Geometric Interpretation of Nonnegative Components

From Additive Profiles to a Cone

In multi-omics data, a sample may have nonnegative measurements across many features, such as expression levels or abundances. Nonnegative matrix factorization (NMF) represents each sample as an additive combination of a small set of nonnegative component profiles. Geometrically, those profiles point in directions in feature space, and nonnegative coefficients allow movement along those directions but not backward along them. The set of all such combinations is a cone.

This geometry gives a direct interpretation of the model introduced in the earlier material on Nonnegative Matrix Factorization as a Data Model. It explains why an observation can be described as a mixture of component profiles, why negative coefficients are incompatible with that interpretation, and how a fixed set of profiles constrains which data vectors can be represented exactly.

Definition: Let $X$ be a nonnegative data matrix with $p$ features and $n$ samples, and write $X \approx WH$, where $W$ is a $p \times r$ nonnegative matrix and $H$ is an $r \times n$ nonnegative matrix. The $k$th column of $W$, denoted $w_k$, is the $k$th component profile. The $j$th column of $X$, denoted $x_j$, is the feature vector for sample $j$, and the $k$th entry of the $j$th column of $H$, denoted $H_{kj}$, is its coefficient for profile $w_k$. For a fixed $W$, the cone generated by its profiles is $$ \mathcal{C}(W)=\left\{\sum_{k=1}^{r} a_k w_k : a_k \geq 0 \text{ for every } k\right\}. $$

An exact factorization places every sample vector inside $\mathcal{C}(W)$: $x_j=\sum_{k=1}^{r} H_{kj}w_k$. In an approximate factorization, $x_j$ may lie outside the cone, and the model represents it by a vector in or near the cone. A profile with coefficient zero contributes nothing; a coefficient cannot be negative to cancel or subtract a profile. A nonzero profile determines a direction or ray: multiplying it by a positive number changes its scale but not its direction. A zero profile contributes nothing to the cone.

This is a geometric statement about the vectors being factorized, not a causal claim about biological mechanisms. A component may summarize a recurring pattern of feature values without identifying a unique biological process.

Two Features: A Wedge of Possible Vectors

With two features, each nonnegative vector lies in the first quadrant. If two component profiles are not pointing in the same direction, their nonnegative combinations fill the wedge between their rays. A vector inside the wedge can be represented exactly using nonnegative coefficients; a vector outside it cannot be represented exactly by those two profiles.

Worked Example: Locating a Two-Feature Sample in a Wedge

Consider hypothetical profiles $w_1=(2,1)$ and $w_2=(1,3)$, measured in the same feature units. A hypothetical sample has $x=(7,11)$. We test whether it is an additive combination of these profiles by solving

a w_1+b w_2=x,\qquad (2a+b,\ a+3b)=(7,11).

The first coordinate gives $b=7-2a$. Substituting into the second gives $a+3(7-2a)=11$, so $-5a=-10$, $a=2$, and $b=7-4=3$. Both coefficients are nonnegative, and direct substitution confirms the reconstruction:

$$ 2(2,1)+3(1,3)=(4,2)+(3,9)=(7,11). $$

Thus $x$ lies in the wedge generated by the two profiles. The rays are distinct, and the coefficient equations have a unique solution here: the determinant of the profile matrix is $2(3)-1(1)=5$, which is nonzero. That uniqueness depends on these profiles being linearly independent; it is not guaranteed for every NMF representation.

The cone is not generally the same as the convex hull of the profiles. In the cone, coefficients may sum to any nonnegative value, so points can extend arbitrarily far along a ray. A convex combination instead requires coefficients to sum to one and therefore stays within the bounded region between the profile points.

Normalized Vectors and Mixture Weights

For nonzero vectors, normalization by the sum of their entries can expose the mixture geometry. This uses the $L_1$ norm, $\lVert v\rVert_1=\sum_i |v_i|$. Because all profiles and coefficients are nonnegative, the total feature sum of an exact reconstruction is the sum of the total feature sums contributed by its profiles.

Result: Suppose $x=\sum_{k=1}^{r} a_k w_k$, with $a_k\geq 0$, $w_k\geq 0$, and $x\ne 0$. Let $I=\{k:a_k>0 \text{ and } \lVert w_k\rVert_1>0\}$. Then $$ \frac{x}{\lVert x\rVert_1} = \sum_{k\in I} \frac{a_k\lVert w_k\rVert_1}{\lVert x\rVert_1} \frac{w_k}{\lVert w_k\rVert_1}. $$ The weights in this sum are nonnegative and sum to one. Indices with $a_k=0$ or $w_k=0$ contribute nothing and are omitted; if $x=0$, normalization is undefined.

