Direction-Averaging in Nonlinear Spectroscopy

• Direction-Averaging (DA) is the process of averaging a molecular tensor’s lab-frame projection over uniformly random orientations in SO(3), yielding an isotropic response.
• The method employs an exponent-matrix representation and parity rules to derive a closed-form combinatorial formula, simplifying traditional invariant-basis approaches.
• DA streamlines nonlinear spectroscopy computations by converting complex Cartesian tensor calculations into efficient, rotationally invariant operations.

Direction-Averaging (DA) is the standard problem of orientational averaging in nonlinear spectroscopy and nonlinear optics: given a molecular tensor defined in a molecule-fixed frame, one seeks the average of its lab-frame projection over uniformly random molecular orientations. Mathematically, this is an average over the rotation group $\mathrm{SO}(3)$ with Haar measure. In this setting, the central object is the averaged direction cosine product tensor,

$$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$

which is the rank-$2n$ rotationally invariant tensor that converts an $n$th-rank molecular tensor into its orientationally averaged lab-frame response. A closed-form, all-ranks formula for this tensor was derived in “A general method for rotational averages” (Nessler et al., 2019).

1. Physical setting and tensorial formulation

The DA problem arises because a Cartesian molecular tensor $T_{\lambda_1\cdots\lambda_n}$ transforms under rotation according to

$$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$

After averaging over random orientations,

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$

Thus the entire DA problem reduces to determining $I^{(n)}$ (Nessler et al., 2019).

In nonlinear spectroscopy, this formulation is universal because many optical observables involve products of transition dipoles, polarizabilities, hyperpolarizabilities, or higher multipole tensors, contracted with polarization vectors from multiple beams. Random molecular orientation implies that such quantities must be averaged over $\mathrm{SO}(3)$. Direction cosine products are therefore the universal building blocks of orientational averaging, and $I^{(n)}$ is the unique isotropic kernel that performs the average.

A practical implication is that DA is not a separate special-purpose construction for each observable. Rather, once the appropriate rank-$$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$0 tensor $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$1 is known, the orientational average of any $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$2th-rank Cartesian molecular quantity follows by contraction.

2. Rotational invariance and isotropic tensor structure

A key structural fact is that $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$3 is rotationally invariant in both the left and right indices because Haar measure is left- and right-invariant. In invariant-tensor language, $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$4 is isotropic under simultaneous independent rotations of the lab and molecular frames (Nessler et al., 2019).

This invariance explains why earlier methods expressed $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$5 in bases built from Kronecker deltas $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$6 and Levi-Civita symbols $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$7, the primitive invariant tensors in three dimensions. For even $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$8, only $$I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n} = \left\langle l_{i_1\lambda_1}\cdots l_{i_n\lambda_n} \right\rangle = \int_{\mathrm{SO}(3)} dg\; l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g),$$9-type structures contribute; for odd $2n$0, $2n$1-type antisymmetry appears. The closed-form method does not build the result by solving for coefficients in such an invariant basis. Instead, it evaluates the Euler-angle integral directly and thereby exposes invariant-theoretic structure through parity and combinatorics.

The paper also gives the rotational average explicitly in Euler angles using the $2n$2-$2n$3-$2n$4 convention: $2n$5 This is the precise DA problem solved for arbitrary tensor rank.

3. Exponent-matrix representation and selection rules

A major structural simplification is the replacement of $2n$6 explicit tensor indices by exponents of the nine direction cosines: $2n$7 where the nonnegative integers

$2n$8

sum to $2n$9. These are arranged as

$n$0

This $n$1-notation reduces bookkeeping from $n$2 separate index strings to only $n$3 exponent patterns, and many symmetry and selection rules become statements about row and column sums of $n$4 (Nessler et al., 2019).

The principal practical selection rule is obtained from parity under special $n$5-rotations. A necessary condition for a nonzero average is

$n$6

This immediately screens many tensor components.

Additional symmetry facts are equally important. The tensor is unchanged under transpose,

$n$7

because Haar measure is invariant under inversion $n$8. Moreover, $n$9 is symmetric in rows and columns when $T_{\lambda_1\cdots\lambda_n}$0 is even, and antisymmetric in rows and columns when $T_{\lambda_1\cdots\lambda_n}$1 is odd. For odd rank, if two rows or two columns of $T_{\lambda_1\cdots\lambda_n}$2 are equal, antisymmetry forces

$T_{\lambda_1\cdots\lambda_n}$3

This is why determinants enter naturally for odd ranks.

