Recursive estimators on drifting streams must forget stale data fast enough to track the target, yet slowly enough to average out noise. In prior-anchored ridge regression, this trade-off admits a closed-form spectral decomposition in which each eigen-direction interpolates between the data solution and the prior with pull weight si=ρ/(μi+ρ). When the anchor is refreshed every k steps, we show that the moving-frame offset converges to a finite-cadence equilibrium that lags the pure-leaky path by an explicit Δk(s) while reducing estimation variance by a factor Ψ(s,rk). Minimizing the cycle-averaged risk gives ω∗=(4ν2Σ/c′)1/3, c′=σ2deff(ρ), and the plug-in strength ρ∗=σ2/(nδ2). A random-matrix deterministic equivalent extends both laws to finite dimension, renormalizing the anchor to ρeff=ρ(eζq−1)/(ζq); it matches simulation within 5% at fixed operating points and within 13% at the method's own, and rank-deficient populations reduce exactly to their active subspace. The resulting method, SPAR, calibrates window and strength online from residuals. On a 27-configuration grid the calibrated window opens on drift without divergence, and the grid defaults transfer to an hourly electricity stream and to BCI IV 2b. The core algebra, refresh chain, and regret bounds are formalized in Lean 4.
VFF-RLS ( Fortescue et al., 1981 ; Paleologu et al., 2008 )
–
–
–
window
–
–
directional forgetting ( Kulhavý and Kárný, 1984 ; Cao and Schwartz, 2000 ; Lai and Bernstein, 2024 )
–
subspace
–
–
–
–
AROW ( Crammer et al., 2009 )
confidence
per-feature
–
–
–
–
Kalman / state space ( Kalman, 1960 ; Sayed and Kailath, 1994 )
process
–
–
q fixed
–
–
WiSE-FT ( Wortsman et al., 2022 )
frozen model
–
–
–
–
–
Table 1: What is new relative to the nearest lines of work. “DE” is the finite-window deterministic-equivalent correction; “Lean” the machine-checked core.
notation
fixed constants
D
input dimension
k
refresh cadence 10
Tw,T
warm-up, stream length
hf
fast residual EMA 30
A,B,n
sufficient statistics
hslow
slow residual EMA 20,000
Aˉt,At
mean, running accumulator
2000
Gref EMA horizon
μi,U
excitation eigenpairs
ρmin
strength clip floor 0.02
si
pull weight ρ/(μi+ρ)
ρmax
strength clip cap 1.0
Table 2: Notation and the fixed runtime constants, shared by every configuration (Online Appendix F ); every other quantity adapts online.
method
mean MSE
geo. mean
worst MSE
baselines (single hyperparameter, tuned post hoc on the evaluation grid)
leaky RLS ( λ=0.997 )
0.555
0.233
2.12
leaky RLS, information form ( λ=0.99 )
0.342
0.188
1.19
SGD ( lr=0.01 )
0.377
0.147
1.29
Adam ( Kingma and Ba, 2015 ) ( lr=0.01 )
0.503
0.291
1.92
directional forgetting
1.81
0.567
7.54
Table 3: Factorial grid (27 main configurations × 5 seeds): mean, geometric-mean, and worst overall MSE (worst single run of 135; the geometric mean is scale-invariant across the three spectra). Baselines use their family’s best single hyperparameter, selected post hoc on the evaluation grid; the causal Kalman instead selects its q from the warm-up alone. The true-noise Kalman is an oracle comparator (best single q , 0.204/0.86 ); the WiSE-FT envelope sweeps α per configuration. VFF-RLS covariance reports the nominal-length calibration’s best setting ( N0=100 ); the single-step calibration ξ=σ02 diverged in all 135 runs (Online Appendix D ). Standard errors of the mean over the 135 runs are 0.026 – 0.050 , so aggregate differences below about 0.05 are not resolved by the mean alone; the paired tests use configuration-level pairing.
