27  Memory Persistence Study

Hurst exponents and aquifer temporal memory

TipFor Newcomers

You will learn:

  • What “memory” means for an aquifer (how long past conditions affect the present)
  • How to measure persistence using autocorrelation and Hurst exponents
  • Why confined aquifers have longer memory than unconfined ones
  • What memory tells us about aquifer connectivity and storage

Aquifers have memory—today’s water level depends on what happened weeks, months, or even years ago. Long memory means the system responds slowly and stores information about past conditions. This chapter quantifies that memory.

27.1 What You Will Learn in This Chapter

By the end of this chapter, you will be able to:

  • Explain what “aquifer memory” means and how it can be quantified using autocorrelation and Hurst exponents.
  • Interpret ACF plots, persistence curves, and H values to distinguish between short‑memory and long‑memory behavior.
  • Relate memory metrics to physical properties such as confinement, hydraulic conductivity, and storage.
  • Use memory insights to reason about prediction horizons, drought recovery times, and appropriate management timescales.
Data loading functions initialized successfully

27.2 Introduction

How long does the aquifer “remember” past conditions? Understanding temporal memory determines predictability horizons, drought resilience, and optimal management timescales.

Source: Analysis adapted from investigation-26-aquifer-memory-persistence.qmd

27.3 Key Concept: Aquifer Memory

Note💻 For Computer Scientists

Hurst Exponent (H) quantifies long-range dependence:

H = 0.5: Random walk (no memory beyond immediate past) - Geometric Brownian motion - Exponential ACF decay - ARMA process

0.5 < H < 1.0: Persistent (long memory) - Fractional Brownian motion - Power-law ACF decay - ARFIMA process - Past trends continue

0 < H < 0.5: Anti-persistent (mean-reverting) - Past trends reverse - Oscillatory behavior

Practical meaning: - H = 0.6: Mild persistence, ACF decays slowly - H = 0.8: Strong persistence, correlations remain for many lags - Higher H = longer prediction horizon

Tip🌍 For Geologists/Hydrologists

Physical interpretation:

Short memory (H ≈ 0.5): - Unconfined aquifer - High hydraulic conductivity - Rapid response to recharge - Fast equilibration (days-weeks) - Example: Shallow sand/gravel

Long memory (H > 0.7): - Confined aquifer - Low hydraulic conductivity - Slow response to forcing - Long equilibration (months-years) - Example: Deep bedrock, clayey aquitards

Aquifer memory determines: - Recovery time from drought (18-24 months for H=0.68) - Prediction horizon (longer memory = more predictable) - Management timescales (short memory = reactive, long memory = proactive) - Drought buffering (wet years buffer against current dry)

27.4 Methods

27.4.1 Hurst Exponent Calculation

NoteUnderstanding Rescaled Range (R/S) Analysis

What Is It?

Rescaled Range (R/S) analysis, developed by Harold Hurst in 1951 while studying Nile River floods, is a method to detect long-term memory in time series data. It measures how the “range of fluctuations” grows as we look at longer time windows.

Why Does It Matter?

R/S analysis reveals:

  • How far system can deviate from mean before returning
  • How long deviations persist (memory timescale)
  • Whether simple models work (H = 0.5) or need long-memory models (H ≠ 0.5)

How Does It Work?

For each window size n:

  1. Subtract mean: Remove average level
  2. Cumulative sum: Track running total of deviations
  3. Find Range (R): Maximum deviation minus minimum (how far series wanders)
  4. Find Std Dev (S): Typical size of fluctuations
  5. Calculate R/S ratio: Normalized range

Plot log(R/S) vs. log(n) → slope = Hurst exponent

Physical Interpretation:

  • H = 0.5: R/S grows as √n (random walk) - no memory
  • H > 0.5: R/S grows faster than √n - system has memory (persistence)
  • H < 0.5: R/S grows slower than √n - system mean-reverts

Rescaled Range (R/S) Analysis:

  1. For time series X of length N
  2. Calculate cumulative deviations: \(Y(t) = \sum(X(i) - \bar{X})\)
  3. Calculate range: \(R(n) = \max(Y) - \min(Y)\) over windows of size n
  4. Calculate standard deviation: \(S(n)\)
  5. Plot \(\log(R(n)/S(n))\) vs \(\log(n)\)
  6. Slope = Hurst exponent H

Interpretation: \[ R/S \propto n^H \]

Alternative methods:

  • Detrended Fluctuation Analysis (DFA)
  • Wavelet-based estimation
  • Maximum Likelihood Estimation

27.5 Analysis Results

27.5.1 1. Autocorrelation Analysis

NoteUnderstanding Autocorrelation Function (ACF)

What Is It?

The Autocorrelation Function (ACF) measures how strongly a time series correlates with lagged (time-shifted) versions of itself. Developed as part of time series analysis in the 1920s-1940s by statisticians like G. Udny Yule and later formalized by Box & Jenkins (1970), ACF reveals temporal dependencies—how much today’s value depends on yesterday’s, last week’s, or last month’s values.

Historical context: ACF emerged from studying economic cycles and agricultural yields. Hydrologists quickly adopted it because groundwater systems naturally exhibit memory—today’s water level depends on past recharge, not just current conditions.

Why Does It Matter?

ACF analysis reveals:

  1. System memory: How long past conditions influence the present
  2. Predictability horizon: How far ahead we can forecast
  3. Aquifer type: Confined (long memory) vs. unconfined (short memory)
  4. Seasonal patterns: Peaks at lag = 12 months indicate annual cycles
  5. Model selection: Helps choose appropriate time series models (AR, MA, ARMA)

How Does It Work?

Mathematical definition:

\[ \rho(k) = \frac{\text{Cov}(X_t, X_{t-k})}{\text{Var}(X_t)} = \frac{\sum_{t=k+1}^{n}(X_t - \bar{X})(X_{t-k} - \bar{X})}{\sum_{t=1}^{n}(X_t - \bar{X})^2} \]

Where: - \(\rho(k)\) = autocorrelation at lag k - \(X_t\) = water level at time t - \(X_{t-k}\) = water level k time steps earlier - \(\bar{X}\) = mean water level

Step-by-step interpretation:

  1. Lag 0: Always = 1.0 (perfect correlation with itself)
  2. Lag 1: Correlation between today and yesterday (or this month and last month)
  3. Lag k: Correlation between today and k time steps ago
  4. Plot ACF vs. lag: See how correlation decays with time

What Will You See?

The figure below shows ACF plots for multiple wells. Look for how quickly the curves drop to zero—slower decay indicates longer aquifer memory.

How to Interpret:

ACF Pattern Physical Meaning Aquifer Characteristics Example
Rapid decay (1-3 lags) Short memory Unconfined, high K, shallow Sandy aquifer, rapid recharge response
Slow decay (6-12 lags) Moderate memory Semi-confined, moderate K Mixed sand/clay system
Very slow decay (>12 lags) Long memory Confined, low K, deep Regional confined aquifer
Oscillatory pattern Seasonal cycles Annual recharge-discharge Peak at lag 12 months = annual cycle
Exponential decay AR(1) process Simple persistence Most natural aquifers
Slow power-law decay Long-range dependence Fractional processes, high Hurst exponent Highly persistent systems

Significance testing:

  • Red dashed lines: 95% confidence bounds = ±1.96/√n
  • ACF values outside these bounds are statistically significant
  • ACF values inside bounds could be due to random chance

Key ACF metrics for groundwater:

  • ACF at lag 1 month: Immediate persistence (typically 0.7-0.95 for aquifers)
  • ACF at lag 12 months: Annual cycle strength (peak indicates seasonality)
  • e-folding time: Lag where ACF drops to 1/e ≈ 0.37 (characteristic memory timescale)
  • First zero crossing: Lag where ACF reaches zero (effective memory duration)

Autocorrelation measures how strongly a time series correlates with lagged versions of itself. For water levels:

  • High ACF at lag 1 month = this month’s level is strongly related to last month’s
  • High ACF at lag 12 months = annual cycles exist
  • Slow decay = long memory system

The figure below shows ACF plots for multiple wells. Look for how quickly the curves drop to zero—slower decay indicates longer aquifer memory.

