Every essay — page 5
Least squares, and the road not to take
The normal equations are taught first and used by nobody, because forming AᵀA squares the condition number and then, below a computable value of ε, breaks outright. Underneath that is a harder fact: a wide range of very different fits explain the data equally well, and no arithmetic can choose between them.
Feasible and wrong
A third constraint that nearly repeats the first takes the best route's answer from 2.96·10⁻¹⁵ to 1.16·10⁻⁴, and the other two routes to no correct digit at all. Every one of those answers satisfies every constraint to 10⁻¹⁵. The quantity a caller checks after a constrained solve is the one quantity here that says nothing.
The condition number that does not know
Two constrained fits with the same size, the same number of constraints and the same κ(A) to twelve figures. One returns 4.7·10⁻¹⁶ and the other 3.0·10⁻⁴. What separates them is the conditioning of A restricted to the constraint's null space — 1.00 against 10¹² — which every solver computes on the way and none reports.
The reference was a method
The optimality conditions of a constrained fit contain AᵀA, so solving them is the road that squares the problem wearing a block structure. At κ(A) = 10¹¹ the route that never forms a cross-product returns 1.89·10⁻⁹ and the route that does returns 4.64·10⁻⁴ — and forming AᵀA and then solving it in exact rationals returns 3.45·10⁻⁴, so nearly all of the loss happens before any elimination begins.
The factor a sparse code keeps anyway
Every deletion diagnostic divides by one minus a leverage, and computing it as a subtraction loses a digit for every decade the leverage is from one. The route that does not subtract needs the orthogonal factor, which a sparse factorisation is supposed not to have. Three repairs that avoid it all fail at exactly a unit of roundoff over the divisor — and the fourth, which reaches the orthogonal factor through the Householder vectors a sparse code keeps in order to solve anything at all, returns the same bits as a stored factor in 900 operations.
The weight the factor met first
The route to one minus a leverage through the orthogonal factor was said to lose a digit for every decade of the condition number, whatever else it does. Put a weight on one row and it does not. With the heavy row first, the complement keeps every digit at κ(A) = 2.5·10⁹ while both subtractions return nothing. With the same row last it loses digits as the row's scale grows. And two heavy rows that leave κ(A) at 3.1 still lose six digits when the light rows come first. The law was about the order the factor met the rows, and the condition number had been standing in for it.
The residual the solution cannot hold
Sorting a weighted fit's rows heaviest first gave every digit of one minus the heavy row's leverage back. It gives nothing back to the heavy row's residual, if that residual is computed the way every textbook computes it — as the datum minus the fitted value. The fitted value is a double, and a double cannot resolve a misfit smaller than its own last digit times the weight: at a weight of 4²⁴ the residual formed from the solution is wrong in its second digit in every order, and forming the subtraction exactly changes nothing. Taken from the same orthogonal factor as the divisor, the residual keeps fifteen digits, and so does Cook's distance at 3.4·10¹⁷.
A multiplier is a force
A third constraint nearly parallel to the first made the multipliers of a constrained fit rise in exact proportion to κ(B), which looked like the conditioning measured a second, dearer way. It was not. Give the third constraint a datum that asks for nothing new and, at the same κ(B) = 4.6·10¹², the multipliers are eighteen thousand times smaller; give it a strain δ and they are 0.0133 δ/ε², a force on a lever of length ε. What they measure is what the constraint asks. What they do not measure is the error of the best route, which sits at the same level whether the constraint asks for nothing or for a displacement of 3·10⁹.
The degree that is safe to overshoot
The rules that choose a Tikhonov parameter miss by factors of millions on one draw in twenty. Transplanted to the degree of a polynomial fit, in a basis orthonormal on the data, the same rules never cost more than 2.7 times the best degree's error in three hundred draws. The reason is the shape of the valley they search: six degrees too few costs from 44 to 16,000 times the best error, forty degrees too many costs about twice it. The one rule with a tail, the discrepancy principle, has its threshold half a standard deviation above the residual it is waiting for.
The scale that only moved a pivot
Multiply the constraint rows of a saddle-point system until its multipliers are the size of its solution, and the extra error the route was blamed for — 4.6·10⁻⁴ against the null-space route's 1.8·10⁻⁸ — falls to 3.3·10⁻⁸. The prediction holds and its reason does not. A scale of ten does what a scale of 6·10⁵ does; hold the elimination's row order fixed and nine decades of scale move the error by less than a factor of five. What the scale changed was which row partial pivoting took at the second step, and taking the constraint rows first does the same job with no scale at all.
Regularisation, and the answer that is chosen
Some problems do not determine their own answer. The data is consistent with a range of solutions that differ by orders of magnitude, no arithmetic can choose between them, and something outside the data has to. That choice is the computation rather than a preliminary to it — and every published rule for making it is a heuristic scored, here, against a truth that exists only because the problem was constructed.
When the answer is a choice
A backward-stable least-squares solve of this problem returns an answer whose relative error is 5.5·10⁸. Nothing went wrong. The singular values decay exponentially with no gap anywhere in them, the data does not determine the answer, and something outside the data has to choose — which is the computation rather than a preliminary to it.
Where the answer stops being in the data
The Picard condition finds the index where a noisy right-hand side stops carrying signal, from the data alone, with no knowledge of the answer. It lands at 32 where the truncation that actually minimises the error is 28 — and at 45 where the best is 38. It overshoots at every stop from 10% noise to 0.0001%, and it overshoots for a reason. The best truncation walks up the spectrum in a straight line, six or seven indices a decade; the crossing climbs in jumps of 11, 0, 8, 5 and 1.
