Chapter 10 Directional Dependence

Chapter 9 showed that classical covariance and correlation arise from aggregations of directional co-partial moments. While correlation summarizes joint variation with a single symmetric statistic, many real-world relationships are nonlinear, asymmetric, or concentrated in extreme events.

A familiar example occurs in financial markets. During ordinary periods, many assets appear weakly correlated. Yet during crises, losses often occur simultaneously across markets. Correlation averages across all observations and therefore may fail to capture this type of asymmetric tail dependence.

Directional statistics addresses this limitation by examining how variables move relative to benchmarks for each variable simultaneously. Instead of collapsing joint behavior into a single number, the directional framework partitions the joint distribution and measures deviations within each region separately.

This chapter develops directional dependence using co-partial moments. These statistics preserve the directional structure of the joint distribution and reveal nonlinear and asymmetric relationships that classical correlation can obscure.


10.1 Directional Benchmarks

Let \(X\) and \(Y\) be random variables with benchmarks \(t_X\) and \(t_Y\).

Benchmarks may be chosen in several ways depending on the application:

  • Internal benchmarks, such as the mean or median.
  • External benchmarks, such as target returns or policy thresholds.
  • Context-specific benchmarks, reflecting operational constraints or decision thresholds.

The benchmarks partition the joint distribution into four directional regions:

\[ X \le t_X, \quad Y \le t_Y \]

\[ X \le t_X, \quad Y > t_Y \]

\[ X > t_X, \quad Y \le t_Y \]

\[ X > t_X, \quad Y > t_Y \]

These four regions represent combinations of directional deviations for the two variables.

\(Y \le t_Y\) \(Y > t_Y\)
\(X \le t_X\) CoLPM region DUPM region
\(X > t_X\) DLPM region CoUPM region

Each quadrant corresponds directly to one of the four directional co-partial moments.


10.2 Co-Partial Moments

Let the positive-part operator be

\[ (x)^+ = \max(x,0). \]

Directional co-partial moments measure joint deviations relative to the benchmarks.

10.2.1 Co-Lower Partial Moment

\[ CoLPM_{r,s}(X,Y) = E[(t_X-X)_+^r (t_Y-Y)_+^s] \]

Joint deviations below both benchmarks.

10.2.2 Co-Upper Partial Moment

\[ CoUPM_{r,s}(X,Y) = E[(X-t_X)_+^r (Y-t_Y)_+^s] \]

Joint deviations above both benchmarks.

10.2.3 Divergent Lower Partial Moment

\[ DLPM_{r,s}(X,Y) = E[(X-t_X)_+^r (t_Y-Y)_+^s] \]

\(X\) above its benchmark while \(Y\) falls below.

10.2.4 Divergent Upper Partial Moment

\[ DUPM_{r,s}(X,Y) = E[(t_X-X)_+^r (Y-t_Y)_+^s] \]

\(X\) below its benchmark while \(Y\) exceeds its benchmark.

Together these four quantities provide a directional decomposition of joint dependence.


10.3 Worked Example

Consider the sample

\[ (X,Y) = (-3,-2), (-1,-1), (0,1), (2,4), (3,5). \]

Let the benchmarks be

\[ t_X = 0, \quad t_Y = 0. \]

Compute first-degree co-partial moments.

10.3.1 CoLPM

\[ CoLPM_{1,1} = \frac{1}{5}(3\cdot2 + 1\cdot1 + 0 + 0 + 0) = \frac{7}{5} = 1.4 \]

10.3.2 CoUPM

\[ CoUPM_{1,1} = \frac{1}{5}(0 + 0 + 0 + 2\cdot4 + 3\cdot5) = \frac{23}{5} = 4.6 \]

10.3.3 DLPM

\[ DLPM_{1,1} = 0 \]

10.3.4 DUPM

\[ DUPM_{1,1} = 0 \]

The interpretation is immediate.

  • Downside dependence exists but is modest.
  • Upside deviations occur together strongly.
  • Divergent movements do not occur.

In this dataset, whenever \(X\) is above its benchmark, \(Y\) is also above its benchmark, and whenever \(X\) is below its benchmark, \(Y\) is also below its benchmark. Consequently observations never fall into divergent regions.

