When Wasserstein Distances Deceive: Landscape Mismatches in Topological Summaries

I watched a staff burn two months chasing a false signal. They had computed persistence landscape from lone-cell RNA data, then used Wasserstein distance to cluster cell types. The clusters looked beautiful — until someone checked the raw diagram. Nearly all the Wasserstein distance came from one noisy feature dimension, drowning out the biological signal. That's when I started collecting postmortems. This article is that collection: five blocks where Wasserstein distance deceive, why they happen, and what you can do about it.

Where This Bites: Three Real-World Ambushes

A community mentor says however confident you feel, rehearse the failure case once before you ship the adjustment.

Shape matching with noisy outliers

I once watched a staff compare two 3D hand scans — one clean, one with a lone spike artifact near the thumb. Wasserstein distance between their persistence diagram: 0.04. Almost identical, by that metric. The landscape told a different story: a 32% mismatch in the second-homology peak. That spike wasn't harmless. It buried the true topological difference under a cheap transport spend. The algorithm ranked the clean scan as a near-match to the noisy one. off group. The catch is that Wasserstein treats the outlier as just another point to shift, not as a structural contaminant. When your outlier count hits 5% of the total diagram point, expect false similarity scores to inflate by 15–20%. I have seen this sink automated craft control pipelines more exact once — one phase was enough.

Sensor network anomaly detection

Thirty temperature sensors across a warehouse floor. The persistence diagram for a normal day shows two concentrated clusters — the cool zone and the warm zone. A refrigerant leak shifts one sensor into a third regime. Wasserstein distance between normal and leak diagram? 0.11. Below the alarm threshold. The crew had tuned that threshold using synthetic outliers, not real creep. What broke primary: the landscape's L2 norm jumped 2.3× while Wasserstein barely twitched. That hurts.

'A metric that cannot see a one-off sensor going rogue is not a metric — it is a liability.'

— field engineer, after the third false negative

The geometry explains why: Wasserstein aligns point by mass transport, so one stray point pairs cheaply with a far cluster. The landscape, by contrast, stacks func values at every birth-death sync. One outlier inflates an entire local peak. The trade-off is clear. You trade sensitivity for stability. But when you volume to catch the one failing sensor, stability is the enemy.

lone-cell genomics group effects

Here the deception cuts deeper. Two batches of cells, same tissue, different sequencing runs. Wasserstein distance between the topological summaries: 0.07. Looks great, correct? The researchers merged the datasets. Then the clustering fell apart — group-specific loops appeared that had zero biological meaning. The landscape mismatch was 41%. Why? Wasserstein averages over all feature. A compact systematic shift in 200 genes gets washed out against 20,000 stable ones. The landscape, however, captures the distribual of those shifts across persistence scales. That modest lot effect concentrates at short lifetimes — exact where biological signal also lives. Not yet separable. Most group skip this diagnostic. They compute Wasserstein, see a low number, and assume the batches are harmonized. They are not. The metric lied, and the downstream analysis paid the price.

What Everyone Gets off: Landscape Geometry vs. Diagram Geometry

The Geometry You Think You Have

Most group treat a persistence landscape like a tidy bag of numbers—a curve, a funcing, something you can feed into an Lp Wasserstein metric without thinking twice. The diagram however lives in a plane: birth on x, death on y. That is a point cloud, not a signal. When I primary saw a staff shove landscape vectors into a Wasserstein-2 computation and call it 'topological distance,' I knew we had a geometry issue. The diagram respects the diagonal—point near y=x matter less, they die fast. landscape flatten that priority into a bump func. off sequence. The Wasserstein distance on diagram measures minimal effort to shift one multiset of point to another, with point allowed to slide into the diagonal at zero overhead. On a landscape? You are measuring the Lp distance between two piecewise-linear func. That is a different universe.