4. Closed-form all-ranks formula

The derivation expands the Euler-angle integrand and reduces each angular factor to Beta integrals. After binomial expansion of the powers involving $T_{\lambda_1\cdots\lambda_n}$4, the parity structure becomes explicit in the angular exponents. Assuming the row/column parity rule holds, the surviving Beta functions reduce to factorial or double-factorial forms, and the paper obtains the all-ranks formula (Nessler et al., 2019): $T_{\lambda_1\cdots\lambda_n}$5 where the prime means

$T_{\lambda_1\cdots\lambda_n}$6

This formula gives every rotational average of a direction-cosine monomial in terms of a finite sum of binomial coefficients and double factorials, with only a parity restriction. No rank-specific invariant-basis matrix inversion is needed. The paper emphasizes that the formula “fits in a few lines of code,” and also notes that all components of $T_{\lambda_1\cdots\lambda_n}$7 are rational numbers.

A plausible implication is that the closed form is not merely a symbolic simplification. It is also an algorithmic replacement for prior invariant-theory approaches whose coefficient matrices and storage demands grow rapidly with tensor rank.

5. Low-rank structure, determinant behavior, and special cases

For odd ranks, determinant structure appears explicitly. The paper proves

$T_{\lambda_1\cdots\lambda_n}$8

and, assuming the parity rule,

$T_{\lambda_1\cdots\lambda_n}$9

These formulas make the odd-rank antisymmetry fully transparent (Nessler et al., 2019).

The paper also gives useful low-rank characterizations. For

$$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$0

if $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$1 is even, then

$$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$2

For

$$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$3

if $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$4 is odd, then

$$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$5

These criteria fail in ranks $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$6 and $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$7, so they are not the general theorem; the general theorem is the explicit all-ranks formula.

The paper also records number-theoretic statements. For even $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$8, if $$T_{i_1\cdots i_n}(g)= l_{i_1\lambda_1}(g)\cdots l_{i_n\lambda_n}(g)\, T_{\lambda_1\cdots\lambda_n}.$$9 is an odd prime, then

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$0

For odd prime $$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$1, it proves the sufficient condition

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$2

Several special families collapse to compact factorial formulas. One particularly useful case is

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$3

Then if $$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$4 is odd,

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$5

If $$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$6 is even,

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$7

A further specialization yields the connection

$$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$8

This links Cartesian DA directly to angular-momentum coupling theory.

6. Computational workflow and significance for nonlinear spectroscopy

For practical DA in nonlinear spectroscopy, the implications are direct. The method replaces invariant-basis constructions by a universal finite sum. The paper states the implementation workflow as follows (Nessler et al., 2019):

1. Express the observable as a polynomial in direction cosines $$\langle T_{i_1\cdots i_n}\rangle = I^{(n)}_{i_1\cdots i_n;\lambda_1\cdots\lambda_n}\, T_{\lambda_1\cdots\lambda_n}.$$9.
2. Collect powers into the $I^{(n)}$0 exponent matrix $I^{(n)}$1.
3. Check the row/column parity rule; if it fails, the average is zero.
4. If needed, exploit transpose, row/column permutation symmetry, and odd-rank antisymmetry to reduce cases.
5. Evaluate the closed-form sum.

This workflow is especially useful because previous methods represented $I^{(n)}$2 in a basis of isotropic tensors built from $I^{(n)}$3 and $I^{(n)}$4, with coefficients determined by solving increasingly large linear systems. The paper’s method instead evaluates a short combinatorial sum over

$I^{(n)}$5

subject only to one parity condition.

The conceptual significance is broader than a computational shortcut. Since any isotropic tensor in three dimensions is built from products of $I^{(n)}$6 and $I^{(n)}$7, the closed-form formulas show how these invariant-theoretic properties emerge from parity, row/column symmetry, antisymmetry, and determinant structure without explicitly constructing an invariant basis. This suggests a reorganization of DA theory around universal direction-cosine kernels rather than basis-specific coefficient solving.

7. Terminological scope and acronym ambiguity

Within this usage, DA means Direction-Averaging: the isotropic orientational average over $I^{(n)}$8 of products of direction cosines, central to nonlinear spectroscopy and nonlinear optics (Nessler et al., 2019). In arXiv literature, however, the acronym “DA” is used for several unrelated constructions, including stochastic and online dual averaging in optimization (Liu et al., 27 May 2025, Fang et al., 2020), dual averaging with non-strongly-convex prox-functions (Zhao, 4 Apr 2025), double-averaging in hierarchical three-body dynamics (Luo et al., 2016), data weighted averaging in sigma-delta DACs (Laguna et al., 15 Dec 2025), and directional variance adjustment in high-dimensional covariance estimation (Bartz et al., 2011).

A common misconception is therefore terminological rather than mathematical: DA in spectroscopy is not an optimization algorithm, not a secular-orbit approximation, and not a dynamic element matching rule. In the spectroscopy literature, the defining object is the rotationally invariant tensor of averaged direction cosine products, and the fundamental problem is the exact orientational average of Cartesian tensor monomials over uniformly random molecular orientations.