dimension
assumption or protocol
consequence outside it
design model
Theorems 2 – 3 use the mean-field accumulator; Theorem 10 retains the random design
at drift-phase ζ≤6.4 the exact random-design variance replaces the mean-field term in the cycle-averaged risk; the runtime keeps the conservative mean-field plug-in
spectrum
Theorem 10 assumes the nonzero part is bounded away from zero, τmin+>0 ; rank-deficient spectra reduce exactly to range(Σ) (Corollary 19 )
null directions drop out of the prediction risk, so the deterministic equivalent restricts cleanly to range(Σ) with effective aspect ratio ζ+=pζ
contraction
Theorem 11 contracts globally on the parameter box [ωmin,ωmax]×[ρlo,ρhi] , with the clip a nonexpansive projection (Corollary 20 )
outside the box the gate and the measured-product hypothesis apply; the basin is the whole box
moving targets
fixed-protocol grid, 27 configurations
unanchored information-form leaky RLS wins 12/18 moving-target configurations, but all 12 are small-margin (isotropic 2.1% , rank-deficient 0.8% tie); SPAR wins the ill-conditioned ones by 2× and the moving-target mean by 9.3% , and all 9 static-target covariance-rotation configurations
strength rule
onset guard; average effect measured by ablation
freezing it lowers mean MSE by 2.2% ; its value is the transient guard, not a continuous tracking gain
detector timescale
two-scale residual EMA, hf=30 , hslow=20,000≈2T ; a scale-free sticky floor (frozen while an excess is present) handles sustained drift
the fixed-horizon detector under-opens once a drift persists beyond hslow , while the sticky-floor variant keeps the window open and halves the sustained-drift tail MSE
Table 4: Scope of the claims. Each row names a condition, the protocol that tests it, and the consequence outside the condition.
Table 5: Map from the paper’s results to the Lean artifact. The differentiation step of Theorem 3 ’s convexity and the probabilistic statements are pen-and-paper; every algebraic identity with a named Lean declaration above is formalized; the remaining rows are pen-and-paper and appear in the list below.
method
mean MSE
worst MSE
Kalman, true noise ( q=10−4 )
1.29
6.26
SPAR-EG +
1.60
7.83
SPAR-EG
1.60
7.77
leaky RLS, information form ( λ=0.99 )
1.61
7.70
SPAR-Pull
1.66
8.10
SGD ( lr=0.01 )
2.07
10.03
Appendix
Table 6: Stress suite (4 configurations × 5 seeds): mean and worst overall MSE, one fixed setting per method. The suite is deliberately outside the main grid’s operating envelope (loss of persistent excitation, oscillating drift, a stationary control, heavy-tailed rotation noise); the true-noise Kalman is an oracle comparator, and the covariance-form recursion diverges in all 20 runs.
axis
value
SPAR-Pull
SPAR-EG +
KF-slow
VFF
leaky-info
spectrum
isotropic
0.673
0.644
0.507
0.562
0.636
decay-3
0.114
0.115
0.065
0.120
0.257
rank-deficient
0.131
0.132
0.071
0.119
0.134
drift
abrupt
0.454
0.440
0.310
0.387
0.487
covariance rotation
0.011
0.011
0.023
0.027
0.054
both
0.454
0.440
0.310
0.387
0.487
Appendix
Table 7: Marginal mean MSE over the 27 main configurations (each row averages 9 configurations × 5 seeds), addressing the aggregate-only view of Table 3 . The two SPAR methods agree within 0.03 on every slice; the Kalman and VFF leaders are prior-independent and lose the covariance-rotation slice. The abrupt and both rows coincide to three decimals because both streams carry the same abrupt target jump, and the added covariance rotation shifts the drift-axis mean by less than 10−3 .
λ ( 1/(1−λ) )
D
standard recursion
symmetrized
0.997 ( 333 )
128
1.15
1.15
0.995 ( 200 )
128
25.1
0.80
0.99 ( 100 )
128
2.8×104
0.61
0.99 ( 100 )
512
7.0×104
10.4
Appendix
Table 8: Mean MSE of the covariance-form leaky recursion on full-rank spectra, with and without per-step symmetrization of the iterated inverse. At the longest window both variants agree to machine precision; at short windows the unsymmetrized recursion diverges at both dimensions while the symmetrized variant remains stable at every window tested.
variant
warm overall
adversarial overall
adversarial early
full
2.26±3.28
1.04±0.21
16.4±3.3
fixed ρ (refresh on)
2.20±3.16
1.05±0.21
16.5±3.3
frozen anchor (no refresh)
4.98±6.49
7.69±1.62
47.0±10.2
no variance floor
2.30±3.33
1.06±0.21
16.7±3.3
Appendix
Table 9: Component dissociation on the 4-configuration ablation suite (mean ± s.d. over 20 runs per cell; early MSE is the first 500 post-jump steps). Freezing the anchor refresh is the only variant that loses prior immunity; freezing the strength estimator is indistinguishable from the full method.
k
strength policy
pre-drift MSE
post-jump MSE [0,50]
overall MSE
10
adaptive
0.0136
9.48
0.166
10
fixed ρ=0.01
0.0116
9.48
0.168
10
fixed ρ=0.1
0.0271
9.49
0.195
10
fixed ρ=1
0.0489
7.72
0.311
50
adaptive
0.0197
9.52
0.254
50
fixed ρ=0.01
0.0146
9.54
0.243
Appendix
Table 10: Transient strength sweep on the abrupt-jump stream, decay-3 spectrum, 5 seeds: pre-drift MSE, MSE over the 50 steps after the jump, and overall MSE. The refresh is on in every row; the adaptive row is the paper’s method and the others pin ρ for the whole stream.