The result follows by taking the $L_1$ norm of the exact sum. Nonnegativity means no cancellation occurs, so $\lVert x\rVert_1=\sum_k a_k\lVert w_k\rVert_1$. Terms with zero coefficient or a zero profile have zero contribution to that sum. For every remaining index, write $a_k w_k=(a_k\lVert w_k\rVert_1)(w_k/\lVert w_k\rVert_1)$, then divide the sum by $\lVert x\rVert_1$. The resulting weights sum to $\sum_{k\in I}a_k\lVert w_k\rVert_1/\lVert x\rVert_1=1$. Thus the normalized vector is a convex combination of the normalized, nonzero profiles that contribute to it.

Worked Example: A Three-Feature Sample as a Normalized Mixture

Suppose two hypothetical three-feature profiles are $w_1=(1,0,1)$ and $w_2=(0,2,1)$, and a sample is $x=2w_1+w_2$. All entries are nonnegative, and both profiles are nonzero. First calculate the sample vector:

x=2(1,0,1)+(0,2,1)=(2,2,3),\qquad \lVert x\rVert_1=2+2+3=7.

The profile sums are $\lVert w_1\rVert_1=2$ and $\lVert w_2\rVert_1=3$. Their contributions to the sample's total are $2(2)=4$ and $1(3)=3$, giving normalized mixture weights $4/7$ and $3/7$. The normalized profiles are $w_1/2=(1/2,0,1/2)$ and $w_2/3=(0,2/3,1/3)$. Therefore:

\frac{x}{\lVert x\rVert_1} = \frac{4}{7}\left(\frac{1}{2},0,\frac{1}{2}\right) + \frac{3}{7}\left(0,\frac{2}{3},\frac{1}{3}\right) = \left(\frac{2}{7},\frac{2}{7},\frac{3}{7}\right).

The normalized sample lies on the line segment between the two normalized profiles in the three-feature simplex. The weights $4/7$ and $3/7$ describe each profile's share of the sample's total feature sum; they are not the original coefficients $2$ and $1$.

Three Features: A Cone Within the Nonnegative Orthant

With three features, all nonnegative vectors lie in the nonnegative orthant of three-dimensional space. Two non-collinear profiles generate a planar wedge in that space; three or more profiles may generate a larger cone. Normalized nonzero vectors lie in the two-dimensional simplex defined by nonnegative coordinates that sum to one. Normalizing each profile therefore turns the cone question into a convex-mixture question on that simplex, with weights determined by the profiles' total contributions.

The normalized view is useful for interpreting composition, but it discards overall magnitude. Two samples can have the same normalized feature proportions and very different total abundance. Consequently, a compositional interpretation should not replace examination of the original scale when total signal is scientifically relevant.

Worked Example: Detecting a Vector Outside a Three-Feature Cone

Consider the hypothetical profiles $w_1=(1,0,1)$ and $w_2=(0,1,1)$. A hypothetical nonnegative sample is $x=(1,2,1)$. If it were an exact combination $a w_1+b w_2$, its first coordinate would require $a=1$ and its second would require $b=2$. Those coefficients would give a third coordinate $a+b=3$, not the observed value $1$. Hence no coefficients $a,b\geq0$ (or even unrestricted real coefficients) produce this vector from these two profiles.

Geometrically, every vector in this cone has the form $(a,b,a+b)$, so its third coordinate equals the sum of its first two. The sample violates that constraint because $1\ne 1+2$. It is nonnegative but outside the cone. An approximate factorization could still represent it by a nearby cone vector, but exact membership has failed. This distinction matters: nonnegative input alone does not ensure that a chosen set of nonnegative profiles can reproduce it.

Assumptions, Diagnostics, and Alternatives

The cone interpretation requires a nonnegative data representation and nonnegative profiles and coefficients. It also assumes that the rows of $W$ and $X$ refer to the same features and in the same order. If preprocessing creates negative values, the ordinary nonnegative cone model no longer directly applies to those values. Centering a feature, for example, can create negative entries and changes the geometric question; silently treating those entries as nonnegative is not valid.