Boundary observations contribute zero to these positive-degree co-partial moments; degree-zero moments instead use the event conventions from Chapter 3. For example, the point \((0,1)\) lies exactly on the \(X\) benchmark, so both \((X-t_X)_+\) and \((t_X-X)_+\) equal zero.

Real datasets rarely exhibit such perfect alignment, and in practice the divergent moments capture regions where one variable rises while the other falls.


10.4 Dependence Versus Correlation

With the benchmarks at the means, \(t_X=\mu_X\) and \(t_Y=\mu_Y\), covariance aggregates the degree-one directional components:

\[ Cov(X,Y) = CoUPM_{1,1} + CoLPM_{1,1} - DLPM_{1,1} - DUPM_{1,1}. \]

For general benchmarks the same sum is \(E[(X-t_X)(Y-t_Y)] = Cov(X,Y) + (\mu_X-t_X)(\mu_Y-t_Y)\). With \(X\) and \(Y\) independent on \(\{9, 11\}\) and both benchmarks at 0, the sum is 100 while the covariance is 0.

Correlation further standardizes covariance.

\[ (CoLPM,CoUPM,DLPM,DUPM) \rightarrow Cov(X,Y) \rightarrow Corr(X,Y) \]

Directional statistics therefore preserves structural information lost through aggregation.


10.5 Nonlinear Dependence Detection

Directional dependence can reveal nonlinear relationships that correlation cannot detect.

Consider

\[ Y = X^2 \]

with \(X\) symmetrically distributed around zero, \(E[X^4]<\infty\), and \(\operatorname{Var}(X^2)>0\). These conditions make both variances and the correlation well-defined.

In this case

\[ Corr(X,Y)=0. \]

Despite zero correlation, \(Y\) is determined by \(X\). This functional relation is a property of the joint law; a finite-resolution dependence statistic need not assign it the numerical value 1.

Directional moments reveal the structure.

When \(X>0\), both \(X\) and \(Y=X^2\) exceed their benchmarks, producing contributions to the CoUPM region.

When \(X<0\), \(X\) lies below its benchmark while \(Y=X^2\) remains positive and therefore above its benchmark. These observations fall into the DUPM region, capturing the mirrored dependence structure.

Thus the directional decomposition exposes dependence that the symmetric aggregation in correlation cancels.


10.6 Asymmetric Dependence

Many systems exhibit asymmetric dependence, where relationships differ between positive and negative deviations.

Financial markets provide a common example.

Assets may behave largely independently during rising markets but move strongly together during market crashes.

In such cases

\[ CoLPM_{1,1} \gg CoUPM_{1,1}. \]

A replicable simulation makes this asymmetry explicit. In the construction below, negative shocks are shared between both variables, while positive-side behavior is generated independently:

library(NNS)
set.seed(42)
n <- 500
shock <- rnorm(n)

x <- ifelse(shock < 0, shock, rnorm(n))
y <- ifelse(shock < 0, shock + rnorm(n, 0, 0.1), rnorm(n))

Co.LPM(1, x, y, mean(x), mean(y))
## [1] 0.2770795
Co.UPM(1, x, y, mean(x), mean(y))
## [1] 0.2103299
D.LPM(1, 1, x, y, mean(x), mean(y))
## [1] 0.06191035
D.UPM(1, 1, x, y, mean(x), mean(y))
## [1] 0.08481611

The same workflow in Python:

import numpy as np
from nns import co_lpm, co_upm, d_lpm, d_upm

rng = np.random.default_rng(42)
n = 500
shock = rng.normal(size=n)

x = np.where(shock < 0, shock, rng.normal(size=n))
y = np.where(shock < 0, shock + rng.normal(0, 0.1, n), rng.normal(size=n))

co_lpm(1, x, y, x.mean(), y.mean())
co_upm(1, x, y, x.mean(), y.mean())
d_lpm(1, 1, x, y, x.mean(), y.mean())
d_upm(1, 1, x, y, x.mean(), y.mean())

In this example, the concordant downside component exceeds the upside component. These raw co-moments combine dependence and marginal severity; the marginal-product comparison developed below is needed to isolate departures from independence.

