Weighing matrix¶

A square 0, +1 and −1 matrix whose rows have equal weight and are mutually orthogonal.

Version: v1 · 2026-10-31 · History
Domain-specific #: 14597
Domain group: Formal Sciences
Origin domain: Mathematics
Subdomain: Combinatorial Matrix Theory → Mathematics

Core Idea¶

A weighing matrix \(W(n,w)\) is a real \(n\times n\) matrix whose entries are \(0,+1,-1\) and whose rows satisfy

\[ WW^{\mathsf T}=wI_n. \]

Here \(n\) is a positive integer and \(w\) is an integer between zero and \(n\). The diagonal of the product says each row has exactly \(w\) nonzero entries; its off-diagonal zeros say distinct rows are orthogonal. The name comes from signed arrangements used to weigh objects together, but the matrix class does not require an actual weighing experiment.^[1]

For \(w>0\), the identity gives \(W^{-1}=W^{\mathsf T}/w\), and the columns satisfy the corresponding orthogonality condition. At \(w=0\), the zero matrix satisfies the equation but has no inverse. Thus recovery of a vector is a positive-weight application, not part of the definition for every weight.^[1]

How would you explain it like I'm…

Plus-Minus Balance Grid

Picture a square grid filled with plus signs, minus signs, and empty spots. Every row has the same number of filled spots. And if you line up any two different rows, the spots where both have signs come out exactly even: as many matching as not matching.

Rows That Cancel Out

A Weighing matrix is a square grid of numbers where every entry is 0, +1 or −1. Each row has exactly the same number of nonzero entries, called the weight w. Also, if you take any two different rows, multiply them spot by spot and add up, you get zero; mathematicians call that being 'orthogonal.' The name comes from using plus and minus patterns to weigh several objects together, but the grid doesn't need an actual scale to count. When w is more than zero, you can use the grid to undo the mix and recover the original values.

Signed Orthogonal-Row Matrix

A Weighing matrix W(n, w) is an n×n matrix with entries 0, +1 and −1 satisfying W Wᵀ = wI, where I is the identity matrix and w is an integer from 0 to n. The diagonal of this product says each row has exactly w nonzero entries, and the zeros off the diagonal say distinct rows are orthogonal (their dot product is zero). The name comes from signed arrangements for weighing objects together, but the definition doesn't require any actual weighing. When w > 0, the inverse is Wᵀ/w, so you can recover a measured vector, and the columns are orthogonal as well. When w = 0, the zero matrix fits the equation but has no inverse, so recovery is an application only for positive weight.

A weighing matrix W(n, w) is a real n×n matrix with entries in {0, +1, −1} satisfying WWᵀ = wI_n, where n is a positive integer and w is an integer with 0 ≤ w ≤ n, called the weight. The diagonal entries of WWᵀ force each row to contain exactly w nonzero entries, and the off-diagonal zeros force distinct rows to be orthogonal. The name derives from designs for weighing objects in signed combinations on a balance, but membership in the class requires no such experiment. For w > 0, the identity implies W⁻¹ = Wᵀ/w, so WᵀW = wI_n as well and the columns satisfy the same orthogonality condition. For w = 0 the zero matrix satisfies the defining equation but is not invertible, so recovering a vector from weighed combinations is an application available only for positive weight.

Structural Signature¶

Square carrier. Positive order \(n\) gives \(n\) row vectors, each with \(n\) coordinates.
Discrete entries. Each coordinate is zero or a unit sign; an arbitrary real orthogonal matrix need not satisfy this restriction.
Common weight. Each row has \(w\) nonzero entries. Because their squares are one, its squared norm is \(w\).
Mutual orthogonality. Distinct rows have inner product zero. Equal numbers of nonzero entries alone do not ensure this condition.
Gram certificate. The single equation \(WW^{\mathsf T}=wI_n\) records both norm and pairwise orthogonality. A proposed matrix can be checked by multiplication.^[1]

What It Is Not¶

A weighing matrix is not any signed square array: two rows with nonzero inner product violate the defining equation. Nor is it any orthogonal matrix, because unrestricted real entries need not lie in \(\{0,\pm1\}\). A matrix can satisfy the formal definition without describing a balance, a code or an experiment.