Choosing without knowing
Three published rules for choosing a regularisation parameter, scored against an oracle that requires the exact answer and is therefore not a method. Generalised cross-validation lands on the oracle's λ exactly; the discrepancy principle costs 6%; the L-curve costs 129%. And told a noise level ten times too small, the discrepancy principle's error goes from 0.112 to 10,449.
The basis decides what a filter is
The vocabulary of regularisation is spectral — a method keeps a component or discards it, and the weights are a function of the singular value. Row-normalising a symmetric blur so that it preserves a constant makes it 8.6% asymmetric, and that is enough to move GMRES's weights from 7·10⁻¹⁴ off a function of σ to 4.4·10⁻².
A second blur, narrower than the first
A regularised answer is not the truth with the noise taken out. It is the truth seen through a second blur, V F Vᵀ, which depends on the operator and λ and on nothing that was measured. At the best λ for 0.1% noise its rows are 2.82 points wide against the instrument's 5.89, they dip to −0.075 on either side, and their width times the number of components kept stays between 1.10n and 1.27n across seven decades of λ. Two spikes four points apart come back as two; three apart, as one.
Noise that spares the answer and fools the rules
Make each noise sample remember the last one, keep its size fixed, and the best answer available gets slightly better — 0.1056 to 0.1010 — because slow noise hides in the directions where dividing by σ costs nothing. The Picard crossing still lands two dozen indices past the best truncation. What breaks is the rules. Generalised cross-validation more than doubles the best error on 14 draws of 48 instead of 3, the discrepancy principle's typical cost triples, and the two miss in opposite directions. Whitening by the covariance takes GCV back to 3.
One draw in twenty
Sixteen draws gave generalised cross-validation a worst case of 12%. A thousand draws at each of five noise levels give it a second answer on four to six in every hundred, ten to seven million times worse than the oracle, while its median stays among the best of five rules. The quasi-optimality criterion, told nothing either, never costs more than 1.41 in five thousand draws. The share settles by a thousand draws, and letting the search look further down more than triples it.
The corner reads the norm it is drawn in
The L-curve was the costliest rule this field scored, and the cost was not the rule's. On the same sixty draws, with the same best achievable error, the corner of ‖x‖ against the residual costs 1.53 times the oracle and the corner of ‖L₁x‖ costs 1.003. Across five signals and three penalties the corner lands wherever amplified noise is between a tenth and a fifth of the norm being plotted, and it finds the oracle only when the oracle happens to sit there — twenty-nine times too costly on a smooth signal under ‖x‖, within half a per cent on four spikes.
The grid was the first filter
A continuous deconvolution discretised on n points and solved with no regularisation at all is not unregularised. Its error against the continuous signal is least at 24, 26 and 34 points for noise of 1%, 0.1% and 0.01% per sample — beside best truncations of 24, 28 and 32 components on a 64-point grid — and within 4 to 16 per cent of their error. The grid's own filter factors sum to n exactly and fall through a half at k = n. Choosing the grid was choosing a truncation, before anybody chose a λ.
Thirty-two coefficients instead of a noise level
The discrepancy principle has to be told the noise, and told too little it does not degrade — it falls off a cliff, at 0.80 of the truth when the noise is 10% and at 0.58 when it is 0.001%, exactly where the understatement forces the filter past its best truncation. The missing number is in the data. The root mean square of the last thirty-two coefficients never sends the rule over the cliff at or below 1% noise in four hundred draws, where eight coefficients with the same median do so thirty-five times.
A better discretisation is a weaker filter
A coarse grid's error has two sources — how well the discrete operator approximates the integral, and how well the grid's function represents the answer — and the grid essay could not separate them. Changed one at a time they separate: integrating the kernel against the hat functions takes a fifth off the 12-point error, reading the answer as a cubic spline takes 15 per cent more, and both roughly double the condition number on every grid. At 0.1% noise the spline discretisation's unregularised solve on 26 points reaches the best truncation of a 64-point grid to 0.3%. At 1% it is worse than the crude grid.
Where the grid hands over to λ
An unregularised solve on a coarse grid comes within a tenth of the best Tikhonov answer on a fine one, and the pair of a grid and a λ was left unmeasured. Measured, the two do not trade. On every grid up to the best unregularised one no λ helps at all. On every grid of 40 points and more the best λ is the same to within a quarter of a decade — 3.2·10⁻² at 1% noise per sample, 10⁻³ at 0.01% — and the 96-point grid with it beats the best coarse grid by 7, 9 and 13 per cent. The grids between the two, given their own λ, land between them.
A rule that has to be told how good its answer will be
The L-curve's corner reads a noise share of 0.10 to 0.21 across fifteen pairings of signal and penalty, and the share the best λ sits at runs from 0.0037 to 0.43 — a factor of a hundred and fifteen. A rule aimed at the right share is within a few per cent of the oracle on every one of them. The right share is about a third to four-fifths of the relative error that λ will achieve, which is the number the answer was wanted for.
A second penalty is not a second parameter
Penalise ‖x‖ and ‖L₁x‖ at once and there are two λ to choose. Over ten draws on five signals the best pair beats the better single penalty by between 0.00% and 3.1%, and one of its two parameters is exactly zero on 30 to 70 per cent of draws. Choosing the wrong one of the two costs up to 54%. The surface is a choice between two curves with a knob nobody needs.
The grid on which the discretisation stops mattering
Without regularisation, integrating the blur's kernel against cubic splines beat sampling it at 0.1% noise and lost to it at 1%. Give each discretisation its own best λ on every grid and the difference shrinks to nothing where grids are fine — 0.17, 0.28 and 0.20 per cent apart on 96 points at the three noise levels, with every discretisation choosing the same λ — and stays at 17 to 18 per cent on 16 points. The choice between them is a choice of how coarse a grid can be: at 0.1% noise the integrated discretisations reach the fine-grid answer on 26 points and the sampled one needs 40.