Correlation averages across all regions and may therefore appear moderate even when downside dependence dominates.


10.7 Tail-Sensitive Dependence

Higher-order co-partial moments emphasize extreme deviations.

Increasing the orders \(r\) and \(s\) increases sensitivity to large observations.

\[ CoLPM_{r,s}, \quad CoUPM_{r,s} \]

summarize joint tail severity. They differ from the limiting conditional tail probabilities defined in Chapter 12 and should be compared with marginal-product benchmarks when the goal is to isolate dependence.

For example, a risk manager concerned with extreme joint losses may examine \(CoLPM_{2,2}\) or higher orders rather than \(CoLPM_{1,1}\), since larger powers place greater weight on large deviations.


10.8 Empirical Estimation

Directional co-partial moments can be estimated from data.

For observations

\[ (x_i,y_i), i=1,\dots,n \]

the empirical estimators are

\[ \widehat{CoLPM}_{r,s} = \frac{1}{n} \sum_{i=1}^{n} (t_X-x_i)_+^r (t_Y-y_i)_+^s \]

\[ \widehat{CoUPM}_{r,s} = \frac{1}{n} \sum_{i=1}^{n} (x_i-t_X)_+^r (y_i-t_Y)_+^s \]

\[ \widehat{DLPM}_{r,s} = \frac{1}{n} \sum_{i=1}^{n} (x_i-t_X)_+^r (t_Y-y_i)_+^s \]

\[ \widehat{DUPM}_{r,s} = \frac{1}{n} \sum_{i=1}^{n} (t_X-x_i)_+^r (y_i-t_Y)_+^s \]

For independent and identically distributed observations, fixed benchmarks, and finite expectations of the displayed products, these sample averages converge to their population values by the law of large numbers. Dependent observations or estimated benchmarks require the corresponding regularity conditions.

Statistical inference for these estimators—including bootstrap procedures—is discussed later in the book when directional dependence measures are applied in empirical analysis.

In practice these quantities are implemented in the Nonlinear Nonparametric Statistics (NNS) framework, which computes empirical co-partial moment matrices and nonlinear dependence measures.


10.9 Directional Dependence Profiles

Directional dependence can be studied across multiple moment orders.

\[ CoLPM_{1,1},CoLPM_{2,2},CoLPM_{3,3},\dots \]

\[ CoUPM_{1,1},CoUPM_{2,2},CoUPM_{3,3},\dots \]

These sequences describe how dependence changes across the distribution, but only relative to what the margins alone produce. Under independence the co-moments factorise, \(CoLPM_{d,d}(t_X,t_Y) = L_d(t_X;X)\,L_d(t_Y;Y)\) and \(CoUPM_{d,d}(t_X,t_Y) = U_d(t_X;X)\,U_d(t_Y;Y)\), so skewed margins alone make the two sequences diverge: for two independent negated Exp(1) variables at their means, \(CoLPM_{d,d}/CoUPM_{d,d}\) is 1, 7.8 and 113 at degrees 1, 2 and 3. A profile is therefore read against these independence products.

Example:

Order CoLPM CoUPM
1 1.2 1.1
2 3.9 1.4
3 8.5 1.8

Moderate deviations appear symmetric, while higher orders put more weight on the lower co-moment. Where the independence products of the margins grow the same way, this is marginal skew; only the excess over them is increasing downside dependence.

Directional profiles therefore show how dependence evolves from ordinary fluctuations to extreme events.


10.10 Finite-Resolution Dependence Summaries

NNS.dep combines directional frequency and area information across a finite set of recursive partition levels. This is a useful summary of the structure resolved by those partitions, not a complete representation of the joint law. Chapter 14 builds its directional statistic from the same cell measure, so the full definition is given here.

Partitions. Three recursive partitions of the \(n\) observations are built with NNS.part at noise.reduction = "off", which splits each cell at the gravity of its observations (Chapter 19, Section 19.7) rather than at its mean: the partition on \(x\) alone, whose cells are intervals of \(x\) and carry \(y\) as a function of \(x\); the partition on \(y\) alone; and the joint quadrant partition of \((x, y)\), which splits each cell on both coordinates and carries relations such as a circle or \(y = x\varepsilon\). Splitting runs to depth \(\lceil \log_2 n \rceil\), and a cell with at most \(\max(8, n/8)\) observations is not split further. Level \(l\) of a partition is its set of cells after \(l\) rounds of splitting, so the levels are nested.

