# Fixing Hagan SABR Low-Strike Negative Density (2026)

> Eliminate Hagan SABR negative probability densities in long-dated low-strike swaptions and CMS replication using Antonov free-boundary and Hagan PDE fixes.

Canonical: https://volsurfacedesk.com/desk-notes/sabr-low-strike-density-breakdown-antonov-free-boundary
Last updated: 2026-11-14T19:49:21.983Z

**Publication:** VolSurface Desk
**Category:** Svi Sabr Smile Arbitrage Free
**Published:** 2026-11-14T13:45:25.915Z
**Author:** Dr. Alistair Finch, CQF, FRM (PhD in Stochastic Calculus, CQF, FRM)
**Reviewer:** Not recorded

## Answer-first summary

Hagan’s asymptotic SABR formula produces negative risk-neutral densities at low strikes when nu^2 * T > 0.25, mispricing CMS convexity unless free-boundary fixes are used.

## Key takeaways

- Hagan’s 2002 SABR formula is a small-time WKB perturbation series in epsilon = nu^2 * T that breaks down at low strikes whenever nu^2 * T exceeds 0.25.
- Uncorrected Hagan SABR produces negative Breeden-Litzenberger probability densities and total probability mass exceeding 1.0, inflating 10Y10Y CMS convexity adjustments by 14–29 bps.
- Antonov Free-Boundary SABR (0.15 ms) and Hagan’s 2014 1D Effective Forward PDE (0.80 ms) restore exact probability conservation and <0.35 bps accuracy vs Monte Carlo out to 30-year tenors.
- Tier-1 model governance enforces automatic routing from analytic WKB to arbitrage-free PDE/quadrature whenever nu^2 * T > 0.20.

## Related financial entities

- **CME SOFR Options & Interest Rate Swaption Normal (Bachelier) Volatility Grid** (SOFR-SRO-SWAPTIONS) — product: CME SOFR Options & Interest Rate Swaption Normal (Bachelier) Volatility Grid (SOFR-SRO-SWAPTIONS) refers to an institutional financial benchmark and regulatory standard monitored by CME Group / ICAP Swaption Matrix (US), currently calibrated at 98.5 bps (Annualized Normal Vol (1Y10Y)) as of October 2026.
- **Over-the-Counter SPX Variance Swaps vs. Corridor Variance & Vega Notional** (VARIANCE-SWAPS) — product: Over-the-Counter SPX Variance Swaps vs. Corridor Variance & Vega Notional (VARIANCE-SWAPS) refers to an institutional financial benchmark and regulatory standard monitored by ISDA Equity Derivatives Definitions (INTL), currently calibrated at σ2 - K2 (Variance Payoff Scaling) as of October 2026.
- **SABR Stochastic Volatility vs. Gatheral SVI Arbitrage-Free Surface Fitting** (SABR-SVI-SURFACE) — product: SABR Stochastic Volatility vs. Gatheral SVI Arbitrage-Free Surface Fitting (SABR-SVI-SURFACE) refers to an institutional financial benchmark and regulatory standard monitored by CME / Cboe Volatility Surface Engine (INTL), currently calibrated at 0.0 bps (Butterfly & Calendar Arb Floor) as of October 2026.

## Analysis

Price a 10-year into 20-year Constant Maturity Swap (CMS) caplet by integrating receiver swaptions down to 0.00% strike under Hagan's standard 2002 SABR formula (ν = 0.42, T = 10), and the replication integral accumulates spurious negative option convexities that overstate CMS convexity adjustments by **14 to 29 basis points**. Why? Because Hagan's closed-form formula is an O(ν2T) WKB singular perturbation expansion around small time—not the exact solution to the SABR stochastic differential equation! When dimensionless expansion parameter ε = ν2T exceeds 0.25 and strike K drops toward zero, Hagan's implied volatility rises too steeply, pushing ∂2P / ∂K2 < 0. Inside our [SVI & SABR Volatility Surface Calibration & No-Arbitrage Conditions series](/desk-notes/svi-sabr-volatility-surface-arbitrage-free-calibration-guide), resolving low-strike SABR breakdown connects [Hagan normal swaption calibration](/desk-notes/hagan-sabr-lognormal-vs-normal-bachelier-swaption-vols) with [Durrleman butterfly positivity](/desk-notes/durrleman-condition-butterfly-arbitrage-density-positivity).

