produces responses too small to measure.
Combined-load test basis	The rig or plant cannot isolate individual inputs, so every executed case is a mixture.	Achieves realistic excitation levels and better signal-to-noise, at the cost of a coordinate solve per prediction and sensitivity to the conditioning of the test matrix.
Orthogonal test design	The test set can be chosen freely and prediction accuracy matters more than physical convenience.	Gives $\kappa(P) = 1$ so noise is not amplified and the coordinate solve is a transpose-multiply, but orthogonal excitations may be unrealisable in a physical rig.
Redundant test set with least-squares fit	Measurement noise is significant and rig time is available.	Averages error and yields a residual that flags non-linearity, at the cost of extra tests and a pseudoinverse rather than a square solve.
Exact rational coordinates	Symbolic verification, small analytic examples and derivation of a standard-basis matrix by hand.	Produces exact coefficients such as $\tfrac{7}{2}$ with no rounding, but is unavailable for measured data and impractical at scale.
IEEE 754 binary64 with a stored LU factorisation	Many predictions from one fixed test basis, as in a design-office load combination workflow.	Reduces each prediction to $O(n^2)$ and is numerically sound, but the factorisation must be invalidated and recomputed whenever the test set changes.

Manufacturing Notes

Implementation notes — how the result is actually produced by hand, by algorithm and by library, including cost and numerical behaviour.

Cost breakdown of a prediction

Stage one, the coordinate solve, costs $\tfrac{2}{3}n^3$ operations for a fresh LU factorisation and $2n^2$ for each subsequent solve reusing it. Stage two, the weighted sum of $n$ measured response vectors of length $m$ , costs $2mn$ operations. For repeated prediction the first factorisation dominates once and never again, which is why the factorisation should be stored with the test data.

Doing it by hand

Row-reduce the augmented array $[\, P \mid \vec{w} \,]$ to read the coordinates directly from the final column. Then multiply each measured response by its coefficient and add componentwise. Reconstructing $\vec{w}$ from the coefficients before proceeding is a cheap and effective check — almost every error in this procedure occurs in the coordinate solve, not in the weighted sum.

Recovering the standard representation

Collect the tested inputs as the columns of $P$ and their responses as the columns of $R$ ; then the standard-basis matrix is $A = RP^{-1}$ , obtained in practice by solving $A P = R$ , that is by solving $P^{t}A^{t} = R^{t}$ with a single factorisation and $m$ right-hand sides. Never form $P^{-1}$ explicitly unless the inverse itself is required.

Library support

The coordinate solve is numpy.linalg.solve(P, w) or scipy.linalg.lu_solve with a stored factorisation; the redundant case is numpy.linalg.lstsq, which also returns residuals that are the natural non-linearity detector. Check numpy.linalg.cond(P) as a matter of routine, since a large value invalidates the prediction long before the solve reports any error.

Detecting a violated hypothesis

If the executed test set contains more cases than the domain dimension, the extra cases give a free consistency test: predict each redundant case from the others and compare against its measurement. A residual growing with load amplitude indicates non-linearity; a residual that is large but amplitude-independent indicates a measurement or bookkeeping error. The two failure modes are distinguishable and should be reported separately.

Failure Modes & Common Mistakes

Failure modes, root causes and prevention
Failure mode / mistake	Impact	Root cause	Prevention & detection
Superposing responses of a non-linear system	high	Applying the identity to a system that saturates, yields, or carries a constant offset.	Establish and record the linear operating range, and validate one high-amplitude case against a prediction built from low-amplitude cases.
Test set that does not span the input space	high	Executed cases are linearly dependent, so some input directions were never excited.	Compute the rank of the assembled test matrix before releasing the rig; a rank below $n$ means the missing directions must be tested, not interpolated.
Ill-conditioned test basis	high	Test cases are technically independent but nearly parallel, so the coordinate solve amplifies measurement noise.	Report $\kappa(P)$ with the results and design the test set for separation; treat a large condition number as a test-planning defect.
Extrapolating outside the tested span	medium	A query input lies outside the subspace spanned by the executed cases, but the arithmetic proceeds regardless.	Store the test matrix with the results and check membership of the span before answering; flag any input requiring extrapolation.
Coordinate solve errors propagating silently	medium	A wrong coefficient produces a plausible-looking prediction with no visible symptom.	Reconstruct the target input from the computed coefficients and compare with the original before forming the weighted sum.
Mixing coordinate systems or sign conventions between tests	medium	Load or gauge sign conventions differ between test sessions, so responses are combined inconsistently.	Fix and document conventions once, and include a repeat of one earlier case in every later session as a convention check.
Assuming coefficients are unique for a merely spanning set	low	The known set spans but is not independent, so different expansions of the same vector exist.	The prediction is still correct, but do not report the coefficients as properties of the input; reduce the set to a basis if the coefficients are to be interpreted.
Treating additional tests as additional accuracy of determination	low	Believing that more than $n$ tests improves the determinacy of the transformation rather than the estimate.	Once the span is achieved, budget further tests explicitly as noise reduction and validation, and process them by least squares rather than by selection.