Cell dependence. In a cell \(c\) with \(m\) observations, deviations are taken about the cell means, \(d_x = x - \bar x_c\) and \(d_y = y - \bar y_c\). Two concordance terms measure the degree-0 and degree-1 co-partial moments of the cell:

\[ T_0 = |2p_0 - 1|, \qquad T_1 = \frac{\left|\sum d_x d_y\right|}{\sum |d_x d_y|}, \]

where \(p_0\) is the share of the cell’s points in the co-lower or co-upper orthant (frequency, degree 0) and \(T_1\) is the corresponding share of co-moment area (degree 1). The divergent shares are \(1 - p_0\) and \(1 - p_1\), so they add nothing. Each term is centred at its expectation under random pairing of the cell’s \(x\) and \(y\) values and rescaled so that a perfectly concordant or discordant cell still scores 1,

\[ T^{*} = \frac{T - E_0[T]}{1 - E_0[T]}, \qquad d_c = \frac{T_0^{*} + T_1^{*}}{2}, \]

where \(d_c\) is negative when the cell is less concordant than chance. \(E_0[T_0]\) is hypergeometric (with a normal approximation above 400 points); \(E_0[T_1]\) enumerates the distinct arrangements when there are at most 720 of them (every cell of up to six points) and otherwise uses the permutation variance of \(\sum d_x d_y\). A cell in which the conditioning variable is constant scores whether the other variable is also constant there (1 or 0), measured against the chance of that under the permutation null; in the joint partition such a cell scores 0.

Levels and selection. The dependence of a level is the cell-size-weighted average \(D_l = \sum_c (m_c / n)\, d_c\), and its correlation is \(\sum_c (m_c / n)\, s_c d_c\), where \(s_c\) is the sign of \(\sum d_x d_y\) in the cell (the sign of the overall covariance for a cell that is constant in the conditioning variable). NNS.dep(x, y, asym = TRUE) reports the largest \(D_l\) over the levels of the partition on \(x\); with asym = FALSE the largest over the levels of all three partitions. Dependence is floored at 0, and Correlation is the signed average at the selected level, kept within \([-\text{Dependence}, \text{Dependence}]\). Chapter 14’s \(\rho_{X \to Y}\) uses the same cells on lagged copula pairs but averages \(D_l\) over the levels of the partition on the candidate cause instead of taking the largest.

A numerical value of 1 corresponds to maximal contrast under the statistic’s cellwise calibration. It is not guaranteed for every noiseless function: a coarse cell can contain several branches or oscillations of \(Y=f(X)\), so considerable response variation remains inside that cell. Finer partitions can resolve more structure, but also change sampling variation and the effect of selecting the strongest level. Conversely, a score near zero does not establish independence against every possible relationship.

The implementation compares cells with their permutation-null expectations. Some small-cell calculations use exact enumeration and larger cells use approximations; this is distinct from an exact universal null law for the final selected score. When a p-value is requested, its permutation interpretation depends on the exchangeability of the observations under the null. Serially dependent data require a null resampling scheme appropriate to their process.

These properties preserve the distinction between a directional profile, a dependence summary, and a calibrated statistical test. Power comparisons must specify the alternative, sample size, and testing procedure rather than claim uniform superiority for one summary.


10.11 Summary

This chapter introduced directional dependence using co-partial moments.

Key ideas:

  1. Joint distributions partition into four directional regions.
  2. Co-partial moments measure deviations within those regions.
  3. Covariance and correlation arise as aggregations of directional components.
  4. Directional statistics reveals nonlinear, asymmetric, and tail-specific dependence.
  5. Empirical estimators can be computed directly from data.
  6. Dependence profiles show how relationships evolve across deviation magnitudes.

Correlation therefore represents only a limited summary of joint behavior.

Directional dependence provides a richer representation of relationships between variables.

Chapter 12 connects this framework to copula interpretation, linking directional partial moments with rank-based dependence structures used in multivariate statistics.