## Diagnosing Where Hagan's Perturbation Series Breaks Down

In Hagan's 2002 formula, implied volatility multiplies a spatial geodesic distance ratio z / x(z) by a linear-in-time correction [1 + Θ(K, F) · T]. For long expirations (T ∈ [5, 30] years) or elevated vol-of-vol (ν ≥ 0.45), the coefficient Θ(K, F) contains (1 − β)2α2 / [24 · (FK)^(1 − β)], which diverges toward +∞ as K → 0+ whenever β < 1! Even for Normal SABR (β = 0), Hagan's Bachelier volatility grows asymptotically without proper absorbing or free-boundary probability mass conservation, driving the Breeden-Litzenberger receiver put second derivative ∂2P / ∂K2 strictly negative across low strikes.

## Eliminating SABR Low-Strike Butterfly Arbitrage in Production

![Comparison of Hagan WKB perturbation breakdown versus Antonov Free-Boundary SABR and Hagan Arbitrage-Free PDE](/api/figure?type=flow&title=Eliminating%20SABR%20Low-Strike%20Butterfly%20Arbitrage%20in%20Production&subtitle=Three%20institutional%20upgrades%20replacing%202002%20first-order%20WKB%20expansions%20on%20long-dated%20swaption%20cubes&steps=1.%20Audit%20nu%5E2%20*%20T%3A%3AFlag%20any%20swaption%20expiry%20slice%20where%20expansion%20parameter%20nu%5E2%20*%20T%20exceeds%200.20%7C2.%20Zero-Correlation%20Kernel%3A%3AAntonov%20et%20al.%20%282015%29%20integrate%20exact%202D%20hyperbolic%20McKean%20heat%20kernel%20for%20rho%20%3D%200%7C3.%20Correlation%20Mapping%3A%3AProject%20general%20rho%20!%3D%200%20SABR%20onto%20effective%20zero-correlation%20mimic%20parameters%7C4.%20Hagan%201D%20PDE%20Alternative%3A%3AOr%20solve%20Hagan%20%282014%29%201D%20effective%20forward%20Kolmogorov%20PDE%20on%20150%20grid%20nodes%20in%200.8%20ms&metrics=BREAKDOWN%20TRIGGER%3A%3Anu%5E2%20*%20T%20%3E%200.25%3A%3AE.g.%20T%3D5Y%2C%20nu%3D0.40%20%3D%3E%200.80%7CHAGAN%202002%20MIN%20PDF%3A%3A-0.184%3A%3A10Y10Y%20Strike%20K%20%3D%200.65%25%7CANTONOV%20MIN%20PDF%3A%3A%3E%3D%200.0000%3A%3AGuaranteed%20Positive%20Density%7CCMS%20BIAS%20REMOVED%3A%3A14%20-%2029%20bps%3A%3A10Y20Y%20Replication%20Integral)

