---
title: Least Squares Principle Value (LSPV)
url: https://www.emergentmind.com/topics/least-squares-principle-value-lspv
type: topic
---

# Least Squares Principle Value (LSPV)

Searching arXiv for the cited papers and topic terminology.
Least Squares Principle Value (LSPV) is not a standard standalone term in the cited papers, but the underlying construction is explicit in several distinct least-squares settings. In the context of Strutinsky smoothing, the natural LSPV is the Strutinsky-smoothed level density itself: the value at energy \(\lambda\) of the optimal local polynomial obtained by minimizing a Gaussian-weighted least-squares functional [1911.12592]. In the context of positive definite matrices and nonpositively curved metric spaces, the analogous object is the least squares mean or weighted least squares mean: the unique minimizer of a weighted sum of squared distances, equivalently a barycenter or Karcher mean [1007.4792]. A different usage arises in nonconvex \(\ell_p\)-regularized least squares for \(0<p<1\), where the relevant “principle values” are organized not by a single smooth solution path but by critical points and critical paths of constrained and penalized least-squares formulations [1304.6591]. Taken together, these works show that an LSPV is best understood not as a universal named object, but as the value selected by a least-squares principle within a specified geometric, functional, or regularized setting.

## 1. Least-squares principle as a value-selection mechanism

The common structure is a minimization principle. In the Strutinsky setting, one starts from the quantum level density
\[
g(\epsilon)=\sum_{i=0}^{\infty}\delta(\epsilon-\epsilon_{i}),
\]
and assumes a decomposition
\[
g(\epsilon) = g_0(\epsilon) + \delta g(\epsilon),
\]
where \(g_0(\epsilon)\) is a smooth, slowly varying part and \(\delta g(\epsilon)\) contains the shell oscillations [1911.12592]. The least-squares principle enters through a local approximation problem: at a fixed reference energy \(\lambda\), one chooses a polynomial \(Q_M(\epsilon)\) of degree \(M\) that minimizes
\[
I(M,\gamma,\lambda) = \int_{-\infty}^{+\infty}
\Big\{ g(\epsilon) - Q_M(\epsilon)\Big\}^2
\exp\!\left[-\left(\frac{\epsilon-\lambda}{\gamma}\right)^2\right] \, d\epsilon.
\]
The resulting central value \(Q_{M,\gamma}(\lambda)\) is the smoothed value selected by the least-squares construction, and the paper shows that the standard Strutinsky smoothing is exactly the outcome of this procedure [1911.12592].

In the matrix and metric-space setting, the same pattern appears in intrinsic form. For positive definite Hermitian matrices \(A_1,\dots,A_n\in P\), the weighted least squares mean is defined by
\[
\Theta_n(w;A_1,\dots,A_n) = \operatorname*{arg\,min}_{X\in P} \sum_{i=1}^n w_i\, d^2(X,A_i),
\]
where \(d\) is the trace metric and \(w\) lies in the probability simplex [1007.4792]. In a general NPC space \(M\), the corresponding definition is
\[
\Theta_n(w_1,\dots,w_n; a_1,\dots,a_n) = \operatorname*{arg\,min}_{z\in M} \sum_{i=1}^n w_i d^2(z,a_i).
\]
Here the LSPV is the unique minimizing point itself: a least-squares-selected representative of the data in the underlying geometry [1007.4792].

In \(\ell_p\)-regularized least squares with \(0<p<1\), the situation is more intricate because nonconvexity destroys the usual one-to-one correspondence between constrained and penalized formulations. The relevant “principle values” are therefore not exhausted by global minimizers; instead, the paper studies the broader set of critical points and critical paths determined by least-squares stationarity under \(\ell_p\) regularization [1304.6591]. This suggests that the meaning of LSPV depends strongly on whether the least-squares landscape is convex, geodesically convex, or nonconvex.