The zero matrix at weight zero is a boundary case. Its Gram identity holds, but it cannot be inverted to recover arbitrary inputs. Conversely, a measured error improvement depends on a model of the observations and is not a consequence of the alphabet alone.^[1]

Scope of Application¶

This account covers the real square \(0,\pm1\) class. The order and weight are part of the statement: verification of one matrix does not establish existence for every requested pair \((n,w)\). Hadamard matrices are the full-weight case \(W(n,n)\); a zero-diagonal real conference matrix is a weight-\(n-1\) case with a particular arrangement of its zeros.^[1]

The same equation can serve different purposes. In a weighing model, signs assign objects to opposite pans. In a ternary-code construction, the signed integer rows are reduced modulo three and used as code generators. The meaning of the entries changes with the application; the square shape and Gram identity do not.^[1]^[2]

Clarity¶

The word “weight” has two different roles in the balance example. The matrix weight \(w\) is the number of nonzero coefficients in a row; the unknown masses are the quantities being estimated. Keeping them separate prevents a property of the design matrix from being mistaken for a measured object weight.

The equation also distinguishes checking from constructing. Given a candidate matrix, multiply it by its transpose and compare with \(wI_n\). If no matrix has been supplied, the same equation states what a construction must satisfy; it does not supply one automatically.

Manages Complexity¶

For \(n\) rows, checking all norms and distinct-pair inner products separately can be summarized in one matrix product. A diagonal entry different from \(w\) exposes a row-weight failure; a nonzero off-diagonal entry exposes a pair that is not orthogonal. The certificate keeps both kinds of failure visible.

When \(w>0\), the same identity supplies an inverse. That algebraic simplification is useful for reconstruction, although any claim about measurement accuracy still depends on how noise enters the observations.^[1]

Abstract Reasoning¶

Let \(w>0\) and suppose a measurement model is \(y=Wx+\varepsilon\), with \(W=W(n,w)\). Then \(\hat{x}=W^{\mathsf T}y/w\) equals \(x+W^{\mathsf T}\varepsilon/w\). If the errors have mean zero and covariance \(\sigma^2I_n\), the estimate has covariance \(\sigma^2I_n/w\). This follows from the matrix identity and those explicit error assumptions.^[1]

Correlated, unequal-variance or load-dependent errors require a different covariance analysis. The equation alone does not promise a factor-\(w\) practical improvement, and fair comparisons must also account for the measurement design and cost.

Knowledge Transfer¶

The square carrier, sign alphabet, common weight and orthogonality recur in both the hypothetical balance model and the code/lattice example below. In the first, matrix entries organize measurement configurations; in the second, integer rows and their reductions organize a mathematical construction. The shared structure is the matrix certificate, not an assertion that the two applications have the same noise model or physical meaning.^[1]^[2]

Examples¶

Four objects on a two-pan balance¶

In a theoretical example, four objects are weighed in four sign arrangements represented by

\[ H=\begin{pmatrix}1&1&1&1\\1&-1&1&-1\\1&1&-1&-1\\1&-1&-1&1\end{pmatrix}. \]

Each row specifies a weighing and each column an object; opposite signs place objects in opposite pans. The matrix has \(HH^{\mathsf T}=4I_4\), so it is \(W(4,4)\). With independent errors of common variance \(\sigma^2\), transpose reconstruction gives each estimated mass variance \(\sigma^2/4\). This is a model calculation, not a report of a physical trial. Each component uses its own signed column in reconstruction.^[1]

A ternary code and an orthogonal lattice¶

Chandrasekaran, Gandikota and Grigorescu use the integer matrix

\[ M=\begin{pmatrix}1&1&1&0\\1&-1&0&1\\1&0&-1&-1\\0&1&-1&1\end{pmatrix},\qquad MM^{\mathsf T}=3I_4. \]