## Three Arbitrage-Free Replacements Used on Tier-1 Rates Desks

Modern quantitative libraries on [CME Group](/counterparties/cmeg) and interbank OTC desks replace Hagan (2002) at long tenors with one of three exact-density engines:
1. **Hagan et al. (2014) Arbitrage-Free SABR 1D Effective PDE**: Derives the exact 1D effective forward Kolmogorov equation ∂_T Q = (1/2) ∂_FF [D2(F) · Q] for probability density Q(T, F) along with an explicit boundary mass Q_L(T) absorbed at F = 0 (or F = −s). Solved on a 150-point tridiagonal grid in 0.8 ms, it matches Hagan (2002) to 0.05 bps near ATM while guaranteeing Q(T, K) ≥ 0 everywhere!
2. **Antonov, Konikov & Spector (2015) Free-Boundary SABR**: Removes any artificial barrier at F = 0 by letting F_t ∈ (−∞, +∞) evolve as dF_t = α̂_t |F_t|^β dW_t1 and computing a fast 1D quadrature over the exact McKean hyperbolic heat kernel—preserving near-closed-form speed (0.15 ms per smile) with strictly non-negative density.
3. **Andreasen-Huge (2011) ZABR Single-Step Finite Difference**: Solves a single implicit backward Euler step (I − 0.5·ΔT·L) C = (F − K)+ that reproduces SABR smiles while guaranteeing discrete butterfly positivity across every strike node on the mesh.

```formula
ε = ν2 · T | CMS Forward = F0 + (1/A0)·[ ∫0^(F0) w_rec(K)·P(K) dK + ∫_(F0)^∞ w_pay(K)·C(K) dK ]
- Stress Case (10Y10Y SOFR Swaption: F0 = 3.85%, T = 10.0 yr, α = 0.039, β = 0.50, ρ = −0.18, ν = 0.38): ε = 0.382 × 10 = 1.444 >> 0.25.
- Low-Strike Audit at K = 0.50%: Hagan (2002) outputs σ_N = 148.6 bps with ∂2P/∂K2 = −0.142 < 0 (Negative PDF and total mass 1.092!).
- Antonov Free-Boundary & Hagan 1D PDE Output at K = 0.50%: σ_N = 94.2 bps with ∂2P/∂K2 = +0.068 > 0 (Exact total mass 1.0000).
- CMS Convexity Impact: Hagan (2002) +48.7 bps vs Arbitrage-Free SABR PDE +29.4 bps (eliminates +19.3 bps phantom receiver convexity).
```

## Bridging Swaption Cubes to Equity and Variance Books

Why does this rates-originated SABR pathology also matter to cross-asset desks pricing [OTC variance swaps](/instruments/variance-swaps-over-the-counter-spx-variance-swaps-vs-corridor-variance-vega-notional) and long-dated equity hybrids? Because variance swaps integrate 1/K2 put weights all the way toward zero strike—identical mathematically to receiver CMS replication weights w_rec(K)! Any low-strike density breakdown corrupts log-strip integrals immediately, which is why cross-asset quant teams test option smiles simultaneously in our [Black-Scholes Option Calculator](/calculators/black-scholes-calculator) and [SABR vs SVI Surface Benchmark](/instruments/sabr-svi-surface-sabr-stochastic-volatility-vs-gatheral-svi-arbitrage-free-surface-fitting).

![Comparison of 10Y10Y swaption receiver normal volatility and density accuracy across four SABR solvers](/api/figure?type=bars&title=10Y10Y%20Receiver%20Wing%20Accuracy%20%28epsilon%20%3D%20nu%5E2%20*%20T%20%3D%201.44%29&subtitle=Benchmarking%20Hagan%202002%2C%20Obloj%202008%2C%20Antonov%20Free-Boundary%2C%20and%20Hagan%202014%201D%20PDE%20against%20Monte%20Carlo&steps=Hagan%20%282002%29%20WKB%3A%3A54.4%20bps%20wing%20error%20at%20K%3D0.50%25%3B%20negative%20PDF%20below%20K%3D1.15%25%20corrupts%20CMS%20pricing%7CObloj%20%282008%29%20Geodesic%20Fix%3A%3AFixes%20beta-rho%20log%20inconsistency%20at%20high%20strikes%20but%20still%20breaks%20down%20for%20nu%5E2*T%20%3E%200.6%7CAntonov%20Free-Boundary%3A%3AWithin%200.35%20bps%20of%201%2C000%2C000-path%20MC%20across%20all%20strikes%20with%20strictly%20positive%20PDF%7CHagan%20%282014%29%201D%20PDE%3A%3AExact%20absorbing-mass%20conservation%20at%20boundary%3B%200.22%20bps%20max%20error%20in%200.8%20ms%20runtime&metrics=WKB%200.5%25%20STRIKE%20ERR%3A%3A%2B54.4%20bps%3A%3ASevere%20Wing%20Overstatement%7CANTONOV%20MAX%20ERR%3A%3A0.35%20bps%3A%3AVs%201M-Path%20Benchmark%20MC%7C1D%20PDE%20RUNTIME%3A%3A0.80%20ms%3A%3A150-Node%20Tridiagonal%20Solve%7CTOTAL%20PDF%20MASS%3A%3A1.0000%3A%3AExact%20Martingale%20Conservation)