(a) Autocorrelation function showing aquifer memory decay over time lags. Slow decay (values above red significance lines for many lags) indicates long memory systems; fast decay indicates short memory.
(b)
Figure 27.1
Note💻 Interpreting ACF Plots

Slow decay (red zone) = Long memory: - ACF remains significant (above dashed red lines) for many lags - Past values continue to influence present for 6-12+ months - Indicates persistent/confined aquifer behavior

Fast decay (drop to zero quickly) = Short memory: - ACF drops below significance threshold within 1-3 lags - Only recent past matters - Indicates unconfined/responsive aquifer

The dashed red lines are 95% confidence bounds - ACF values outside indicate significant correlation.

27.5.2 2. Persistence Metrics Over Time Lags

This visualization combines two related metrics: the ACF curve (blue) showing correlation at each lag, and a cumulative memory curve (orange) showing what fraction of total influence comes from progressively longer time windows.

Key question: At what lag does the aquifer “forget” its past? The green vertical line marks the effective memory duration—beyond this point, correlations are no longer statistically significant.

Tip🌍 Physical Meaning of Memory Duration

Effective memory = 18 months: - Aquifer “remembers” conditions from 1.5 years ago - Wet period in 2020 still influences levels in mid-2021 - Drought recovery takes 18-24 months after rain returns

Cumulative memory curve shows how much total influence comes from the past: - 50% at ~6 months = Half of current level explained by last 6 months - 90% at ~18 months = Most influence from last 1.5 years - After 24 months, additional past has minimal impact

This is confined aquifer behavior - slow response, long memory.

Important🔗 Connecting ACF Decay to Hurst Exponent

The Bridge Between Metrics:

ACF decay rate and Hurst exponent measure the SAME underlying property (memory/persistence) using different mathematical frameworks:

ACF Decay Pattern What It Looks Like Hurst Exponent Memory Type
Rapid decay Falls to 0 within 3-6 months H ≈ 0.50-0.55 Short memory (random walk)
Moderate decay Falls to 0 in 6-12 months H ≈ 0.55-0.65 Moderate persistence
Slow decay Takes 12-24 months to reach 0 H ≈ 0.65-0.75 Strong persistence
Very slow decay Still >0.2 after 24 months H ≈ 0.75-0.85 Long-range dependence

Why Both Metrics?

  • ACF is intuitive: “How long until the system forgets?”
  • Hurst is predictive: Directly indicates forecast horizon
  • Together they validate each other—if ACF shows slow decay but H ≈ 0.5, investigate data quality

For This Aquifer:

If you observe ACF remaining above 0.3 at lag=18 months, expect H ≈ 0.68-0.72. This means: - Past conditions influence present for 1.5+ years - Forecasts can extend further than typical weather-driven systems - Drought recovery takes longer than the drought itself

27.5.3 3. Hurst Exponent Distribution Across Wells

NoteUnderstanding the Hurst Exponent

What Is It?

The Hurst exponent (H), named after British hydrologist Harold Edwin Hurst (1951), is a single number (0 to 1) that quantifies long-term memory in a time series. Unlike ACF that measures correlation at specific time lags, H captures the overall persistence pattern across all timescales.

Why Does It Matter?

H reveals fundamental aquifer behavior:

  • Predictability: Higher H = more predictable (past trends continue)
  • Recovery time: Higher H = slower recovery from drought/pumping
  • Aquifer type: H indicates confinement and storage characteristics
  • Forecast horizon: Higher H = can use longer historical windows

How Does It Work?

H is calculated using Rescaled Range (R/S) analysis (explained in section 4). The value indicates:

  • H = 0.5: Random walk (no memory beyond immediate past)
  • H > 0.5: Persistent (trends tend to continue)
  • H < 0.5: Anti-persistent (trends tend to reverse)

What Will You See?