Landscape Encoding and Information Loss

Here is the catch: a persistence landscape compresses a 2D point set into a sequence of layered 1D funcing. The initial layer captures the largest loops, the second layer the next, and so on. That sounds elegant until you realize you have thrown away the correlation between birth and death timing. A point (1.0, 2.0) and a point (10.0, 11.0) produce more exact the same landscape—a bump of height 0.5. The metric cannot tell if the feature was a short-lived noise at tight growth or a durable structure at substantial capacity. That hurts. I have watched analysts compare two landscape, get a distance near zero, and declare the underlying diagram 'similar.' Meanwhile the diagram had one outlier far out on the birth axis. The landscape metric punished the shape mismatch, not the align mismatch.

A Wasserstein distance on landscape measures funcal shape. A Wasserstein distance on diagram measures point displacement. Those are different metrics wearing the same name.

— contrast drawn from a debugging session where a 0.3 landscape distance hid a 12.4 diagram distance

Why Wasserstein Distance on landscape Is Not Wasserstein Distance on diagram

The math cuts clean. On a persistence diagram, the p-Wasserstein distance is the solution to an optimal transport issue between two point sets, plus penalty for unmatched point snapped to the diagonal. On a landscape, the Lp Wasserstein distance reduces to an integral of absolute differences between two func raised to p, integrated over t. That is an Lp norm—not an optimal transport overhead. Many group skip this: the name 'Wasserstein' seduces them into thinking they are comparing topologies. They are comparing align curves. If your pipeline computes Wasserstein on landscape and expects diagram-like discrimination, you lose a day every slot a medium-lived feature gets buried under a long-lived one. What usually breaks primary is sensitivity to noise. diagram naturally handle noisy short bars near the diagonal—the transport spend is low. landscape amplify those short bars into cusps that shift the whole integral. One noisy point can spike the landscape distance by 0.5 while the true diagram distance stays under 0.1.

Worth flagging—you can craft this effort if your feature are uniformly scaled and your landscape layers are truncated. But most group do not truncate. They take the primary five layers and call it done. That is not Wasserstein on diagram. That is a fragile heuristic dressed in borrowed math. The real fix? Match your metric to your encoding. If you insist on landscape, use Lp distance and admit you left the diagram geometry behind. If you orders the diagonal sliding property, retain the point. Do not cross the streams.

repeats That Usually Work — When the Metric Behaves

A shop-floor trainer explained that the pitfall is treating symptoms while the root cause stays in the checklist.

Low-noise, low-dimensional settings

Wasserstein distance on persistence landscape behave themselves when the data is clean and the ambient area is compact. I mean genuinely modest — embedding dimension under ten, birth-death pairs that cluster tightly rather than smear across the diagram plane. In these conditions the landscape representation preserves the metric structure nearly intact. The p=2 Wasserstein distance between two landscape will track the usual diagram distance within three to five percent relative error, based on what my own group have observed across roughly forty synthetic benchmarks. The catch: this only holds when the noise floor sits below 0.1 times the persistence of your strongest feature. Cross that threshold and the landscape starts warping the geometry — not catastrophically at initial, but the creep is measurable.

What usually breaks primary is the ordering of distance. Two diagram that are closer than a third in pure Wasserstein area may flip positions in landscape area once the birth-death coordinates scatter. Worth flagging — this rarely matters if you only require relative rankings for a tight set of classes (say, three vs. five). But for nearest-neighbor search or clustering, those flips compound.

Dominant topological feature

When your data has one or two feature that absolutely tower over the rest — a persistent H₁ loop that lives three times longer than the next best bar — the Wasserstein-landscape combination stabilizes. The tall feature anchors the distance calculation; the rest become rounding error. I have seen this hold in materials-science pore networks where one dominant channel accounts for seventy percent of the total persistence mass. The landscape essentially collapses to a comparison of that lone peak, and the Wasserstein distance behaves like a univariate difference — predictable, monotonic, boring in a good way.

The trade-off arrives when your so-called dominant feature shares a align range with medium-strength survivors. Then the landscape superposition smears the peak height, mixing which topological attribute contributes what. Most group skip this check: they compute the landscape norm, declare one feature dominant, and proceed. But dominance must be measured in landscape amplitude — not in raw persistence. A bar that lives long but lies far from the diagonal can inject more landscape mass than a short fat bar near the diagonal.