## 2. LSPV in Strutinsky smoothing

The most explicit identification of LSPV appears in the derivation of the Strutinsky method from the least squares principle. At each energy \(\lambda\), the polynomial approximant is expanded in Hermite polynomials,
\[
Q_{M,\gamma,\lambda}(\epsilon) = \sum_{m=0}^{M} d_m(M,\gamma,\lambda)\, H_m\!\left( \frac{\epsilon-\lambda}{\gamma} \right),
\]
with
\[
H_n(x) = (-1)^n e^{x^2} \frac{d^n}{dx^n} e^{-x^2},
\]
and orthogonality
\[
\int_{-\infty}^{+\infty} H_n(x) H_m(x) e^{-x^2} dx = 2^n\,n!\sqrt{\pi}\,\delta_{nm}.
\]
Minimization of the least-squares functional yields the normal equations
\[
d_k(\gamma,\lambda) = \frac{1}{2^{k} k! \sqrt{\pi}\,\gamma} \int_{-\infty}^{+\infty} g(\epsilon)\, H_k\!\left(\frac{\epsilon-\lambda}{\gamma}\right)\, \exp\!\left[-\left(\frac{\epsilon-\lambda}{\gamma}\right)^2\right] d\epsilon.
\]
The smoothed level density is then defined as the value of the fitted polynomial at its center,
\[
Q_{M,\gamma}(\lambda) \equiv Q_{M,\gamma,\lambda}(\lambda)
= \sum_{k=0}^M d_k(\gamma,\lambda)\, H_k(0).
\]
This is the natural LSPV in this setting: the optimal smoothed value at \(\lambda\) implied by the least-squares principle [1911.12592].

The paper further shows that this value admits the integral representation
\[
Q_{M,\gamma}(\lambda) = \int_{-\infty}^{+\infty} g(\lambda+\gamma x)\, F_M(x)\, dx,
\]
or equivalently
\[
Q_{M,\gamma}(\lambda) = \int_{-\infty}^{+\infty} g(\epsilon)\,\frac{1}{\gamma}\, F_M\!\left(\frac{\epsilon-\lambda}{\gamma}\right) d\epsilon,
\]
with averaging function
\[
F_M(x) = \left\{ \sum_{m=0}^{M} B_m\, H_m(x) \right\} \frac{e^{-x^2}}{\sqrt{\pi}}, \quad
B_m = \frac{H_m(0)}{2^m m!}.
\]
This is exactly the usual Strutinsky “Gaussian + curvature correction” convolution, now derived from the least-squares principle rather than introduced as an ad hoc prescription [1911.12592].

The paper explicitly states that the phrase “Least Squares Principle Value (LSPV)” does not appear in the paper, but identifies the natural quantity embodying the concept as \(Q_{M,\gamma}(\lambda)\). More concretely, the optimal coefficients \(d_k(\gamma,\lambda)\) can be seen as least-squares principle values of the projections of \(g\) onto the Hermite basis with Gaussian weight, while the optimal level density \(Q_{M,\gamma}(\lambda)\) is the LSPV of \(g(\lambda)\) under the polynomial moving-average scheme [1911.12592]. Since the Strutinsky method uses these values as the smoothed input for shell-correction calculations, the LSPV is, in practice, the Strutinsky density itself.

## 3. Averaging functions, polynomial reproduction, and asymptotic interpretation

A central feature of the Strutinsky LSPV is the structure of the averaging function \(F_M(x)\). For even \(M\), the paper establishes parity,
\[
F_M(-x)=F_M(x),
\]
normalization,
\[
\int_{-\infty}^{+\infty} F_M(x)\,dx = 1,
\]
and vanishing moments up to order \(M\),
\[
\int_{-\infty}^{+\infty} x^k\, F_M(x)\,dx = 0,
\qquad 1\le k\le M.
\]
These properties imply exact reproduction of polynomials of degree \(\le M\): if \(g(\epsilon)\) is any polynomial of degree \(\le M\), then
\[
\tilde g(\lambda)
=
\int_{-\infty}^{+\infty}
g(\epsilon)\,\frac{1}{\gamma}\,F_M\!\left(\frac{\epsilon-\lambda}{\gamma}\right)d\epsilon
=
g(\lambda).
\]
For a general smooth function, the transform has the expansion
\[
\int_{-\infty}^{+\infty}
g(\epsilon)\,\frac{1}{\gamma}\,
F_M\!\left(\frac{\epsilon-\lambda}{\gamma}\right)\,d\epsilon
=
g(\lambda) +
\sum_{k=M+2}^{\infty} D_{kM}\,\gamma^{k}\,
\frac{d^{k}g(\lambda)}{d\lambda^{k}},
\]
with
\[
D_{kM} =
\frac{(-1)^{M/2}\,(M/2)!}{[(k-M-2)/2]!}
\frac{1}{2^{k-1}k},
\qquad k,M \text{ even}.
\]
The first nonzero contribution arises at order \(k=M+2\) because lower moments vanish [1911.12592].

