Chapter 2 Directional Deviation Operators
Chapter 1 argued that many failures of classical statistics arise from symmetric aggregation.
Classical moments, covariance, and correlation collapse directional information into symmetric summaries before analysis begins.
This chapter introduces the mathematical primitive that avoids that collapse:
directional deviation operators.
These operators measure deviations relative to a benchmark separately above and below the reference point. From this simple construction we will derive many familiar objects of statistics.
The framework begins with a simple observation:
any deviation relative to a benchmark has a direction.
2.1 Deviations Relative to a Benchmark
Let \(X\) be a real-valued random variable and let \(t \in \mathbb{R}\) denote a benchmark.
Classical statistics measures deviations using
\[ X - t \]
which mixes positive and negative deviations together.
Directional statistics separates them.
Define the positive-part operator
\[ x^{+} = \max(x,0). \]
Using this operator we define two directional deviations:
\[ (X-t)^+ = \max(X-t,0) \]
\[ (t-X)^+ = \max(t-X,0). \]
These represent
- deviations above the benchmark
- deviations below the benchmark
Both quantities are nonnegative.
Together they fully characterize the magnitude of deviation relative to \(t\).
2.2 Directional Decomposition of Deviations
Every deviation can be decomposed into directional components.
For any real number \(x\),
\[ x = x^+ - (-x)^+. \]
Applying this identity to \(X-t\) yields
\[ X-t = (X-t)^+ - (t-X)^+. \]
Thus the classical deviation can be expressed as the difference between two directional magnitudes.
The directional operators also reconstruct the magnitude of deviation:
\[ |X-t| = (X-t)^+ + (t-X)^+. \]
Thus directional components fully determine both the signed deviation and its magnitude.
The key implication is structural:
the symmetric deviation is an aggregation of directional components.
Classical statistics begins with the aggregate.
Directional statistics begins with the components.
2.3 Directional Operators
The functions
\[ (X-t)^+ , \quad (t-X)^+ \]
are called directional deviation operators.
They induce a natural partition of the sample space:
- \(X > t\)
- \(X \le t\)
Within each region the operators measure the magnitude of deviation from the benchmark.
This partition is fundamental. Many real-world systems evaluate outcomes relative to targets:
- profits relative to costs
- returns relative to required benchmarks
- losses relative to liabilities
- forecast errors relative to expected demand
Directional deviation operators formalize this benchmark-relative measurement.
2.4 Partial Moments
Once directional deviations are defined, their magnitudes can be summarized through expectations.
For integer \(r \ge 0\), define the lower partial moment
\[ L_r(t;X) = E[(t-X)_+^r] \]
and the upper partial moment
\[ U_r(t;X) = E[(X-t)_+^r]. \]
These quantities measure directional deviation magnitudes relative to the benchmark.
For these expectations to be finite, it is sufficient that the corresponding directional powers are integrable (for example, \(E[|X|^r]<\infty\) for fixed \(r\)).
The parameter \(r\) determines the type of deviation measured.
| Degree \(r\) | Interpretation |
|---|---|
| 0 | probability mass |
| 1 | directional mean deviation |
| 2 | directional variance |
| \(r>2\) | higher-order tail structure |
Partial moments therefore generalize classical moments while preserving directional structure.
2.5 Notation Bridge: Theory to R Implementation
The manuscript uses theoretical notation in proofs and function-style notation in implementation examples. The mapping is direct:
| Theoretical object | Meaning | R implementation pattern (Using NNS Package) |
|---|---|---|
| \(L_r(t;X)\) | lower partial moment of degree \(r\) at benchmark \(t\) | LPM(r, t, X) |
| \(U_r(t;X)\) | upper partial moment of degree \(r\) at benchmark \(t\) | UPM(r, t, X) |
| \(L_r(t;X)_{\text{ratio}}\) | normalized lower share \(L_r/(L_r+U_r)\) | LPM.ratio(r, t, X) |
| \(U_r(t;X)_{\text{ratio}}\) | normalized upper share \(U_r/(L_r+U_r)\) | UPM.ratio(r, t, X) |
| \(CoLPM\), \(CoUPM\), \(DLPM\), \(DUPM\) | concordant/divergent co-partial moments | Co.LPM(...), Co.UPM(...), D.LPM(...), D.UPM(...) |
Unless otherwise stated, later chapters use the mathematical form for derivations and the function-call form for reproducible examples.
2.6 Benchmarks
A distinctive feature of partial moments is the benchmark \(t\).
In classical statistics, reference points are usually determined by the distribution itself.
The mean, median, and variance are all defined internally.
Partial moments differ in an important way: the benchmark need not be determined by the distribution.
Instead, \(t\) may represent an externally meaningful reference point chosen by the analyst or by the decision context.
Examples include:
- target returns in finance
- policy thresholds in economics
- forecast baselines in operations
- safety limits in engineering
- aspiration levels in behavioral decision theory
The benchmark therefore embeds the context in which deviations matter.
Directional statistics evaluates distributions relative to those contexts rather than purely distributional averages.
2.7 Relationship to Classical Moments
Classical moments arise as aggregations of partial moments.
For integer \(r \ge 1\),
\[ E[(X-t)^r] = U_r(t;X) + (-1)^r L_r(t;X). \]
This identity shows that symmetric moments are signed combinations of directional components.
Several familiar quantities follow immediately.
2.7.2 Variance
\[ Var(X) = U_2(\mu;X) + L_2(\mu;X) \]
This identity is the population variance decomposition. In R, var(x) returns the sample variance, so numerical checks against UPM(2, mean(x), x) + LPM(2, mean(x), x) should include the Bessel correction factor \(n/(n-1)\).
2.7.3 Third Central Moment
\[ E[(X-\mu)^3] = U_3(\mu;X) - L_3(\mu;X). \]
Thus classical symmetric statistics do not introduce fundamentally new objects.
They aggregate directional ones.
Importantly, the mapping from partial moments to symmetric moments is many-to-one.
Directional components determine the symmetric moment uniquely, but the symmetric moment does not generally determine the directional components.
Directional moments therefore contain strictly more information.
2.8 Information Loss in Symmetric Aggregation
Consider two distributions with identical variance.
Distribution A may exhibit
- large upside deviations
- small downside deviations.
Distribution B may exhibit
- small upside deviations
- large downside deviations.
Both can produce the same value of
\[ Var(X). \]
Variance alone cannot distinguish them.
However the directional quantities
\[ U_2(\mu;X), \quad L_2(\mu;X) \]
immediately reveal the asymmetry.
Symmetric statistics therefore represent projections of directional structure.
Once this projection occurs, the original directional information cannot generally be recovered.
2.9 Directional Operators as a Statistical Primitive
Directional deviation operators provide a foundation from which many statistical constructs can be derived.
Rather than beginning with symmetric statistics and imposing directional interpretation afterward, the directional framework reverses the order: deviations relative to a benchmark are measured first, and symmetric statistics emerge only as aggregations of those directional components.
The implications of this perspective are surprisingly broad.
The next chapter begins with a result that illustrates the power of the framework:
the cumulative distribution function itself is a partial moment.