One concrete anecdote: a collaborator's cytometry dataset had three persistent loops. The longest-lived one (birth 0.2, death 0.8) dominated the persistence diagram but contributed less to the landscape L² norm than a shorter feature at coordinates (0.4, 0.9). Why? Because the landscape funcal encodes not just lifespan but the slope of the tent func, and that second feature's vertical asymmetry inflated its integral. The Wasserstein distance on landscape then over-weighted the shorter feature — more exact the off signal for the biological question.

Pre-filtered or denoised diagram

This is the block that actually works in manufacturing settings. Filter the diagram before building landscape: remove point below a persistence threshold, cluster birth-death pairs into barycenters, or apply a Gaussian blur to the landscape itself. The histogram of effective errors from our internal regression benchmarks shows a sharp drop once you filter out point with persistence less than 0.15 × max_persistence. Below that threshold the landscape geometry distorts the Wasserstein distance by up to forty percent. Above it, the mismatch falls to lone-digit percentages.

Filter primary, landscape second. The metric only returns honest distance when the diagram is surgically clean.

— rule of thumb from a structural-biology group that ran 200+ benchmark trials, shared during a workshop sidebar

But filtering too aggressively kills sensitivity. If you cut at 0.3 times max persistence, you will miss feature that separate two classes by a narrow margin — the very templates topological data analysis is supposed to catch. The sweet spot around 0.12–0.18 emerges from cross-validated loss curves, not from theory. That said, there is a mechanical reason: the landscape Lp topology and the diagram p-Wasserstein topology coincide when both spaces are restricted to a compact set of point whose persistence exceeds a uniform bound. Below that bound, the mapping becomes merely Lipschitz, not bi-Lipschitz — distance compress or expand unpredictably. Denoising forces the bound upward, restoring metric faithfulness.

One last detail. The denoising must happen before landscape construction, not after. Smoothing the landscape kernel post hoc removes high-frequency oscillations but leaves the structural bias intact. That hurts more than it helps. Fix the diagram initial — then let the landscape do what it does best, which is to embed persistence into a Hilbert area where you can compute means and variances. off sequence. Not yet. Do the hard pruning upfront and the Wasserstein distance will cooperate for weeks of subsequent analysis.

Anti-Patterns: Six Ways group Get Burned

Ignoring feature dimension imbalance

compact persistence often hides big trouble. I watched a materials-science staff compare two PDs: one with a one-off dominant H₁ loop, the other with five tiny ones. Wasserstein-2 scored them nearly identical — 0.07 apart. The catch: their landscape norms differed by 7×. The crew had normalized feature counts, not dimensions. That second diagram carried five noisy feature, each barely alive, yet the metric treated them as one big loop plus scattered background. Diagnostic: compute sum of lifetimes per homological dimension separately. If ratios exceed 2:1 but Wasserstein says 'identical,' your distance are hallucinating.

Using landscape as drop-in replacements

Slap a landscape transform on raw diagram coordinates and plug it into a vector pipeline — that's the transition. off queue. landscape reweight point positions by the vertical stacking of piecewise-linear tent funcal: a topological feature's influence spreads laterally at a constant slope, so a high-persistence point dominates far beyond its own birth–death window, notes a computational-geometry colleague who debugged an ROC plot that flatlined. Most group skip this: a landscape isn't a distance-preserving map. Two diagram far apart in constraint area can collapse into nearly identical L² integrals if their tall feature align vertically. The staff I mentioned earlier switched from sliced Wasserstein to landscape L₂ and saw their validation accuracy drop 12 point overnight. Trade-off — you gain differentiability, you lose discriminative power. Don't swap unless you've tested both on your specific null distribu primary.

Trusting p-values from permutation tests

Permutation tests assume exchangeability under the null. That breaks when your landscape have built-in boundary artifacts from finite sampl. One staff permuted group labels across 500 subjects, computed Wasserstein between landscape vectors, and reported p = 0.003. Beautiful result. They ran it five times — p-values danced between 0.001 and 0.19. The culprit? landscape truncated at the maximum death value introduced correlation across permutations. The trial was permuting dependent units. Indicator: your p-value distribu under null isn't uniform.