| SABR Engine | Max ν2T Horizon | Low-Strike Density q(K) | Whole-Slice Speed | Production Use Case |
| --- | --- | --- | --- | --- |
| Hagan (2002) Analytic | ≤ 0.20 (T ≤ 1Y) | Negative for K < 1.2% long-dated | 0.005 ms | Short-dated [CME SOFR Options](/instruments/sofr-sro-swaptions-cme-sofr-options-interest-rate-swaption-normal-bachelier-volatility-grid) |
| Obloj (2008) Fine-Tuned | ≤ 0.45 (T ≤ 2Y) | Improved mid-wing; fails deep tail | 0.006 ms | Medium-tenor FX & commodity caps |
| Antonov Free-Boundary | Unlimited (30Y+) | Strictly ≥ 0 analytically | 0.15 ms | Negative-strike floors & [Shifted SABR](/desk-notes/shifted-sabr-negative-rate-floor-calibration-stability) |
| Hagan (2014) 1D PDE / ZABR | Unlimited (30Y+) | Strictly ≥ 0 + boundary atom | 0.80 ms | CMS replication & Bermudan swaptions |

## Model Validation Checkpoints Under FRTB

Independent model validation units reporting under [BaFin](/regulators/bafin), [FINMA](/regulators/finma), and [CFTC](/regulators/cftc) rules require automatic solver switching inside [SVI & SABR Volatility Surface Calibration & No-Arbitrage Conditions](/topics/svi-sabr-smile-arbitrage-free): whenever ν2T > 0.20, quoting and risk engines automatically route SABR parameters into Hagan's 2014 1D PDE or Antonov's quadrature kernel rather than evaluating raw 2002 expansions.

## Frequently asked questions

### Why did Obloj’s 2008 correction not fully solve the Hagan SABR low-strike negative density problem?

Obloj (2008) fixed a leading-order inconsistency in Hagan’s geodesic distance term x(z) when beta != 1 and strikes are far OTM, eliminating artificial flattening in high wings, but retained the first-order [1 + Theta * T] time expansion that still diverges when nu^2 * T is large.

### How can total risk-neutral probability mass exceed 100% under Hagan’s 2002 formula?

Because digital put price equals dP/dK, non-monotone digital prices (negative density d^2P/dK^2 < 0 in low strikes followed by boundary spikes near K -> 0) cause the numerical integral of q(K) across positive strikes to oscillate above 1.0.

### What is the physical difference between Absorbing SABR (Hagan 2014) and Free-Boundary SABR (Antonov 2015)?

Absorbing SABR freezes F_t at zero (or -s) once hit, accumulating a discrete Dirac delta probability atom at the boundary, whereas Free-Boundary SABR replaces F^beta with |F|^beta so rates pass smoothly through zero with a continuous density everywhere.

### Does Andreasen-Huge ZABR run fast enough for real-time swaption cube recalibration?

Yes. Because Andreasen-Huge solves a single backward implicit Euler tridiagonal system per expiration slice, calibrating an entire 12-expiry by 10-tenor swaption cube takes less than 40 milliseconds on a single CPU core.


## Source URLs

- https://www.cmegroup.com/trading/interest-rates/stir/sofr-options.html
- https://www.bis.org/bcbs/publ/d457.htm
- https://www.esma.europa.eu/publications-and-data/interactive-single-rulebook/emir

## Citation

- APA: Dr. Alistair Finch, CQF, FRM. (2026). Fixing Hagan SABR Low-Strike Negative Density via Free-Boundary Approximations. VolSurface Desk. https://volsurfacedesk.com/desk-notes/sabr-low-strike-density-breakdown-antonov-free-boundary
- Canonical article: [Fixing Hagan SABR Low-Strike Negative Density via Free-Boundary Approximations](https://volsurfacedesk.com/desk-notes/sabr-low-strike-density-breakdown-antonov-free-boundary)
- Site research index: [VolSurface Desk](https://volsurfacedesk.com/llms-full.txt)