FAQs

Why does knowing a linear transformation on a basis determine it everywhere?

Because every vector in the domain is a unique linear combination of the basis vectors, and linearity moves the transformation inside that combination. The image is therefore the same combination of the known images. No other information about the transformation is needed or exists.

How many tests do I need to characterise a linear system?

Exactly the dimension of the input space, provided the test inputs are linearly independent. Fewer leaves some input directions undetermined and no processing can recover them. More adds no new determinacy, though redundant tests are valuable for averaging noise and for detecting non-linearity.

Can I use a spanning set that is not a basis?

Yes for prediction, no for interpretation. A spanning set determines the transformation, since every vector is some combination of it, and linearity guarantees all valid expansions give the same image. What is lost is uniqueness of the coefficients, so the coefficients themselves carry no intrinsic meaning.

Why is prediction easier with the standard unit vectors?

Because the coordinates of a vector with respect to the standard basis are simply its components, so the coordinate solve disappears and prediction is a single matrix-vector product. For any other basis the coordinates must be found by solving $P\vec{a} = \vec{w}$ , which is where the extra arithmetic goes.

Can I choose the values on the basis freely?

Yes. Given a basis of the domain and any list of vectors in the codomain of matching length, there is exactly one linear transformation taking each basis vector to the corresponding value. The chosen values need not be independent, need not span, and may repeat — each choice defines a different but perfectly valid linear transformation.

What happens to my prediction if the system is slightly non-linear?

The error grows with the size of the coefficients, since a mild non-linearity in each elementary response is amplified by the weight applied to it. A prediction assembled from large positive and negative coefficients is therefore far more exposed than one with small weights, which is an argument for choosing a test basis close to the operating region of interest.

How do I recover unit responses from combined-load tests?

Assemble the tested inputs as the columns of $P$ and their measured responses as the columns of $R$ . The standard-basis matrix, whose columns are the unit responses, is $A = RP^{-1}$ , computed in practice by solving $AP = R$ rather than by inverting $P$ . Its condition number sets the reliability of the recovered unit responses.

References

Beezer, R. A. A First Course in Linear Algebra, Version 0.70. University of Puget Sound, 2006. Section LT, subsection LTLC. Licensed under the GNU Free Documentation License v1.2.
Axler, S. Linear Algebra Done Right, 3rd edition. Springer, 2015. Chapter 3 on linear maps and their determination on a basis.
Strang, G. Introduction to Linear Algebra, 5th edition. Wellesley-Cambridge Press, 2016.
EN 1990:2002, Eurocode — Basis of structural design. European Committee for Standardization.
ISO 80000-2:2019, Quantities and units — Part 2: Mathematics. International Organization for Standardization.

AI Suggested Questions

My three executed load cases have a condition number of 400. How much does that degrade a predicted response, given 2% measurement noise?
Show that any assignment of values on a basis extends to exactly one linear transformation, and where uniqueness of coordinates is used in the proof.
Design a set of four excitation vectors for a three-input system that minimises noise amplification in the recovered unit responses.
Work an example where the known set spans the domain but is not independent, and show that two different expansions give the same image.
How does superposition of point spread functions in imaging relate to the finite-dimensional identity on this page?
Given combined-load tests and their responses, derive the standard influence matrix without ever forming an explicit inverse.

Related Calculators

Superposition Predictor

Enter tested inputs with their measured responses and any target input; returns the coordinates, the predicted response and the condition number of the test matrix.

Coordinate Vector Calculator

Express any vector in terms of a chosen basis by row reduction, with the reconstruction check shown.

Transformation Matrix Builder

Recover the standard-basis matrix $A = RP^{-1}$ from an arbitrary independent set of tested inputs and responses.