These facts give the Strutinsky LSPV a precise approximation-theoretic meaning. The kernel is designed so that the least-squares averaging leaves all polynomials up to degree \(M\) unchanged and modifies more general functions only through controlled higher-order terms. The paper therefore concludes that, when suitable asymptotic conditions are met,
\[
\lambda \gg \hbar\omega,\qquad \gamma \sim \hbar\omega,\qquad \lambda \gg \gamma,
\]
Strutinsky’s method is nothing more than a polynomial moving average of the semi-classical level density [1911.12592]. In this regime, the quantum level density becomes essentially continuous and close to the semiclassical density \(g_{\text{sc}}(\epsilon)\), and the LSPV \(Q_{M,\gamma}(\lambda)\) approximates that smooth density.

A particularly sharp consequence is noted for systems whose semiclassical density is exactly a polynomial of degree 2. The paper cites the three-dimensional harmonic oscillator, where \(g_{\text{sc}}\) is parabolic; in that case, the Strutinsky method with \(M\ge 2\) reproduces \(g_{\text{sc}}\) exactly, because the kernel leaves polynomials of degree \(\le M\) unchanged [1911.12592]. This does not define a universal property of all LSPVs, but it shows that in this setting the least-squares-selected value can coincide exactly with the target smooth density.

## 4. LSPV as least squares mean, barycenter, and Karcher mean

A second rigorous realization of LSPV is the least squares mean for positive definite matrices and, more generally, barycenters in NPC spaces. The geometric setting is the cone \(P=P_m\) of \(m\times m\) positive definite Hermitian matrices equipped with the trace metric
\[
d(A,B) = \left( \sum_{i=1}^m \log^2 \lambda_i(A^{-1}B) \right)^{1/2},
\]
under which \(P\) is a complete Riemannian manifold of nonpositive curvature and hence a CAT(0) space [1007.4792]. In this geometry, the two-variable geometric mean is the trace-metric midpoint
\[
A\#B := A^{1/2}\bigl(A^{-1/2}BA^{-1/2}\bigr)^{1/2}A^{1/2},
\]
and the weighted mean
\[
A\#_t B := A^{1/2}\bigl(A^{-1/2}BA^{-1/2}\bigr)^t A^{1/2},\quad t\in[0,1],
\]
is the point at proportional distance \(t\) along the unique geodesic from \(A\) to \(B\) [1007.4792].

For data \(A_1,\dots,A_n\in P\), the unweighted least squares mean is
\[
\Theta_n(A_1,\dots,A_n) = \operatorname*{arg\,min}_{X\in P} \sum_{i=1}^n d^2(X,A_i),
\]
and the weighted least squares mean is
\[
\Theta_n(w;A_1,\dots,A_n) = \operatorname*{arg\,min}_{X\in P} \sum_{i=1}^n w_i\, d^2(X,A_i),
\]
with intrinsic characterization
\[
\sum_{i=1}^n w_i \,\log(XA_i^{-1}) = 0.
\]
The paper identifies this as the Karcher equation [1007.4792]. In this setting, the LSPV is the unique matrix \(X\) satisfying the least-squares minimization problem or, equivalently, the logarithmic balance equation.

The same construction extends to general NPC spaces. For a square-integrable probability measure \(\mu\), the barycenter is
\[
b(\mu) = \operatorname*{arg\,min}_{z\in M} \int_M d^2(z,x)\,\mu(dx),
\]
and when \(\mu=\sum_{i=1}^n w_i \delta_{a_i}\), this reduces exactly to the weighted least squares mean
\[
b(\mu)=\Theta_n(w;a_1,\dots,a_n).
\]
The paper explicitly states that the Lawson–Lim least squares mean is exactly the barycenter, Fréchet mean, or Karcher mean of a discrete probability measure in an NPC space [1007.4792]. This establishes a geometric interpretation of LSPV as a canonical center of mass determined by least-squares minimization.

The terminology is therefore partly contextual. What may be described informally as an LSPV corresponds, in the paper’s terminology, to the least squares mean or weighted least squares mean; in adjacent literatures, the same construction is called the Karcher mean, Fréchet mean, barycenter, or matrix Riemannian center of mass [1007.4792]. This suggests that “LSPV” is best regarded as a descriptive umbrella rather than a field-wide standard label.