Several checks help reveal when the geometric interpretation is strained:

Inspect feature and sample conventions. Confirm nonnegativity, feature alignment, and the meaning of the scale before interpreting a component as an additive profile.
Look for samples with large reconstruction discrepancies. If many observations cannot be represented closely by the fitted profiles, the cone may not capture their patterns. A residual is the difference between the observed vector and its reconstruction.
Check for nearly parallel profiles. If two profiles point in almost the same direction, their rays make a narrow wedge. Coefficients can then be hard to distinguish: small changes in a sample may cause large changes in the allocation between the two profiles.
Separate direction from magnitude. Normalizing vectors can clarify mixture proportions, but it hides total signal and is undefined for the zero vector.

A signed linear decomposition, such as the geometry used in principal component analysis, permits movement in both positive and negative directions along a component. Its components span a linear subspace rather than only a nonnegative cone. That flexibility can represent patterns involving subtraction, but signed loadings are not additive nonnegative profiles. A convex-mixture model, by contrast, constrains coefficients to sum to one and describes mixtures of normalized profiles; the NMF coefficients themselves need not sum to one. These are different geometric constraints, not interchangeable labels for the same model.

A cone is also not necessarily a unique explanation. Profiles can be redundant or linearly dependent, and different nonnegative coefficient vectors may generate the same sample. Even when coefficients are unique for a fixed collection of independent profiles, changing the profiles changes the cone and may change the interpretation. The geometry describes what a particular set of factors can represent; it does not by itself establish that those factors are uniquely determined or scientifically true.

Common pitfall: “The sample is nonnegative, so any nonnegative profiles can explain it.” This is false. Nonnegative data occupy the nonnegative orthant, while a fixed set of profiles generates only the cone inside that orthant. Check whether the vector satisfies the constraints imposed by those profiles; nonnegativity is necessary for this interpretation, but it does not guarantee cone membership.

Worked Example: Omitting Zero Contributions Safely

Let hypothetical profiles be $w_1=(1,2)$, $w_2=(2,1)$, and $w_3=(0,0)$, with coefficients $a_1=2$, $a_2=0$, and $a_3=4$. The resulting sample is $x=2w_1+0w_2+4w_3=(2,4)$, so $\lVert x\rVert_1=6$. Only profile $w_1$ contributes a nonzero vector; $w_2$ has a zero coefficient and $w_3$ is itself zero. Thus $I=\{1\}$, and the normalized representation is

\frac{x}{\lVert x\rVert_1} = \frac{2\lVert w_1\rVert_1}{6}\frac{w_1}{\lVert w_1\rVert_1} = \frac{2(3)}{6}\left(\frac{1}{3},\frac{2}{3}\right) = \left(\frac{1}{3},\frac{2}{3}\right).

The weights sum to one, as required. The zero profile contributes nothing even though its coefficient is positive, and it must not be normalized: $w_3/\lVert w_3\rVert_1$ is undefined. Omitting zero contributions keeps the formula defined without changing the reconstructed vector.

Check Your Understanding

Answer each question, then compare with the solutions below.

For $w_1=(1,0)$, $w_2=(0,1)$, and $x=(4,3)$, find the nonnegative coefficients and state whether $x$ lies in the generated cone.
What condition must hold for a vector $x=(u,v,t)$ to lie in the cone generated by $(1,0,1)$ and $(0,1,1)$?
Why are the normalized mixture weights generally not the same as the original NMF coefficients?
What goes wrong if a zero profile is included in a formula that normalizes each profile by its $L_1$ norm?
How does a cone generated by profiles differ from the subspace used by a signed linear decomposition?

Solutions. (1) $x=4(1,0)+3(0,1)$, so the coefficients are $4$ and $3$, both nonnegative; the vector lies in the cone. (2) A representation requires $u=a$, $v=b$, and $t=a+b$ for some $a,b\geq0$. Thus $u,v\geq0$ and $t=u+v$. (3) Normalized weights are proportional to each coefficient times the profile's $L_1$ sum: $a_k\lVert w_k\rVert_1/\lVert x\rVert_1$. They equal the coefficients only under special scaling conditions. (4) The denominator is zero, so the normalized profile is undefined; omit that zero contribution. (5) Cone coefficients must be nonnegative, allowing movement only forward along each profile ray. A signed linear decomposition permits negative coefficients too, so it spans a subspace with directions on both sides of its component axes.