Here is a concrete fix — we started using block permutations that preserve diagram-level structure. Even then, monitor the tail of the permuted distribuing. If it spikes near zero, your check statistic is leaking information from the landscape's implicit boundary. That hurts.

Sixth anti-template: ignoring death-tail artifacts

Short paragraph this slot. landscape decay linearly from the diagonal. feature near death = 0.1 get the same slope as a loop lasting ten units. So noisy low-persistence point stretch horizontally just as far as signal — they just sit lower. Result: Wasserstein-1 on landscape is dominated by thousands of near-diagonal points. You lose a day debugging separation that isn't there.

Maintenance creep: How landscape Degrade Over Repeated Use

An experienced technician says the trade-off is speed now versus rework later — most shops lose on rework.

Cumulative noise from multiple diagram extractions

You run the pipeline once, the landscape looks clean. You run it again, same data, same parameters—a slight shift. By the hundredth iteration, your Wasserstein distance have drifted by 12% and nobody noticed. This is not random. landscape constructed from persistence diagram accumulate boundary artifacts each slot you extract them—tiny coordinate wobbles from floating-point rounding in vertex assignments, triangle flips in mesh retesselation, or solver non-determinism in gradient flows. Extracted alone, each artifact is harmless. A pixel off. A birth-death pair nudged by 0.003. But manufacturing pipelines re-extract diagram dozens of times per deployment: model retraining, data refresh cycles, A/B experiment rollbacks. Every extraction stamps fresh noise into the landscape grid. I have seen group chase false positives in their slippage monitoring dashboard for three weeks before realizing the metric itself was slowly inflating.

Sensitivity to samplion density changes

The real ambush: your data distribual stays constant, but your sampl rate changes. A sensor that once recorded 10k points per lot now records 8k due to a hardware throttle. The Wasserstein distance between two landscape jumps—not because the topology changed, but because the landscape construction algorithm interprets sparser sampl as structural loss. Birth-death pairs get pulled apart in the metric area when they should stay put. Most group skip this: they validate their pipeline on fixed-sample benchmarks, then deploy into environments where samplion density fluctuates naturally. The catch is that landscape Wasserstein distance punishes density variation more harshly than diagram-level distance do—it integrates over a continuous funcal that becomes increasingly jagged as points thin out. One concrete anecdote: a crew at a logistics venture watched their wander metric alarm every Tuesday evening for two months. Tuesday was the day their upstream data source ran batch compressions. Not a topological shift. A sampling artifact.

'We were monitoring landscape slippage as a proxy for data finish. Turned out the proxy was hallucinating the quality issue.'

— principal engineer, industrial IoT monitoring stack, after unwinding six false incident pages

expense of re-landscaping after pipeline updates

Recomputing the landscape after every code adjustment is expensive—not computationally impossible, but organizationally draining. A modest update to your filtration strategy, say switching from Vietoris-Rips to alpha complexes, requires rebuilding every historical landscape from scratch to maintain comparability. You cannot mix landscape from different filtration types in the same Wasserstein comparison pipeline; the results become meaningless. So group defer updates. They patch around the old landscape representation instead of fixing it. The metric degrades slowly, imperceptibly, like a seam pulling apart under repeated stress. Worth flagging: this maintenance drift is not visible in any lone comparison—it emerges across the timeline of comparisons. The primary re-landscape expenses one afternoon. The tenth expenses a sprint. By the phase the crew decides to drop Wasserstein altogether, they have sunk weeks into preserving compatibility with a representation they no longer trust. That hurts. The next section addresses when you should just let the metric go.

When to Drop Wasserstein Altogether

When constraint distance wins

The simplest case for dropping Wasserstein: your analysis only cares about the largest feature. Not the cluster of medium-sized loops, not the distribution of compact cavities—just the lone most persistent structure. constraint distance looks at exact that: the supremum over matched points. It is brutally plain. I have watched group wrestle with Wasserstein landscape for weeks, tweaking bin sizes and smoothing parameters, when every question they actually asked was constraint-answerable. The catch: constraint is brittle under noise. A one-off outlier point—one spurious feature that barely lives—can shift the whole distance if it aligns badly. But if your data has strong, dominant signals and you orders speed, constraint runs in near-linear slot on persistence diagram.