## 5. Structural properties in ordered metric geometry

Within the matrix setting, the least-squares-selected value has strong order-theoretic and metric properties. The relevant order is the Loewner order,
\[
A\leq B \quad\Longleftrightarrow\quad B-A\ \text{is positive semidefinite}.
\]
Lawson and Lim prove that the weighted least squares mean is monotone in Loewner-Heinz NPC spaces: if \(x_i\le y_i\) for all \(i\), then
\[
\Theta_n(w;x_1,\dots,x_n)\le \Theta_n(w;y_1,\dots,y_n).
\]
Specializing to positive definite matrices gives monotonicity of the weighted least squares mean on \(P\) in the Loewner order, thereby settling the long-standing open problem described as the Bhatia–Holbrook conjecture or Zhan’s problem 19 [1007.4792].

The same work establishes continuity and non-expansiveness. For
\[
\mu_1 = \sum_{i=1}^n w_i \delta_{x_i},\quad \mu_2 = \sum_{i=1}^n w_i \delta_{y_i},
\]
the least squares means satisfy
\[
d\bigl(\Theta_n(w;x_1,\dots,x_n),\,\Theta_n(w;y_1,\dots,y_n)\bigr) \le W(\mu_1,\mu_2) \le \sum_{i=1}^n w_i d(x_i,y_i),
\]
where \(W\) is the Wasserstein \(L^1\)-distance [1007.4792]. For positive definite matrices, the paper also proves joint concavity:
\[
\Theta_n\bigl(w; tA_1+(1-t)B_1,\dots,tA_n+(1-t)B_n\bigr)
\ge
t\Theta_n(w; A_1,\dots,A_n) + (1-t)\Theta_n(w; B_1,\dots,B_n),
\]
for \(0<t<1\) [1007.4792].

The paper further states that the weighted least squares mean satisfies the standard geometric mean axioms P1–P10 for positive definite matrices, including consistency with scalars, joint homogeneity, permutation invariance, monotonicity, continuity, congruence invariance, joint concavity, self-duality, the determinantal identity,
\[
\det\Theta_n(w;A_1,\dots,A_n)
=
\prod_{i=1}^n (\det A_i)^{w_i},
\]
and the AGH weighted mean inequalities
\[
\left(\sum_{i=1}^n w_i A_i^{-1}\right)^{-1}
\le
\Theta_n(w;A_1,\dots,A_n)
\le
\sum_{i=1}^n w_i A_i.
\]
In this context, the LSPV is not merely a minimizer; it is a multivariate mean with an extensive structural theory [1007.4792].

A plausible implication is that this matrix-geometric realization of LSPV is more canonical than the term itself: the minimizer is uniquely determined by the geometry, respects the ambient order, and behaves stably under perturbations. The cited paper does not formulate this as a general philosophy of LSPV, but its results support that interpretation.

## 6. Nonconvex \(\ell_p\)-regularized least squares and critical-path interpretations

A third setting complicates any naive idea that a least-squares principle should always select a single smooth path of values. The paper on \(\ell_p\)-regularized least squares with \(0<p<1\) considers the constrained problem
\[
P^p:\quad
\min_{\boldsymbol{\beta}\in\mathbb{R}^n} \ \varphi(\boldsymbol{\beta})
\quad
\text{subject to } F_p(\boldsymbol{\beta}) \le c,
\]
with
\[
\varphi(\boldsymbol{\beta}) := \frac{1}{2}\|X^{\sf T}\boldsymbol{\beta} - \mathbf{y}\|_2^2,
\qquad
F_p(\boldsymbol{\beta}) := \frac{1}{p}\|\boldsymbol{\beta}\|_p^p
= \sum_{i=1}^n \psi_p(\beta_i),
\]
and the penalized problem
\[
L^p:\quad \min_{\boldsymbol{\beta}\in\mathbb{R}^n}
f_\lambda(\boldsymbol{\beta}) := \varphi(\boldsymbol{\beta}) + \lambda F_p(\boldsymbol{\beta}).
\]
For \(p\ge 1\), these problems are convex and equivalent in the usual sense. For \(0<p<1\), this equivalence breaks down because the \(\ell_p\) quasi-norm is nonconvex [1304.6591].

