Coefficient Restrictions as C B M = Theta

Why Write Restrictions in Matrix Form?

In a multivariate regression, each predictor has a coefficient for each response. A hypothesis about one coefficient is therefore only one possible restriction: we may instead compare predictors, combine responses, or impose several such requirements at once. The form $C B M=\Theta$ organizes these restrictions by separating the predictor-side combinations from the response-side combinations and recording their required values in one target matrix.

This notation is useful only when its dimensions and meaning are clear. The matrix $C$ acts on the predictor rows of the coefficient matrix; $M$ acts on its response columns. Their product specifies which combinations are being restricted, while $\Theta$ says what those combinations must equal. The equation describes a hypothesis, not by itself a test statistic or a conclusion about the data.

Dimensions and Interpretation

Let $Y$ be an $n\times q$ response matrix, $X$ an $n\times p$ design matrix, and $B$ the $p\times q$ coefficient matrix in the linear model $Y=XB+U$. Each row of $B$ corresponds to one design column, and each column corresponds to one response. For example, a row might contain the coefficients for an intercept or for a particular predictor across all $q$ responses.

Definition: For fixed matrices $C\in\mathbb{R}^{a\times p}$, $M\in\mathbb{R}^{q\times b}$, and a target matrix $\Theta\in\mathbb{R}^{a\times b}$, the coefficient restriction is $$ C B M=\Theta. $$ The rows of $C$ specify $a$ linear combinations of predictor coefficients. The columns of $M$ specify $b$ linear combinations of responses. The entry $\Theta_{ij}$ is the required value of the combination formed by row $i$ of $C$ and column $j$ of $M$.

The dimensions confirm that the multiplication is defined: $CB$ is $a\times q$, and $CBM$ is $a\times b$, the same size as $\Theta$. If $c_i^T$ denotes row $i$ of $C$, and $m_j$ denotes column $j$ of $M$, then entry $(i,j)$ of the restriction is

(CBM)_{ij}=c_i^T Bm_j=\Theta_{ij}.

Thus the notation combines two operations. First, $c_i^T B$ forms a response-wise summary of coefficient rows. Then multiplying by $m_j$ combines the responses in that summary. A row of $C$ may select one predictor coefficient row, or it may compare or combine several rows. A column of $M$ may select one response or form a response contrast. The word “contrast” is often used for a difference or other combination that cancels a common component; the matrices can also simply select coefficients or responses.

The target need not be zero. The special case $\Theta=0$ states that every specified combination vanishes. A nonzero target states exact values for those combinations. Both are linear restrictions on $B$.

Worked Example: Comparing a Treatment Effect Across Two Responses

Suppose a Gaussian multivariate regression has three design columns: an intercept, a treatment indicator, and a continuous covariate. There are two responses, and the fitted coefficient matrix is

B= \begin{pmatrix} 10&8\\ 2&1\\ 0.5&0.8 \end{pmatrix}.

The second row gives the treatment effects, $2$ for response 1 and $1$ for response 2. To ask whether those treatment effects are equal, choose $C=(0,1,0)$ to select the treatment row and $M=(1,-1)^T$ to subtract response 2 from response 1. The target is the scalar zero:

CBM = (0,1,0) \begin{pmatrix} 10&8\\ 2&1\\ 0.5&0.8 \end{pmatrix} \begin{pmatrix} 1\\-1 \end{pmatrix} = (2,1) \begin{pmatrix} 1\\-1 \end{pmatrix} =2-1=1.

The proposed restriction is $CBM=0$, but this coefficient matrix gives $CBM=1$, so it does not satisfy the restriction. The calculation compares treatment effects, not response means: the intercept row was excluded by $C$.

Reading Multiple Restrictions

When $C$ has several rows and $M$ has several columns, the matrix equation specifies all pairings between them. Every row of $C$ is paired with every column of $M$; the corresponding target entry gives the required value. This structure is important: a single product $CBM$ does not select an arbitrary list of unrelated combinations. It creates a rectangular array of paired restrictions.