Thus \(M\) is \(W(4,3)\): every row has three unit signs and different rows are orthogonal. Reducing its rows modulo three gives a length-four ternary linear code; the rows generate an orthogonal integer lattice. The paper uses such blocks, including specified column sign changes, in its characterization of orthogonal Construction-A lattices from ternary linear codes. The code generated by this displayed \(M\) has rank two over \(\mathbb F_3\); length four describes its coordinates, not four independent generators.^[2]

The theorem concerns its stated ternary construction and coordinate equivalence. It does not claim that every weighing matrix or arbitrary modulus yields the same lattice characterization.^[2]

Structural Tensions¶

No universal tradeoff between sparsity, noise and usefulness follows from the matrix definition alone. The balance model shows one conditional variance calculation; another application may have different constraints. A separate question is whether a matrix exists for given \((n,w)\): the ease of checking a supplied candidate does not by itself settle that existence question.

Structural–Framed Character¶

The identity is structural within discrete matrix theory. It has no built-in judgment that a matrix is useful or optimal: the defining claim is a checkable equation over a restricted alphabet. Applications and researchers supply goals such as estimating masses or constructing codes.

Its mathematical representation belongs to human formal practice, while the implications of the stated equation do not depend on an institution choosing to use it. The class is not defined by a particular institution's authority: a balance design helped motivate its name, but membership follows from the entries and equation, wherever the matrix is studied. Words such as Matrix, orthogonal and weight travel among fields, but importing a similar-looking array into another field is not enough. Recognition requires its actual entries and Gram identity.

Its character: a reusable formal matrix class whose role in an application must be established separately from its mathematical membership.

Structural Core vs. Domain Accent¶

The skeletal relationship is mutual orthogonality with equal row norms, certified by a matrix product. Such relationships recur beyond any one application. A future-prime question is whether this equal-norm orthogonality relationship has an independently reusable identity across domains, including carriers other than the discrete arrays defined here.

The domain-bound identity fixes a square real array and the exact alphabet \(0,\pm1\), with weight given by its common number of nonzero entries. Removing those restrictions can leave an orthogonal system, but not necessarily a weighing matrix.

Why the named class does not itself clear the prime bar. Its exact discrete matrix constraints belong to a mathematical formalism, rather than a relation that recurs unchanged in every domain. Broader ideas of orthogonality, constraint and representation may travel further and describe aspects of the construction without capturing all of these restrictions. The two applications show reuse of this mathematical object, not that all orthogonality is a weighing matrix.

Constraint is a conceptual neighbor: the entry alphabet and Gram equation restrict which square matrices belong to this class. A general constraint need not involve matrices or orthogonality, and the presence of restrictions alone does not decide a typed parent relationship.

Neighborhood in Abstraction Space¶

Weighing matrix sits in a sparse region of the domain-specific corpus (79^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2835 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-31

Not to Be Confused With¶

Hadamard matrix. Every entry is a sign, so a real Hadamard matrix is the full-weight case. A weighing matrix with zero entries need not be Hadamard.^[1]
Real conference matrix. In the zero-diagonal convention, the off-diagonal entries are signs and the weight is \(n-1\). A weighing matrix may have other weights and arrangements of zeros.
Orthogonal real matrix. Scaling a positive-weight weighing matrix by \(1/\sqrt w\) gives an orthogonal real matrix, but general orthogonal matrices need not have entries in the weighing-matrix alphabet.
A weighing design. A design may use signs to schedule physical observations; the abstract matrix definition does not guarantee an experiment, independent noise or an accuracy gain.

References¶

[1] Thomas Pender, Balanced Group Matrices: Theory and Applications, University of Lethbridge, 2020, Chapter 2, printed pp. 16–17. The four-object example is a model with stated independent equal-variance errors; the matrix equation supplies the component-specific signs. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k

[2] Karthekeyan Chandrasekaran, Venkata Gandikota and Elena Grigorescu, “Deciding Orthogonality in Construction-A Lattices”, SIAM Journal on Discrete Mathematics 31 (2017), 1244–1262, DOI 10.1137/15M1054766. Matrix \(M\) and Theorem 1.3, p. 1246; Lemma 2.6, p. 1249; proof of Theorem 1.3, pp. 1255–1258. registry ↩a ↩b ↩c ↩d