Critical points satisfy an active-coordinate stationarity condition. If \(\mathcal{I}=\operatorname{supp}(\boldsymbol{\beta})\), then \(\widetilde{\boldsymbol{\beta}}\) is a critical point when
\[
\nabla_{\mathcal{I}}\varphi(\widetilde{\boldsymbol{\beta}})
=
-\widetilde{\lambda}\,\nabla_{\mathcal{I}} F_p(\widetilde{\boldsymbol{\beta}})
\quad\text{for some}\quad \widetilde{\lambda}\ge 0.
\]
The paper emphasizes that critical points include global or local minima, saddle points, and local maxima [1304.6591]. Because the global solution paths of \(P^p\) and \(L^p\) need not be continuous, the paper introduces critical paths: maximal continuous curves of critical points parametrized by \(c\) or \(\lambda\). These curves are piecewise smooth and generally contain non-optimal points such as saddle points and local maxima as well as global or local minima [1304.6591].

In this nonconvex setting, the most direct analogue of LSPV is therefore not a unique minimizer defined for each regularization parameter in a monotone way. The paper proves instead that, for \(p<1\), the \(L^p\)-solution path is a proper subset of the \(P^p\)-solution path, that the \(L^p\)-solution path is always discontinuous at the origin, and that the relationship between \(c\) and \(\lambda\) along a critical path is non-monotonic and has multiplicity [1304.6591]. This undermines any universal assumption that a least-squares principle automatically yields a single globally coherent value-selection map.

The same paper identifies two structured families of critical points: the main path, which starts at an OLS solution and moves backward toward the origin, and the greedy path, which starts at the origin and activates variables step by step. Part of the greedy path can be constructed with a generalized Minkowskian gradient, and the breakpoints of the greedy path coincide with the step-by-step solutions generated by Orthogonal Matching Pursuit [1304.6591]. This suggests that, in nonconvex least-squares problems, “principle values” may be better understood as organized pathwise objects rather than as a single canonical estimator.

## 7. Scope, terminology, and limitations

The cited papers support three different but related uses of a least-squares principle. In Strutinsky theory, the least-squares-selected value is a local polynomial estimate evaluated at the center of a Gaussian window, and the paper shows rigorously that standard Strutinsky smoothing has exactly this form [1911.12592]. In matrix analysis and NPC geometry, the least-squares-selected value is a barycenter or Karcher mean determined by minimizing a sum of squared distances, with monotonicity, continuity, and joint concavity available in the positive definite matrix case [1007.4792]. In nonconvex \(\ell_p\)-regularized least squares, least-squares selection generates a stratified landscape of critical points and critical paths rather than a unique regularization path [1304.6591].

A common misconception would be to treat LSPV as the name of a single established object across these areas. The data do not support that conclusion. One paper explicitly states that the phrase “Least Squares Principle Value (LSPV)” does not appear in the paper, even though a natural quantity embodies the concept [1911.12592]. Another paper uses the terminology least squares mean, weighted least squares mean, barycenter, and Karcher mean rather than LSPV [1007.4792]. The \(\ell_p\)-regularization paper likewise does not define LSPV as a named concept, instead distinguishing global solution paths from critical paths and emphasizing the main path and greedy path as especially informative trajectories [1304.6591].

The limitations are likewise setting-dependent. In the Strutinsky derivation, validity depends on the assumption that the level density can be reasonably approximated by a polynomial of degree \(M\) over a window of width \(\gamma\), with \(\gamma\) of order \(\hbar\omega\) and \(M\) even and not too large [1911.12592]. In the matrix setting, the clean theory relies on NPC geometry and a compatible order structure; outside globally nonpositively curved contexts, uniqueness of the minimizer can fail and some properties are not guaranteed [1007.4792]. In \(\ell_p\)-regularized least squares for \(0<p<1\), nonconvexity implies multiple local minima, saddle points, discontinuous global solution paths, and non-monotone parameter correspondences [1304.6591].

Across these settings, the most precise encyclopedic characterization is therefore contextual. “Least Squares Principle Value” denotes the value selected by minimizing an appropriate least-squares criterion, but the mathematical nature of that value depends on the ambient structure: a smoothed central estimate in Strutinsky theory, a barycentric mean in matrix or NPC geometry, or a point on a critical-path manifold in nonconvex sparse regularization.

Source: https://www.emergentmind.com/topics/least-squares-principle-value-lspv