Add an upper bound function for the operator norm.

[MaxenceGollier]

#270 (comment)

@MohamedLaghdafHABIBOULLAH @d1Lab,

Can you fetch my branch and make some benchmarks ? I think we should try to see if it works because it is very cheap.

[codecov]

Codecov Report

❌ Patch coverage is 77.41935% with 7 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.60%. Comparing base (e0f214d) to head (4732ee7).
⚠️ Report is 259 commits behind head on master.

Files with missing lines	Patch %	Lines
src/utils.jl	80.00%	5 Missing ⚠️
src/R2N.jl	50.00%	2 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           master     #275       +/-   ##
===========================================
+ Coverage   61.53%   84.60%   +23.07%     
===========================================
  Files          11       13        +2     
  Lines        1292     1650      +358     
===========================================
+ Hits          795     1396      +601     
+ Misses        497      254      -243     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MohamedLaghdafHABIBOULLAH]

Suggested change

	# Compute upper bounds for μ‖B‖₂, where μ ∈ (0, 1].
	# Compute upper bounds for ‖B‖₂.

[MohamedLaghdafHABIBOULLAH]

@MaxenceGollier @dpo @d1Lab, I conducted some benchmarks with the joss example, with solve! being more constraining:

  solve!(solver, reg_nlp, stats;
       atol=1e-6,
       rtol=1e-6,
       verbose=1,
       sub_kwargs=(max_iter=1000,),
       opnorm_maxiter= ...  % see below
   )

Below are my remarks concerning different operator–norm estimation
strategies when L-SR1 is used, which becomes problematic because $B_k$ may
be indefinite with a very small norm. Tests were performed on Julia
1.11.7 and Julia 1.12.2.

1. Behavior with opnorm_maxiter in {5,100}
   The solver fails on both versions. (with $$\xi < 0$$).

2. Behavior with opnorm_maxiter = 10

It works on both versions, showing that the instability is not simply due to using too few or too many power-method iterations.

3. Behavior with opnorm_maxiter = -1 (upper bound)

fails on both versions. (with $$\xi < 0$$).

4. Behavior with ARPACK

Using ARPACK to compute the dominant eigenvalue leads to instability:

• repeated runs do not produce identical results,
• sometimes the method converges,
• sometimes it fails with xi < 0,
• both Julia versions exhibit this behavior.

ARPACK introduces non-determinism and rounding differences that R2N does
not tolerate well.

Even for a small value such as $\theta = 0.1$, all the failures above persist. This suggests that choosing an appropriate $\theta$ is delicate: the solver succeeds with $\theta$ close to~1 when opnorm_maxiter = 10, but fails with $\theta = 0.1$

L-BFGS showed no errors with any operator-norm option.
So the instabilities above might be caused by the potentially indefinite LSR1 matrix rather than by the operator-norm estimation.

[MaxenceGollier]

Wouldn't it be more numerically stable to have $$\sigma‖s_{cp}‖$$ as stationarity measure ? It is also a generalization of the norm of the gradient for smooth optimization and does not involve potential catastrophic cancellation which might be the cause for the $$\xi &lt; 0$$.
Moreover, the inequality $$\sigma‖s_{cp}‖ \leq \sigma^{-1/2}\xi^{1/2}$$ keeps the convergence proof valid with that stationarity measure.
Just a suggestion... @dpo @d1Lab.

[MohamedLaghdafHABIBOULLAH]

Hey @MaxenceGollier — I apologize, most of the errors mentioned earlier
(specifically when opnorm = -1) are actually due to R2 and not R2N. The
core issue comes from the fact that:

    ERROR: LoadError: R2: prox-gradient step should produce a decrease but ξ = -7.653729494313895e288

In this situation, the proximal operator is not computed accurately for
the RootNormLhalf (which contains cos and arcos...), which leads to an incorrect (and extremely large
negative) value of ξ in contrast to the L0 or L1.

I ran additional experiments using:

    sub_kwargs = (max_iter = 0,)
    atol = rtol = 1e-4

to force the selection of $$s_{k,cp}$$ exactly.
Under these conditions, if the norm of Bk is not well approximated, we may still encounter an error.

Specifically:

• When opnorm = -1, we do not observe errors of the form
  ERROR: LoadError: R2N: failed to compute a step: ξ
  which confirms that R2N is robust in this case.

• These errors do appear when opnorm ∈ {10, 50, 100}, which suggests
  that inaccurate spectral‑norm estimates can destabilize the step
  computation.

So overall, this is good news:

• R2N is not the source of the problematic behavior if opnorm_maxiter = -1
• The main issue is linked to R2 combined with inaccurate norm
  estimation for Bk and the proximal operator of RootNormLhalf we should address this issue

[MohamedLaghdafHABIBOULLAH]

Wouldn't it be more numerically stable to have σ ‖ s c p ‖ as stationarity measure ? It is also a generalization of the norm of the gradient for smooth optimization and does not involve potential catastrophic cancellation which might be the cause for the ξ < 0 . Moreover, the inequality σ ‖ s c p ‖ ≤ σ − 1 / 2 ξ 1 / 2 keeps the convergence proof valid with that stationarity measure. Just a suggestion... @dpo @d1Lab.

@MaxenceGollier

I am not entirely sure whether this will make the problem disappear, because it is not
$$\xi$$ that is used as the stationarity measure, but rather $$\xi_1$$, which is related to
the step $$s_{\mathrm{cp}}$$ as you mentioned above.

The check on the condition $$\xi &lt; 0$$ is mainly for debugging purposes and to make the
implementation consistent with the theoretical framework (sufficient decrease).

In fact, as far as I can see in the benchmarks $$\xi_1$$ is always positive as long as you compute your prox exactly.

[MaxenceGollier]

Ok. Still, can we compare in term of time or number of iterations opnorm_max_iter ?
Actually, it would be most interesting to see the upper_bound approach vs the approach with arpack.