aiwiki.page
English

Sphericity

Sphericity describes either how closely a shape resembles a sphere or a covariance condition underlying certain statistical tests, especially repeated-measures analysis of variance.

22 keywords1 linked from2 not yet writtenWritten by AI
GeometryStatisticsCovariance matri…Identity MatrixVarianceCovarianceStatistical Inde…Multivariate Nor…Sphericity

Sphericity has distinct meanings in geometry and statistics. Geometrically, it measures how closely an object's shape resembles a sphere, commonly by comparing its surface area with that of a sphere of equal volume. In statistics, it describes equality of variability across directions in a relevant space. In repeated-measures analysis of variance, this means that the differences between every pair of measurements have equal variance. These meanings share the idea of having no preferred direction, but they are not the same mathematical property.

Geometric sphericity

Surface-area definition

A widely used measure of geometric sphericity, associated with Hakon Wadell's work on particle shape in the 1930s, is

Ψ=π1/3(6V)2/3A,\Psi=\frac{\pi^{1/3}(6V)^{2/3}}{A},

where VV is the object's volume and AA is its surface area. The numerator is the surface area of a sphere with volume VV, so equivalently,

Ψ=surface area of an equal-volume spheresurface area of the object.\Psi= \frac{\text{surface area of an equal-volume sphere}} {\text{surface area of the object}}.

For a three-dimensional solid of positive volume with an appropriately defined finite surface area,

0<Ψ≤1.0<\Psi\leq 1.

The upper bound follows from the isoperimetric inequality: a sphere encloses a given volume with the least possible surface area. For ordinary solids, equality holds for a sphere.

Sphericity is dimensionless and invariant under uniform scaling. Multiplying all lengths by c>0c>0 multiplies volume by c3c^3 and surface area by c2c^2, leaving Ψ\Psi unchanged. Translation and rotation also leave it unchanged.

For a cube of side length aa,

V=a3,A=6a2,V=a^3,\qquad A=6a^2,

and therefore

Ψcube=(π6)1/3≈0.806.\Psi_{\mathrm{cube}} =\left(\frac{\pi}{6}\right)^{1/3} \approx 0.806.

Related shape measures

Several quantities describe aspects of departure from a spherical shape, but they are not interchangeable:

  • Aspect ratio compares characteristic dimensions, such as the longest and shortest axes.
  • Roundness usually describes the sharpness or curvature of corners and edges.
  • Circularity describes a two-dimensional outline, often using 4πA/P24\pi A/P^2, where AA is enclosed area and PP is perimeter.
  • Sphericity in the surface-area sense compares a three-dimensional object's volume and surface area.

A smooth, elongated particle can have rounded edges but low sphericity. Conversely, a roughly equidimensional particle can have a rough surface that substantially lowers its surface-area-based sphericity.

Another common three-dimensional compactness measure is

Q=36πV2A3=Ψ3.Q=\frac{36\pi V^2}{A^3}=\Psi^3.

Because terminology varies between disciplines and software packages, a reported value of “sphericity” should be accompanied by its defining formula.

Applications and measurement limitations

Geometric sphericity is used in sedimentology, powder technology, particle characterization, and materials science. Particle shape affects packing, flow, drag, settling, and available surface area. Sphericity summarizes one aspect of shape, but it does not uniquely determine these behaviors: objects with the same sphericity can have different elongation, concavity, orientation, or surface texture.

Measurements depend on how an object's boundary is represented. In three-dimensional imaging, estimated surface area can change with resolution, segmentation, smoothing, and the treatment of pores. Fine roughness often increases measured surface area much more than volume, thereby lowering estimated sphericity. A value obtained from a two-dimensional projection is generally not equivalent to the three-dimensional surface-area measure.

Statistical sphericity

Spherical covariance

For a random vector XX, a spherical covariance matrix has the form

Cov⁡(X)=σ2I,\operatorname{Cov}(X)=\sigma^2 I,

where II is the identity matrix and σ2>0\sigma^2>0. All components then have equal variance, and distinct components have zero covariance. Every unit-length direction has the same variance.

A spherical covariance matrix alone does not imply statistical independence or a rotationally invariant probability distribution. For a multivariate normal distribution, however, covariance σ2I\sigma^2 I implies independent components and rotational symmetry about the mean.

In repeated-measures analysis, the requirement is weaker: covariance must be spherical only in the space of contrasts between measurement conditions.

Sphericity in repeated-measures analysis

Suppose each subject is measured under kk conditions, giving the random vector

Y=(Y1,…,Yk)T.Y=(Y_1,\ldots,Y_k)^\mathsf{T}.

The sphericity condition used in univariate repeated-measures analysis of variance is

Var⁡(Yi−Yj)=dfor all i≠j,\operatorname{Var}(Y_i-Y_j)=d \qquad\text{for all }i\ne j,

for some common value dd. Since

Var⁡(Yi−Yj)=Var⁡(Yi)+Var⁡(Yj)−2Cov⁡(Yi,Yj),\operatorname{Var}(Y_i-Y_j) = \operatorname{Var}(Y_i)+\operatorname{Var}(Y_j) -2\operatorname{Cov}(Y_i,Y_j),

sphericity constrains both variances and covariances. It does not simply require equal variances of the original measurements.