Two visualizations:

  1. Histogram: Distribution of H values across wells (shows aquifer variability)
  2. Scatter Plot: H vs. record duration (tests if estimate is stable)

How to Interpret:

H Value Classification Physical System Recovery Time
< 0.5 Anti-persistent Unusual - strong local forcing Fast (weeks)
0.5 - 0.6 Short memory Unconfined aquifer, high K 3-6 months
0.6 - 0.7 Moderate memory Semi-confined, moderate K 6-12 months
0.7 - 0.8 Long memory Confined aquifer, low K 18-24 months
> 0.8 Very long memory Highly confined, very low K >3 years

Key Insight: Most natural aquifers have H = 0.55-0.75. Values far outside this range suggest measurement issues, pumping interference, or unusual hydrogeology.

The Hurst exponent (H) provides a single number summarizing memory strength across all time scales. Unlike ACF (which shows correlation at specific lags), H captures the overall persistence pattern.

This two-panel figure shows:

  • Left: Distribution of H values across all analyzed wells (histogram)
  • Right: Relationship between H and record duration (do longer records give different estimates?)

Wells with H > 0.5 (above the red dashed line) exhibit persistent behavior where past trends tend to continue.

Hurst exponent distribution reveals aquifer memory characteristics


**Hurst Exponent Classification:**
- Anti-persistent (H < 0.5): 0 wells (0.0%)
- Short memory (0.5 ≤ H < 0.6): 0 wells (0.0%)
- Moderate memory (0.6 ≤ H < 0.7): 0 wells (0.0%)
- Long memory (H ≥ 0.7): 8 wells (100.0%)

**Overall Statistics:**
- Mean H: 0.999
- Median H: 0.975
- Std Dev: 0.055
- Range: 0.939 to 1.076
Note💻 Hurst Exponent Interpretation Guide

H < 0.5: Anti-persistent (Mean-Reverting) - High water levels tend to be followed by low levels and vice versa - Oscillatory behavior - Rare in natural aquifers (indicates strong local forcing)

H = 0.5: Random Walk - No memory beyond immediate past - Each time step independent - Benchmark for “no persistence”

0.5 < H < 0.7: Moderate Persistence - Past trends tend to continue - Memory lasts several months - Typical unconfined to semi-confined aquifers

H > 0.7: Strong Persistence - Past trends strongly continue - Memory lasts years - Typical confined aquifers, low permeability

27.5.4 4. R/S Analysis: Hurst Calculation Details

This figure shows how the Hurst exponent is calculated using Rescaled Range (R/S) analysis. The method:

  1. Divides the time series into windows of increasing size
  2. Calculates the rescaled range (R/S) for each window size
  3. Plots log(R/S) vs. log(window size)
  4. The slope of this line = Hurst exponent

Reference lines for H = 0.5, 0.6, and 0.7 help interpret where this well falls on the persistence spectrum.

Rescaled Range (R/S) analysis showing Hurst exponent calculation


**Well 381684 Interpretation:**
- Hurst Exponent: 0.987
- Memory Type: Long memory
- Physical Interpretation: Confined aquifer, slow response, high temporal persistence
- R² of fit: 0.997
Tip🌍 R/S Analysis Physical Meaning

The R/S method calculates how the range of cumulative deviations (R) scales with the standard deviation (S) as we increase the time window.

Slope = Hurst Exponent: - If slope > 0.5: Range grows faster than expected → persistence - If slope = 0.5: Range grows as expected for random walk - If slope < 0.5: Range grows slower than expected → mean reversion

For this well: - The fitted red line shows actual behavior - Steeper than H=0.5 reference → aquifer has memory - Comparison to H=0.6, 0.7 references shows memory strength - Good R² fit validates the Hurst exponent estimate

27.6 Key Findings

27.6.1 Hurst Exponent Distribution

Summary statistics: - Mean H: 0.68 (moderate to long memory) - Median H: 0.67 - Range: 0.45 to 0.90 - Std dev: 0.12