'We switched from Wasserstein-2 to constraint after three failed peer reviews. The results barely changed—but the review cycle did.'

— Data scientist at a biotech startup, internal post-mortem, 2023

Trade-off: constraint is stable under modest perturbations of large feature, but blind to the bulk of the diagram. Use it when comparing two noisy point clouds where only the top homology group matters—say, detecting a solo persistent void in a CT scan. off sequence: reaching for constraint when you demand a metric that respects the full geometry of medium-ranked feature. That hurts.

Sliced Wasserstein for high-dimensional persistence

Persistence diagram live in the plane—birth versus death—but modern topological summaries often push into higher dimensions: persistence landscape, persistence images, or even learned coordinates via neural nets. Standard Wasserstein on raw landscape scales poorly here. The distance computation becomes a linear programming headache. Enter sliced Wasserstein: project both landscape onto random lines, compute one-dimensional Wasserstein on those projections, and average. It stabilizes variance. More importantly, sliced Wasserstein is differentiable—you can backprop through it for topological loss function in deep learning. I have seen groups adopt this after their landscape Wasserstein pipeline took twelve hours per hyperparameter sweep. Sliced brought it to forty minutes.

The anti-pattern: sliced Wasserstein loses exactness for unbalanced landscape—if one summary has many more dominant feature than the other, random projections smear the mismatch. Use it when you have hundreds of high-dimensional persistence vectors and require a fast, smooth, and stable proxy. Not for precise pairwise comparisons where the exact geometry of peaks matters. Most groups skip this: they hold using vanilla Wasserstein on landscape without checking whether the landscape representation itself is the constraint—pun intended.

Kernel methods for stability

Sometimes the right shift is not a distance at all. Kernel methods—like the persistence scale-area kernel or the sliced Wasserstein kernel—map landscape implicitly into reproducing kernel Hilbert spaces where inner products swap distance. Why would you do this? Two reasons. initial, kernels can be made provably stable to noise under the same perturbations that break Wasserstein on landscape—explicitly controlled via the kernel bandwidth. Second, you get access to Gaussian process regression, support vector machines, or kernel PCA directly on topological summaries. We fixed a persistent misalignment in a materials-science workflow by swapping Wasserstein for a Laplacian kernel on persistence images. The Wasserstein results had been drifting across experimental batches; the kernel method held at a fixed probe error for six months.

The trade-off is interpretability. A Wasserstein distance between landscape tells you how much you would call to deform one summary into another. A kernel similarity score? It is a number in a black-box feature area. That said, if your downstream task is classification or regression, and the metric is just a component in a larger model, kernel methods often outperform Wasserstein with less hyperparameter tuning. One rhetorical question: do you actually require to interpret the distance, or do you just require it to separate classes correctly? If the latter, drop Wasserstein. Replace it with a kernel that matches your noise model. Not yet convinced? Run both on a held-out validation set. The difference is often stark.

Open Questions and Reader FAQ

An experienced operator says the trade-off is speed now versus rework later — most shops lose on rework.

Can we fix the landscape mismatch?

Yes — but the fix costs you more than most groups expect. You can slap a multiscale smoothing on the persistence landscape, or embed it into a kernel that reweights slopes. That sounds fine until you realize every smoothing shift erases the very signal you were trying to rank. I have seen one group spend three months tuning a landscape kernel only to discover raw constraint distance on the original diagram outperformed their smoothed version. The trade-off is brutal: you either accept Wasserstein's geometric blind spots or you throw out the metric's computational speed. There is no free lunch — the landscape's vertical geometry is structurally different from the diagram's horizontal pairs, and no straightforward preprocessing bridge fully reconciles them.

What about entropy-regularized Wasserstein?

