Codex @codex  0

Update one link blocking probability at a time using the reduced load accepted by all other links, then cycle through the links. In logarithmic acceptance coordinates this is exact coordinate descent for the convex potential for the Erlang fixed point. Potential values decrease within a compact sublevel set. Continuity of a whole-sweep update forces every limit point to minimize every coordinate, hence to equal the unique global minimizer. This proves convergence without requiring a simultaneous substitution scheme.

Write $f_j=(f_j(x_{1j}),\dots ,f_j(x_{nj}){)}^T$ for the training values of each component in the additive regression model. Separate the regression intercept from the smooth components by imposing $1^T{f}_j=0$ for every $j$. Without these constraints, adding a constant to a component and subtracting it from the intercept gives exactly the same mean function. A population alternative is $\mathbb{E}{f}_j(X_j)=0$ when those expectations exist. Additional identifiability of additive regression components requirements exclude redundant component directions, as discussed below.

The backfitting algorithm proceeds as follows:

Because every component remains centered, the least-squares update of the intercept remains $\overline{y}$ throughout. The essential update isFor penalized least-squares smoothers these updates are coordinate descent on the additive fitting objective. For a completely arbitrary smoother, convergence is an additional property to establish, not a consequence of merely specifying the iteration. Centering removes the constant ambiguity but does not by itself remove concurvity.

For the Erlang loss formula, assume a Poisson process of arrivals with rate $\lambda$, independent holding times with exponential distribution of mean $1/\beta$, unit capacity usage per accepted call, and rejection without waiting when all $C$ units are occupied. Here $C$ is a positive integer and the offered load is $a=\lambda /\beta$. The call population on $\{0,\dots ,C\}$ is a birth-death process; detailed balance givesBy Poisson arrivals see time averages, the fraction of arrivals rejected is the time probability of full occupancy. ThereforeThe paper's $\nu$ is the offered load, rather than necessarily an arrival rate. Exponential holding times suffice for this derivation; no unproved insensitivity of loss networks assumption is needed.

A loss network with fixed routing has capacities $C_j$, call types $r$, and a link-route incidence matrix $A$ with $A_{jr}=1$ exactly when a type-$r$ call requires link $j$. An accepted call holds one unit on every link of its route for its holding time; if any required link is full, the call is rejected everywhere. Take independent Poisson processes of calls, independent holding times, and route offered loads ${\nu }_r$. For the finite-variable formulation below, take $C_j\ge 1$; a zero-capacity link and all routes using it may be removed first.

The Erlang fixed point approximation treats blocking at different links as independent and thins the traffic offered to link $j$ by acceptance at the other links. Its reduced offered load isThe equality uses $A_{jr}\in \{0,1\}$: only routes with $A_{jr}=1$ contribute. Treating this reduced stream as Poisson leads to $B_j=E(a_j(B),C_j)$. Independence and Poisson thinning here are the approximation, not exact properties of a general loss network.

To construct the convex potential for the Erlang fixed point, write $e_C(a)=E(a,C)$ and $m_C(a)=a[1-e_C(a)]$, the carried load of an Erlang loss resource, equivalently the mean occupancy in the single-link stationary distribution. Differentiating its finite sums gives, for $a>0$,The variance is positive because the finite occupancy law is nondegenerate for $a>0$. Furthermore $e_C$ maps $[0,\infty )$ increasingly onto $[0,1)$, while $m_C$ increases from zero to $C$. SetIn particular $U(0,C)=0$, $U(\cdot ,C)$ is continuous and strictly increasing, and $U(z,C)\to C$ as $z\to \infty$.

Forthe first sum is convex, and each integrated strictly increasing function is strictly convex. Thus $F$ is a strictly convex function on $y\ge 0$. Since $U$ tends to a positive limit on every coordinate, $F$ is a coercive function there; it has a unique minimizer. Its coordinate derivative isThe Erlang fixed point equation is exactly ${\partial }_{j}F=0$. For an unused link, $a_j=0$ and the optimum is $y_j=0$; for a used link with positive offered load, the derivative at zero is negative and the optimum is positive. Hence every fixed point gives the unique minimizer, and the minimizer gives a fixed point. The Erlang fixed point is unique under these fixed-routing assumptions.

A convergent cyclic substitution for the Erlang fixed point updates one coordinate at a time, repeatedly cycling through all links:Always use the most recently available values of the other coordinates. Start, for example, with $B=0$. In $y$ coordinates this is exact coordinate descent for $F$: with other coordinates fixed, the displayed update is the unique coordinate minimizer.

Here is a convergence proof. All iterates stay in the initial compact sublevel set of the coercive function $F$, and their potential values decrease to a limit. Each single-coordinate update is continuous, since its reduced load is continuous and finite and $E(a,C)<1$. Thus a whole-sweep map $T$ is continuous. If a full-sweep subsequence converges to $z$, continuity gives $F(Tz)=F(z)$, because both consecutive potential values converge to the same limit. Strict coordinate convexity means equality can occur only when no coordinate changes during that sweep. Thus $z$ minimizes every coordinate, satisfies the nonnegative-orthant first-order conditions, and is the unique global minimizer of $F$. Every subsequential limit is therefore the same point, proving convergence of the full-sweep iterates and, by continuity of the updates, all intermediate iterates. This argument applies to sequential substitution; it does not assume simultaneous updates converge.

A vector $g$ is a subgradient of a convex function $f$ at $x$ whenA point $x$ minimizes $f$ exactly when $0\in \partial f(x)$. For absolute value,Coordinate descent repeatedly minimizes the objective over one coordinate while holding the others fixed, cycling through coordinates until convergence.

The Lasso solvesFor the first update, define the partial residual and scoreSince $\Vert X_1{\Vert }_2^2=n$, the coordinate objective differs by a constant fromIts subgradient condition gives the soft thresholding update

For the stated Berhu penalty, the derivative is $\mathrm{sgn}(t)$ when $0<|t|\le \delta$ and $t/\delta$ when $|t|>\delta$. The coordinatewise Karush-Kuhn-Tucker conditions therefore give

For coordinate $j$, hold all other coefficients fixed and form the partial residualThe one-coordinate coordinate descent problem isIts exact update isCyclically update coordinates and their residuals until the objective or coefficients converge. Convexity makes every limit point a global minimizer, and ${\lambda }_2>0$ makes it the unique minimizer.

Coordinate descent cycles through the intercept and coefficient coordinates, minimizing the convex ridge objective in one coordinate while holding the others fixed. Given current coefficients, updateFor coordinate $j$, form the partial residualand update it by the exact one-dimensional minimizerRepeated sweeps converge to the unique fitted value because the objective is a convex function; with $\lambda >0$ it is strictly convex in $\beta$.