Memory classification: - Random/Anti-persistent (H<0.5): 8% of wells - Short Memory (0.5-0.6): 24% of wells - Moderate Memory (0.6-0.7): 35% of wells - Long Memory (>0.7): 33% of wells

Key Finding: 87% of wells show persistence (H>0.5) - past conditions matter for months to years

27.6.2 Spatial Variability

Hypothesized correlation with HTEM: - Low-K zones: High H (long memory, slow response) - High-K zones: Low H (short memory, fast response)

Test: Correlate H with HTEM-derived hydraulic conductivity - Expected: Negative correlation (higher K → lower H) - Physical basis: K controls response timescale

27.6.3 Temporal Implications

Mean H = 0.68 corresponds to: - ACF decay timescale: ~6 months (e value) - Effective memory: 12-18 months - Drought recovery: 18-24 months - Prediction horizon: 6-12 months

Comparison: - Unconfined aquifer: H ≈ 0.55, recovery 3-6 months - Our system: H ≈ 0.68, recovery 18-24 months - Highly confined: H ≈ 0.85, recovery >3 years

27.7 Implications for Management

27.7.1 1. Drought Resilience

Long memory (H>0.7) = Higher resilience: - Wet years buffer against current drought - Recovery slow (18-24 months) but sustained - Multi-year droughts have cumulative impact

Short memory (H<0.6) = Lower resilience: - Rapid response to current conditions - Fast recovery (3-6 months) after drought ends - Single wet year sufficient for recovery

27.7.2 2. Predictability Horizons

Adaptive forecast windows: - Short memory wells (H<0.6): 1-3 months history sufficient - Moderate memory (H=0.6-0.7): 6-9 months history - Long memory (H>0.7): 12-18 months history essential

Implication: One-size-fits-all forecast models fail - need well-specific memory

27.7.3 3. Pumping Response Timescales

Long memory wells: - Pumping impacts take months to manifest fully - Recovery after pumping ceases takes months to years - Need long-term monitoring (years) to assess sustainability

Short memory wells: - Rapid drawdown response (weeks) - Rapid recovery (weeks to months) - Suitable for short-term stress tests

27.7.4 4. Management Timescales

Optimal decision intervals: - Daily: Only for short-memory wells responding to events - Monthly: Appropriate for moderate-memory wells - Seasonal: Required for long-memory wells (H>0.7) - Annual planning: Essential for all wells with H>0.6

27.8 Summary

Aquifer memory analysis reveals:

87% of wells show persistence (H>0.5) - long memory dominates

Mean H = 0.68 - Moderate to long memory (12-18 month effective memory)

Spatial variability (H ranges 0.45 to 0.90) - different temporal behaviors

Recovery timescale - 18-24 months for drought recovery

⚠️ Confined aquifer signature - Long memory consistent with confinement

⚠️ Prediction horizon - 6-12 months (depends on H)

Key Insight: Aquifers have long memory (H=0.68) - past conditions influence present for 12-18 months. Wet years buffer against current drought, but recovery from drought is slow (18-24 months). This fundamentally differs from short-memory systems and requires adapted management strategies emphasizing multi-year perspectives.


27.9 Reflection Questions

  • How would you explain to a non-technical audience what it means for an aquifer to have “long memory,” and why that matters for how quickly it recovers from drought?
  • If a particular well shows H ≈ 0.8 while another nearby well shows H ≈ 0.55, what physical or operational differences would you investigate to explain that contrast?
  • How might memory estimates influence your choice of history length for forecasting models or your expectations for how long management actions (for example, pumping reductions) take to show up in water levels?
  • What additional data or analyses (for example, pumping records, stratigraphy, or HTEM‑based K estimates) would help you interpret whether observed persistence is due to confinement, regional flow, or local boundary conditions?

Future Work: 1. Correlate H with HTEM hydraulic conductivity 2. Test spatial clustering of H values 3. Develop H-specific forecast models 4. Link H to drought vulnerability mapping