Worked Example: Several Predictor Combinations for Each Response

Consider three design columns: an intercept, predictor $x_1$, and predictor $x_2$. There are two responses, with coefficient matrix

B= \begin{pmatrix} 5&7\\ 2&3\\ 1&4 \end{pmatrix}.

Choose

C= \begin{pmatrix} 0&1&-1\\ 0&0&1 \end{pmatrix}, \qquad M=I_2, \qquad \Theta= \begin{pmatrix} 1&-1\\ 0&0 \end{pmatrix}.

The first row of $C$ compares the coefficients of $x_1$ and $x_2$; the second selects the coefficient of $x_2$. Since $M=I_2$, each restriction is imposed separately for both responses. Compute the product:

CB= \begin{pmatrix} 0&1&-1\\ 0&0&1 \end{pmatrix} \begin{pmatrix} 5&7\\ 2&3\\ 1&4 \end{pmatrix} = \begin{pmatrix} 2-1&3-4\\ 1&4 \end{pmatrix} = \begin{pmatrix} 1&-1\\ 1&4 \end{pmatrix}.

The first row matches the target $(1,-1)$, but the second row is $(1,4)$, not $(0,0)$. Therefore $CBM\ne\Theta$. In words, the $x_1$-minus-$x_2$ coefficients meet their specified values for both responses, while the $x_2$ coefficients do not both equal zero.

The Scalar and Vectorized Forms

A matrix restriction can always be unpacked into scalar equations. This gives a direct way to check what each row and column contributes.

Result: The equation $CBM=\Theta$ holds if and only if $$ c_i^T Bm_j=\Theta_{ij} \quad\text{for every }i=1,\ldots,a\text{ and }j=1,\ldots,b. $$

Proof. By the definition of matrix multiplication, entry $(i,j)$ of $CBM$ is the product of row $i$ of $C$, the matrix $B$, and column $j$ of $M$, namely $c_i^T Bm_j$. Two $a\times b$ matrices are equal exactly when all corresponding entries are equal. Thus $CBM=\Theta$ holds exactly when every displayed scalar equation holds. $\square$

There is also a standard vectorized representation. Use column stacking for $\operatorname{vec}(B)$: the first column of $B$ is followed by the second, and so on. The Kronecker product $A\otimes D$ is the block matrix whose blocks are the entries of $A$ multiplied by $D$.

Result: With column-stacking vectorization, $$ CBM=\Theta \quad\Longleftrightarrow\quad (M^T\otimes C)\operatorname{vec}(B)=\operatorname{vec}(\Theta). $$

Proof. The $(i,j)$ entry of $CBM$ is $$ (CBM)_{ij}=\sum_{r=1}^{p}\sum_{s=1}^{q} C_{ir}B_{rs}M_{sj}. $$ In $\operatorname{vec}(CBM)$, that entry occupies position $(j-1)a+i$. In row $(j-1)a+i$ of $M^T\otimes C$, the coefficient multiplying $B_{rs}$, whose position in $\operatorname{vec}(B)$ is $(s-1)p+r$, is $M_{sj}C_{ir}$. The resulting component of $(M^T\otimes C)\operatorname{vec}(B)$ is therefore the displayed double sum. This is true for every $i,j$, so $$ \operatorname{vec}(CBM)=(M^T\otimes C)\operatorname{vec}(B). $$ Vectorization preserves matrix equality entry by entry, giving the stated equivalence. $\square$

The dimensions provide a useful check: $M^T\otimes C$ has $ab$ rows and $pq$ columns, while $\operatorname{vec}(B)$ has length $pq$ and $\operatorname{vec}(\Theta)$ has length $ab$. This form makes explicit that $CBM=\Theta$ is a set of linear equations in the entries of $B$.

Worked Example: Combining Predictor and Response Differences

Let there be two design columns and three responses, with coefficient matrix

B= \begin{pmatrix} 4&6&5\\ 3&2&1 \end{pmatrix}.

Take $C=(1,-1)$, which compares the two coefficient rows, and let the two columns of $M$ compare response 1 with response 3 and response 2 with response 3:

M= \begin{pmatrix} 1&0\\ 0&1\\ -1&-1 \end{pmatrix}.