Entropy regularization smooths the optimization landscape — it makes Wasserstein differentiable and faster to approximate. The catch is that it also blurs the very persistence pairs you care about. Sinkhorn iterations redistribute mass between diagram in ways that can hide the topological birth-death discontinuities. One concrete anecdote: a colleague ran entropy-regularized Wasserstein on two functionally identical point-cloud samples and got a distance that was 40% larger than between a clean sample and one with injected random noise. The regularization was smoothing away the topological structure and measuring geometric noise instead. Entropy helps with convergence but it does not help with ranking truthfulness — it may actually make the deception worse.

Regularization is a patch, not a proof. You fix the solver, but the representation still lies.

— comment from a computational topology workshop, 2023

How to check if your Wasserstein ranking is trustworthy

Most units skip this: you need a calibration procedure. Take your full dataset of persistence diagram, compute the Wasserstein distance matrix, then deliberately perturb one diagram by adding a lone high-persistence point. Recompute. If the rank order of nearest neighbors shifts drastically, your Wasserstein ranking is brittle. The basic trial takes twenty minutes. I have run this on three real-world datasets — in two cases, adding one outlier point caused a 60% reshuffling of the top-ten nearest diagram. That hurts. The deeper open problem here is whether any metric on the space of persistence landscape can avoid this sensitivity while remaining computationally tractable. Nobody has a proof yet. Next time you publish a pipeline using Wasserstein, run the shake probe initial.

Next Experiments: Three Things to Try Tomorrow

Run a landscape alignment plot

Grab your two most recent persistence diagram—before and after a parameter change, or two runs on different data chunks. Compute their landscape representations. Then overlay the first two landscape functions on the same axes. What do you see? If the peaks align within a tight horizontal shift, the Wasserstein distance probably reflects real topological similarity. If the landscapes are vertically stacked like plates in a rack—same birth times, different death magnitudes—you are looking at a structural mismatch that the metric alone will hide. The catch: alignment plots expense almost nothing to generate. We fixed this in one project by running the plot before any formal distance computation. Saved us three days of chasing phantom dissimilarities.

Most groups skip this step. They compute Wasserstein, get a number, and move on. That hurts. A one-off visual check can reveal whether the metric is responding to noise or to genuine feature migration. Run the plot on three pairs, not just one.

Compute rank signature stability

Take a single diagram and generate ten noisy bootstrap replicates—add modest jitter to birth and death coordinates. Compute the Wasserstein distance between the original and each replicate. Now do the same for constraint distance. Rank the distances from smallest to largest for both metrics. Are the orderings consistent? Not always.

Worth flagging—Wasserstein tends to smooth out modest fluctuations; constraint catches them sharply. When the rank signatures diverge, your metric is unstable. The practical impact: you cannot trust downstream clustering or classification that depends on relative distances. I have seen groups commit weeks of tuning to a pipeline that collapsed once they ran this basic stability check. The fix is not to abandon Wasserstein, but to know exactly where it wobbles. Compute rank stability on a holdout set before any serious modeling.

Three lines of code. One scatter plot. You will either relax or pivot.

— verbatim advice from a production MLOps staff, after their third metric misadventure

Compare with limiter distance on a holdout set

Here is the sharpest test. Divide your diagram collection into a training set A and a holdout set B. Compute both Wasserstein and limiter distances for every pair across the split. Train a simple classifier—k-nearest neighbors, no tuning—using each metric separately. Compare accuracy on the holdout set.

That sounds fine until you see the results flip. constraint might dominate when diagrams are sparse and features are few; Wasserstein can win when density is high and small shifts matter. The pitfall: teams often default to Wasserstein because it is 'richer.' Richer is not better when the geometry is wrong. If bottleneck achieves equal or better performance on the holdout set, you have a clear signal that the cost of transporting mass is adding noise, not insight. Drop Wasserstein for that dataset. Or keep both and stack them—but only after you measure.

Do this tomorrow. Not next sprint.

Shrinkage, skew, bowing, spirality, pilling, crocking, and color migration show up weeks after a rushed approval.

Silhouettes, darts, pleats, yokes, plackets, gussets, facings, and linings punish vague instructions during size runs.

Preproduction, top-of-production, inline, midline, final, and pre-shipment audits catch different classes of drift.

Thread cones, bobbin spools, needle kits, oil cartridges, cleaning brushes, and lint traps belong on distinct reorder triggers.