method
ms/step at D=2048
leaky RLS
32.9
leaky RLS, information form
189
AROW
30.6
Kalman
50.5
directional forgetting
58.6
SPAR-Pull-D (cached basis, O(D2) /step)
51.3
Appendix
Table 11: Per-step cost at D=2048 (fp64 CPU). The cached-basis per-direction head projects each feature into the last refresh’s eigenbasis and solves the diagonal system ( O(D2) /step, O(D3) every k refreshes); the scalar method performs a Cholesky factorization plus triangular solves per step ( O(D3) ), which motivates the per-direction path at large D .
Figure 6: (a) Pre-drift MSE remains at the irreducible floor σ2=0.01 across all three spectra (dots show 5-seed means ±1 s.d. against the best tuned- q Kalman with true σ2 ), illustrating Proposition 5 at the optimum. (b) Overall MSE (log scale) across oracle, warm, and adversarial priors (means over all runs). While the WiSE-FT best- α envelope sits near 3.2 under every prior, SPAR-Pull varies by only 4.1% across all three priors (bracket): Proposition 6 .
method
full horizon
final 2,000 h
leaky RLS, information form ( λ=0.999 )
0.0209
0.0134
leaky RLS ( λ=0.995 )
0.0219
0.0160
leaky RLS ( λ=0.99 )
0.0228
0.0174
Kalman, slow estimated R ( q=10−5 )
0.0660
0.0170
Kalman, slow estimated R ( q=10−4 )
0.3054
0.0183
SPAR-EG + (grid defaults)
0.0212
0.0143
Appendix
Table 12: Real stream (hourly electricity load, D=324 , T=6,000 evaluation hours, 25th-percentile held-out client): MSE of the next-hour log-load over the full horizon and over the final 2,000 hours. SPAR methods use the factorial-grid defaults with no stream-specific tuning; baseline hyperparameters are swept on the same stream and reported at their best.
held-out client
best oracle leaky
SPAR-EG +
SPAR-Pull
25th percentile
0.0209
0.0212
0.0232
50th percentile
0.0086
0.0083
0.0090
75th percentile
0.0112
0.0113
0.0130
three-client mean
0.0136
0.0136
0.0151
Appendix
Table 13: Real stream, held-out client sweep: MSE of the next-hour log-load over the full 6,000 -hour horizon for the 25th-, 50th-, and 75th-percentile clients by mean load. The oracle column is the best of the three leaky windows ( λ∈{0.99,0.995,0.999} ) chosen per client ( λ=0.999,0.995,0.995 at the 25th, 50th, 75th percentiles), tuning in the baseline’s favour, while the SPAR methods use the grid defaults with no stream-specific tuning. Bold marks the best entry in each row.
method
BCI IV 2a
BCI IV 2b
classical (canonical implementations)
MDM (pyriemann)
0.566±0.160
0.728±0.148
FgMDM (pyriemann)
0.641±0.171
0.755±0.149
LDA (raw features)
0.480±0.069
0.651±0.107
LDA (raw) + session EA T
0.563±0.076
0.677±0.133
CSP-LDA (mne)
0.607±0.151
0.749±0.141
Appendix
Table 14: BCI IV cross-session, balanced accuracy (mean ± s.d. over 9 subjects per data set). Supervised online protocol for adaptive methods; classical methods are static. T Session-level Euclidean alignment from the unlabeled evaluation session (transductive); the online-EA row is strictly causal. Highest value per column in bold.
parameter
grid default
EEG head
refresh cadence k
10
5 trials
fast-EMA horizon hf
30
40
warm-up anchor strength ρ0
0.1
0.1
stationary ρ floor
10−5
10−5
drift-active ρ floor
0.02
—
anchor-strength cap ρmax
1.0
1.0
Appendix
Table 15: Fixed default control parametrization (grid: T=104 ). The EEG head uses a reduced estimator range and shorter horizons for the short evaluation streams, hand-set once and disclosed here; no per-configuration search was performed for any SPAR method.
Research Center for Computer Science and Information Technologies, Macedonian Academy of Sciences and Arts, Bul. Krste Misirkov 2, 1000 Skopje, Macedonia
School of Science for Open and Environmental Systems, Graduate School of Science and Technology, Keio University, Japan · Department of System Design, Faculty of Science and Technology, Keio University, Japan