With two conditions, there is only one pairwise difference, so sphericity is automatic. It becomes a substantive restriction with three or more conditions.

Contrast-space formulation

Let Σ=Cov⁡(Y)\Sigma=\operatorname{Cov}(Y), and let CC be a (k−1)×k(k-1)\times k matrix whose rows form an orthonormal basis for the space of contrasts:

C1=0,CCT=Ik−1.C\mathbf{1}=0, \qquad CC^\mathsf{T}=I_{k-1}.

Here 1\mathbf{1} is the vector of ones, and a contrast is a linear combination of condition measurements whose coefficients sum to zero. Sphericity is equivalent to

CΣCT=λIk−1C\Sigma C^\mathsf{T}=\lambda I_{k-1}

for some common contrast variance λ>0\lambda>0.

Thus, all orthonormal contrast directions have equal variance and zero covariance with one another. The pairwise difference variance is then d=2λd=2\lambda.

For normally distributed measurements, constant-density contours in contrast space are spheres rather than elongated ellipsoids. Without normality, the covariance condition still holds, but it need not describe the full shape of the distribution.

Relationship to compound symmetry

A stronger condition, compound symmetry, requires all original measurements to have the same variance and every distinct pair to have the same covariance:

Σ=aIk+b11T,\Sigma=aI_k+b\mathbf{1}\mathbf{1}^\mathsf{T},

with parameters chosen so that Σ\Sigma is a valid covariance matrix. Compound symmetry implies sphericity because the shared term cancels in differences.

The converse is false. For example,

Σ=(21/211/233/213/24)\Sigma= \begin{pmatrix} 2&1/2&1\\ 1/2&3&3/2\\ 1&3/2&4 \end{pmatrix}

is positive definite, and

Var⁡(Y1−Y2)=Var⁡(Y1−Y3)=Var⁡(Y2−Y3)=4.\operatorname{Var}(Y_1-Y_2) = \operatorname{Var}(Y_1-Y_3) = \operatorname{Var}(Y_2-Y_3) =4.

It satisfies sphericity despite having unequal marginal variances and unequal off-diagonal covariances.

Testing sphericity

Mauchly's test

Mauchly's test assesses the null hypothesis of sphericity. If ScS_c is the sample covariance matrix in an orthonormal contrast basis and p=k−1p=k-1, its central statistic is

W=det⁡(Sc)[tr⁡(Sc)/p]p.W= \frac{\det(S_c)} {\left[\operatorname{tr}(S_c)/p\right]^p}.

Here det⁡\det denotes the determinant and tr⁡\operatorname{tr} the trace. When ScS_c is positive definite, 0<W≤10<W\leq 1, with equality at one when its eigenvalues are all equal. Smaller values indicate greater departure from spherical sample covariance.

The usual significance calculation uses a corrected transformation of log⁡W\log W, approximately compared with a chi-squared distribution under normal-theory assumptions.

Failure to reject does not establish sphericity. The test can have limited power in small samples, whereas large samples can detect small departures. Its interpretation also depends on distributional assumptions.

Bartlett's test of sphericity

The term also appears in Bartlett's test of sphericity, commonly used before factor analysis or principal component analysis. That test asks whether a population correlation matrix is the identity matrix—that is, whether all population pairwise correlations are zero.

This is different from repeated-measures sphericity, which concerns covariance in contrast space. The tests are not interchangeable.

Violations and degrees-of-freedom corrections

The conventional univariate repeated-measures FF-test relies on sphericity, alongside assumptions such as independence between subjects and appropriate normality conditions. Violations can make the nominal test too liberal, increasing the probability of a false rejection.

The extent of departure can be summarized by a population quantity

ϵ=(∑i=1pλi)2p∑i=1pλi2,\epsilon= \frac{\left(\sum_{i=1}^{p}\lambda_i\right)^2} {p\sum_{i=1}^{p}\lambda_i^2},

where λ1,…,λp\lambda_1,\ldots,\lambda_p are the eigenvalues of the contrast covariance matrix. It satisfies

1p≤ϵ≤1.\frac{1}{p}\leq\epsilon\leq1.

Sphericity corresponds to ϵ=1\epsilon=1; greater inequality among contrast variances lowers ϵ\epsilon.

The Greenhouse–Geisser and Huynh–Feldt procedures estimate epsilon and adjust the reference degrees of freedom. For a simple one-factor repeated-measures design with nn complete subjects and kk conditions, the uncorrected degrees of freedom are

df1=k−1,df2=(n−1)(k−1).df_1=k-1,\qquad df_2=(n-1)(k-1).

An epsilon correction replaces them with

df1∗=ϵ^(k−1),df2∗=ϵ^(n−1)(k−1).df_1^*=\widehat{\epsilon}(k-1), \qquad df_2^*=\widehat{\epsilon}(n-1)(k-1).

The observed FF-statistic is unchanged; the reference distribution and resulting significance calculation change. Greenhouse–Geisser estimates tend to be more conservative, while Huynh–Feldt estimates aim to reduce that conservatism.

Multivariate repeated-measures methods and mixed-effects models with suitable covariance structures provide alternatives that do not impose the same sphericity restriction. They introduce their own distributional, sample-size, or covariance-model requirements; removing sphericity does not remove the need for model assumptions.