First compute the predictor-side combination, then the two response differences:

CB=(4-3,\;6-2,\;5-1)=(1,4,4),

CBM=(1,4,4) \begin{pmatrix} 1&0\\ 0&1\\ -1&-1 \end{pmatrix} = (1-4,\;4-4)=(-3,0).

Thus the target for these combinations is $\Theta=(-3,0)$. The first entry says that the predictor-row difference for response 1 minus response 3 equals $-3$; the second says that the corresponding difference for response 2 minus response 3 equals zero. The two entries are not a single omnibus scalar restriction: they are two specified equations.

Choosing the Matrices from a Verbal Hypothesis

A reliable construction begins with the verbal statement, not with guessed matrix dimensions. Identify the coefficient rows being combined, write their weights as a row of $C$, identify the response combination and write its weights as a column of $M$, then place the required value in the corresponding entry of $\Theta$. Repeat for each intended combination and check that all row-column pairings are genuinely part of the hypothesis.

Index the coefficient matrix.
Record which design column corresponds to each row of $B$, and which response corresponds to each column.

Encode predictor-side combinations.
Put the weights for each selected or compared coefficient row into a row of $C$.

Encode response-side combinations.
Put the weights for each selected or compared response into a column of $M$.

Specify the target and verify.
Use the required values in $\Theta$, then check dimensions and expand at least one entry as $c_i^T Bm_j$.

The predictor-side weights are tied to the chosen design columns. For example, a row of $C$ can select the coefficient of a treatment indicator, or subtract one predictor's coefficient from another's. The response-side weights are tied to the ordered response columns. A response difference must therefore use the correct signs and column order. If a factor has been represented through a particular coding scheme, the meaning of its coefficient rows depends on that scheme; the restriction must be written for the actual $B$ in the model.

The algebraic equation also does not guarantee that a restriction is identifiable from the design. If $X$ is rank deficient, some coefficient matrices can give the same fitted means. For a nontrivial response combination, the predictor-side rows used in $C$ must correspond to estimable coefficient combinations for the restriction to be identifiable. This is distinct from writing down the restriction: an equation can be algebraically well-formed even when the data design cannot determine its left-hand side.

Common Mistakes and Why the Form Matters

Reversing the roles of $C$ and $M$. $C$ multiplies on the left and combines coefficient rows; $M$ multiplies on the right and combines response columns. Their dimensions make this distinction explicit.
Assuming the target is always zero. A null hypothesis may set combinations to zero, but the matrix form also represents equality to nonzero specified values.
Checking only one entry of a matrix equation. If $CBM$ and $\Theta$ are $a\times b$, all $ab$ corresponding entries must match.
Using an unintended pairing. With multiple rows in $C$ and columns in $M$, the product imposes every row-column pairing. If the verbal hypothesis calls for a different collection, the matrices must be redesigned or the restrictions written as separate equations.
Ignoring response order or predictor coding. A sign or index error changes the linear combination. State the order of responses and the meaning of design columns before constructing weights.

The central value of $CBM=\Theta$ is organizational: it makes a multivariate coefficient hypothesis readable as predictor combinations crossed with response combinations. Its scalar and vectorized forms also provide mechanical checks on interpretation, dimensions, and signs. Later calculations for a test can then start from a clearly specified hypothesis rather than an ambiguous verbal description.

Check Your Understanding

For each question, identify the relevant dimensions or write the restriction in words before calculating.

If $B$ is $4\times 3$, $C$ is $2\times 4$, and $M$ is $3\times 1$, what are the dimensions of $CBM$ and $\Theta$?
What coefficient combination is selected by $C=(0,1,0)$, and what response combination is formed by $M=(1,-1)^T$?
In $CBM=\Theta$, how many scalar equations are specified when $C$ has three rows and $M$ has two columns?
Why does $M=I_q$ impose the predictor-side restrictions separately for each response?
Explain why a well-formed equation $CBM=\Theta$ does not, by itself, guarantee that the restriction is estimable under a rank-deficient design.